1 /*
2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm.h"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/nmethod.hpp"
32 #include "code/pcDesc.hpp"
33 #include "code/scopeDesc.hpp"
34 #include "interpreter/bytecode.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "interpreter/oopMapCache.hpp"
37 #include "memory/allocation.inline.hpp"
38 #include "memory/oopFactory.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "memory/universe.hpp"
41 #include "oops/constantPool.hpp"
42 #include "oops/method.hpp"
43 #include "oops/objArrayKlass.hpp"
44 #include "oops/objArrayOop.inline.hpp"
45 #include "oops/oop.inline.hpp"
46 #include "oops/fieldStreams.hpp"
47 #include "oops/typeArrayOop.inline.hpp"
48 #include "oops/valueArrayKlass.hpp"
49 #include "oops/valueArrayOop.hpp"
50 #include "oops/valueKlass.hpp"
51 #include "oops/verifyOopClosure.hpp"
52 #include "prims/jvmtiThreadState.hpp"
53 #include "runtime/biasedLocking.hpp"
54 #include "runtime/compilationPolicy.hpp"
55 #include "runtime/deoptimization.hpp"
56 #include "runtime/fieldDescriptor.hpp"
57 #include "runtime/fieldDescriptor.inline.hpp"
58 #include "runtime/frame.inline.hpp"
59 #include "runtime/jniHandles.inline.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/interfaceSupport.inline.hpp"
62 #include "runtime/safepointVerifiers.hpp"
63 #include "runtime/sharedRuntime.hpp"
64 #include "runtime/signature.hpp"
65 #include "runtime/stubRoutines.hpp"
66 #include "runtime/thread.hpp"
67 #include "runtime/threadSMR.hpp"
68 #include "runtime/vframe.hpp"
69 #include "runtime/vframeArray.hpp"
70 #include "runtime/vframe_hp.hpp"
71 #include "utilities/events.hpp"
72 #include "utilities/preserveException.hpp"
73 #include "utilities/xmlstream.hpp"
74
75
76 bool DeoptimizationMarker::_is_active = false;
77
78 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
79 int caller_adjustment,
80 int caller_actual_parameters,
81 int number_of_frames,
82 intptr_t* frame_sizes,
83 address* frame_pcs,
84 BasicType return_type,
85 int exec_mode) {
86 _size_of_deoptimized_frame = size_of_deoptimized_frame;
87 _caller_adjustment = caller_adjustment;
88 _caller_actual_parameters = caller_actual_parameters;
89 _number_of_frames = number_of_frames;
90 _frame_sizes = frame_sizes;
91 _frame_pcs = frame_pcs;
92 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
93 _return_type = return_type;
94 _initial_info = 0;
95 // PD (x86 only)
96 _counter_temp = 0;
97 _unpack_kind = exec_mode;
98 _sender_sp_temp = 0;
99
100 _total_frame_sizes = size_of_frames();
101 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
102 }
103
104
105 Deoptimization::UnrollBlock::~UnrollBlock() {
106 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
107 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
108 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
109 }
110
111
112 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
113 assert(register_number < RegisterMap::reg_count, "checking register number");
114 return &_register_block[register_number * 2];
115 }
116
117
118
119 int Deoptimization::UnrollBlock::size_of_frames() const {
120 // Acount first for the adjustment of the initial frame
121 int result = _caller_adjustment;
122 for (int index = 0; index < number_of_frames(); index++) {
123 result += frame_sizes()[index];
124 }
125 return result;
126 }
127
128
129 void Deoptimization::UnrollBlock::print() {
130 ttyLocker ttyl;
131 tty->print_cr("UnrollBlock");
132 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
133 tty->print( " frame_sizes: ");
134 for (int index = 0; index < number_of_frames(); index++) {
135 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
136 }
137 tty->cr();
138 }
139
140
141 // In order to make fetch_unroll_info work properly with escape
142 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
143 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
144 // of previously eliminated objects occurs in realloc_objects, which is
145 // called from the method fetch_unroll_info_helper below.
146 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
147 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
148 // but makes the entry a little slower. There is however a little dance we have to
149 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
150
151 // fetch_unroll_info() is called at the beginning of the deoptimization
152 // handler. Note this fact before we start generating temporary frames
153 // that can confuse an asynchronous stack walker. This counter is
154 // decremented at the end of unpack_frames().
155 if (TraceDeoptimization) {
156 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
157 }
158 thread->inc_in_deopt_handler();
159
160 return fetch_unroll_info_helper(thread, exec_mode);
161 JRT_END
162
163
164 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
165 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
166
167 // Note: there is a safepoint safety issue here. No matter whether we enter
168 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
169 // the vframeArray is created.
170 //
171
172 // Allocate our special deoptimization ResourceMark
173 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
174 assert(thread->deopt_mark() == NULL, "Pending deopt!");
175 thread->set_deopt_mark(dmark);
176
177 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
178 RegisterMap map(thread, true);
179 RegisterMap dummy_map(thread, false);
180 // Now get the deoptee with a valid map
181 frame deoptee = stub_frame.sender(&map);
182 // Set the deoptee nmethod
183 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
184 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
185 thread->set_deopt_compiled_method(cm);
186
187 if (VerifyStack) {
188 thread->validate_frame_layout();
189 }
190
191 // Create a growable array of VFrames where each VFrame represents an inlined
192 // Java frame. This storage is allocated with the usual system arena.
193 assert(deoptee.is_compiled_frame(), "Wrong frame type");
194 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
195 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
196 while (!vf->is_top()) {
197 assert(vf->is_compiled_frame(), "Wrong frame type");
198 chunk->push(compiledVFrame::cast(vf));
199 vf = vf->sender();
200 }
201 assert(vf->is_compiled_frame(), "Wrong frame type");
202 chunk->push(compiledVFrame::cast(vf));
203
204 bool realloc_failures = false;
205
206 #if COMPILER2_OR_JVMCI
207 // Reallocate the non-escaping objects and restore their fields. Then
208 // relock objects if synchronization on them was eliminated.
209 #if !INCLUDE_JVMCI
210 if (DoEscapeAnalysis || EliminateNestedLocks) {
211 if (EliminateAllocations) {
212 #endif // INCLUDE_JVMCI
213 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
214 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
215
216 // The flag return_oop() indicates call sites which return oop
217 // in compiled code. Such sites include java method calls,
218 // runtime calls (for example, used to allocate new objects/arrays
219 // on slow code path) and any other calls generated in compiled code.
220 // It is not guaranteed that we can get such information here only
221 // by analyzing bytecode in deoptimized frames. This is why this flag
222 // is set during method compilation (see Compile::Process_OopMap_Node()).
223 // If the previous frame was popped or if we are dispatching an exception,
224 // we don't have an oop result.
225 ScopeDesc* scope = chunk->at(0)->scope();
226 bool save_oop_result = scope->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Unpack_deopt);
227 // In case of the return of multiple values, we must take care
228 // of all oop return values.
229 GrowableArray<Handle> return_oops;
230 ValueKlass* vk = NULL;
231 if (save_oop_result && scope->return_vt()) {
232 vk = ValueKlass::returned_value_klass(map);
233 if (vk != NULL) {
234 vk->save_oop_fields(map, return_oops);
235 save_oop_result = false;
236 }
237 }
238 if (save_oop_result) {
239 // Reallocation may trigger GC. If deoptimization happened on return from
240 // call which returns oop we need to save it since it is not in oopmap.
241 oop result = deoptee.saved_oop_result(&map);
242 assert(oopDesc::is_oop_or_null(result), "must be oop");
243 return_oops.push(Handle(thread, result));
244 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
245 if (TraceDeoptimization) {
246 ttyLocker ttyl;
247 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
248 }
249 }
250 if (objects != NULL || vk != NULL) {
251 bool skip_internal = (cm != NULL) && !cm->is_compiled_by_jvmci();
252 JRT_BLOCK
253 if (vk != NULL) {
254 realloc_failures = realloc_value_type_result(vk, map, return_oops, THREAD);
255 }
256 if (objects != NULL) {
257 realloc_failures = realloc_failures || realloc_objects(thread, &deoptee, &map, objects, THREAD);
258 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD);
259 }
260 JRT_END
261 #ifndef PRODUCT
262 if (TraceDeoptimization) {
263 ttyLocker ttyl;
264 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
265 print_objects(objects, realloc_failures);
266 }
267 #endif
268 }
269 if (save_oop_result || vk != NULL) {
270 // Restore result.
271 assert(return_oops.length() == 1, "no value type");
272 deoptee.set_saved_oop_result(&map, return_oops.pop()());
273 }
274 #if !INCLUDE_JVMCI
275 }
276 if (EliminateLocks) {
277 #endif // INCLUDE_JVMCI
278 #ifndef PRODUCT
279 bool first = true;
280 #endif
281 for (int i = 0; i < chunk->length(); i++) {
282 compiledVFrame* cvf = chunk->at(i);
283 assert (cvf->scope() != NULL,"expect only compiled java frames");
284 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
285 if (monitors->is_nonempty()) {
286 relock_objects(monitors, thread, realloc_failures);
287 #ifndef PRODUCT
288 if (PrintDeoptimizationDetails) {
289 ttyLocker ttyl;
290 for (int j = 0; j < monitors->length(); j++) {
291 MonitorInfo* mi = monitors->at(j);
292 if (mi->eliminated()) {
293 if (first) {
294 first = false;
295 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
296 }
297 if (mi->owner_is_scalar_replaced()) {
298 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
299 tty->print_cr(" failed reallocation for klass %s", k->external_name());
300 } else {
301 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
302 }
303 }
304 }
305 }
306 #endif // !PRODUCT
307 }
308 }
309 #if !INCLUDE_JVMCI
310 }
311 }
312 #endif // INCLUDE_JVMCI
313 #endif // COMPILER2_OR_JVMCI
314
315 ScopeDesc* trap_scope = chunk->at(0)->scope();
316 Handle exceptionObject;
317 if (trap_scope->rethrow_exception()) {
318 if (PrintDeoptimizationDetails) {
319 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());
320 }
321 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
322 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
323 ScopeValue* topOfStack = expressions->top();
324 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
325 guarantee(exceptionObject() != NULL, "exception oop can not be null");
326 }
327
328 // Ensure that no safepoint is taken after pointers have been stored
329 // in fields of rematerialized objects. If a safepoint occurs from here on
330 // out the java state residing in the vframeArray will be missed.
