1
2
3 /*
4 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "precompiled.hpp"
28 #include "jvm.h"
29 #include "classfile/javaClasses.inline.hpp"
30 #include "classfile/symbolTable.hpp"
31 #include "classfile/systemDictionary.hpp"
32 #include "code/codeCache.hpp"
33 #include "code/debugInfoRec.hpp"
34 #include "code/nmethod.hpp"
35 #include "code/pcDesc.hpp"
36 #include "code/scopeDesc.hpp"
37 #include "compiler/compilationPolicy.hpp"
38 #include "interpreter/bytecode.hpp"
39 #include "interpreter/interpreter.hpp"
40 #include "interpreter/oopMapCache.hpp"
41 #include "memory/allocation.inline.hpp"
42 #include "memory/oopFactory.hpp"
43 #include "memory/resourceArea.hpp"
44 #include "memory/universe.hpp"
45 #include "oops/constantPool.hpp"
46 #include "oops/method.hpp"
47 #include "oops/objArrayKlass.hpp"
48 #include "oops/objArrayOop.inline.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "oops/fieldStreams.inline.hpp"
51 #include "oops/typeArrayOop.inline.hpp"
52 #include "oops/verifyOopClosure.hpp"
53 #include "prims/jvmtiThreadState.hpp"
54 #include "runtime/atomic.hpp"
55 #include "runtime/biasedLocking.hpp"
56 #include "runtime/deoptimization.hpp"
57 #include "runtime/fieldDescriptor.hpp"
58 #include "runtime/fieldDescriptor.inline.hpp"
59 #include "runtime/frame.inline.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/interfaceSupport.inline.hpp"
62 #include "runtime/jniHandles.inline.hpp"
63 #include "runtime/safepointVerifiers.hpp"
64 #include "runtime/sharedRuntime.hpp"
65 #include "runtime/signature.hpp"
66 #include "runtime/stubRoutines.hpp"
67 #include "runtime/thread.hpp"
68 #include "runtime/threadSMR.hpp"
69 #include "runtime/vframe.hpp"
70 #include "runtime/vframeArray.hpp"
71 #include "runtime/vframe_hp.hpp"
72 #include "utilities/events.hpp"
73 #include "utilities/macros.hpp"
74 #include "utilities/preserveException.hpp"
75 #include "utilities/xmlstream.hpp"
76 #if INCLUDE_JFR
77 #include "jfr/jfrEvents.hpp"
78 #include "jfr/metadata/jfrSerializer.hpp"
79 #endif
80
81 bool DeoptimizationMarker::_is_active = false;
82
83 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
84 int caller_adjustment,
85 int caller_actual_parameters,
86 int number_of_frames,
87 intptr_t* frame_sizes,
88 address* frame_pcs,
89 BasicType return_type,
90 int exec_mode) {
91 _size_of_deoptimized_frame = size_of_deoptimized_frame;
92 _caller_adjustment = caller_adjustment;
93 _caller_actual_parameters = caller_actual_parameters;
94 _number_of_frames = number_of_frames;
95 _frame_sizes = frame_sizes;
96 _frame_pcs = frame_pcs;
97 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
98 _return_type = return_type;
99 _initial_info = 0;
100 // PD (x86 only)
101 _counter_temp = 0;
102 _unpack_kind = exec_mode;
103 _sender_sp_temp = 0;
104
105 _total_frame_sizes = size_of_frames();
106 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
107 }
108
109
110 Deoptimization::UnrollBlock::~UnrollBlock() {
111 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
112 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
113 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
114 }
115
116
117 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
118 assert(register_number < RegisterMap::reg_count, "checking register number");
119 return &_register_block[register_number * 2];
120 }
121
122
123
124 int Deoptimization::UnrollBlock::size_of_frames() const {
125 // Acount first for the adjustment of the initial frame
126 int result = _caller_adjustment;
127 for (int index = 0; index < number_of_frames(); index++) {
128 result += frame_sizes()[index];
129 }
130 return result;
131 }
132
133
134 void Deoptimization::UnrollBlock::print() {
135 ttyLocker ttyl;
136 tty->print_cr("UnrollBlock");
137 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
138 tty->print( " frame_sizes: ");
139 for (int index = 0; index < number_of_frames(); index++) {
140 tty->print(INTX_FORMAT " ", frame_sizes()[index]);
141 }
142 tty->cr();
143 }
144
145
146 // In order to make fetch_unroll_info work properly with escape
147 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
148 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
149 // of previously eliminated objects occurs in realloc_objects, which is
150 // called from the method fetch_unroll_info_helper below.
151 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
152 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
153 // but makes the entry a little slower. There is however a little dance we have to
154 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
155
156 // fetch_unroll_info() is called at the beginning of the deoptimization
157 // handler. Note this fact before we start generating temporary frames
158 // that can confuse an asynchronous stack walker. This counter is
159 // decremented at the end of unpack_frames().
160 if (TraceDeoptimization) {
161 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
162 }
163 thread->inc_in_deopt_handler();
164
165 return fetch_unroll_info_helper(thread, exec_mode);
166 JRT_END
167
168 #if COMPILER2_OR_JVMCI
169 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method,
170 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk) {
171 bool realloc_failures = false;
172 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
173
174 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
175
176 // The flag return_oop() indicates call sites which return oop
177 // in compiled code. Such sites include java method calls,
178 // runtime calls (for example, used to allocate new objects/arrays
179 // on slow code path) and any other calls generated in compiled code.
180 // It is not guaranteed that we can get such information here only
181 // by analyzing bytecode in deoptimized frames. This is why this flag
182 // is set during method compilation (see Compile::Process_OopMap_Node()).
183 // If the previous frame was popped or if we are dispatching an exception,
184 // we don't have an oop result.
185 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
186 Handle return_value;
187 if (save_oop_result) {
188 // Reallocation may trigger GC. If deoptimization happened on return from
189 // call which returns oop we need to save it since it is not in oopmap.
190 oop result = deoptee.saved_oop_result(&map);
191 assert(oopDesc::is_oop_or_null(result), "must be oop");
192 return_value = Handle(thread, result);
193 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
194 if (TraceDeoptimization) {
195 ttyLocker ttyl;
196 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
197 }
198 }
199 if (objects != NULL) {
200 JRT_BLOCK
201 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
202 JRT_END
203 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci();
204 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
205 #ifndef PRODUCT
206 if (TraceDeoptimization) {
207 ttyLocker ttyl;
208 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
209 Deoptimization::print_objects(objects, realloc_failures);
210 }
211 #endif
212 }
213 if (save_oop_result) {
214 // Restore result.
215 deoptee.set_saved_oop_result(&map, return_value());
216 }
217 return realloc_failures;
218 }
219
220 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
221 #ifndef PRODUCT
222 bool first = true;
223 #endif
224 for (int i = 0; i < chunk->length(); i++) {
225 compiledVFrame* cvf = chunk->at(i);
226 assert (cvf->scope() != NULL,"expect only compiled java frames");
227 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
228 if (monitors->is_nonempty()) {
229 Deoptimization::relock_objects(monitors, thread, realloc_failures);
230 #ifndef PRODUCT
231 if (PrintDeoptimizationDetails) {
232 ttyLocker ttyl;
233 for (int j = 0; j < monitors->length(); j++) {
234 MonitorInfo* mi = monitors->at(j);
235 if (mi->eliminated()) {
236 if (first) {
237 first = false;
238 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
239 }
240 if (mi->owner_is_scalar_replaced()) {
241 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
242 tty->print_cr(" failed reallocation for klass %s", k->external_name());
243 } else {
244 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
245 }
246 }
247 }
248 }
249 #endif // !PRODUCT
250 }
251 }
252 }
253 #endif // COMPILER2_OR_JVMCI
254
255 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
256 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
257
258 // Note: there is a safepoint safety issue here. No matter whether we enter
259 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
260 // the vframeArray is created.
261 //
262
263 // Allocate our special deoptimization ResourceMark
264 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
265 assert(thread->deopt_mark() == NULL, "Pending deopt!");
266 thread->set_deopt_mark(dmark);
267
268 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
269 RegisterMap map(thread, true);
270 RegisterMap dummy_map(thread, false);
271 // Now get the deoptee with a valid map
272 frame deoptee = stub_frame.sender(&map);
273 // Set the deoptee nmethod
274 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
275 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
276 thread->set_deopt_compiled_method(cm);
277
278 if (VerifyStack) {
279 thread->validate_frame_layout();
280 }
281
282 // Create a growable array of VFrames where each VFrame represents an inlined
283 // Java frame. This storage is allocated with the usual system arena.
284 assert(deoptee.is_compiled_frame(), "Wrong frame type");
285 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
286 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
287 while (!vf->is_top()) {
288 assert(vf->is_compiled_frame(), "Wrong frame type");
289 chunk->push(compiledVFrame::cast(vf));
290 vf = vf->sender();
291 }
292 assert(vf->is_compiled_frame(), "Wrong frame type");
293 chunk->push(compiledVFrame::cast(vf));
294
295 bool realloc_failures = false;
296
297 #if COMPILER2_OR_JVMCI
298 #if INCLUDE_JVMCI
299 bool jvmci_enabled = true;
300 #else
301 bool jvmci_enabled = false;
302 #endif
303
304 // Reallocate the non-escaping objects and restore their fields. Then
305 // relock objects if synchronization on them was eliminated.
306 if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) )) {
307 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk);
308 }
309 #endif // COMPILER2_OR_JVMCI
310
311 // Revoke biases, done with in java state.
312 // No safepoints allowed after this
313 revoke_from_deopt_handler(thread, deoptee, &map);
314
315 // Ensure that no safepoint is taken after pointers have been stored
316 // in fields of rematerialized objects. If a safepoint occurs from here on
317 // out the java state residing in the vframeArray will be missed.
318 // Locks may be rebaised in a safepoint.
319 NoSafepointVerifier no_safepoint;
320
321 #if COMPILER2_OR_JVMCI
322 if (jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) )) {
323 eliminate_locks(thread, chunk, realloc_failures);
324 }
325 #endif // COMPILER2_OR_JVMCI
326
327 ScopeDesc* trap_scope = chunk->at(0)->scope();
328 Handle exceptionObject;
329 if (trap_scope->rethrow_exception()) {
330 if (PrintDeoptimizationDetails) {
331 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());
332 }
333 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
334 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
335 ScopeValue* topOfStack = expressions->top();
336 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
337 guarantee(exceptionObject() != NULL, "exception oop can not be null");
338 }
339
340 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
341 #if COMPILER2_OR_JVMCI
342 if (realloc_failures) {
343 pop_frames_failed_reallocs(thread, array);
344 }
345 #endif
346
347 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
348 thread->set_vframe_array_head(array);
349
350 // Now that the vframeArray has been created if we have any deferred local writes
351 // added by jvmti then we can free up that structure as the data is now in the
352 // vframeArray
353
354 if (thread->deferred_locals() != NULL) {
355 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
356 int i = 0;
357 do {
358 // Because of inlining we could have multiple vframes for a single frame
359 // and several of the vframes could have deferred writes. Find them all.
