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