331 NoSafepointVerifier no_safepoint;
332
333 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
334 #if COMPILER2_OR_JVMCI
335 if (realloc_failures) {
336 pop_frames_failed_reallocs(thread, array);
337 }
338 #endif
339
340 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
341 thread->set_vframe_array_head(array);
342
343 // Now that the vframeArray has been created if we have any deferred local writes
344 // added by jvmti then we can free up that structure as the data is now in the
345 // vframeArray
346
347 if (thread->deferred_locals() != NULL) {
348 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
349 int i = 0;
350 do {
351 // Because of inlining we could have multiple vframes for a single frame
352 // and several of the vframes could have deferred writes. Find them all.
353 if (list->at(i)->id() == array->original().id()) {
354 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
355 list->remove_at(i);
356 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
357 delete dlv;
358 } else {
359 i++;
360 }
361 } while ( i < list->length() );
362 if (list->length() == 0) {
363 thread->set_deferred_locals(NULL);
364 // free the list and elements back to C heap.
365 delete list;
366 }
367
368 }
369
370 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
371 CodeBlob* cb = stub_frame.cb();
372 // Verify we have the right vframeArray
373 assert(cb->frame_size() >= 0, "Unexpected frame size");
374 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
375
376 // If the deopt call site is a MethodHandle invoke call site we have
377 // to adjust the unpack_sp.
378 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
379 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
380 unpack_sp = deoptee.unextended_sp();
381
382 #ifdef ASSERT
383 assert(cb->is_deoptimization_stub() ||
384 cb->is_uncommon_trap_stub() ||
385 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
386 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
387 "unexpected code blob: %s", cb->name());
388 #endif
389
390 // This is a guarantee instead of an assert because if vframe doesn't match
391 // we will unpack the wrong deoptimized frame and wind up in strange places
392 // where it will be very difficult to figure out what went wrong. Better
393 // to die an early death here than some very obscure death later when the
394 // trail is cold.
395 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
396 // in that it will fail to detect a problem when there is one. This needs
397 // more work in tiger timeframe.
398 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
399
400 int number_of_frames = array->frames();
401
402 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
403 // virtual activation, which is the reverse of the elements in the vframes array.
404 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
405 // +1 because we always have an interpreter return address for the final slot.
406 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
407 int popframe_extra_args = 0;
408 // Create an interpreter return address for the stub to use as its return
409 // address so the skeletal frames are perfectly walkable
410 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
411
412 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
413 // activation be put back on the expression stack of the caller for reexecution
414 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
415 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
416 }
417
418 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
419 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
420 // than simply use array->sender.pc(). This requires us to walk the current set of frames
421 //
422 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
423 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
424
425 // It's possible that the number of parameters at the call site is
426 // different than number of arguments in the callee when method
427 // handles are used. If the caller is interpreted get the real
428 // value so that the proper amount of space can be added to it's
429 // frame.
430 bool caller_was_method_handle = false;
431 if (deopt_sender.is_interpreted_frame()) {
432 methodHandle method = deopt_sender.interpreter_frame_method();
433 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
434 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
435 // Method handle invokes may involve fairly arbitrary chains of
436 // calls so it's impossible to know how much actual space the
437 // caller has for locals.
438 caller_was_method_handle = true;
439 }
440 }
441
442 //
443 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
444 // frame_sizes/frame_pcs[1] next oldest frame (int)
445 // frame_sizes/frame_pcs[n] youngest frame (int)
446 //
447 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
448 // owns the space for the return address to it's caller). Confusing ain't it.
449 //
450 // The vframe array can address vframes with indices running from
451 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
452 // When we create the skeletal frames we need the oldest frame to be in the zero slot
453 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
454 // so things look a little strange in this loop.
455 //
456 int callee_parameters = 0;
457 int callee_locals = 0;
458 for (int index = 0; index < array->frames(); index++ ) {
459 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
460 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
461 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
462 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
463 callee_locals,
464 index == 0,
465 popframe_extra_args);
466 // This pc doesn't have to be perfect just good enough to identify the frame
467 // as interpreted so the skeleton frame will be walkable
468 // The correct pc will be set when the skeleton frame is completely filled out
469 // The final pc we store in the loop is wrong and will be overwritten below
470 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
471
472 callee_parameters = array->element(index)->method()->size_of_parameters();
473 callee_locals = array->element(index)->method()->max_locals();
474 popframe_extra_args = 0;
475 }
476
477 // Compute whether the root vframe returns a float or double value.
478 BasicType return_type;
479 {
480 methodHandle method(thread, array->element(0)->method());
481 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
482 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
483 }
484
485 // Compute information for handling adapters and adjusting the frame size of the caller.
486 int caller_adjustment = 0;
487
488 // Compute the amount the oldest interpreter frame will have to adjust
489 // its caller's stack by. If the caller is a compiled frame then
490 // we pretend that the callee has no parameters so that the
491 // extension counts for the full amount of locals and not just
492 // locals-parms. This is because without a c2i adapter the parm
493 // area as created by the compiled frame will not be usable by
494 // the interpreter. (Depending on the calling convention there
495 // may not even be enough space).
496
497 // QQQ I'd rather see this pushed down into last_frame_adjust
498 // and have it take the sender (aka caller).
499
500 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
501 caller_adjustment = last_frame_adjust(0, callee_locals);
502 } else if (callee_locals > callee_parameters) {
503 // The caller frame may need extending to accommodate
504 // non-parameter locals of the first unpacked interpreted frame.
505 // Compute that adjustment.
506 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
507 }
508
509 // If the sender is deoptimized we must retrieve the address of the handler
510 // since the frame will "magically" show the original pc before the deopt
511 // and we'd undo the deopt.
512
513 frame_pcs[0] = deopt_sender.raw_pc();
514
515 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
516
517 #if INCLUDE_JVMCI
518 if (exceptionObject() != NULL) {
519 thread->set_exception_oop(exceptionObject());
520 exec_mode = Unpack_exception;
521 }
522 #endif
523
524 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
525 assert(thread->has_pending_exception(), "should have thrown OOME");
526 thread->set_exception_oop(thread->pending_exception());
527 thread->clear_pending_exception();
528 exec_mode = Unpack_exception;
529 }
530
531 #if INCLUDE_JVMCI
532 if (thread->frames_to_pop_failed_realloc() > 0) {
533 thread->set_pending_monitorenter(false);
534 }
535 #endif
536
537 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
538 caller_adjustment * BytesPerWord,
539 caller_was_method_handle ? 0 : callee_parameters,
540 number_of_frames,
541 frame_sizes,
542 frame_pcs,
543 return_type,
544 exec_mode);
545 // On some platforms, we need a way to pass some platform dependent
546 // information to the unpacking code so the skeletal frames come out
547 // correct (initial fp value, unextended sp, ...)
548 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
549
550 if (array->frames() > 1) {
551 if (VerifyStack && TraceDeoptimization) {
552 ttyLocker ttyl;
553 tty->print_cr("Deoptimizing method containing inlining");
554 }
555 }
556
557 array->set_unroll_block(info);
558 return info;
559 }
560
561 // Called to cleanup deoptimization data structures in normal case
562 // after unpacking to stack and when stack overflow error occurs
563 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
564 vframeArray *array) {
565
566 // Get array if coming from exception
567 if (array == NULL) {
568 array = thread->vframe_array_head();
569 }
570 thread->set_vframe_array_head(NULL);
571
572 // Free the previous UnrollBlock
573 vframeArray* old_array = thread->vframe_array_last();
574 thread->set_vframe_array_last(array);
575
576 if (old_array != NULL) {
577 UnrollBlock* old_info = old_array->unroll_block();
578 old_array->set_unroll_block(NULL);
579 delete old_info;
580 delete old_array;
581 }
582
583 // Deallocate any resource creating in this routine and any ResourceObjs allocated
584 // inside the vframeArray (StackValueCollections)
585
586 delete thread->deopt_mark();
587 thread->set_deopt_mark(NULL);
588 thread->set_deopt_compiled_method(NULL);
589
590
591 if (JvmtiExport::can_pop_frame()) {
592 #ifndef CC_INTERP
593 // Regardless of whether we entered this routine with the pending
594 // popframe condition bit set, we should always clear it now
595 thread->clear_popframe_condition();
596 #else
597 // C++ interpreter will clear has_pending_popframe when it enters
598 // with method_resume. For deopt_resume2 we clear it now.
599 if (thread->popframe_forcing_deopt_reexecution())
600 thread->clear_popframe_condition();
601 #endif /* CC_INTERP */
602 }
603
604 // unpack_frames() is called at the end of the deoptimization handler
605 // and (in C2) at the end of the uncommon trap handler. Note this fact
606 // so that an asynchronous stack walker can work again. This counter is
607 // incremented at the beginning of fetch_unroll_info() and (in C2) at
608 // the beginning of uncommon_trap().
609 thread->dec_in_deopt_handler();
610 }
611
612 // Moved from cpu directories because none of the cpus has callee save values.
613 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
614 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
615
616 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
617 // the days we had adapter frames. When we deoptimize a situation where a
618 // compiled caller calls a compiled caller will have registers it expects
619 // to survive the call to the callee. If we deoptimize the callee the only
620 // way we can restore these registers is to have the oldest interpreter
621 // frame that we create restore these values. That is what this routine
622 // will accomplish.
623
624 // At the moment we have modified c2 to not have any callee save registers
625 // so this problem does not exist and this routine is just a place holder.
626
627 assert(f->is_interpreted_frame(), "must be interpreted");
628 }
629
630 // Return BasicType of value being returned
631 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
632
633 // We are already active in the special DeoptResourceMark any ResourceObj's we
634 // allocate will be freed at the end of the routine.
635
636 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
637 // but makes the entry a little slower. There is however a little dance we have to
638 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
639 ResetNoHandleMark rnhm; // No-op in release/product versions
640 HandleMark hm;
641
642 frame stub_frame = thread->last_frame();
643
644 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
645 // must point to the vframeArray for the unpack frame.