360 if (list->at(i)->id() == array->original().id()) {
361 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
362 list->remove_at(i);
363 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
364 delete dlv;
365 } else {
366 i++;
367 }
368 } while ( i < list->length() );
369 if (list->length() == 0) {
370 thread->set_deferred_locals(NULL);
371 // free the list and elements back to C heap.
372 delete list;
373 }
374
375 }
376
377 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
378 CodeBlob* cb = stub_frame.cb();
379 // Verify we have the right vframeArray
380 assert(cb->frame_size() >= 0, "Unexpected frame size");
381 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
382
383 // If the deopt call site is a MethodHandle invoke call site we have
384 // to adjust the unpack_sp.
385 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
386 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
387 unpack_sp = deoptee.unextended_sp();
388
389 #ifdef ASSERT
390 assert(cb->is_deoptimization_stub() ||
391 cb->is_uncommon_trap_stub() ||
392 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
393 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
394 "unexpected code blob: %s", cb->name());
395 #endif
396
397 // This is a guarantee instead of an assert because if vframe doesn't match
398 // we will unpack the wrong deoptimized frame and wind up in strange places
399 // where it will be very difficult to figure out what went wrong. Better
400 // to die an early death here than some very obscure death later when the
401 // trail is cold.
402 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
403 // in that it will fail to detect a problem when there is one. This needs
404 // more work in tiger timeframe.
405 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
406
407 int number_of_frames = array->frames();
408
409 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
410 // virtual activation, which is the reverse of the elements in the vframes array.
411 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
412 // +1 because we always have an interpreter return address for the final slot.
413 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
414 int popframe_extra_args = 0;
415 // Create an interpreter return address for the stub to use as its return
416 // address so the skeletal frames are perfectly walkable
417 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
418
419 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
420 // activation be put back on the expression stack of the caller for reexecution
421 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
422 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
423 }
424
425 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
426 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
427 // than simply use array->sender.pc(). This requires us to walk the current set of frames
428 //
429 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
430 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
431
432 // It's possible that the number of parameters at the call site is
433 // different than number of arguments in the callee when method
434 // handles are used. If the caller is interpreted get the real
435 // value so that the proper amount of space can be added to it's
436 // frame.
437 bool caller_was_method_handle = false;
438 if (deopt_sender.is_interpreted_frame()) {
439 methodHandle method(thread, deopt_sender.interpreter_frame_method());
440 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
441 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
442 // Method handle invokes may involve fairly arbitrary chains of
443 // calls so it's impossible to know how much actual space the
444 // caller has for locals.
445 caller_was_method_handle = true;
446 }
447 }
448
449 //
450 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
451 // frame_sizes/frame_pcs[1] next oldest frame (int)
452 // frame_sizes/frame_pcs[n] youngest frame (int)
453 //
454 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
455 // owns the space for the return address to it's caller). Confusing ain't it.
456 //
457 // The vframe array can address vframes with indices running from
458 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
459 // When we create the skeletal frames we need the oldest frame to be in the zero slot
460 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
461 // so things look a little strange in this loop.
462 //
463 int callee_parameters = 0;
464 int callee_locals = 0;
465 for (int index = 0; index < array->frames(); index++ ) {
466 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
467 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
468 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
469 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
470 callee_locals,
471 index == 0,
472 popframe_extra_args);
473 // This pc doesn't have to be perfect just good enough to identify the frame
474 // as interpreted so the skeleton frame will be walkable
475 // The correct pc will be set when the skeleton frame is completely filled out
476 // The final pc we store in the loop is wrong and will be overwritten below
477 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
478
479 callee_parameters = array->element(index)->method()->size_of_parameters();
480 callee_locals = array->element(index)->method()->max_locals();
481 popframe_extra_args = 0;
482 }
483
484 // Compute whether the root vframe returns a float or double value.
485 BasicType return_type;
486 {
487 methodHandle method(thread, array->element(0)->method());
488 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
489 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
490 }
491
492 // Compute information for handling adapters and adjusting the frame size of the caller.
493 int caller_adjustment = 0;
494
495 // Compute the amount the oldest interpreter frame will have to adjust
496 // its caller's stack by. If the caller is a compiled frame then
497 // we pretend that the callee has no parameters so that the
498 // extension counts for the full amount of locals and not just
499 // locals-parms. This is because without a c2i adapter the parm
500 // area as created by the compiled frame will not be usable by
501 // the interpreter. (Depending on the calling convention there
502 // may not even be enough space).
503
504 // QQQ I'd rather see this pushed down into last_frame_adjust
505 // and have it take the sender (aka caller).
506
507 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
508 caller_adjustment = last_frame_adjust(0, callee_locals);
509 } else if (callee_locals > callee_parameters) {
510 // The caller frame may need extending to accommodate
511 // non-parameter locals of the first unpacked interpreted frame.
512 // Compute that adjustment.
513 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
514 }
515
516 // If the sender is deoptimized the we must retrieve the address of the handler
517 // since the frame will "magically" show the original pc before the deopt
518 // and we'd undo the deopt.
519
520 frame_pcs[0] = deopt_sender.raw_pc();
521
522 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
523
524 #if INCLUDE_JVMCI
525 if (exceptionObject() != NULL) {
526 thread->set_exception_oop(exceptionObject());
527 exec_mode = Unpack_exception;
528 }
529 #endif
530
531 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
532 assert(thread->has_pending_exception(), "should have thrown OOME");
533 thread->set_exception_oop(thread->pending_exception());
534 thread->clear_pending_exception();
535 exec_mode = Unpack_exception;
536 }
537
538 #if INCLUDE_JVMCI
539 if (thread->frames_to_pop_failed_realloc() > 0) {
540 thread->set_pending_monitorenter(false);
541 }
542 #endif
543
544 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
545 caller_adjustment * BytesPerWord,
546 caller_was_method_handle ? 0 : callee_parameters,
547 number_of_frames,
548 frame_sizes,
549 frame_pcs,
550 return_type,
551 exec_mode);
552 // On some platforms, we need a way to pass some platform dependent
553 // information to the unpacking code so the skeletal frames come out
554 // correct (initial fp value, unextended sp, ...)
555 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
556
557 if (array->frames() > 1) {
558 if (VerifyStack && TraceDeoptimization) {
559 ttyLocker ttyl;
560 tty->print_cr("Deoptimizing method containing inlining");
561 }
562 }
563
564 array->set_unroll_block(info);
565 return info;
566 }
567
568 // Called to cleanup deoptimization data structures in normal case
569 // after unpacking to stack and when stack overflow error occurs
570 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
571 vframeArray *array) {
572
573 // Get array if coming from exception
574 if (array == NULL) {
575 array = thread->vframe_array_head();
576 }
577 thread->set_vframe_array_head(NULL);
578
579 // Free the previous UnrollBlock
580 vframeArray* old_array = thread->vframe_array_last();
581 thread->set_vframe_array_last(array);
582
583 if (old_array != NULL) {
584 UnrollBlock* old_info = old_array->unroll_block();
585 old_array->set_unroll_block(NULL);
586 delete old_info;
587 delete old_array;
588 }
589
590 // Deallocate any resource creating in this routine and any ResourceObjs allocated
591 // inside the vframeArray (StackValueCollections)
592
593 delete thread->deopt_mark();
594 thread->set_deopt_mark(NULL);
595 thread->set_deopt_compiled_method(NULL);
596
597
598 if (JvmtiExport::can_pop_frame()) {
599 #ifndef CC_INTERP
600 // Regardless of whether we entered this routine with the pending
601 // popframe condition bit set, we should always clear it now
602 thread->clear_popframe_condition();
603 #else
604 // C++ interpreter will clear has_pending_popframe when it enters
605 // with method_resume. For deopt_resume2 we clear it now.
606 if (thread->popframe_forcing_deopt_reexecution())
607 thread->clear_popframe_condition();
608 #endif /* CC_INTERP */
609 }
610
611 // unpack_frames() is called at the end of the deoptimization handler
612 // and (in C2) at the end of the uncommon trap handler. Note this fact
613 // so that an asynchronous stack walker can work again. This counter is
614 // incremented at the beginning of fetch_unroll_info() and (in C2) at
615 // the beginning of uncommon_trap().
616 thread->dec_in_deopt_handler();
617 }
618
619 // Moved from cpu directories because none of the cpus has callee save values.
620 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
621 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
622
623 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
624 // the days we had adapter frames. When we deoptimize a situation where a
625 // compiled caller calls a compiled caller will have registers it expects
626 // to survive the call to the callee. If we deoptimize the callee the only
627 // way we can restore these registers is to have the oldest interpreter
628 // frame that we create restore these values. That is what this routine
629 // will accomplish.
630
631 // At the moment we have modified c2 to not have any callee save registers
632 // so this problem does not exist and this routine is just a place holder.
633
634 assert(f->is_interpreted_frame(), "must be interpreted");
635 }
636
637 // Return BasicType of value being returned
638 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
639
640 // We are already active in the special DeoptResourceMark any ResourceObj's we
641 // allocate will be freed at the end of the routine.
642
643 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
644 // but makes the entry a little slower. There is however a little dance we have to
645 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
646 ResetNoHandleMark rnhm; // No-op in release/product versions
647 HandleMark hm;
648
649 frame stub_frame = thread->last_frame();
650
651 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
652 // must point to the vframeArray for the unpack frame.
653 vframeArray* array = thread->vframe_array_head();
654
655 #ifndef PRODUCT
656 if (TraceDeoptimization) {
657 ttyLocker ttyl;
658 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
659 p2i(thread), p2i(array), exec_mode);
660 }
661 #endif
662 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
663 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
664
665 UnrollBlock* info = array->unroll_block();
666
667 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
668 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
669
670 BasicType bt = info->return_type();
671
672 // If we have an exception pending, claim that the return type is an oop
673 // so the deopt_blob does not overwrite the exception_oop.
