1
2
3 /*
4 * Copyright (c) 1997, 2019, 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 // restore elements of an eliminated type array
1036 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1037 int index = 0;
1038 intptr_t val;
1039
1040 for (int i = 0; i < sv->field_size(); i++) {
1041 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1042 switch(type) {
1043 case T_LONG: case T_DOUBLE: {
1044 assert(value->type() == T_INT, "Agreement.");
1045 StackValue* low =
1046 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1047 #ifdef _LP64
1048 jlong res = (jlong)low->get_int();
1049 #else
1050 #ifdef SPARC
1051 // For SPARC we have to swap high and low words.
1052 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1053 #else
1054 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1055 #endif //SPARC
1056 #endif
1057 obj->long_at_put(index, res);
1058 break;
1059 }
1060
1061 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1062 case T_INT: case T_FLOAT: { // 4 bytes.
1063 assert(value->type() == T_INT, "Agreement.");
1064 bool big_value = false;
1065 if (i + 1 < sv->field_size() && type == T_INT) {
1066 if (sv->field_at(i)->is_location()) {
1067 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1068 if (type == Location::dbl || type == Location::lng) {
1069 big_value = true;
1070 }
1071 } else if (sv->field_at(i)->is_constant_int()) {
1072 ScopeValue* next_scope_field = sv->field_at(i + 1);
1073 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1074 big_value = true;
1075 }
1076 }
1077 }
1078
1079 if (big_value) {
1080 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1081 #ifdef _LP64
1082 jlong res = (jlong)low->get_int();
1083 #else
1084 #ifdef SPARC
1085 // For SPARC we have to swap high and low words.
1086 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1087 #else
1088 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1089 #endif //SPARC
1090 #endif
1091 obj->int_at_put(index, (jint)*((jint*)&res));
1092 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1093 } else {
1094 val = value->get_int();
1095 obj->int_at_put(index, (jint)*((jint*)&val));
1096 }
1097 break;
1098 }
1099
1100 case T_SHORT:
1101 assert(value->type() == T_INT, "Agreement.");
1102 val = value->get_int();
1103 obj->short_at_put(index, (jshort)*((jint*)&val));
1104 break;
1105
1106 case T_CHAR:
1107 assert(value->type() == T_INT, "Agreement.");
1108 val = value->get_int();
1109 obj->char_at_put(index, (jchar)*((jint*)&val));
1110 break;
1111
1112 case T_BYTE:
1113 assert(value->type() == T_INT, "Agreement.");
1114 val = value->get_int();
1115 obj->byte_at_put(index, (jbyte)*((jint*)&val));
1116 break;
1117
1118 case T_BOOLEAN:
1119 assert(value->type() == T_INT, "Agreement.");
1120 val = value->get_int();
1121 obj->bool_at_put(index, (jboolean)*((jint*)&val));
1122 break;
1123
1124 default:
1125 ShouldNotReachHere();
1126 }
1127 index++;
1128 }
1129 }
1130
1131
1132 // restore fields of an eliminated object array
1133 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1134 for (int i = 0; i < sv->field_size(); i++) {
1135 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1136 assert(value->type() == T_OBJECT, "object element expected");
1137 obj->obj_at_put(i, value->get_obj()());
1138 }
1139 }
1140
1141 class ReassignedField {
1142 public:
1143 int _offset;
1144 BasicType _type;
1145 public:
1146 ReassignedField() {
1147 _offset = 0;
1148 _type = T_ILLEGAL;
1149 }
1150 };
1151
1152 int compare(ReassignedField* left, ReassignedField* right) {
1153 return left->_offset - right->_offset;
1154 }
1155
1156 // Restore fields of an eliminated instance object using the same field order
1157 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1158 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1159 if (klass->superklass() != NULL) {
1160 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal);
1161 }
1162
1163 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1164 for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1165 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1166 ReassignedField field;
1167 field._offset = fs.offset();
1168 field._type = FieldType::basic_type(fs.signature());
1169 fields->append(field);
1170 }
1171 }
1172 fields->sort(compare);
1173 for (int i = 0; i < fields->length(); i++) {
1174 intptr_t val;
1175 ScopeValue* scope_field = sv->field_at(svIndex);
1176 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1177 int offset = fields->at(i)._offset;
1178 BasicType type = fields->at(i)._type;
1179 switch (type) {
1180 case T_OBJECT: case T_ARRAY:
1181 assert(value->type() == T_OBJECT, "Agreement.");
1182 obj->obj_field_put(offset, value->get_obj()());
1183 break;
1184
1185 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1186 case T_INT: case T_FLOAT: { // 4 bytes.