646 vframeArray* array = thread->vframe_array_head();
647
648 #ifndef PRODUCT
649 if (TraceDeoptimization) {
650 ttyLocker ttyl;
651 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
652 p2i(thread), p2i(array), exec_mode);
653 }
654 #endif
655 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
656 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
657
658 UnrollBlock* info = array->unroll_block();
659
660 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
661 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
662
663 BasicType bt = info->return_type();
664
665 // If we have an exception pending, claim that the return type is an oop
666 // so the deopt_blob does not overwrite the exception_oop.
667
668 if (exec_mode == Unpack_exception)
669 bt = T_OBJECT;
670
671 // Cleanup thread deopt data
672 cleanup_deopt_info(thread, array);
673
674 #ifndef PRODUCT
675 if (VerifyStack) {
676 ResourceMark res_mark;
677 // Clear pending exception to not break verification code (restored afterwards)
678 PRESERVE_EXCEPTION_MARK;
679
680 thread->validate_frame_layout();
681
682 // Verify that the just-unpacked frames match the interpreter's
683 // notions of expression stack and locals
684 vframeArray* cur_array = thread->vframe_array_last();
685 RegisterMap rm(thread, false);
686 rm.set_include_argument_oops(false);
687 bool is_top_frame = true;
688 int callee_size_of_parameters = 0;
689 int callee_max_locals = 0;
690 for (int i = 0; i < cur_array->frames(); i++) {
691 vframeArrayElement* el = cur_array->element(i);
692 frame* iframe = el->iframe();
693 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
694
695 // Get the oop map for this bci
696 InterpreterOopMap mask;
697 int cur_invoke_parameter_size = 0;
698 bool try_next_mask = false;
699 int next_mask_expression_stack_size = -1;
700 int top_frame_expression_stack_adjustment = 0;
701 methodHandle mh(thread, iframe->interpreter_frame_method());
702 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
703 BytecodeStream str(mh, iframe->interpreter_frame_bci());
704 int max_bci = mh->code_size();
705 // Get to the next bytecode if possible
706 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
707 // Check to see if we can grab the number of outgoing arguments
708 // at an uncommon trap for an invoke (where the compiler
709 // generates debug info before the invoke has executed)
710 Bytecodes::Code cur_code = str.next();
711 if (Bytecodes::is_invoke(cur_code)) {
712 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
713 cur_invoke_parameter_size = invoke.size_of_parameters();
714 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
715 callee_size_of_parameters++;
716 }
717 }
718 if (str.bci() < max_bci) {
719 Bytecodes::Code next_code = str.next();
720 if (next_code >= 0) {
721 // The interpreter oop map generator reports results before
722 // the current bytecode has executed except in the case of
723 // calls. It seems to be hard to tell whether the compiler
724 // has emitted debug information matching the "state before"
725 // a given bytecode or the state after, so we try both
726 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
727 // Get expression stack size for the next bytecode
728 InterpreterOopMap next_mask;
729 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
730 next_mask_expression_stack_size = next_mask.expression_stack_size();
731 if (Bytecodes::is_invoke(next_code)) {
732 Bytecode_invoke invoke(mh, str.bci());
733 next_mask_expression_stack_size += invoke.size_of_parameters();
734 }
735 // Need to subtract off the size of the result type of
736 // the bytecode because this is not described in the
737 // debug info but returned to the interpreter in the TOS
738 // caching register
739 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
740 if (bytecode_result_type != T_ILLEGAL) {
741 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
742 }
743 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
744 try_next_mask = true;
745 }
746 }
747 }
748
749 // Verify stack depth and oops in frame
750 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
751 if (!(
752 /* SPARC */
753 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
754 /* x86 */
755 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
756 (try_next_mask &&
757 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
758 top_frame_expression_stack_adjustment))) ||
759 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
760 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
761 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
762 )) {
763 {
764 ttyLocker ttyl;
765
766 // Print out some information that will help us debug the problem
767 tty->print_cr("Wrong number of expression stack elements during deoptimization");
768 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
769 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
770 iframe->interpreter_frame_expression_stack_size());
771 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
772 tty->print_cr(" try_next_mask = %d", try_next_mask);
773 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
774 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
775 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
776 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
777 tty->print_cr(" exec_mode = %d", exec_mode);
778 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
779 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
780 tty->print_cr(" Interpreted frames:");
781 for (int k = 0; k < cur_array->frames(); k++) {
782 vframeArrayElement* el = cur_array->element(k);
783 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
784 }
785 cur_array->print_on_2(tty);
786 } // release tty lock before calling guarantee
787 guarantee(false, "wrong number of expression stack elements during deopt");
788 }
789 VerifyOopClosure verify;
790 iframe->oops_interpreted_do(&verify, &rm, false);
791 callee_size_of_parameters = mh->size_of_parameters();
792 callee_max_locals = mh->max_locals();
793 is_top_frame = false;
794 }
795 }
796 #endif /* !PRODUCT */
797
798
799 return bt;
800 JRT_END
801
802
803 int Deoptimization::deoptimize_dependents() {
804 Threads::deoptimized_wrt_marked_nmethods();
805 return 0;
806 }
807
808 Deoptimization::DeoptAction Deoptimization::_unloaded_action
809 = Deoptimization::Action_reinterpret;
810
811
812
813 #if INCLUDE_JVMCI || INCLUDE_AOT
814 template<typename CacheType>
815 class BoxCacheBase : public CHeapObj<mtCompiler> {
816 protected:
817 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) {
818 ResourceMark rm;
819 char* klass_name_str = klass_name->as_C_string();
820 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD);
821 guarantee(k != NULL, "%s must be loaded", klass_name_str);
822 InstanceKlass* ik = InstanceKlass::cast(k);
823 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
824 CacheType::compute_offsets(ik);
825 return ik;
826 }
827 };
828
829 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
830 PrimitiveType _low;
831 PrimitiveType _high;
832 jobject _cache;
833 protected:
834 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
835 BoxCache(Thread* thread) {
836 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread);
837 objArrayOop cache = CacheType::cache(ik);
838 assert(cache->length() > 0, "Empty cache");
839 _low = BoxType::value(cache->obj_at(0));
840 _high = _low + cache->length() - 1;
841 _cache = JNIHandles::make_global(Handle(thread, cache));
842 }
843 ~BoxCache() {
844 JNIHandles::destroy_global(_cache);
845 }
846 public:
847 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
848 if (_singleton == NULL) {
849 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
850 if (!Atomic::replace_if_null(s, &_singleton)) {
851 delete s;
852 }
853 }
854 return _singleton;
855 }
856 oop lookup(PrimitiveType value) {
857 if (_low <= value && value <= _high) {
858 int offset = value - _low;
859 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
860 }
861 return NULL;
862 }
863 oop lookup_raw(intptr_t raw_value) {
864 // Have to cast to avoid little/big-endian problems.
865 if (sizeof(PrimitiveType) > sizeof(jint)) {
866 jlong value = (jlong)raw_value;
867 return lookup(value);
868 }
869 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
870 return lookup(value);
871 }
872 };
873
874 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
875 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
876 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
877 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
878 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
879
880 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL;
881 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL;
882 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL;
883 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL;
884 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL;
885
886 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
887 jobject _true_cache;
888 jobject _false_cache;
889 protected:
890 static BooleanBoxCache *_singleton;
891 BooleanBoxCache(Thread *thread) {
892 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread);
893 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
894 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
895 }
896 ~BooleanBoxCache() {
897 JNIHandles::destroy_global(_true_cache);
898 JNIHandles::destroy_global(_false_cache);
899 }
900 public:
901 static BooleanBoxCache* singleton(Thread* thread) {
902 if (_singleton == NULL) {
903 BooleanBoxCache* s = new BooleanBoxCache(thread);
904 if (!Atomic::replace_if_null(s, &_singleton)) {
905 delete s;
906 }
907 }
908 return _singleton;
909 }
910 oop lookup_raw(intptr_t raw_value) {
911 // Have to cast to avoid little/big-endian problems.
912 jboolean value = (jboolean)*((jint*)&raw_value);
913 return lookup(value);
914 }
915 oop lookup(jboolean value) {
916 if (value != 0) {
917 return JNIHandles::resolve_non_null(_true_cache);
918 }
919 return JNIHandles::resolve_non_null(_false_cache);
920 }
921 };
922
923 BooleanBoxCache* BooleanBoxCache::_singleton = NULL;
924
925 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) {
926 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
927 BasicType box_type = SystemDictionary::box_klass_type(k);
928 if (box_type != T_OBJECT) {
929 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
930 switch(box_type) {
931 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
932 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
933 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
934 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
935 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
936 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
937 default:;
938 }
939 }
940 return NULL;
941 }
942 #endif // INCLUDE_JVMCI || INCLUDE_AOT
943
944 #if COMPILER2_OR_JVMCI
945 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
946 Handle pending_exception(THREAD, thread->pending_exception());
947 const char* exception_file = thread->exception_file();
948 int exception_line = thread->exception_line();
949 thread->clear_pending_exception();
950
951 bool failures = false;
952
953 for (int i = 0; i < objects->length(); i++) {
954 assert(objects->at(i)->is_object(), "invalid debug information");
955 ObjectValue* sv = (ObjectValue*) objects->at(i);
956
957 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
958 oop obj = NULL;
959
960 if (k->is_instance_klass()) {
961 #if INCLUDE_JVMCI || INCLUDE_AOT
962 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null();
963 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) {
964 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
965 obj = get_cached_box(abv, fr, reg_map, THREAD);
966 if (obj != NULL) {
967 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
968 abv->set_cached(true);
969 }
970 }
971 #endif // INCLUDE_JVMCI || INCLUDE_AOT
972 InstanceKlass* ik = InstanceKlass::cast(k);
973 if (obj == NULL) {
974 obj = ik->allocate_instance(THREAD);
975 }
976 } else if (k->is_valueArray_klass()) {
977 ValueArrayKlass* ak = ValueArrayKlass::cast(k);
978 // Value type array must be zeroed because not all memory is reassigned
979 obj = ak->allocate(sv->field_size(), THREAD);
980 } else if (k->is_typeArray_klass()) {
981 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
982 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
983 int len = sv->field_size() / type2size[ak->element_type()];
984 obj = ak->allocate(len, THREAD);
985 } else if (k->is_objArray_klass()) {
986 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
987 obj = ak->allocate(sv->field_size(), THREAD);
988 }
989
990 if (obj == NULL) {
991 failures = true;
992 }
993
994 assert(sv->value().is_null(), "redundant reallocation");
995 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
996 CLEAR_PENDING_EXCEPTION;
997 sv->set_value(obj);
998 }
999
1000 if (failures) {
1001 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1002 } else if (pending_exception.not_null()) {
1003 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1004 }
1005
1006 return failures;
1007 }
1008
1009 // We're deoptimizing at the return of a call, value type fields are
1010 // in registers. When we go back to the interpreter, it will expect a
1011 // reference to a value type instance. Allocate and initialize it from
1012 // the register values here.