674
675 if (exec_mode == Unpack_exception)
676 bt = T_OBJECT;
677
678 // Cleanup thread deopt data
679 cleanup_deopt_info(thread, array);
680
681 #ifndef PRODUCT
682 if (VerifyStack) {
683 ResourceMark res_mark;
684 // Clear pending exception to not break verification code (restored afterwards)
685 PRESERVE_EXCEPTION_MARK;
686
687 thread->validate_frame_layout();
688
689 // Verify that the just-unpacked frames match the interpreter's
690 // notions of expression stack and locals
691 vframeArray* cur_array = thread->vframe_array_last();
692 RegisterMap rm(thread, false);
693 rm.set_include_argument_oops(false);
694 bool is_top_frame = true;
695 int callee_size_of_parameters = 0;
696 int callee_max_locals = 0;
697 for (int i = 0; i < cur_array->frames(); i++) {
698 vframeArrayElement* el = cur_array->element(i);
699 frame* iframe = el->iframe();
700 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
701
702 // Get the oop map for this bci
703 InterpreterOopMap mask;
704 int cur_invoke_parameter_size = 0;
705 bool try_next_mask = false;
706 int next_mask_expression_stack_size = -1;
707 int top_frame_expression_stack_adjustment = 0;
708 methodHandle mh(thread, iframe->interpreter_frame_method());
709 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
710 BytecodeStream str(mh, iframe->interpreter_frame_bci());
711 int max_bci = mh->code_size();
712 // Get to the next bytecode if possible
713 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
714 // Check to see if we can grab the number of outgoing arguments
715 // at an uncommon trap for an invoke (where the compiler
716 // generates debug info before the invoke has executed)
717 Bytecodes::Code cur_code = str.next();
718 if (Bytecodes::is_invoke(cur_code)) {
719 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
720 cur_invoke_parameter_size = invoke.size_of_parameters();
721 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
722 callee_size_of_parameters++;
723 }
724 }
725 if (str.bci() < max_bci) {
726 Bytecodes::Code next_code = str.next();
727 if (next_code >= 0) {
728 // The interpreter oop map generator reports results before
729 // the current bytecode has executed except in the case of
730 // calls. It seems to be hard to tell whether the compiler
731 // has emitted debug information matching the "state before"
732 // a given bytecode or the state after, so we try both
733 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
734 // Get expression stack size for the next bytecode
735 InterpreterOopMap next_mask;
736 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
737 next_mask_expression_stack_size = next_mask.expression_stack_size();
738 if (Bytecodes::is_invoke(next_code)) {
739 Bytecode_invoke invoke(mh, str.bci());
740 next_mask_expression_stack_size += invoke.size_of_parameters();
741 }
742 // Need to subtract off the size of the result type of
743 // the bytecode because this is not described in the
744 // debug info but returned to the interpreter in the TOS
745 // caching register
746 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
747 if (bytecode_result_type != T_ILLEGAL) {
748 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
749 }
750 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
751 try_next_mask = true;
752 }
753 }
754 }
755
756 // Verify stack depth and oops in frame
757 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
758 if (!(
759 /* SPARC */
760 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
761 /* x86 */
762 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
763 (try_next_mask &&
764 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
765 top_frame_expression_stack_adjustment))) ||
766 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
767 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
768 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
769 )) {
770 {
771 ttyLocker ttyl;
772
773 // Print out some information that will help us debug the problem
774 tty->print_cr("Wrong number of expression stack elements during deoptimization");
775 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
776 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
777 iframe->interpreter_frame_expression_stack_size());
778 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
779 tty->print_cr(" try_next_mask = %d", try_next_mask);
780 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
781 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
782 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
783 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
784 tty->print_cr(" exec_mode = %d", exec_mode);
785 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
786 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
787 tty->print_cr(" Interpreted frames:");
788 for (int k = 0; k < cur_array->frames(); k++) {
789 vframeArrayElement* el = cur_array->element(k);
790 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
791 }
792 cur_array->print_on_2(tty);
793 } // release tty lock before calling guarantee
794 guarantee(false, "wrong number of expression stack elements during deopt");
795 }
796 VerifyOopClosure verify;
797 iframe->oops_interpreted_do(&verify, &rm, false);
798 callee_size_of_parameters = mh->size_of_parameters();
799 callee_max_locals = mh->max_locals();
800 is_top_frame = false;
801 }
802 }
803 #endif /* !PRODUCT */
804
805
806 return bt;
807 JRT_END
808
809 class DeoptimizeMarkedClosure : public HandshakeClosure {
810 public:
811 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
812 void do_thread(Thread* thread) {
813 JavaThread* jt = (JavaThread*)thread;
814 jt->deoptimize_marked_methods();
815 }
816 };
817
818 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) {
819 ResourceMark rm;
820 DeoptimizationMarker dm;
821
822 // Make the dependent methods not entrant
823 if (nmethod_only != NULL) {
824 nmethod_only->mark_for_deoptimization();
825 nmethod_only->make_not_entrant();
826 } else {
827 MutexLocker mu(SafepointSynchronize::is_at_safepoint() ? NULL : CodeCache_lock, Mutex::_no_safepoint_check_flag);
828 CodeCache::make_marked_nmethods_not_entrant();
829 }
830
831 DeoptimizeMarkedClosure deopt;
832 if (SafepointSynchronize::is_at_safepoint()) {
833 Threads::java_threads_do(&deopt);
834 } else {
835 Handshake::execute(&deopt);
836 }
837 }
838
839 Deoptimization::DeoptAction Deoptimization::_unloaded_action
840 = Deoptimization::Action_reinterpret;
841
842
843
844 #if INCLUDE_JVMCI || INCLUDE_AOT
845 template<typename CacheType>
846 class BoxCacheBase : public CHeapObj<mtCompiler> {
847 protected:
848 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) {
849 ResourceMark rm;
850 char* klass_name_str = klass_name->as_C_string();
851 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD);
852 guarantee(k != NULL, "%s must be loaded", klass_name_str);
853 InstanceKlass* ik = InstanceKlass::cast(k);
854 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
855 CacheType::compute_offsets(ik);
856 return ik;
857 }
858 };
859
860 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> {
861 PrimitiveType _low;
862 PrimitiveType _high;
863 jobject _cache;
864 protected:
865 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
866 BoxCache(Thread* thread) {
867 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread);
868 objArrayOop cache = CacheType::cache(ik);
869 assert(cache->length() > 0, "Empty cache");
870 _low = BoxType::value(cache->obj_at(0));
871 _high = _low + cache->length() - 1;
872 _cache = JNIHandles::make_global(Handle(thread, cache));
873 }
874 ~BoxCache() {
875 JNIHandles::destroy_global(_cache);
876 }
877 public:
878 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
879 if (_singleton == NULL) {
880 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
881 if (!Atomic::replace_if_null(&_singleton, s)) {
882 delete s;
883 }
884 }
885 return _singleton;
886 }
887 oop lookup(PrimitiveType value) {
888 if (_low <= value && value <= _high) {
889 int offset = value - _low;
890 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
891 }
892 return NULL;
893 }
894 oop lookup_raw(intptr_t raw_value) {
895 // Have to cast to avoid little/big-endian problems.
896 if (sizeof(PrimitiveType) > sizeof(jint)) {
897 jlong value = (jlong)raw_value;
898 return lookup(value);
899 }
900 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
901 return lookup(value);
902 }
903 };
904
905 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
906 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
907 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
908 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
909 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
910
911 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL;
912 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL;
913 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL;
914 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL;
915 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL;
916
917 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
918 jobject _true_cache;
919 jobject _false_cache;
920 protected:
921 static BooleanBoxCache *_singleton;
922 BooleanBoxCache(Thread *thread) {
923 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread);
924 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
925 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
926 }
927 ~BooleanBoxCache() {
928 JNIHandles::destroy_global(_true_cache);
929 JNIHandles::destroy_global(_false_cache);
930 }
931 public:
932 static BooleanBoxCache* singleton(Thread* thread) {
933 if (_singleton == NULL) {
934 BooleanBoxCache* s = new BooleanBoxCache(thread);
935 if (!Atomic::replace_if_null(&_singleton, s)) {
936 delete s;
937 }
938 }
939 return _singleton;
940 }
941 oop lookup_raw(intptr_t raw_value) {
942 // Have to cast to avoid little/big-endian problems.
943 jboolean value = (jboolean)*((jint*)&raw_value);
944 return lookup(value);
945 }
946 oop lookup(jboolean value) {
947 if (value != 0) {
948 return JNIHandles::resolve_non_null(_true_cache);
949 }
950 return JNIHandles::resolve_non_null(_false_cache);
951 }
952 };
953
954 BooleanBoxCache* BooleanBoxCache::_singleton = NULL;
955
956 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) {
957 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
958 BasicType box_type = SystemDictionary::box_klass_type(k);
959 if (box_type != T_OBJECT) {
960 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
961 switch(box_type) {
962 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
963 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
964 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
965 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
966 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
967 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
968 default:;
969 }
970 }
971 return NULL;
972 }
973 #endif // INCLUDE_JVMCI || INCLUDE_AOT
974
975 #if COMPILER2_OR_JVMCI
976 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
977 Handle pending_exception(THREAD, thread->pending_exception());
978 const char* exception_file = thread->exception_file();
979 int exception_line = thread->exception_line();
980 thread->clear_pending_exception();
981
982 bool failures = false;
983
984 for (int i = 0; i < objects->length(); i++) {
985 assert(objects->at(i)->is_object(), "invalid debug information");
986 ObjectValue* sv = (ObjectValue*) objects->at(i);
987
988 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
989 oop obj = NULL;
990
991 if (k->is_instance_klass()) {
992 #if INCLUDE_JVMCI || INCLUDE_AOT
993 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null();
994 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) {
995 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
996 obj = get_cached_box(abv, fr, reg_map, THREAD);
997 if (obj != NULL) {
998 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
999 abv->set_cached(true);
1000 }
1001 }
1002 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1003 InstanceKlass* ik = InstanceKlass::cast(k);
1004 if (obj == NULL) {
1005 obj = ik->allocate_instance(THREAD);
1006 }
1007 } else if (k->is_typeArray_klass()) {
1008 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1009 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1010 int len = sv->field_size() / type2size[ak->element_type()];
1011 obj = ak->allocate(len, THREAD);
1012 } else if (k->is_objArray_klass()) {
1013 ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1014 obj = ak->allocate(sv->field_size(), THREAD);
1015 }
1016
1017 if (obj == NULL) {
1018 failures = true;
1019 }
1020
1021 assert(sv->value().is_null(), "redundant reallocation");
1022 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
1023 CLEAR_PENDING_EXCEPTION;
1024 sv->set_value(obj);
1025 }
1026
1027 if (failures) {
1028 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1029 } else if (pending_exception.not_null()) {
1030 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1031 }
1032
1033 return failures;
1034 }
1035
1036 #if INCLUDE_JVMCI
1037 /**
1038 * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1039 * we need to somehow be able to recover the actual kind to be able to write the correct
1040 * amount of bytes.