1187 assert(value->type() == T_INT, "Agreement.");
1188 bool big_value = false;
1189 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1190 if (scope_field->is_location()) {
1191 Location::Type type = ((LocationValue*) scope_field)->location().type();
1192 if (type == Location::dbl || type == Location::lng) {
1193 big_value = true;
1194 }
1195 }
1196 if (scope_field->is_constant_int()) {
1197 ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1198 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1199 big_value = true;
1200 }
1201 }
1202 }
1203
1204 if (big_value) {
1205 i++;
1206 assert(i < fields->length(), "second T_INT field needed");
1207 assert(fields->at(i)._type == T_INT, "T_INT field needed");
1208 } else {
1209 val = value->get_int();
1210 obj->int_field_put(offset, (jint)*((jint*)&val));
1211 break;
1212 }
1213 }
1214 /* no break */
1215
1216 case T_LONG: case T_DOUBLE: {
1217 assert(value->type() == T_INT, "Agreement.");
1218 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1219 #ifdef _LP64
1220 jlong res = (jlong)low->get_int();
1221 #else
1222 #ifdef SPARC
1223 // For SPARC we have to swap high and low words.
1224 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
1225 #else
1226 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1227 #endif //SPARC
1228 #endif
1229 obj->long_field_put(offset, res);
1230 break;
1231 }
1232
1233 case T_SHORT:
1234 assert(value->type() == T_INT, "Agreement.");
1235 val = value->get_int();
1236 obj->short_field_put(offset, (jshort)*((jint*)&val));
1237 break;
1238
1239 case T_CHAR:
1240 assert(value->type() == T_INT, "Agreement.");
1241 val = value->get_int();
1242 obj->char_field_put(offset, (jchar)*((jint*)&val));
1243 break;
1244
1245 case T_BYTE:
1246 assert(value->type() == T_INT, "Agreement.");
1247 val = value->get_int();
1248 obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1249 break;
1250
1251 case T_BOOLEAN:
1252 assert(value->type() == T_INT, "Agreement.");
1253 val = value->get_int();
1254 obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1255 break;
1256
1257 default:
1258 ShouldNotReachHere();
1259 }
1260 svIndex++;
1261 }
1262 return svIndex;
1263 }
1264
1265 // restore fields of all eliminated objects and arrays
1266 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1267 for (int i = 0; i < objects->length(); i++) {
1268 ObjectValue* sv = (ObjectValue*) objects->at(i);
1269 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1270 Handle obj = sv->value();
1271 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1272 if (PrintDeoptimizationDetails) {
1273 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1274 }
1275 if (obj.is_null()) {
1276 continue;
1277 }
1278 #if INCLUDE_JVMCI || INCLUDE_AOT
1279 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1280 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1281 continue;
1282 }
1283 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1284 if (k->is_instance_klass()) {
1285 InstanceKlass* ik = InstanceKlass::cast(k);
1286 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1287 } else if (k->is_typeArray_klass()) {
1288 TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1289 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1290 } else if (k->is_objArray_klass()) {
1291 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1292 }
1293 }
1294 }
1295
1296
1297 // relock objects for which synchronization was eliminated
1298 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
1299 for (int i = 0; i < monitors->length(); i++) {
1300 MonitorInfo* mon_info = monitors->at(i);
1301 if (mon_info->eliminated()) {
1302 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1303 if (!mon_info->owner_is_scalar_replaced()) {
1304 Handle obj(thread, mon_info->owner());
1305 markWord mark = obj->mark();
1306 if (UseBiasedLocking && mark.has_bias_pattern()) {
1307 // New allocated objects may have the mark set to anonymously biased.
1308 // Also the deoptimized method may called methods with synchronization
1309 // where the thread-local object is bias locked to the current thread.
1310 assert(mark.is_biased_anonymously() ||
1311 mark.biased_locker() == thread, "should be locked to current thread");
1312 // Reset mark word to unbiased prototype.
1313 markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1314 obj->set_mark(unbiased_prototype);
1315 }
1316 BasicLock* lock = mon_info->lock();
1317 ObjectSynchronizer::enter(obj, lock, thread);
1318 assert(mon_info->owner()->is_locked(), "object must be locked now");
1319 }
1320 }
1321 }
1322 }
1323
1324
1325 #ifndef PRODUCT
1326 // print information about reallocated objects
1327 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1328 fieldDescriptor fd;
1329
1330 for (int i = 0; i < objects->length(); i++) {
1331 ObjectValue* sv = (ObjectValue*) objects->at(i);
1332 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1333 Handle obj = sv->value();
1334
1335 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1336 k->print_value();
1337 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1338 if (obj.is_null()) {
1339 tty->print(" allocation failed");
1340 } else {
1341 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1342 }
1343 tty->cr();
1344
1345 if (Verbose && !obj.is_null()) {
1346 k->oop_print_on(obj(), tty);
1347 }
1348 }
1349 }
1350 #endif
1351 #endif // COMPILER2_OR_JVMCI
1352
1353 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1354 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1355
1356 #ifndef PRODUCT
1357 if (PrintDeoptimizationDetails) {
1358 ttyLocker ttyl;
1359 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1360 fr.print_on(tty);
1361 tty->print_cr(" Virtual frames (innermost first):");
1362 for (int index = 0; index < chunk->length(); index++) {
1363 compiledVFrame* vf = chunk->at(index);
1364 tty->print(" %2d - ", index);
1365 vf->print_value();
1366 int bci = chunk->at(index)->raw_bci();
1367 const char* code_name;
1368 if (bci == SynchronizationEntryBCI) {
1369 code_name = "sync entry";
1370 } else {
1371 Bytecodes::Code code = vf->method()->code_at(bci);
1372 code_name = Bytecodes::name(code);
1373 }
1374 tty->print(" - %s", code_name);
1375 tty->print_cr(" @ bci %d ", bci);
1376 if (Verbose) {
1377 vf->print();
1378 tty->cr();
1379 }
1380 }
1381 }
1382 #endif
1383
1384 // Register map for next frame (used for stack crawl). We capture
1385 // the state of the deopt'ing frame's caller. Thus if we need to
1386 // stuff a C2I adapter we can properly fill in the callee-save
1387 // register locations.