1013 bool Deoptimization::realloc_value_type_result(ValueKlass* vk, const RegisterMap& map, GrowableArray<Handle>& return_oops, TRAPS) {
1014 oop new_vt = vk->realloc_result(map, return_oops, THREAD);
1015 if (new_vt == NULL) {
1016 CLEAR_PENDING_EXCEPTION;
1017 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), true);
1018 }
1019 return_oops.clear();
1020 return_oops.push(Handle(THREAD, new_vt));
1021 return false;
1022 }
1023
1024 // restore elements of an eliminated type array
1025 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1026 int index = 0;
1027 intptr_t val;
1028
1029 for (int i = 0; i < sv->field_size(); i++) {
1030 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1031 switch(type) {
1032 case T_LONG: case T_DOUBLE: {
1033 assert(value->type() == T_INT, "Agreement.");
1034 StackValue* low =
1035 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1036 #ifdef _LP64
1037 jlong res = (jlong)low->get_int();
1038 #else
1039 #ifdef SPARC
1040 // For SPARC we have to swap high and low words.
1041 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1042 #else
1043 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1044 #endif //SPARC
1045 #endif
1046 obj->long_at_put(index, res);
1047 break;
1048 }
1049
1050 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1051 case T_INT: case T_FLOAT: { // 4 bytes.
1052 assert(value->type() == T_INT, "Agreement.");
1053 bool big_value = false;
1054 if (i + 1 < sv->field_size() && type == T_INT) {
1055 if (sv->field_at(i)->is_location()) {
1056 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1057 if (type == Location::dbl || type == Location::lng) {
1058 big_value = true;
1059 }
1060 } else if (sv->field_at(i)->is_constant_int()) {
1061 ScopeValue* next_scope_field = sv->field_at(i + 1);
1062 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1063 big_value = true;
1064 }
1065 }
1066 }
1067
1068 if (big_value) {
1069 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1070 #ifdef _LP64
1071 jlong res = (jlong)low->get_int();
1072 #else
1073 #ifdef SPARC
1074 // For SPARC we have to swap high and low words.
1075 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1076 #else
1077 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1078 #endif //SPARC
1079 #endif
1080 obj->int_at_put(index, (jint)*((jint*)&res));
1081 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1082 } else {
1083 val = value->get_int();
1084 obj->int_at_put(index, (jint)*((jint*)&val));
1085 }
1086 break;
1087 }
1088
1089 case T_SHORT:
1090 assert(value->type() == T_INT, "Agreement.");
1091 val = value->get_int();
1092 obj->short_at_put(index, (jshort)*((jint*)&val));
1093 break;
1094
1095 case T_CHAR:
1096 assert(value->type() == T_INT, "Agreement.");
1097 val = value->get_int();
1098 obj->char_at_put(index, (jchar)*((jint*)&val));
1099 break;
1100
1101 case T_BYTE:
1102 assert(value->type() == T_INT, "Agreement.");
1103 val = value->get_int();
1104 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1105 break;
1106
1107 case T_BOOLEAN:
1108 assert(value->type() == T_INT, "Agreement.");
1109 val = value->get_int();
1110 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1111 break;
1112
1113 default:
1114 ShouldNotReachHere();
1115 }
1116 index++;
1117 }
1118 }
1119
1120
1121 // restore fields of an eliminated object array
1122 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1123 for (int i = 0; i < sv->field_size(); i++) {
1124 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1125 assert(value->type() == T_OBJECT, "object element expected");
1126 obj->obj_at_put(i, value->get_obj()());
1127 }
1128 }
1129
1130 class ReassignedField {
1131 public:
1132 int _offset;
1133 BasicType _type;
1134 InstanceKlass* _klass;
1135 public:
1136 ReassignedField() {
1137 _offset = 0;
1138 _type = T_ILLEGAL;
1139 _klass = NULL;
1140 }
1141 };
1142
1143 int compare(ReassignedField* left, ReassignedField* right) {
1144 return left->_offset - right->_offset;
1145 }
1146
1147 // Restore fields of an eliminated instance object using the same field order
1148 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1149 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) {
1150 if (klass->superklass() != NULL) {
1151 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0);
1152 }
1153
1154 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1155 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1156 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1157 ReassignedField field;
1158 field._offset = fs.offset();
1159 field._type = FieldType::basic_type(fs.signature());
1160 if (field._type == T_VALUETYPE) {
1161 field._type = T_OBJECT;
1162 }
1163 if (fs.is_flattened()) {
1164 // Resolve klass of flattened value type field
1165 Klass* vk = klass->get_value_field_klass(fs.index());
1166 field._klass = ValueKlass::cast(vk);
1167 field._type = T_VALUETYPE;
1168 }
1169 fields->append(field);
1170 }
1171 }
1172 fields->sort(compare);
1173 for (int i = 0; i < fields->length(); i++) {
1174 intptr_t val;
1175 ScopeValue* scope_field = sv->field_at(svIndex);
1176 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1177 int offset = base_offset + fields->at(i)._offset;
1178 BasicType type = fields->at(i)._type;
1179 switch (type) {
1180 case T_OBJECT:
1181 case T_ARRAY:
1182 assert(value->type() == T_OBJECT, "Agreement.");
1183 obj->obj_field_put(offset, value->get_obj()());
1184 break;
1185
1186 case T_VALUETYPE: {
1187 // Recursively re-assign flattened value type fields
1188 InstanceKlass* vk = fields->at(i)._klass;
1189 assert(vk != NULL, "must be resolved");
1190 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header
1191 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0);
1192 continue; // Continue because we don't need to increment svIndex
1193 }
1194
1195 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1196 case T_INT: case T_FLOAT: { // 4 bytes.
1197 assert(value->type() == T_INT, "Agreement.");
1198 bool big_value = false;
1199 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1200 if (scope_field->is_location()) {
1201 Location::Type type = ((LocationValue*) scope_field)->location().type();
1202 if (type == Location::dbl || type == Location::lng) {
1203 big_value = true;
1204 }
1205 }
1206 if (scope_field->is_constant_int()) {
1207 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1208 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1209 big_value = true;
1210 }
1211 }
1212 }
1213
1214 if (big_value) {
1215 i++;
1216 assert(i < fields->length(), "second T_INT field needed");
1217 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1218 } else {
1219 val = value->get_int();
1220 obj->int_field_put(offset, (jint)*((jint*)&val));
1221 break;
1222 }
1223 }
1224 /* no break */
1225
1226 case T_LONG: case T_DOUBLE: {
1227 assert(value->type() == T_INT, "Agreement.");
1228 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1229 #ifdef _LP64
1230 jlong res = (jlong)low->get_int();
1231 #else
1232 #ifdef SPARC
1233 // For SPARC we have to swap high and low words.
1234 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1235 #else
1236 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1237 #endif //SPARC
1238 #endif
1239 obj->long_field_put(offset, res);
1240 break;
1241 }
1242
1243 case T_SHORT:
1244 assert(value->type() == T_INT, "Agreement.");
1245 val = value->get_int();
1246 obj->short_field_put(offset, (jshort)*((jint*)&val));
1247 break;
1248
1249 case T_CHAR:
1250 assert(value->type() == T_INT, "Agreement.");
1251 val = value->get_int();
1252 obj->char_field_put(offset, (jchar)*((jint*)&val));
1253 break;
1254
1255 case T_BYTE:
1256 assert(value->type() == T_INT, "Agreement.");
1257 val = value->get_int();
1258 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1259 break;
1260
1261 case T_BOOLEAN:
1262 assert(value->type() == T_INT, "Agreement.");
1263 val = value->get_int();
1264 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1265 break;
1266
1267 default:
1268 ShouldNotReachHere();
1269 }
1270 svIndex++;
1271 }
1272 return svIndex;
1273 }
1274
1275 // restore fields of an eliminated value type array
1276 void Deoptimization::reassign_value_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, valueArrayOop obj, ValueArrayKlass* vak, TRAPS) {
1277 ValueKlass* vk = vak->element_klass();
1278 assert(vk->flatten_array(), "should only be used for flattened value type arrays");
1279 // Adjust offset to omit oop header
1280 int base_offset = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE) - ValueKlass::cast(vk)->first_field_offset();
1281 // Initialize all elements of the flattened value type array
1282 for (int i = 0; i < sv->field_size(); i++) {
1283 ScopeValue* val = sv->field_at(i);
1284 int offset = base_offset + (i << Klass::layout_helper_log2_element_size(vak->layout_helper()));
1285 reassign_fields_by_klass(vk, fr, reg_map, val->as_ObjectValue(), 0, (oop)obj, false /* skip_internal */, offset, CHECK);
1286 }
1287 }
1288
1289 // restore fields of all eliminated objects and arrays
1290 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) {
1291 for (int i = 0; i < objects->length(); i++) {
1292 ObjectValue* sv = (ObjectValue*) objects->at(i);
1293 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1294 Handle obj = sv->value();
1295 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1296 if (PrintDeoptimizationDetails) {
1297 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1298 }
1299 if (obj.is_null()) {
1300 continue;
1301 }
1302 #if INCLUDE_JVMCI || INCLUDE_AOT
1303 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1304 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1305 continue;
1306 }
1307 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1308 if (k->is_instance_klass()) {
1309 InstanceKlass* ik = InstanceKlass::cast(k);
1310 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK);
1311 } else if (k->is_valueArray_klass()) {
1312 ValueArrayKlass* vak = ValueArrayKlass::cast(k);
1313 reassign_value_array_elements(fr, reg_map, sv, (valueArrayOop) obj(), vak, CHECK);
1314 } else if (k->is_typeArray_klass()) {
1315 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1316 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1317 } else if (k->is_objArray_klass()) {
1318 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1319 }
1320 }
1321 }
1322
1323
1324 // relock objects for which synchronization was eliminated
1325 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1326 for (int i = 0; i < monitors->length(); i++) {
1327 MonitorInfo* mon_info = monitors->at(i);
1328 if (mon_info->eliminated()) {
1329 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1330 if (!mon_info->owner_is_scalar_replaced()) {
1331 Handle obj(thread, mon_info->owner());
1332 markWord mark = obj->mark();
1333 if (UseBiasedLocking && mark.has_bias_pattern()) {
1334 // New allocated objects may have the mark set to anonymously biased.