1041 * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1042 * the entries at index n + 1 to n + i are 'markers'.
1043 * For example, if we were writing a short at index 4 of a byte array of size 8, the
1044 * expected form of the array would be:
1045 *
1046 * {b0, b1, b2, b3, INT, marker, b6, b7}
1047 *
1048 * Thus, in order to get back the size of the entry, we simply need to count the number
1049 * of marked entries
1050 *
1051 * @param virtualArray the virtualized byte array
1052 * @param i index of the virtual entry we are recovering
1053 * @return The number of bytes the entry spans
1054 */
1055 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1056 int index = i;
1057 while (++index < virtualArray->field_size() &&
1058 virtualArray->field_at(index)->is_marker()) {}
1059 return index - i;
1060 }
1061
1062 /**
1063 * If there was a guarantee for byte array to always start aligned to a long, we could
1064 * do a simple check on the parity of the index. Unfortunately, that is not always the
1065 * case. Thus, we check alignment of the actual address we are writing to.
1066 * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1067 * write a long to index 3.
1068 */
1069 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1070 jbyte* res = obj->byte_at_addr(index);
1071 assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1072 return res;
1073 }
1074
1075 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) {
1076 switch (byte_count) {
1077 case 1:
1078 obj->byte_at_put(index, (jbyte) *((jint *) &val));
1079 break;
1080 case 2:
1081 *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val);
1082 break;
1083 case 4:
1084 *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val);
1085 break;
1086 case 8: {
1087 #ifdef _LP64
1088 jlong res = (jlong) *((jlong *) &val);
1089 #else
1090 #ifdef SPARC
1091 // For SPARC we have to swap high and low words.
1092 jlong v = (jlong) *((jlong *) &val);
1093 jlong res = 0;
1094 res |= ((v & (jlong) 0xffffffff) << 32);
1095 res |= ((v >> 32) & (jlong) 0xffffffff);
1096 #else
1097 jlong res = (jlong) *((jlong *) &val);
1098 #endif // SPARC
1099 #endif
1100 *((jlong *) check_alignment_get_addr(obj, index, 8)) = res;
1101 break;
1102 }
1103 default:
1104 ShouldNotReachHere();
1105 }
1106 }
1107 #endif // INCLUDE_JVMCI
1108
1109
1110 // restore elements of an eliminated type array
1111 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1112 int index = 0;
1113 intptr_t val;
1114
1115 for (int i = 0; i < sv->field_size(); i++) {
1116 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1117 switch(type) {
1118 case T_LONG: case T_DOUBLE: {
1119 assert(value->type() == T_INT, "Agreement.");
1120 StackValue* low =
1121 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1122 #ifdef _LP64
1123 jlong res = (jlong)low->get_int();
1124 #else
1125 #ifdef SPARC
1126 // For SPARC we have to swap high and low words.
1127 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1128 #else
1129 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1130 #endif //SPARC
1131 #endif
1132 obj->long_at_put(index, res);
1133 break;
1134 }
1135
1136 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1137 case T_INT: case T_FLOAT: { // 4 bytes.
1138 assert(value->type() == T_INT, "Agreement.");
1139 bool big_value = false;
1140 if (i + 1 < sv->field_size() && type == T_INT) {
1141 if (sv->field_at(i)->is_location()) {
1142 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1143 if (type == Location::dbl || type == Location::lng) {
1144 big_value = true;
1145 }
1146 } else if (sv->field_at(i)->is_constant_int()) {
1147 ScopeValue* next_scope_field = sv->field_at(i + 1);
1148 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1149 big_value = true;
1150 }
1151 }
1152 }
1153
1154 if (big_value) {
1155 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1156 #ifdef _LP64
1157 jlong res = (jlong)low->get_int();
1158 #else
1159 #ifdef SPARC
1160 // For SPARC we have to swap high and low words.
1161 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1162 #else
1163 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1164 #endif //SPARC
1165 #endif
1166 obj->int_at_put(index, (jint)*((jint*)&res));
1167 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1168 } else {
1169 val = value->get_int();
1170 obj->int_at_put(index, (jint)*((jint*)&val));
1171 }
1172 break;
1173 }
1174
1175 case T_SHORT:
1176 assert(value->type() == T_INT, "Agreement.");
1177 val = value->get_int();
1178 obj->short_at_put(index, (jshort)*((jint*)&val));
1179 break;
1180
1181 case T_CHAR:
1182 assert(value->type() == T_INT, "Agreement.");
1183 val = value->get_int();
1184 obj->char_at_put(index, (jchar)*((jint*)&val));
1185 break;
1186
1187 case T_BYTE: {
1188 assert(value->type() == T_INT, "Agreement.");
1189 // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1190 val = value->get_int();
1191 #if INCLUDE_JVMCI
1192 int byte_count = count_number_of_bytes_for_entry(sv, i);
1193 byte_array_put(obj, val, index, byte_count);
1194 // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1195 i += byte_count - 1; // Balance the loop counter.
1196 index += byte_count;
1197 // index has been updated so continue at top of loop
1198 continue;
1199 #else
1200 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1201 break;
1202 #endif // INCLUDE_JVMCI
1203 }
1204
1205 case T_BOOLEAN: {
1206 assert(value->type() == T_INT, "Agreement.");
1207 val = value->get_int();
1208 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1209 break;
1210 }
1211
1212 default:
1213 ShouldNotReachHere();
1214 }
1215 index++;
1216 }
1217 }
1218
1219 // restore fields of an eliminated object array
1220 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1221 for (int i = 0; i < sv->field_size(); i++) {
1222 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1223 assert(value->type() == T_OBJECT, "object element expected");
1224 obj->obj_at_put(i, value->get_obj()());
1225 }
1226 }
1227
1228 class ReassignedField {
1229 public:
1230 int _offset;
1231 BasicType _type;
1232 public:
1233 ReassignedField() {
1234 _offset = 0;
1235 _type = T_ILLEGAL;
1236 }
1237 };
1238
1239 int compare(ReassignedField* left, ReassignedField* right) {
1240 return left->_offset - right->_offset;
1241 }
1242
1243 // Restore fields of an eliminated instance object using the same field order
1244 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1245 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1246 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1247 InstanceKlass* ik = klass;
1248 while (ik != NULL) {
1249 for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1250 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1251 ReassignedField field;
1252 field._offset = fs.offset();
1253 field._type = Signature::basic_type(fs.signature());
1254 fields->append(field);
1255 }
1256 }
1257 ik = ik->superklass();
1258 }
1259 fields->sort(compare);
1260 for (int i = 0; i < fields->length(); i++) {
1261 intptr_t val;
1262 ScopeValue* scope_field = sv->field_at(svIndex);
1263 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1264 int offset = fields->at(i)._offset;
1265 BasicType type = fields->at(i)._type;
1266 switch (type) {
1267 case T_OBJECT: case T_ARRAY:
1268 assert(value->type() == T_OBJECT, "Agreement.");
1269 obj->obj_field_put(offset, value->get_obj()());
1270 break;
1271
1272 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1273 case T_INT: case T_FLOAT: { // 4 bytes.
1274 assert(value->type() == T_INT, "Agreement.");
1275 bool big_value = false;
1276 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1277 if (scope_field->is_location()) {
1278 Location::Type type = ((LocationValue*) scope_field)->location().type();
1279 if (type == Location::dbl || type == Location::lng) {
1280 big_value = true;
1281 }
1282 }
1283 if (scope_field->is_constant_int()) {
1284 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1285 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1286 big_value = true;
1287 }
1288 }
1289 }
1290
1291 if (big_value) {
1292 i++;
1293 assert(i < fields->length(), "second T_INT field needed");
1294 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1295 } else {
1296 val = value->get_int();
1297 obj->int_field_put(offset, (jint)*((jint*)&val));
1298 break;
1299 }
1300 }
1301 /* no break */
1302
1303 case T_LONG: case T_DOUBLE: {
1304 assert(value->type() == T_INT, "Agreement.");
1305 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1306 #ifdef _LP64
1307 jlong res = (jlong)low->get_int();
1308 #else
1309 #ifdef SPARC
1310 // For SPARC we have to swap high and low words.
1311 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1312 #else
1313 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1314 #endif //SPARC
1315 #endif
1316 obj->long_field_put(offset, res);
1317 break;
1318 }
1319
1320 case T_SHORT:
1321 assert(value->type() == T_INT, "Agreement.");
1322 val = value->get_int();
1323 obj->short_field_put(offset, (jshort)*((jint*)&val));
1324 break;
1325
1326 case T_CHAR:
1327 assert(value->type() == T_INT, "Agreement.");
1328 val = value->get_int();
1329 obj->char_field_put(offset, (jchar)*((jint*)&val));
1330 break;
1331
1332 case T_BYTE:
1333 assert(value->type() == T_INT, "Agreement.");
1334 val = value->get_int();
1335 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1336 break;
1337
1338 case T_BOOLEAN:
1339 assert(value->type() == T_INT, "Agreement.");
1340 val = value->get_int();
1341 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1342 break;
1343
1344 default:
1345 ShouldNotReachHere();
1346 }
1347 svIndex++;
1348 }
1349 return svIndex;
1350 }
1351
1352 // restore fields of all eliminated objects and arrays
1353 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1354 for (int i = 0; i < objects->length(); i++) {
1355 ObjectValue* sv = (ObjectValue*) objects->at(i);
1356 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1357 Handle obj = sv->value();
1358 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1359 if (PrintDeoptimizationDetails) {
1360 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1361 }
1362 if (obj.is_null()) {
1363 continue;
1364 }
1365 #if INCLUDE_JVMCI || INCLUDE_AOT
1366 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1367 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1368 continue;
1369 }
1370 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1371 if (k->is_instance_klass()) {
1372 InstanceKlass* ik = InstanceKlass::cast(k);
1373 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1374 } else if (k->is_typeArray_klass()) {
1375 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1376 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1377 } else if (k->is_objArray_klass()) {
1378 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1379 }
1380 }
1381 }
1382
1383
1384 // relock objects for which synchronization was eliminated
1385 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1386 for (int i = 0; i < monitors->length(); i++) {
1387 MonitorInfo* mon_info = monitors->at(i);
1388 if (mon_info->eliminated()) {
1389 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1390 if (!mon_info->owner_is_scalar_replaced()) {
1391 Handle obj(thread, mon_info->owner());
1392 markWord mark = obj->mark();
1393 if (UseBiasedLocking && mark.has_bias_pattern()) {
1394 // New allocated objects may have the mark set to anonymously biased.