1335 // Also the deoptimized method may called methods with synchronization
1336 // where the thread-local object is bias locked to the current thread.
1337 assert(mark.is_biased_anonymously() ||
1338 mark.biased_locker() == thread, "should be locked to current thread");
1339 // Reset mark word to unbiased prototype.
1340 markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1341 obj->set_mark(unbiased_prototype);
1342 }
1343 BasicLock* lock = mon_info->lock();
1344 ObjectSynchronizer::enter(obj, lock, thread);
1345 assert(mon_info->owner()->is_locked(), "object must be locked now");
1346 }
1347 }
1348 }
1349 }
1350
1351
1352 #ifndef PRODUCT
1353 // print information about reallocated objects
1354 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1355 fieldDescriptor fd;
1356
1357 for (int i = 0; i < objects->length(); i++) {
1358 ObjectValue* sv = (ObjectValue*) objects->at(i);
1359 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1360 Handle obj = sv->value();
1361
1362 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1363 k->print_value();
1364 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1365 if (obj.is_null()) {
1366 tty->print(" allocation failed");
1367 } else {
1368 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1369 }
1370 tty->cr();
1371
1372 if (Verbose && !obj.is_null()) {
1373 k->oop_print_on(obj(), tty);
1374 }
1375 }
1376 }
1377 #endif
1378 #endif // COMPILER2_OR_JVMCI
1379
1380 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1381 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1382
1383 #ifndef PRODUCT
1384 if (PrintDeoptimizationDetails) {
1385 ttyLocker ttyl;
1386 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1387 fr.print_on(tty);
1388 tty->print_cr(" Virtual frames (innermost first):");
1389 for (int index = 0; index < chunk->length(); index++) {
1390 compiledVFrame* vf = chunk->at(index);
1391 tty->print(" %2d - ", index);
1392 vf->print_value();
1393 int bci = chunk->at(index)->raw_bci();
1394 const char* code_name;
1395 if (bci == SynchronizationEntryBCI) {
1396 code_name = "sync entry";
1397 } else {
1398 Bytecodes::Code code = vf->method()->code_at(bci);
1399 code_name = Bytecodes::name(code);
1400 }
1401 tty->print(" - %s", code_name);
1402 tty->print_cr(" @ bci %d ", bci);
1403 if (Verbose) {
1404 vf->print();
1405 tty->cr();
1406 }
1407 }
1408 }
1409 #endif
1410
1411 // Register map for next frame (used for stack crawl). We capture
1412 // the state of the deopt'ing frame's caller. Thus if we need to
1413 // stuff a C2I adapter we can properly fill in the callee-save
1414 // register locations.
1415 frame caller = fr.sender(reg_map);
1416 int frame_size = caller.sp() - fr.sp();
1417
1418 frame sender = caller;
1419
1420 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1421 // the vframeArray containing the unpacking information is allocated in the C heap.
1422 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1423 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1424
1425 // Compare the vframeArray to the collected vframes
1426 assert(array->structural_compare(thread, chunk), "just checking");
1427
1428 #ifndef PRODUCT
1429 if (PrintDeoptimizationDetails) {
1430 ttyLocker ttyl;
1431 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1432 }
1433 #endif // PRODUCT
1434
1435 return array;
1436 }
1437
1438 #if COMPILER2_OR_JVMCI
1439 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1440 // Reallocation of some scalar replaced objects failed. Record
1441 // that we need to pop all the interpreter frames for the
1442 // deoptimized compiled frame.
1443 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1444 thread->set_frames_to_pop_failed_realloc(array->frames());
1445 // Unlock all monitors here otherwise the interpreter will see a
1446 // mix of locked and unlocked monitors (because of failed
1447 // reallocations of synchronized objects) and be confused.
1448 for (int i = 0; i < array->frames(); i++) {
1449 MonitorChunk* monitors = array->element(i)->monitors();
1450 if (monitors != NULL) {
1451 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1452 BasicObjectLock* src = monitors->at(j);
1453 if (src->obj() != NULL) {
1454 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1455 }
1456 }
1457 array->element(i)->free_monitors(thread);
1458 #ifdef ASSERT
1459 array->element(i)->set_removed_monitors();
1460 #endif
1461 }
1462 }
1463 }
1464 #endif
1465
1466 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1467 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1468 Thread* thread = Thread::current();
1469 for (int i = 0; i < monitors->length(); i++) {
1470 MonitorInfo* mon_info = monitors->at(i);
1471 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1472 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1473 }
1474 }
1475 }
1476
1477
1478 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1479 if (!UseBiasedLocking) {
1480 return;
1481 }
1482
1483 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1484
1485 // Unfortunately we don't have a RegisterMap available in most of
1486 // the places we want to call this routine so we need to walk the
1487 // stack again to update the register map.
1488 if (map == NULL || !map->update_map()) {
1489 StackFrameStream sfs(thread, true);
1490 bool found = false;
1491 while (!found && !sfs.is_done()) {
1492 frame* cur = sfs.current();
1493 sfs.next();
1494 found = cur->id() == fr.id();
1495 }
1496 assert(found, "frame to be deoptimized not found on target thread's stack");
1497 map = sfs.register_map();
1498 }
1499
1500 vframe* vf = vframe::new_vframe(&fr, map, thread);
1501 compiledVFrame* cvf = compiledVFrame::cast(vf);
1502 // Revoke monitors' biases in all scopes
1503 while (!cvf->is_top()) {
1504 collect_monitors(cvf, objects_to_revoke);
1505 cvf = compiledVFrame::cast(cvf->sender());
1506 }
1507 collect_monitors(cvf, objects_to_revoke);
1508
1509 if (SafepointSynchronize::is_at_safepoint()) {
1510 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1511 } else {
1512 BiasedLocking::revoke(objects_to_revoke, thread);
1513 }
1514 }
1515
1516
1517 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1518 assert(fr.can_be_deoptimized(), "checking frame type");
1519
1520 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1521
1522 if (LogCompilation && xtty != NULL) {
1523 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1524 assert(cm != NULL, "only compiled methods can deopt");
1525
1526 ttyLocker ttyl;
1527 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1528 cm->log_identity(xtty);
1529 xtty->end_head();
1530 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1531 xtty->begin_elem("jvms bci='%d'", sd->bci());
1532 xtty->method(sd->method());
1533 xtty->end_elem();
1534 if (sd->is_top()) break;
1535 }
1536 xtty->tail("deoptimized");
1537 }
1538
1539 // Patch the compiled method so that when execution returns to it we will
1540 // deopt the execution state and return to the interpreter.
1541 fr.deoptimize(thread);
1542 }
1543
1544 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1545 deoptimize(thread, fr, map, Reason_constraint);
1546 }
1547
1548 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1549 // Deoptimize only if the frame comes from compile code.
1550 // Do not deoptimize the frame which is already patched
1551 // during the execution of the loops below.
1552 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1553 return;
1554 }
1555 ResourceMark rm;
1556 DeoptimizationMarker dm;
1557 if (UseBiasedLocking) {
1558 revoke_biases_of_monitors(thread, fr, map);
1559 }
1560 deoptimize_single_frame(thread, fr, reason);
1561
1562 }
1563
1564 #if INCLUDE_JVMCI
1565 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1566 // there is no exception handler for this pc => deoptimize
1567 cm->make_not_entrant();
1568
1569 // Use Deoptimization::deoptimize for all of its side-effects:
1570 // revoking biases of monitors, gathering traps statistics, logging...
1571 // it also patches the return pc but we do not care about that
1572 // since we return a continuation to the deopt_blob below.
1573 JavaThread* thread = JavaThread::current();
1574 RegisterMap reg_map(thread, UseBiasedLocking);
1575 frame runtime_frame = thread->last_frame();
1576 frame caller_frame = runtime_frame.sender(®_map);
1577 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1578 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1579
1580 MethodData* trap_mdo = get_method_data(thread, cm->method(), true);
1581 if (trap_mdo != NULL) {
1582 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1583 }
1584
1585 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1586 }
1587 #endif
1588
1589 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1590 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1591 "can only deoptimize other thread at a safepoint");
1592 // Compute frame and register map based on thread and sp.
1593 RegisterMap reg_map(thread, UseBiasedLocking);
1594 frame fr = thread->last_frame();
1595 while (fr.id() != id) {
1596 fr = fr.sender(®_map);
1597 }
1598 deoptimize(thread, fr, ®_map, reason);
1599 }
1600
1601
1602 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1603 if (thread == Thread::current()) {
1604 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1605 } else {
1606 VM_DeoptimizeFrame deopt(thread, id, reason);
1607 VMThread::execute(&deopt);
1608 }
1609 }
1610
1611 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1612 deoptimize_frame(thread, id, Reason_constraint);
1613 }
1614
1615 // JVMTI PopFrame support
1616 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1617 {
1618 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1619 }
1620 JRT_END
1621
1622 MethodData*
1623 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1624 bool create_if_missing) {
1625 Thread* THREAD = thread;
1626 MethodData* mdo = m()->method_data();
1627 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1628 // Build an MDO. Ignore errors like OutOfMemory;
1629 // that simply means we won't have an MDO to update.
1630 Method::build_interpreter_method_data(m, THREAD);
1631 if (HAS_PENDING_EXCEPTION) {
1632 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1633 CLEAR_PENDING_EXCEPTION;
1634 }
1635 mdo = m()->method_data();
1636 }
1637 return mdo;
1638 }
1639
1640 #if COMPILER2_OR_JVMCI
1641 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1642 // in case of an unresolved klass entry, load the class.
1643 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1644 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1645 return;
1646 }
1647
1648 if (!constant_pool->tag_at(index).is_symbol()) return;
1649
1650 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1651 Symbol* symbol = constant_pool->symbol_at(index);
1652
1653 // class name?
1654 if (symbol->char_at(0) != '(') {
1655 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1656 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1657 return;
1658 }
1659
1660 // then it must be a signature!
1661 ResourceMark rm(THREAD);
1662 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1663 if (ss.is_object()) {
1664 Symbol* class_name = ss.as_symbol();
1665 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1666 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1667 }
1668 }
1669 }
1670
1671
1672 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1673 EXCEPTION_MARK;
1674 load_class_by_index(constant_pool, index, THREAD);
1675 if (HAS_PENDING_EXCEPTION) {
1676 // Exception happened during classloading. We ignore the exception here, since it
1677 // is going to be rethrown since the current activation is going to be deoptimized and
1678 // the interpreter will re-execute the bytecode.
1679 CLEAR_PENDING_EXCEPTION;
1680 // Class loading called java code which may have caused a stack
1681 // overflow. If the exception was thrown right before the return
1682 // to the runtime the stack is no longer guarded. Reguard the
1683 // stack otherwise if we return to the uncommon trap blob and the
1684 // stack bang causes a stack overflow we crash.
1685 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1686 JavaThread* thread = (JavaThread*)THREAD;
1687 bool guard_pages_enabled = thread->stack_guards_enabled();
1688 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1689 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1690 }
1691 }
1692
1693 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1694 HandleMark hm;
1695
1696 // uncommon_trap() is called at the beginning of the uncommon trap
1697 // handler. Note this fact before we start generating temporary frames
1698 // that can confuse an asynchronous stack walker. This counter is
1699 // decremented at the end of unpack_frames().
1700 thread->inc_in_deopt_handler();
1701
1702 // We need to update the map if we have biased locking.
1703 #if INCLUDE_JVMCI
1704 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1705 RegisterMap reg_map(thread, true);
1706 #else
1707 RegisterMap reg_map(thread, UseBiasedLocking);
1708 #endif
1709 frame stub_frame = thread->last_frame();
1710 frame fr = stub_frame.sender(®_map);
1711 // Make sure the calling nmethod is not getting deoptimized and removed
1712 // before we are done with it.
1713 nmethodLocker nl(fr.pc());
1714
1715 // Log a message
1716 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1717 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1718
1719 {
1720 ResourceMark rm;
1721
1722 // Revoke biases of any monitors in the frame to ensure we can migrate them
1723 revoke_biases_of_monitors(thread, fr, ®_map);
1724
1725 DeoptReason reason = trap_request_reason(trap_request);
1726 DeoptAction action = trap_request_action(trap_request);
1727 #if INCLUDE_JVMCI
1728 int debug_id = trap_request_debug_id(trap_request);
1729 #endif
1730 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1731
1732 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1733 compiledVFrame* cvf = compiledVFrame::cast(vf);
1734
1735 CompiledMethod* nm = cvf->code();
1736
1737 ScopeDesc* trap_scope = cvf->scope();
1738
1739 if (TraceDeoptimization) {
1740 ttyLocker ttyl;
1741 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()
1742 #if INCLUDE_JVMCI
1743 , debug_id
1744 #endif
1745 );
1746 }
1747
1748 methodHandle trap_method = trap_scope->method();
1749 int trap_bci = trap_scope->bci();
1750 #if INCLUDE_JVMCI
1751 jlong speculation = thread->pending_failed_speculation();
1752 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1753 nm->as_nmethod()->update_speculation(thread);
1754 } else {
1755 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1756 }
1757
1758 if (trap_bci == SynchronizationEntryBCI) {
1759 trap_bci = 0;
1760 thread->set_pending_monitorenter(true);
1761 }
1762
1763 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1764 thread->set_pending_transfer_to_interpreter(true);
1765 }
1766 #endif
1767
1768 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1769 // Record this event in the histogram.
1770 gather_statistics(reason, action, trap_bc);
1771
1772 // Ensure that we can record deopt. history:
1773 // Need MDO to record RTM code generation state.
1774 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1775
1776 methodHandle profiled_method;
1777 #if INCLUDE_JVMCI
1778 if (nm->is_compiled_by_jvmci()) {
1779 profiled_method = nm->method();
1780 } else {
1781 profiled_method = trap_method;
1782 }
1783 #else
1784 profiled_method = trap_method;
1785 #endif
1786
1787 MethodData* trap_mdo =
1788 get_method_data(thread, profiled_method, create_if_missing);
1789
1790 // Log a message
1791 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1792 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1793 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1794
1795 // Print a bunch of diagnostics, if requested.
1796 if (TraceDeoptimization || LogCompilation) {
1797 ResourceMark rm;
1798 ttyLocker ttyl;
1799 char buf[100];
1800 if (xtty != NULL) {
1801 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1802 os::current_thread_id(),
1803 format_trap_request(buf, sizeof(buf), trap_request));
1804 #if INCLUDE_JVMCI
1805 if (speculation != 0) {
1806 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1807 }
1808 #endif
1809 nm->log_identity(xtty);
1810 }
1811 Symbol* class_name = NULL;
1812 bool unresolved = false;
1813 if (unloaded_class_index >= 0) {
1814 constantPoolHandle constants (THREAD, trap_method->constants());
1815 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1816 class_name = constants->klass_name_at(unloaded_class_index);
1817 unresolved = true;
1818 if (xtty != NULL)
1819 xtty->print(" unresolved='1'");
1820 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1821 class_name = constants->symbol_at(unloaded_class_index);
1822 }
1823 if (xtty != NULL)
1824 xtty->name(class_name);
1825 }
1826 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1827 // Dump the relevant MDO state.
1828 // This is the deopt count for the current reason, any previous
1829 // reasons or recompiles seen at this point.
1830 int dcnt = trap_mdo->trap_count(reason);
1831 if (dcnt != 0)
1832 xtty->print(" count='%d'", dcnt);
1833 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1834 int dos = (pdata == NULL)? 0: pdata->trap_state();
1835 if (dos != 0) {
1836 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1837 if (trap_state_is_recompiled(dos)) {
1838 int recnt2 = trap_mdo->overflow_recompile_count();
1839 if (recnt2 != 0)
1840 xtty->print(" recompiles2='%d'", recnt2);
1841 }
1842 }
1843 }
1844 if (xtty != NULL) {
1845 xtty->stamp();
1846 xtty->end_head();
1847 }
1848 if (TraceDeoptimization) { // make noise on the tty
1849 tty->print("Uncommon trap occurred in");
1850 nm->method()->print_short_name(tty);
1851 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1852 #if INCLUDE_JVMCI
1853 if (nm->is_nmethod()) {
1854 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
1855 if (installed_code_name != NULL) {
1856 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1857 }
1858 }
1859 #endif
1860 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1861 p2i(fr.pc()),
1862 os::current_thread_id(),
1863 trap_reason_name(reason),
1864 trap_action_name(action),
1865 unloaded_class_index
1866 #if INCLUDE_JVMCI
1867 , debug_id
1868 #endif
1869 );
1870 if (class_name != NULL) {
1871 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1872 class_name->print_symbol_on(tty);
1873 }
1874 tty->cr();
1875 }
1876 if (xtty != NULL) {
1877 // Log the precise location of the trap.
1878 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1879 xtty->begin_elem("jvms bci='%d'", sd->bci());
1880 xtty->method(sd->method());
1881 xtty->end_elem();
1882 if (sd->is_top()) break;
1883 }
1884 xtty->tail("uncommon_trap");
1885 }
1886 }
1887 // (End diagnostic printout.)
1888
1889 // Load class if necessary
1890 if (unloaded_class_index >= 0) {
1891 constantPoolHandle constants(THREAD, trap_method->constants());
1892 load_class_by_index(constants, unloaded_class_index);
1893 }
1894
1895 // Flush the nmethod if necessary and desirable.
1896 //
1897 // We need to avoid situations where we are re-flushing the nmethod
1898 // because of a hot deoptimization site. Repeated flushes at the same
1899 // point need to be detected by the compiler and avoided. If the compiler
1900 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1901 // module must take measures to avoid an infinite cycle of recompilation
1902 // and deoptimization. There are several such measures:
1903 //
1904 // 1. If a recompilation is ordered a second time at some site X
1905 // and for the same reason R, the action is adjusted to 'reinterpret',
1906 // to give the interpreter time to exercise the method more thoroughly.
1907 // If this happens, the method's overflow_recompile_count is incremented.
1908 //
1909 // 2. If the compiler fails to reduce the deoptimization rate, then
1910 // the method's overflow_recompile_count will begin to exceed the set
1911 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1912 // is adjusted to 'make_not_compilable', and the method is abandoned
1913 // to the interpreter. This is a performance hit for hot methods,
1914 // but is better than a disastrous infinite cycle of recompilations.
1915 // (Actually, only the method containing the site X is abandoned.)
1916 //
1917 // 3. In parallel with the previous measures, if the total number of
1918 // recompilations of a method exceeds the much larger set limit
1919 // PerMethodRecompilationCutoff, the method is abandoned.
1920 // This should only happen if the method is very large and has
1921 // many "lukewarm" deoptimizations. The code which enforces this
1922 // limit is elsewhere (class nmethod, class Method).
1923 //
1924 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1925 // to recompile at each bytecode independently of the per-BCI cutoff.
1926 //
1927 // The decision to update code is up to the compiler, and is encoded
1928 // in the Action_xxx code. If the compiler requests Action_none
1929 // no trap state is changed, no compiled code is changed, and the
1930 // computation suffers along in the interpreter.
1931 //
1932 // The other action codes specify various tactics for decompilation
1933 // and recompilation. Action_maybe_recompile is the loosest, and
1934 // allows the compiled code to stay around until enough traps are seen,
1935 // and until the compiler gets around to recompiling the trapping method.
1936 //
1937 // The other actions cause immediate removal of the present code.
1938
1939 // Traps caused by injected profile shouldn't pollute trap counts.
1940 bool injected_profile_trap = trap_method->has_injected_profile() &&
1941 (reason == Reason_intrinsic || reason == Reason_unreached);
1942
1943 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1944 bool make_not_entrant = false;
1945 bool make_not_compilable = false;
1946 bool reprofile = false;
1947 switch (action) {
1948 case Action_none:
1949 // Keep the old code.