1395 // Also the deoptimized method may called methods with synchronization
1396 // where the thread-local object is bias locked to the current thread.
1397 assert(mark.is_biased_anonymously() ||
1398 mark.biased_locker() == thread, "should be locked to current thread");
1399 // Reset mark word to unbiased prototype.
1400 markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1401 obj->set_mark(unbiased_prototype);
1402 }
1403 BasicLock* lock = mon_info->lock();
1404 ObjectSynchronizer::enter(obj, lock, thread);
1405 assert(mon_info->owner()->is_locked(), "object must be locked now");
1406 }
1407 }
1408 }
1409 }
1410
1411
1412 #ifndef PRODUCT
1413 // print information about reallocated objects
1414 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1415 fieldDescriptor fd;
1416
1417 for (int i = 0; i < objects->length(); i++) {
1418 ObjectValue* sv = (ObjectValue*) objects->at(i);
1419 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1420 Handle obj = sv->value();
1421
1422 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1423 k->print_value();
1424 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1425 if (obj.is_null()) {
1426 tty->print(" allocation failed");
1427 } else {
1428 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1429 }
1430 tty->cr();
1431
1432 if (Verbose && !obj.is_null()) {
1433 k->oop_print_on(obj(), tty);
1434 }
1435 }
1436 }
1437 #endif
1438 #endif // COMPILER2_OR_JVMCI
1439
1440 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1441 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1442
1443 #ifndef PRODUCT
1444 if (PrintDeoptimizationDetails) {
1445 ttyLocker ttyl;
1446 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1447 fr.print_on(tty);
1448 tty->print_cr(" Virtual frames (innermost first):");
1449 for (int index = 0; index < chunk->length(); index++) {
1450 compiledVFrame* vf = chunk->at(index);
1451 tty->print(" %2d - ", index);
1452 vf->print_value();
1453 int bci = chunk->at(index)->raw_bci();
1454 const char* code_name;
1455 if (bci == SynchronizationEntryBCI) {
1456 code_name = "sync entry";
1457 } else {
1458 Bytecodes::Code code = vf->method()->code_at(bci);
1459 code_name = Bytecodes::name(code);
1460 }
1461 tty->print(" - %s", code_name);
1462 tty->print_cr(" @ bci %d ", bci);
1463 if (Verbose) {
1464 vf->print();
1465 tty->cr();
1466 }
1467 }
1468 }
1469 #endif
1470
1471 // Register map for next frame (used for stack crawl). We capture
1472 // the state of the deopt'ing frame's caller. Thus if we need to
1473 // stuff a C2I adapter we can properly fill in the callee-save
1474 // register locations.
1475 frame caller = fr.sender(reg_map);
1476 int frame_size = caller.sp() - fr.sp();
1477
1478 frame sender = caller;
1479
1480 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1481 // the vframeArray containing the unpacking information is allocated in the C heap.
1482 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1483 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1484
1485 // Compare the vframeArray to the collected vframes
1486 assert(array->structural_compare(thread, chunk), "just checking");
1487
1488 #ifndef PRODUCT
1489 if (PrintDeoptimizationDetails) {
1490 ttyLocker ttyl;
1491 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1492 }
1493 #endif // PRODUCT
1494
1495 return array;
1496 }
1497
1498 #if COMPILER2_OR_JVMCI
1499 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1500 // Reallocation of some scalar replaced objects failed. Record
1501 // that we need to pop all the interpreter frames for the
1502 // deoptimized compiled frame.
1503 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1504 thread->set_frames_to_pop_failed_realloc(array->frames());
1505 // Unlock all monitors here otherwise the interpreter will see a
1506 // mix of locked and unlocked monitors (because of failed
1507 // reallocations of synchronized objects) and be confused.
1508 for (int i = 0; i < array->frames(); i++) {
1509 MonitorChunk* monitors = array->element(i)->monitors();
1510 if (monitors != NULL) {
1511 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1512 BasicObjectLock* src = monitors->at(j);
1513 if (src->obj() != NULL) {
1514 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1515 }
1516 }
1517 array->element(i)->free_monitors(thread);
1518 #ifdef ASSERT
1519 array->element(i)->set_removed_monitors();
1520 #endif
1521 }
1522 }
1523 }
1524 #endif
1525
1526 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1527 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1528 Thread* thread = Thread::current();
1529 for (int i = 0; i < monitors->length(); i++) {
1530 MonitorInfo* mon_info = monitors->at(i);
1531 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1532 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1533 }
1534 }
1535 }
1536
1537 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) {
1538 // Unfortunately we don't have a RegisterMap available in most of
1539 // the places we want to call this routine so we need to walk the
1540 // stack again to update the register map.
1541 if (map == NULL || !map->update_map()) {
1542 StackFrameStream sfs(thread, true);
1543 bool found = false;
1544 while (!found && !sfs.is_done()) {
1545 frame* cur = sfs.current();
1546 sfs.next();
1547 found = cur->id() == fr.id();
1548 }
1549 assert(found, "frame to be deoptimized not found on target thread's stack");
1550 map = sfs.register_map();
1551 }
1552
1553 vframe* vf = vframe::new_vframe(&fr, map, thread);
1554 compiledVFrame* cvf = compiledVFrame::cast(vf);
1555 // Revoke monitors' biases in all scopes
1556 while (!cvf->is_top()) {
1557 collect_monitors(cvf, objects_to_revoke);
1558 cvf = compiledVFrame::cast(cvf->sender());
1559 }
1560 collect_monitors(cvf, objects_to_revoke);
1561 }
1562
1563 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) {
1564 if (!UseBiasedLocking) {
1565 return;
1566 }
1567 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1568 get_monitors_from_stack(objects_to_revoke, thread, fr, map);
1569
1570 int len = objects_to_revoke->length();
1571 for (int i = 0; i < len; i++) {
1572 oop obj = (objects_to_revoke->at(i))();
1573 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread);
1574 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1575 }
1576 }
1577
1578
1579 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1580 assert(fr.can_be_deoptimized(), "checking frame type");
1581
1582 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1583
1584 if (LogCompilation && xtty != NULL) {
1585 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1586 assert(cm != NULL, "only compiled methods can deopt");
1587
1588 ttyLocker ttyl;
1589 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1590 cm->log_identity(xtty);
1591 xtty->end_head();
1592 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1593 xtty->begin_elem("jvms bci='%d'", sd->bci());
1594 xtty->method(sd->method());
1595 xtty->end_elem();
1596 if (sd->is_top()) break;
1597 }
1598 xtty->tail("deoptimized");
1599 }
1600
1601 // Patch the compiled method so that when execution returns to it we will
1602 // deopt the execution state and return to the interpreter.
1603 fr.deoptimize(thread);
1604 }
1605
1606 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1607 // Deoptimize only if the frame comes from compile code.
1608 // Do not deoptimize the frame which is already patched
1609 // during the execution of the loops below.
1610 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1611 return;
1612 }
1613 ResourceMark rm;
1614 DeoptimizationMarker dm;
1615 deoptimize_single_frame(thread, fr, reason);
1616 }
1617
1618 #if INCLUDE_JVMCI
1619 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1620 // there is no exception handler for this pc => deoptimize
1621 cm->make_not_entrant();
1622
1623 // Use Deoptimization::deoptimize for all of its side-effects:
1624 // gathering traps statistics, logging...
1625 // it also patches the return pc but we do not care about that
1626 // since we return a continuation to the deopt_blob below.
1627 JavaThread* thread = JavaThread::current();
1628 RegisterMap reg_map(thread, false);
1629 frame runtime_frame = thread->last_frame();
1630 frame caller_frame = runtime_frame.sender(®_map);
1631 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1632 Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1633
1634 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1635 if (trap_mdo != NULL) {
1636 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1637 }
1638
1639 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1640 }
1641 #endif
1642
1643 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1644 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1645 "can only deoptimize other thread at a safepoint");
1646 // Compute frame and register map based on thread and sp.
1647 RegisterMap reg_map(thread, false);
1648 frame fr = thread->last_frame();
1649 while (fr.id() != id) {
1650 fr = fr.sender(®_map);
1651 }
1652 deoptimize(thread, fr, reason);
1653 }
1654
1655
1656 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1657 if (thread == Thread::current()) {
1658 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1659 } else {
1660 VM_DeoptimizeFrame deopt(thread, id, reason);
1661 VMThread::execute(&deopt);
1662 }
1663 }
1664
1665 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1666 deoptimize_frame(thread, id, Reason_constraint);
1667 }
1668
1669 // JVMTI PopFrame support
1670 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1671 {
1672 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1673 }
1674 JRT_END
1675
1676 MethodData*
1677 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1678 bool create_if_missing) {
1679 Thread* THREAD = thread;
1680 MethodData* mdo = m()->method_data();
1681 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1682 // Build an MDO. Ignore errors like OutOfMemory;
1683 // that simply means we won't have an MDO to update.
1684 Method::build_interpreter_method_data(m, THREAD);
1685 if (HAS_PENDING_EXCEPTION) {
1686 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1687 CLEAR_PENDING_EXCEPTION;
1688 }
1689 mdo = m()->method_data();
1690 }
1691 return mdo;
1692 }
1693
1694 #if COMPILER2_OR_JVMCI
1695 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1696 // In case of an unresolved klass entry, load the class.
1697 // This path is exercised from case _ldc in Parse::do_one_bytecode,
1698 // and probably nowhere else.
1699 // Even that case would benefit from simply re-interpreting the
1700 // bytecode, without paying special attention to the class index.
1701 // So this whole "class index" feature should probably be removed.
1702
1703 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1704 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1705 return;
1706 }
1707
1708 assert(!constant_pool->tag_at(index).is_symbol(),
1709 "no symbolic names here, please");
1710 }
1711
1712
1713 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1714 EXCEPTION_MARK;
1715 load_class_by_index(constant_pool, index, THREAD);
1716 if (HAS_PENDING_EXCEPTION) {
1717 // Exception happened during classloading. We ignore the exception here, since it
1718 // is going to be rethrown since the current activation is going to be deoptimized and
1719 // the interpreter will re-execute the bytecode.
1720 CLEAR_PENDING_EXCEPTION;
1721 // Class loading called java code which may have caused a stack
1722 // overflow. If the exception was thrown right before the return
1723 // to the runtime the stack is no longer guarded. Reguard the
1724 // stack otherwise if we return to the uncommon trap blob and the
1725 // stack bang causes a stack overflow we crash.
1726 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1727 JavaThread* thread = (JavaThread*)THREAD;
1728 bool guard_pages_enabled = thread->stack_guards_enabled();
1729 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1730 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1731 }
1732 }
1733
1734 #if INCLUDE_JFR
1735
1736 class DeoptReasonSerializer : public JfrSerializer {
1737 public:
1738 void serialize(JfrCheckpointWriter& writer) {
1739 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1740 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1741 writer.write_key((u8)i);
1742 writer.write(Deoptimization::trap_reason_name(i));
1743 }
1744 }
1745 };
1746
1747 class DeoptActionSerializer : public JfrSerializer {
1748 public:
1749 void serialize(JfrCheckpointWriter& writer) {
1750 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1751 writer.write_count(nof_actions);
1752 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1753 writer.write_key(i);
1754 writer.write(Deoptimization::trap_action_name((int)i));
1755 }
1756 }
1757 };
1758
1759 static void register_serializers() {
1760 static int critical_section = 0;
1761 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1762 return;
1763 }
1764 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1765 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1766 }
1767
1768 static void post_deoptimization_event(CompiledMethod* nm,
1769 const Method* method,
1770 int trap_bci,
1771 int instruction,
1772 Deoptimization::DeoptReason reason,
1773 Deoptimization::DeoptAction action) {
1774 assert(nm != NULL, "invariant");
1775 assert(method != NULL, "invariant");
1776 if (EventDeoptimization::is_enabled()) {
1777 static bool serializers_registered = false;
1778 if (!serializers_registered) {
1779 register_serializers();
1780 serializers_registered = true;
1781 }
1782 EventDeoptimization event;
1783 event.set_compileId(nm->compile_id());
1784 event.set_compiler(nm->compiler_type());
1785 event.set_method(method);
1786 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1787 event.set_bci(trap_bci);
1788 event.set_instruction(instruction);
1789 event.set_reason(reason);
1790 event.set_action(action);
1791 event.commit();
1792 }
1793 }
1794
1795 #endif // INCLUDE_JFR
1796
1797 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1798 HandleMark hm;
1799
1800 // uncommon_trap() is called at the beginning of the uncommon trap
1801 // handler. Note this fact before we start generating temporary frames
1802 // that can confuse an asynchronous stack walker. This counter is
1803 // decremented at the end of unpack_frames().
1804 thread->inc_in_deopt_handler();
1805
1806 // We need to update the map if we have biased locking.
1807 #if INCLUDE_JVMCI
1808 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1809 RegisterMap reg_map(thread, true);
1810 #else
1811 RegisterMap reg_map(thread, UseBiasedLocking);
1812 #endif
1813 frame stub_frame = thread->last_frame();
1814 frame fr = stub_frame.sender(®_map);
1815 // Make sure the calling nmethod is not getting deoptimized and removed
1816 // before we are done with it.
1817 nmethodLocker nl(fr.pc());
1818
1819 // Log a message
1820 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1821 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1822
1823 {
1824 ResourceMark rm;
1825
1826 DeoptReason reason = trap_request_reason(trap_request);
1827 DeoptAction action = trap_request_action(trap_request);
1828 #if INCLUDE_JVMCI
1829 int debug_id = trap_request_debug_id(trap_request);
1830 #endif
1831 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1832
1833 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1834 compiledVFrame* cvf = compiledVFrame::cast(vf);
1835
1836 CompiledMethod* nm = cvf->code();
1837
1838 ScopeDesc* trap_scope = cvf->scope();
1839
1840 if (TraceDeoptimization) {
1841 ttyLocker ttyl;
1842 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()
1843 #if INCLUDE_JVMCI
1844 , debug_id
1845 #endif
1846 );
1847 }
1848
1849 methodHandle trap_method(THREAD, trap_scope->method());
1850 int trap_bci = trap_scope->bci();
1851 #if INCLUDE_JVMCI
1852 jlong speculation = thread->pending_failed_speculation();
1853 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1854 nm->as_nmethod()->update_speculation(thread);
1855 } else {
1856 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1857 }
1858
1859 if (trap_bci == SynchronizationEntryBCI) {
1860 trap_bci = 0;
1861 thread->set_pending_monitorenter(true);
1862 }
1863
1864 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1865 thread->set_pending_transfer_to_interpreter(true);
1866 }
1867 #endif
1868
1869 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1870 // Record this event in the histogram.
1871 gather_statistics(reason, action, trap_bc);
1872
1873 // Ensure that we can record deopt. history:
1874 // Need MDO to record RTM code generation state.
1875 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1876
1877 methodHandle profiled_method;
1878 #if INCLUDE_JVMCI
1879 if (nm->is_compiled_by_jvmci()) {
1880 profiled_method = methodHandle(THREAD, nm->method());
1881 } else {
1882 profiled_method = trap_method;
1883 }
1884 #else
1885 profiled_method = trap_method;
1886 #endif
1887
1888 MethodData* trap_mdo =
1889 get_method_data(thread, profiled_method, create_if_missing);
1890
1891 JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);)
1892
1893 // Log a message
1894 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1895 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1896 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1897
1898 // Print a bunch of diagnostics, if requested.
1899 if (TraceDeoptimization || LogCompilation) {
1900 ResourceMark rm;
1901 ttyLocker ttyl;
1902 char buf[100];
1903 if (xtty != NULL) {
1904 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1905 os::current_thread_id(),
1906 format_trap_request(buf, sizeof(buf), trap_request));
1907 #if INCLUDE_JVMCI
1908 if (speculation != 0) {
1909 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1910 }
1911 #endif
1912 nm->log_identity(xtty);
1913 }
1914 Symbol* class_name = NULL;
1915 bool unresolved = false;
1916 if (unloaded_class_index >= 0) {
1917 constantPoolHandle constants (THREAD, trap_method->constants());
1918 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1919 class_name = constants->klass_name_at(unloaded_class_index);
1920 unresolved = true;
1921 if (xtty != NULL)
1922 xtty->print(" unresolved='1'");
1923 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1924 class_name = constants->symbol_at(unloaded_class_index);
1925 }
1926 if (xtty != NULL)
1927 xtty->name(class_name);
1928 }
1929 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1930 // Dump the relevant MDO state.
1931 // This is the deopt count for the current reason, any previous
1932 // reasons or recompiles seen at this point.
1933 int dcnt = trap_mdo->trap_count(reason);
1934 if (dcnt != 0)
1935 xtty->print(" count='%d'", dcnt);
1936 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1937 int dos = (pdata == NULL)? 0: pdata->trap_state();
1938 if (dos != 0) {
1939 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1940 if (trap_state_is_recompiled(dos)) {
1941 int recnt2 = trap_mdo->overflow_recompile_count();
1942 if (recnt2 != 0)
1943 xtty->print(" recompiles2='%d'", recnt2);
1944 }
1945 }
1946 }
1947 if (xtty != NULL) {
1948 xtty->stamp();
1949 xtty->end_head();
1950 }
1951 if (TraceDeoptimization) { // make noise on the tty
1952 tty->print("Uncommon trap occurred in");
1953 nm->method()->print_short_name(tty);
1954 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1955 #if INCLUDE_JVMCI
1956 if (nm->is_nmethod()) {
1957 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
1958 if (installed_code_name != NULL) {
1959 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1960 }
1961 }
1962 #endif
1963 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1964 p2i(fr.pc()),
1965 os::current_thread_id(),
1966 trap_reason_name(reason),
1967 trap_action_name(action),
1968 unloaded_class_index
1969 #if INCLUDE_JVMCI
1970 , debug_id
1971 #endif
1972 );
1973 if (class_name != NULL) {
1974 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1975 class_name->print_symbol_on(tty);
1976 }
1977 tty->cr();
1978 }
1979 if (xtty != NULL) {
1980 // Log the precise location of the trap.
1981 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1982 xtty->begin_elem("jvms bci='%d'", sd->bci());
1983 xtty->method(sd->method());
1984 xtty->end_elem();
1985 if (sd->is_top()) break;
1986 }
1987 xtty->tail("uncommon_trap");
1988 }
1989 }
1990 // (End diagnostic printout.)
1991
1992 // Load class if necessary
1993 if (unloaded_class_index >= 0) {
1994 constantPoolHandle constants(THREAD, trap_method->constants());
1995 load_class_by_index(constants, unloaded_class_index);
1996 }
1997
1998 // Flush the nmethod if necessary and desirable.
1999 //
2000 // We need to avoid situations where we are re-flushing the nmethod
2001 // because of a hot deoptimization site. Repeated flushes at the same
2002 // point need to be detected by the compiler and avoided. If the compiler
2003 // cannot avoid them (or has a bug and "refuses" to avoid them), this
2004 // module must take measures to avoid an infinite cycle of recompilation
2005 // and deoptimization. There are several such measures:
2006 //
2007 // 1. If a recompilation is ordered a second time at some site X
2008 // and for the same reason R, the action is adjusted to 'reinterpret',
2009 // to give the interpreter time to exercise the method more thoroughly.
2010 // If this happens, the method's overflow_recompile_count is incremented.
2011 //
2012 // 2. If the compiler fails to reduce the deoptimization rate, then
2013 // the method's overflow_recompile_count will begin to exceed the set
2014 // limit PerBytecodeRecompilationCutoff. If this happens, the action
2015 // is adjusted to 'make_not_compilable', and the method is abandoned
2016 // to the interpreter. This is a performance hit for hot methods,
2017 // but is better than a disastrous infinite cycle of recompilations.
2018 // (Actually, only the method containing the site X is abandoned.)
2019 //
2020 // 3. In parallel with the previous measures, if the total number of
2021 // recompilations of a method exceeds the much larger set limit
2022 // PerMethodRecompilationCutoff, the method is abandoned.
2023 // This should only happen if the method is very large and has
2024 // many "lukewarm" deoptimizations. The code which enforces this
2025 // limit is elsewhere (class nmethod, class Method).
2026 //
2027 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2028 // to recompile at each bytecode independently of the per-BCI cutoff.
2029 //
2030 // The decision to update code is up to the compiler, and is encoded
2031 // in the Action_xxx code. If the compiler requests Action_none
2032 // no trap state is changed, no compiled code is changed, and the
2033 // computation suffers along in the interpreter.