1950 update_trap_state = false;
1951 break;
1952 case Action_maybe_recompile:
1953 // Do not need to invalidate the present code, but we can
1954 // initiate another
1955 // Start compiler without (necessarily) invalidating the nmethod.
1956 // The system will tolerate the old code, but new code should be
1957 // generated when possible.
1958 break;
1959 case Action_reinterpret:
1960 // Go back into the interpreter for a while, and then consider
1961 // recompiling form scratch.
1962 make_not_entrant = true;
1963 // Reset invocation counter for outer most method.
1964 // This will allow the interpreter to exercise the bytecodes
1965 // for a while before recompiling.
1966 // By contrast, Action_make_not_entrant is immediate.
1967 //
1968 // Note that the compiler will track null_check, null_assert,
1969 // range_check, and class_check events and log them as if they
1970 // had been traps taken from compiled code. This will update
1971 // the MDO trap history so that the next compilation will
1972 // properly detect hot trap sites.
1973 reprofile = true;
1974 break;
1975 case Action_make_not_entrant:
1976 // Request immediate recompilation, and get rid of the old code.
1977 // Make them not entrant, so next time they are called they get
1978 // recompiled. Unloaded classes are loaded now so recompile before next
1979 // time they are called. Same for uninitialized. The interpreter will
1980 // link the missing class, if any.
1981 make_not_entrant = true;
1982 break;
1983 case Action_make_not_compilable:
1984 // Give up on compiling this method at all.
1985 make_not_entrant = true;
1986 make_not_compilable = true;
1987 break;
1988 default:
1989 ShouldNotReachHere();
1990 }
1991
1992 // Setting +ProfileTraps fixes the following, on all platforms:
1993 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1994 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1995 // recompile relies on a MethodData* to record heroic opt failures.
1996
1997 // Whether the interpreter is producing MDO data or not, we also need
1998 // to use the MDO to detect hot deoptimization points and control
1999 // aggressive optimization.
2000 bool inc_recompile_count = false;
2001 ProfileData* pdata = NULL;
2002 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
2003 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
2004 uint this_trap_count = 0;
2005 bool maybe_prior_trap = false;
2006 bool maybe_prior_recompile = false;
2007 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2008 #if INCLUDE_JVMCI
2009 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2010 #endif
2011 nm->method(),
2012 //outputs:
2013 this_trap_count,
2014 maybe_prior_trap,
2015 maybe_prior_recompile);
2016 // Because the interpreter also counts null, div0, range, and class
2017 // checks, these traps from compiled code are double-counted.
2018 // This is harmless; it just means that the PerXTrapLimit values
2019 // are in effect a little smaller than they look.
2020
2021 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2022 if (per_bc_reason != Reason_none) {
2023 // Now take action based on the partially known per-BCI history.
2024 if (maybe_prior_trap
2025 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2026 // If there are too many traps at this BCI, force a recompile.
2027 // This will allow the compiler to see the limit overflow, and
2028 // take corrective action, if possible. The compiler generally
2029 // does not use the exact PerBytecodeTrapLimit value, but instead
2030 // changes its tactics if it sees any traps at all. This provides
2031 // a little hysteresis, delaying a recompile until a trap happens
2032 // several times.
2033 //
2034 // Actually, since there is only one bit of counter per BCI,
2035 // the possible per-BCI counts are {0,1,(per-method count)}.
2036 // This produces accurate results if in fact there is only
2037 // one hot trap site, but begins to get fuzzy if there are
2038 // many sites. For example, if there are ten sites each
2039 // trapping two or more times, they each get the blame for
2040 // all of their traps.
2041 make_not_entrant = true;
2042 }
2043
2044 // Detect repeated recompilation at the same BCI, and enforce a limit.
2045 if (make_not_entrant && maybe_prior_recompile) {
2046 // More than one recompile at this point.
2047 inc_recompile_count = maybe_prior_trap;
2048 }
2049 } else {
2050 // For reasons which are not recorded per-bytecode, we simply
2051 // force recompiles unconditionally.
2052 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2053 make_not_entrant = true;
2054 }
2055
2056 // Go back to the compiler if there are too many traps in this method.
2057 if (this_trap_count >= per_method_trap_limit(reason)) {
2058 // If there are too many traps in this method, force a recompile.
2059 // This will allow the compiler to see the limit overflow, and
2060 // take corrective action, if possible.
2061 // (This condition is an unlikely backstop only, because the
2062 // PerBytecodeTrapLimit is more likely to take effect first,
2063 // if it is applicable.)
2064 make_not_entrant = true;
2065 }
2066
2067 // Here's more hysteresis: If there has been a recompile at
2068 // this trap point already, run the method in the interpreter
2069 // for a while to exercise it more thoroughly.
2070 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2071 reprofile = true;
2072 }
2073 }
2074
2075 // Take requested actions on the method:
2076
2077 // Recompile
2078 if (make_not_entrant) {
2079 if (!nm->make_not_entrant()) {
2080 return; // the call did not change nmethod's state
2081 }
2082
2083 if (pdata != NULL) {
2084 // Record the recompilation event, if any.
2085 int tstate0 = pdata->trap_state();
2086 int tstate1 = trap_state_set_recompiled(tstate0, true);
2087 if (tstate1 != tstate0)
2088 pdata->set_trap_state(tstate1);
2089 }
2090
2091 #if INCLUDE_RTM_OPT
2092 // Restart collecting RTM locking abort statistic if the method
2093 // is recompiled for a reason other than RTM state change.
2094 // Assume that in new recompiled code the statistic could be different,
2095 // for example, due to different inlining.
2096 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2097 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2098 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2099 }
2100 #endif
2101 // For code aging we count traps separately here, using make_not_entrant()
2102 // as a guard against simultaneous deopts in multiple threads.
2103 if (reason == Reason_tenured && trap_mdo != NULL) {
2104 trap_mdo->inc_tenure_traps();
2105 }
2106 }
2107
2108 if (inc_recompile_count) {
2109 trap_mdo->inc_overflow_recompile_count();
2110 if ((uint)trap_mdo->overflow_recompile_count() >
2111 (uint)PerBytecodeRecompilationCutoff) {
2112 // Give up on the method containing the bad BCI.
2113 if (trap_method() == nm->method()) {
2114 make_not_compilable = true;
2115 } else {
2116 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2117 // But give grace to the enclosing nm->method().
2118 }
2119 }
2120 }
2121
2122 // Reprofile
2123 if (reprofile) {
2124 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2125 }
2126
2127 // Give up compiling
2128 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2129 assert(make_not_entrant, "consistent");
2130 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2131 }
2132
2133 } // Free marked resources
2134
2135 }
2136 JRT_END
2137
2138 ProfileData*
2139 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2140 int trap_bci,
2141 Deoptimization::DeoptReason reason,
2142 bool update_total_trap_count,
2143 #if INCLUDE_JVMCI
2144 bool is_osr,
2145 #endif
2146 Method* compiled_method,
2147 //outputs:
2148 uint& ret_this_trap_count,
2149 bool& ret_maybe_prior_trap,
2150 bool& ret_maybe_prior_recompile) {
2151 bool maybe_prior_trap = false;
2152 bool maybe_prior_recompile = false;
2153 uint this_trap_count = 0;
2154 if (update_total_trap_count) {
2155 uint idx = reason;
2156 #if INCLUDE_JVMCI
2157 if (is_osr) {
2158 idx += Reason_LIMIT;
2159 }
2160 #endif
2161 uint prior_trap_count = trap_mdo->trap_count(idx);
2162 this_trap_count = trap_mdo->inc_trap_count(idx);
2163
2164 // If the runtime cannot find a place to store trap history,
2165 // it is estimated based on the general condition of the method.
2166 // If the method has ever been recompiled, or has ever incurred
2167 // a trap with the present reason , then this BCI is assumed
2168 // (pessimistically) to be the culprit.
2169 maybe_prior_trap = (prior_trap_count != 0);
2170 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2171 }
2172 ProfileData* pdata = NULL;
2173
2174
2175 // For reasons which are recorded per bytecode, we check per-BCI data.
2176 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2177 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2178 if (per_bc_reason != Reason_none) {
2179 // Find the profile data for this BCI. If there isn't one,
2180 // try to allocate one from the MDO's set of spares.
2181 // This will let us detect a repeated trap at this point.
2182 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2183
2184 if (pdata != NULL) {
2185 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2186 if (LogCompilation && xtty != NULL) {
2187 ttyLocker ttyl;
2188 // no more room for speculative traps in this MDO
2189 xtty->elem("speculative_traps_oom");
2190 }
2191 }
2192 // Query the trap state of this profile datum.
2193 int tstate0 = pdata->trap_state();
2194 if (!trap_state_has_reason(tstate0, per_bc_reason))
2195 maybe_prior_trap = false;
2196 if (!trap_state_is_recompiled(tstate0))
2197 maybe_prior_recompile = false;
2198
2199 // Update the trap state of this profile datum.
2200 int tstate1 = tstate0;
2201 // Record the reason.
2202 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2203 // Store the updated state on the MDO, for next time.
2204 if (tstate1 != tstate0)
2205 pdata->set_trap_state(tstate1);
2206 } else {
2207 if (LogCompilation && xtty != NULL) {
2208 ttyLocker ttyl;
2209 // Missing MDP? Leave a small complaint in the log.
2210 xtty->elem("missing_mdp bci='%d'", trap_bci);
2211 }
2212 }
2213 }
2214
2215 // Return results:
2216 ret_this_trap_count = this_trap_count;
2217 ret_maybe_prior_trap = maybe_prior_trap;
2218 ret_maybe_prior_recompile = maybe_prior_recompile;
2219 return pdata;
2220 }
2221
2222 void
2223 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2224 ResourceMark rm;
2225 // Ignored outputs:
2226 uint ignore_this_trap_count;
2227 bool ignore_maybe_prior_trap;
2228 bool ignore_maybe_prior_recompile;
2229 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2230 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2231 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2232 query_update_method_data(trap_mdo, trap_bci,
2233 (DeoptReason)reason,
2234 update_total_counts,
2235 #if INCLUDE_JVMCI
2236 false,
2237 #endif
2238 NULL,
2239 ignore_this_trap_count,
2240 ignore_maybe_prior_trap,
2241 ignore_maybe_prior_recompile);
2242 }
2243
2244 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2245 if (TraceDeoptimization) {
2246 tty->print("Uncommon trap ");
2247 }
2248 // Still in Java no safepoints
2249 {
2250 // This enters VM and may safepoint
2251 uncommon_trap_inner(thread, trap_request);
2252 }
2253 return fetch_unroll_info_helper(thread, exec_mode);
2254 }
2255
2256 // Local derived constants.