2034 //
2035 // The other action codes specify various tactics for decompilation
2036 // and recompilation. Action_maybe_recompile is the loosest, and
2037 // allows the compiled code to stay around until enough traps are seen,
2038 // and until the compiler gets around to recompiling the trapping method.
2039 //
2040 // The other actions cause immediate removal of the present code.
2041
2042 // Traps caused by injected profile shouldn't pollute trap counts.
2043 bool injected_profile_trap = trap_method->has_injected_profile() &&
2044 (reason == Reason_intrinsic || reason == Reason_unreached);
2045
2046 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2047 bool make_not_entrant = false;
2048 bool make_not_compilable = false;
2049 bool reprofile = false;
2050 switch (action) {
2051 case Action_none:
2052 // Keep the old code.
2053 update_trap_state = false;
2054 break;
2055 case Action_maybe_recompile:
2056 // Do not need to invalidate the present code, but we can
2057 // initiate another
2058 // Start compiler without (necessarily) invalidating the nmethod.
2059 // The system will tolerate the old code, but new code should be
2060 // generated when possible.
2061 break;
2062 case Action_reinterpret:
2063 // Go back into the interpreter for a while, and then consider
2064 // recompiling form scratch.
2065 make_not_entrant = true;
2066 // Reset invocation counter for outer most method.
2067 // This will allow the interpreter to exercise the bytecodes
2068 // for a while before recompiling.
2069 // By contrast, Action_make_not_entrant is immediate.
2070 //
2071 // Note that the compiler will track null_check, null_assert,
2072 // range_check, and class_check events and log them as if they
2073 // had been traps taken from compiled code. This will update
2074 // the MDO trap history so that the next compilation will
2075 // properly detect hot trap sites.
2076 reprofile = true;
2077 break;
2078 case Action_make_not_entrant:
2079 // Request immediate recompilation, and get rid of the old code.
2080 // Make them not entrant, so next time they are called they get
2081 // recompiled. Unloaded classes are loaded now so recompile before next
2082 // time they are called. Same for uninitialized. The interpreter will
2083 // link the missing class, if any.
2084 make_not_entrant = true;
2085 break;
2086 case Action_make_not_compilable:
2087 // Give up on compiling this method at all.
2088 make_not_entrant = true;
2089 make_not_compilable = true;
2090 break;
2091 default:
2092 ShouldNotReachHere();
2093 }
2094
2095 // Setting +ProfileTraps fixes the following, on all platforms:
2096 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2097 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2098 // recompile relies on a MethodData* to record heroic opt failures.
2099
2100 // Whether the interpreter is producing MDO data or not, we also need
2101 // to use the MDO to detect hot deoptimization points and control
2102 // aggressive optimization.
2103 bool inc_recompile_count = false;
2104 ProfileData* pdata = NULL;
2105 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
2106 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
2107 uint this_trap_count = 0;
2108 bool maybe_prior_trap = false;
2109 bool maybe_prior_recompile = false;
2110 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2111 #if INCLUDE_JVMCI
2112 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2113 #endif
2114 nm->method(),
2115 //outputs:
2116 this_trap_count,
2117 maybe_prior_trap,
2118 maybe_prior_recompile);
2119 // Because the interpreter also counts null, div0, range, and class
2120 // checks, these traps from compiled code are double-counted.
2121 // This is harmless; it just means that the PerXTrapLimit values
2122 // are in effect a little smaller than they look.
2123
2124 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2125 if (per_bc_reason != Reason_none) {
2126 // Now take action based on the partially known per-BCI history.
2127 if (maybe_prior_trap
2128 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2129 // If there are too many traps at this BCI, force a recompile.
2130 // This will allow the compiler to see the limit overflow, and
2131 // take corrective action, if possible. The compiler generally
2132 // does not use the exact PerBytecodeTrapLimit value, but instead
2133 // changes its tactics if it sees any traps at all. This provides
2134 // a little hysteresis, delaying a recompile until a trap happens
2135 // several times.
2136 //
2137 // Actually, since there is only one bit of counter per BCI,
2138 // the possible per-BCI counts are {0,1,(per-method count)}.
2139 // This produces accurate results if in fact there is only
2140 // one hot trap site, but begins to get fuzzy if there are
2141 // many sites. For example, if there are ten sites each
2142 // trapping two or more times, they each get the blame for
2143 // all of their traps.
2144 make_not_entrant = true;
2145 }
2146
2147 // Detect repeated recompilation at the same BCI, and enforce a limit.
2148 if (make_not_entrant && maybe_prior_recompile) {
2149 // More than one recompile at this point.
2150 inc_recompile_count = maybe_prior_trap;
2151 }
2152 } else {
2153 // For reasons which are not recorded per-bytecode, we simply
2154 // force recompiles unconditionally.
2155 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2156 make_not_entrant = true;
2157 }
2158
2159 // Go back to the compiler if there are too many traps in this method.
2160 if (this_trap_count >= per_method_trap_limit(reason)) {
2161 // If there are too many traps in this method, force a recompile.
2162 // This will allow the compiler to see the limit overflow, and
2163 // take corrective action, if possible.
2164 // (This condition is an unlikely backstop only, because the
2165 // PerBytecodeTrapLimit is more likely to take effect first,
2166 // if it is applicable.)
2167 make_not_entrant = true;
2168 }
2169
2170 // Here's more hysteresis: If there has been a recompile at
2171 // this trap point already, run the method in the interpreter
2172 // for a while to exercise it more thoroughly.
2173 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2174 reprofile = true;
2175 }
2176 }
2177
2178 // Take requested actions on the method:
2179
2180 // Recompile
2181 if (make_not_entrant) {
2182 if (!nm->make_not_entrant()) {
2183 return; // the call did not change nmethod's state
2184 }
2185
2186 if (pdata != NULL) {
2187 // Record the recompilation event, if any.
2188 int tstate0 = pdata->trap_state();
2189 int tstate1 = trap_state_set_recompiled(tstate0, true);
2190 if (tstate1 != tstate0)
2191 pdata->set_trap_state(tstate1);
2192 }
2193
2194 #if INCLUDE_RTM_OPT
2195 // Restart collecting RTM locking abort statistic if the method
2196 // is recompiled for a reason other than RTM state change.
2197 // Assume that in new recompiled code the statistic could be different,
2198 // for example, due to different inlining.
2199 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2200 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2201 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2202 }
2203 #endif
2204 // For code aging we count traps separately here, using make_not_entrant()
2205 // as a guard against simultaneous deopts in multiple threads.
2206 if (reason == Reason_tenured && trap_mdo != NULL) {
2207 trap_mdo->inc_tenure_traps();
2208 }
2209 }
2210
2211 if (inc_recompile_count) {
2212 trap_mdo->inc_overflow_recompile_count();
2213 if ((uint)trap_mdo->overflow_recompile_count() >
2214 (uint)PerBytecodeRecompilationCutoff) {
2215 // Give up on the method containing the bad BCI.
2216 if (trap_method() == nm->method()) {
2217 make_not_compilable = true;
2218 } else {
2219 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2220 // But give grace to the enclosing nm->method().
2221 }
2222 }
2223 }
2224
2225 // Reprofile
2226 if (reprofile) {
2227 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2228 }
2229
2230 // Give up compiling
2231 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2232 assert(make_not_entrant, "consistent");
2233 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2234 }
2235
2236 } // Free marked resources
2237
2238 }
2239 JRT_END
2240
2241 ProfileData*
2242 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2243 int trap_bci,
2244 Deoptimization::DeoptReason reason,
2245 bool update_total_trap_count,
2246 #if INCLUDE_JVMCI
2247 bool is_osr,
2248 #endif
2249 Method* compiled_method,
2250 //outputs:
2251 uint& ret_this_trap_count,
2252 bool& ret_maybe_prior_trap,
2253 bool& ret_maybe_prior_recompile) {
2254 bool maybe_prior_trap = false;
2255 bool maybe_prior_recompile = false;
2256 uint this_trap_count = 0;
2257 if (update_total_trap_count) {
2258 uint idx = reason;
2259 #if INCLUDE_JVMCI
2260 if (is_osr) {
2261 idx += Reason_LIMIT;
2262 }
2263 #endif
2264 uint prior_trap_count = trap_mdo->trap_count(idx);
2265 this_trap_count = trap_mdo->inc_trap_count(idx);
2266
2267 // If the runtime cannot find a place to store trap history,
2268 // it is estimated based on the general condition of the method.
2269 // If the method has ever been recompiled, or has ever incurred
2270 // a trap with the present reason , then this BCI is assumed
2271 // (pessimistically) to be the culprit.
2272 maybe_prior_trap = (prior_trap_count != 0);
2273 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2274 }
2275 ProfileData* pdata = NULL;
2276
2277
2278 // For reasons which are recorded per bytecode, we check per-BCI data.
2279 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2280 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2281 if (per_bc_reason != Reason_none) {
2282 // Find the profile data for this BCI. If there isn't one,
2283 // try to allocate one from the MDO's set of spares.
2284 // This will let us detect a repeated trap at this point.
2285 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2286
2287 if (pdata != NULL) {
2288 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2289 if (LogCompilation && xtty != NULL) {
2290 ttyLocker ttyl;
2291 // no more room for speculative traps in this MDO
2292 xtty->elem("speculative_traps_oom");
2293 }
2294 }
2295 // Query the trap state of this profile datum.
2296 int tstate0 = pdata->trap_state();
2297 if (!trap_state_has_reason(tstate0, per_bc_reason))
2298 maybe_prior_trap = false;
2299 if (!trap_state_is_recompiled(tstate0))
2300 maybe_prior_recompile = false;
2301
2302 // Update the trap state of this profile datum.
2303 int tstate1 = tstate0;
2304 // Record the reason.
2305 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2306 // Store the updated state on the MDO, for next time.
2307 if (tstate1 != tstate0)
2308 pdata->set_trap_state(tstate1);
2309 } else {
2310 if (LogCompilation && xtty != NULL) {
2311 ttyLocker ttyl;
2312 // Missing MDP? Leave a small complaint in the log.
2313 xtty->elem("missing_mdp bci='%d'", trap_bci);
2314 }
2315 }
2316 }
2317
2318 // Return results:
2319 ret_this_trap_count = this_trap_count;
2320 ret_maybe_prior_trap = maybe_prior_trap;
2321 ret_maybe_prior_recompile = maybe_prior_recompile;
2322 return pdata;
2323 }
2324
2325 void
2326 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2327 ResourceMark rm;
2328 // Ignored outputs:
2329 uint ignore_this_trap_count;
2330 bool ignore_maybe_prior_trap;
2331 bool ignore_maybe_prior_recompile;
2332 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2333 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2334 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2335 query_update_method_data(trap_mdo, trap_bci,
2336 (DeoptReason)reason,
2337 update_total_counts,
2338 #if INCLUDE_JVMCI
2339 false,
2340 #endif
2341 NULL,
2342 ignore_this_trap_count,
2343 ignore_maybe_prior_trap,
2344 ignore_maybe_prior_recompile);
2345 }
2346
2347 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2348 if (TraceDeoptimization) {
2349 tty->print("Uncommon trap ");
2350 }
2351 // Still in Java no safepoints
2352 {
2353 // This enters VM and may safepoint
2354 uncommon_trap_inner(thread, trap_request);
2355 }
2356 return fetch_unroll_info_helper(thread, exec_mode);
2357 }
2358
2359 // Local derived constants.
2360 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2361 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2362 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2363
2364 //---------------------------trap_state_reason---------------------------------
2365 Deoptimization::DeoptReason
2366 Deoptimization::trap_state_reason(int trap_state) {
2367 // This assert provides the link between the width of DataLayout::trap_bits
2368 // and the encoding of "recorded" reasons. It ensures there are enough
2369 // bits to store all needed reasons in the per-BCI MDO profile.
2370 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2371 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2372 trap_state -= recompile_bit;
2373 if (trap_state == DS_REASON_MASK) {
2374 return Reason_many;
2375 } else {
2376 assert((int)Reason_none == 0, "state=0 => Reason_none");
2377 return (DeoptReason)trap_state;
2378 }
2379 }
2380 //-------------------------trap_state_has_reason-------------------------------
2381 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2382 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2383 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2384 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2385 trap_state -= recompile_bit;
2386 if (trap_state == DS_REASON_MASK) {
2387 return -1; // true, unspecifically (bottom of state lattice)
2388 } else if (trap_state == reason) {
2389 return 1; // true, definitely
2390 } else if (trap_state == 0) {
2391 return 0; // false, definitely (top of state lattice)
2392 } else {
2393 return 0; // false, definitely
2394 }
2395 }
2396 //-------------------------trap_state_add_reason-------------------------------
2397 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2398 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2399 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2400 trap_state -= recompile_bit;
2401 if (trap_state == DS_REASON_MASK) {
2402 return trap_state + recompile_bit; // already at state lattice bottom
2403 } else if (trap_state == reason) {
2404 return trap_state + recompile_bit; // the condition is already true
2405 } else if (trap_state == 0) {
2406 return reason + recompile_bit; // no condition has yet been true
2407 } else {
2408 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2409 }
2410 }
2411 //-----------------------trap_state_is_recompiled------------------------------
2412 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2413 return (trap_state & DS_RECOMPILE_BIT) != 0;
2414 }
2415 //-----------------------trap_state_set_recompiled-----------------------------
2416 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2417 if (z) return trap_state | DS_RECOMPILE_BIT;
2418 else return trap_state & ~DS_RECOMPILE_BIT;
2419 }
2420 //---------------------------format_trap_state---------------------------------
2421 // This is used for debugging and diagnostics, including LogFile output.
2422 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2423 int trap_state) {
2424 assert(buflen > 0, "sanity");
2425 DeoptReason reason = trap_state_reason(trap_state);
2426 bool recomp_flag = trap_state_is_recompiled(trap_state);
2427 // Re-encode the state from its decoded components.
2428 int decoded_state = 0;
2429 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2430 decoded_state = trap_state_add_reason(decoded_state, reason);
2431 if (recomp_flag)
2432 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2433 // If the state re-encodes properly, format it symbolically.
2434 // Because this routine is used for debugging and diagnostics,
2435 // be robust even if the state is a strange value.
2436 size_t len;
2437 if (decoded_state != trap_state) {
2438 // Random buggy state that doesn't decode??
2439 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2440 } else {
2441 len = jio_snprintf(buf, buflen, "%s%s",
2442 trap_reason_name(reason),
2443 recomp_flag ? " recompiled" : "");
2444 }
2445 return buf;
2446 }
2447
2448
2449 //--------------------------------statics--------------------------------------
2450 const char* Deoptimization::_trap_reason_name[] = {
2451 // Note: Keep this in sync. with enum DeoptReason.
2452 "none",
2453 "null_check",
2454 "null_assert" JVMCI_ONLY("_or_unreached0"),
2455 "range_check",
2456 "class_check",
2457 "array_check",
2458 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2459 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2460 "profile_predicate",
2461 "unloaded",
2462 "uninitialized",
2463 "initialized",
2464 "unreached",
2465 "unhandled",
2466 "constraint",
2467 "div0_check",
2468 "age",
2469 "predicate",
2470 "loop_limit_check",
2471 "speculate_class_check",
2472 "speculate_null_check",
2473 "speculate_null_assert",
2474 "rtm_state_change",
2475 "unstable_if",
2476 "unstable_fused_if",
2477 #if INCLUDE_JVMCI
2478 "aliasing",
2479 "transfer_to_interpreter",
2480 "not_compiled_exception_handler",
2481 "unresolved",
2482 "jsr_mismatch",
2483 #endif
2484 "tenured"
2485 };
2486 const char* Deoptimization::_trap_action_name[] = {
2487 // Note: Keep this in sync. with enum DeoptAction.
2488 "none",
2489 "maybe_recompile",
2490 "reinterpret",
2491 "make_not_entrant",
2492 "make_not_compilable"
2493 };
2494
2495 const char* Deoptimization::trap_reason_name(int reason) {
2496 // Check that every reason has a name
2497 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2498
2499 if (reason == Reason_many) return "many";
2500 if ((uint)reason < Reason_LIMIT)
2501 return _trap_reason_name[reason];
2502 static char buf[20];
2503 sprintf(buf, "reason%d", reason);
2504 return buf;
2505 }
2506 const char* Deoptimization::trap_action_name(int action) {
2507 // Check that every action has a name
2508 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2509
2510 if ((uint)action < Action_LIMIT)
2511 return _trap_action_name[action];
2512 static char buf[20];
2513 sprintf(buf, "action%d", action);
2514 return buf;
2515 }
2516
2517 // This is used for debugging and diagnostics, including LogFile output.
2518 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2519 int trap_request) {
2520 jint unloaded_class_index = trap_request_index(trap_request);
2521 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2522 const char* action = trap_action_name(trap_request_action(trap_request));
2523 #if INCLUDE_JVMCI
2524 int debug_id = trap_request_debug_id(trap_request);
2525 #endif
2526 size_t len;
2527 if (unloaded_class_index < 0) {
2528 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2529 reason, action
2530 #if INCLUDE_JVMCI
2531 ,debug_id
2532 #endif
2533 );
2534 } else {
2535 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2536 reason, action, unloaded_class_index
2537 #if INCLUDE_JVMCI
2538 ,debug_id
2539 #endif
2540 );
2541 }
2542 return buf;
2543 }
2544
2545 juint Deoptimization::_deoptimization_hist
2546 [Deoptimization::Reason_LIMIT]
2547 [1 + Deoptimization::Action_LIMIT]
2548 [Deoptimization::BC_CASE_LIMIT]
2549 = {0};
2550
2551 enum {
2552 LSB_BITS = 8,
2553 LSB_MASK = right_n_bits(LSB_BITS)
2554 };
2555
2556 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2557 Bytecodes::Code bc) {
2558 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2559 assert(action >= 0 && action < Action_LIMIT, "oob");
2560 _deoptimization_hist[Reason_none][0][0] += 1; // total
2561 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2562 juint* cases = _deoptimization_hist[reason][1+action];
2563 juint* bc_counter_addr = NULL;
2564 juint bc_counter = 0;
2565 // Look for an unused counter, or an exact match to this BC.
2566 if (bc != Bytecodes::_illegal) {
2567 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2568 juint* counter_addr = &cases[bc_case];
2569 juint counter = *counter_addr;
2570 if ((counter == 0 && bc_counter_addr == NULL)
2571 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2572 // this counter is either free or is already devoted to this BC
2573 bc_counter_addr = counter_addr;
2574 bc_counter = counter | bc;
2575 }
2576 }
2577 }
2578 if (bc_counter_addr == NULL) {
2579 // Overflow, or no given bytecode.
2580 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2581 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2582 }
2583 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2584 }
2585
2586 jint Deoptimization::total_deoptimization_count() {
2587 return _deoptimization_hist[Reason_none][0][0];
2588 }
2589
2590 void Deoptimization::print_statistics() {
2591 juint total = total_deoptimization_count();
2592 juint account = total;
2593 if (total != 0) {
2594 ttyLocker ttyl;
2595 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2596 tty->print_cr("Deoptimization traps recorded:");
2597 #define PRINT_STAT_LINE(name, r) \
2598 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2599 PRINT_STAT_LINE("total", total);
2600 // For each non-zero entry in the histogram, print the reason,
2601 // the action, and (if specifically known) the type of bytecode.
2602 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2603 for (int action = 0; action < Action_LIMIT; action++) {
2604 juint* cases = _deoptimization_hist[reason][1+action];
2605 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2606 juint counter = cases[bc_case];
2607 if (counter != 0) {
2608 char name[1*K];
2609 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2610 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2611 bc = Bytecodes::_illegal;
2612 sprintf(name, "%s/%s/%s",
2613 trap_reason_name(reason),
2614 trap_action_name(action),
2615 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2616 juint r = counter >> LSB_BITS;
2617 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2618 account -= r;
2619 }
2620 }
2621 }
2622 }
2623 if (account != 0) {
2624 PRINT_STAT_LINE("unaccounted", account);
2625 }
2626 #undef PRINT_STAT_LINE
2627 if (xtty != NULL) xtty->tail("statistics");
2628 }
2629 }
2630 #else // COMPILER2_OR_JVMCI
2631
2632
2633 // Stubs for C1 only system.
2634 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2635 return false;
2636 }
2637
2638 const char* Deoptimization::trap_reason_name(int reason) {
2639 return "unknown";
2640 }
2641
2642 void Deoptimization::print_statistics() {
2643 // no output
2644 }
2645
2646 void
2647 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2648 // no udpate
2649 }
2650
2651 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2652 return 0;
2653 }
2654
2655 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2656 Bytecodes::Code bc) {
2657 // no update
2658 }
2659
2660 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2661 int trap_state) {
2662 jio_snprintf(buf, buflen, "#%d", trap_state);
2663 return buf;
2664 }
2665
2666 #endif // COMPILER2_OR_JVMCI