2257 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2258 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2259 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2260
2261 //---------------------------trap_state_reason---------------------------------
2262 Deoptimization::DeoptReason
2263 Deoptimization::trap_state_reason(int trap_state) {
2264 // This assert provides the link between the width of DataLayout::trap_bits
2265 // and the encoding of "recorded" reasons. It ensures there are enough
2266 // bits to store all needed reasons in the per-BCI MDO profile.
2267 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2268 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2269 trap_state -= recompile_bit;
2270 if (trap_state == DS_REASON_MASK) {
2271 return Reason_many;
2272 } else {
2273 assert((int)Reason_none == 0, "state=0 => Reason_none");
2274 return (DeoptReason)trap_state;
2275 }
2276 }
2277 //-------------------------trap_state_has_reason-------------------------------
2278 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2279 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2280 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2281 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2282 trap_state -= recompile_bit;
2283 if (trap_state == DS_REASON_MASK) {
2284 return -1; // true, unspecifically (bottom of state lattice)
2285 } else if (trap_state == reason) {
2286 return 1; // true, definitely
2287 } else if (trap_state == 0) {
2288 return 0; // false, definitely (top of state lattice)
2289 } else {
2290 return 0; // false, definitely
2291 }
2292 }
2293 //-------------------------trap_state_add_reason-------------------------------
2294 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2295 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2296 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2297 trap_state -= recompile_bit;
2298 if (trap_state == DS_REASON_MASK) {
2299 return trap_state + recompile_bit; // already at state lattice bottom
2300 } else if (trap_state == reason) {
2301 return trap_state + recompile_bit; // the condition is already true
2302 } else if (trap_state == 0) {
2303 return reason + recompile_bit; // no condition has yet been true
2304 } else {
2305 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2306 }
2307 }
2308 //-----------------------trap_state_is_recompiled------------------------------
2309 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2310 return (trap_state & DS_RECOMPILE_BIT) != 0;
2311 }
2312 //-----------------------trap_state_set_recompiled-----------------------------
2313 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2314 if (z) return trap_state | DS_RECOMPILE_BIT;
2315 else return trap_state & ~DS_RECOMPILE_BIT;
2316 }
2317 //---------------------------format_trap_state---------------------------------
2318 // This is used for debugging and diagnostics, including LogFile output.
2319 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2320 int trap_state) {
2321 assert(buflen > 0, "sanity");
2322 DeoptReason reason = trap_state_reason(trap_state);
2323 bool recomp_flag = trap_state_is_recompiled(trap_state);
2324 // Re-encode the state from its decoded components.
2325 int decoded_state = 0;
2326 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2327 decoded_state = trap_state_add_reason(decoded_state, reason);
2328 if (recomp_flag)
2329 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2330 // If the state re-encodes properly, format it symbolically.
2331 // Because this routine is used for debugging and diagnostics,
2332 // be robust even if the state is a strange value.
2333 size_t len;
2334 if (decoded_state != trap_state) {
2335 // Random buggy state that doesn't decode??
2336 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2337 } else {
2338 len = jio_snprintf(buf, buflen, "%s%s",
2339 trap_reason_name(reason),
2340 recomp_flag ? " recompiled" : "");
2341 }
2342 return buf;
2343 }
2344
2345
2346 //--------------------------------statics--------------------------------------
2347 const char* Deoptimization::_trap_reason_name[] = {
2348 // Note: Keep this in sync. with enum DeoptReason.
2349 "none",
2350 "null_check",
2351 "null_assert" JVMCI_ONLY("_or_unreached0"),
2352 "range_check",
2353 "class_check",
2354 "array_check",
2355 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2356 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2357 "profile_predicate",
2358 "unloaded",
2359 "uninitialized",
2360 "initialized",
2361 "unreached",
2362 "unhandled",
2363 "constraint",
2364 "div0_check",
2365 "age",
2366 "predicate",
2367 "loop_limit_check",
2368 "speculate_class_check",
2369 "speculate_null_check",
2370 "speculate_null_assert",
2371 "rtm_state_change",
2372 "unstable_if",
2373 "unstable_fused_if",
2374 #if INCLUDE_JVMCI
2375 "aliasing",
2376 "transfer_to_interpreter",
2377 "not_compiled_exception_handler",
2378 "unresolved",
2379 "jsr_mismatch",
2380 #endif
2381 "tenured"
2382 };
2383 const char* Deoptimization::_trap_action_name[] = {
2384 // Note: Keep this in sync. with enum DeoptAction.
2385 "none",
2386 "maybe_recompile",
2387 "reinterpret",
2388 "make_not_entrant",
2389 "make_not_compilable"
2390 };
2391
2392 const char* Deoptimization::trap_reason_name(int reason) {
2393 // Check that every reason has a name
2394 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2395
2396 if (reason == Reason_many) return "many";
2397 if ((uint)reason < Reason_LIMIT)
2398 return _trap_reason_name[reason];
2399 static char buf[20];
2400 sprintf(buf, "reason%d", reason);
2401 return buf;
2402 }
2403 const char* Deoptimization::trap_action_name(int action) {
2404 // Check that every action has a name
2405 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2406
2407 if ((uint)action < Action_LIMIT)
2408 return _trap_action_name[action];
2409 static char buf[20];
2410 sprintf(buf, "action%d", action);
2411 return buf;
2412 }
2413
2414 // This is used for debugging and diagnostics, including LogFile output.
2415 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2416 int trap_request) {
2417 jint unloaded_class_index = trap_request_index(trap_request);
2418 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2419 const char* action = trap_action_name(trap_request_action(trap_request));
2420 #if INCLUDE_JVMCI
2421 int debug_id = trap_request_debug_id(trap_request);
2422 #endif
2423 size_t len;
2424 if (unloaded_class_index < 0) {
2425 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2426 reason, action
2427 #if INCLUDE_JVMCI
2428 ,debug_id
2429 #endif
2430 );
2431 } else {
2432 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2433 reason, action, unloaded_class_index
2434 #if INCLUDE_JVMCI
2435 ,debug_id
2436 #endif
2437 );
2438 }
2439 return buf;
2440 }
2441
2442 juint Deoptimization::_deoptimization_hist
2443 [Deoptimization::Reason_LIMIT]
2444 [1 + Deoptimization::Action_LIMIT]
2445 [Deoptimization::BC_CASE_LIMIT]
2446 = {0};
2447
2448 enum {
2449 LSB_BITS = 8,
2450 LSB_MASK = right_n_bits(LSB_BITS)
2451 };
2452
2453 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2454 Bytecodes::Code bc) {
2455 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2456 assert(action >= 0 && action < Action_LIMIT, "oob");
2457 _deoptimization_hist[Reason_none][0][0] += 1; // total
2458 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2459 juint* cases = _deoptimization_hist[reason][1+action];
2460 juint* bc_counter_addr = NULL;
2461 juint bc_counter = 0;
2462 // Look for an unused counter, or an exact match to this BC.
2463 if (bc != Bytecodes::_illegal) {
2464 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2465 juint* counter_addr = &cases[bc_case];
2466 juint counter = *counter_addr;
2467 if ((counter == 0 && bc_counter_addr == NULL)
2468 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2469 // this counter is either free or is already devoted to this BC
2470 bc_counter_addr = counter_addr;
2471 bc_counter = counter | bc;
2472 }
2473 }
2474 }
2475 if (bc_counter_addr == NULL) {
2476 // Overflow, or no given bytecode.
2477 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2478 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2479 }
2480 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2481 }
2482
2483 jint Deoptimization::total_deoptimization_count() {
2484 return _deoptimization_hist[Reason_none][0][0];
2485 }
2486
2487 void Deoptimization::print_statistics() {
2488 juint total = total_deoptimization_count();
2489 juint account = total;
2490 if (total != 0) {
2491 ttyLocker ttyl;
2492 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2493 tty->print_cr("Deoptimization traps recorded:");
2494 #define PRINT_STAT_LINE(name, r) \
2495 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2496 PRINT_STAT_LINE("total", total);
2497 // For each non-zero entry in the histogram, print the reason,
2498 // the action, and (if specifically known) the type of bytecode.
2499 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2500 for (int action = 0; action < Action_LIMIT; action++) {
2501 juint* cases = _deoptimization_hist[reason][1+action];
2502 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2503 juint counter = cases[bc_case];
2504 if (counter != 0) {
2505 char name[1*K];
2506 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2507 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2508 bc = Bytecodes::_illegal;
2509 sprintf(name, "%s/%s/%s",
2510 trap_reason_name(reason),
2511 trap_action_name(action),
2512 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2513 juint r = counter >> LSB_BITS;
2514 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2515 account -= r;
2516 }
2517 }
2518 }
2519 }
2520 if (account != 0) {
2521 PRINT_STAT_LINE("unaccounted", account);
2522 }
2523 #undef PRINT_STAT_LINE
2524 if (xtty != NULL) xtty->tail("statistics");
2525 }
2526 }
2527 #else // COMPILER2_OR_JVMCI
2528
2529
2530 // Stubs for C1 only system.
2531 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2532 return false;
2533 }
2534
2535 const char* Deoptimization::trap_reason_name(int reason) {
2536 return "unknown";
2537 }
2538
2539 void Deoptimization::print_statistics() {
2540 // no output
2541 }
2542
2543 void
2544 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2545 // no udpate
2546 }
2547
2548 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2549 return 0;
2550 }
2551
2552 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2553 Bytecodes::Code bc) {
2554 // no update
2555 }
2556
2557 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2558 int trap_state) {
2559 jio_snprintf(buf, buflen, "#%d", trap_state);
2560 return buf;
2561 }
2562
2563 #endif // COMPILER2_OR_JVMCI