1388 frame caller = fr.sender(reg_map);
1389 int frame_size = caller.sp() - fr.sp();
1390
1391 frame sender = caller;
1392
1393 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1394 // the vframeArray containing the unpacking information is allocated in the C heap.
1395 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1396 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1397
1398 // Compare the vframeArray to the collected vframes
1399 assert(array->structural_compare(thread, chunk), "just checking");
1400
1401 #ifndef PRODUCT
1402 if (PrintDeoptimizationDetails) {
1403 ttyLocker ttyl;
1404 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array));
1405 }
1406 #endif // PRODUCT
1407
1408 return array;
1409 }
1410
1411 #if COMPILER2_OR_JVMCI
1412 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1413 // Reallocation of some scalar replaced objects failed. Record
1414 // that we need to pop all the interpreter frames for the
1415 // deoptimized compiled frame.
1416 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1417 thread->set_frames_to_pop_failed_realloc(array->frames());
1418 // Unlock all monitors here otherwise the interpreter will see a
1419 // mix of locked and unlocked monitors (because of failed
1420 // reallocations of synchronized objects) and be confused.
1421 for (int i = 0; i < array->frames(); i++) {
1422 MonitorChunk* monitors = array->element(i)->monitors();
1423 if (monitors != NULL) {
1424 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1425 BasicObjectLock* src = monitors->at(j);
1426 if (src->obj() != NULL) {
1427 ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1428 }
1429 }
1430 array->element(i)->free_monitors(thread);
1431 #ifdef ASSERT
1432 array->element(i)->set_removed_monitors();
1433 #endif
1434 }
1435 }
1436 }
1437 #endif
1438
1439 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1440 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1441 Thread* thread = Thread::current();
1442 for (int i = 0; i < monitors->length(); i++) {
1443 MonitorInfo* mon_info = monitors->at(i);
1444 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1445 objects_to_revoke->append(Handle(thread, mon_info->owner()));
1446 }
1447 }
1448 }
1449
1450 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) {
1451 // Unfortunately we don't have a RegisterMap available in most of
1452 // the places we want to call this routine so we need to walk the
1453 // stack again to update the register map.
1454 if (map == NULL || !map->update_map()) {
1455 StackFrameStream sfs(thread, true);
1456 bool found = false;
1457 while (!found && !sfs.is_done()) {
1458 frame* cur = sfs.current();
1459 sfs.next();
1460 found = cur->id() == fr.id();
1461 }
1462 assert(found, "frame to be deoptimized not found on target thread's stack");
1463 map = sfs.register_map();
1464 }
1465
1466 vframe* vf = vframe::new_vframe(&fr, map, thread);
1467 compiledVFrame* cvf = compiledVFrame::cast(vf);
1468 // Revoke monitors' biases in all scopes
1469 while (!cvf->is_top()) {
1470 collect_monitors(cvf, objects_to_revoke);
1471 cvf = compiledVFrame::cast(cvf->sender());
1472 }
1473 collect_monitors(cvf, objects_to_revoke);
1474 }
1475
1476 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) {
1477 if (!UseBiasedLocking) {
1478 return;
1479 }
1480 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1481 get_monitors_from_stack(objects_to_revoke, thread, fr, map);
1482
1483 int len = objects_to_revoke->length();
1484 for (int i = 0; i < len; i++) {
1485 oop obj = (objects_to_revoke->at(i))();
1486 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread);
1487 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1488 }
1489 }
1490
1491
1492 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1493 assert(fr.can_be_deoptimized(), "checking frame type");
1494
1495 gather_statistics(reason, Action_none, Bytecodes::_illegal);
1496
1497 if (LogCompilation && xtty != NULL) {
1498 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1499 assert(cm != NULL, "only compiled methods can deopt");
1500
1501 ttyLocker ttyl;
1502 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1503 cm->log_identity(xtty);
1504 xtty->end_head();
1505 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1506 xtty->begin_elem("jvms bci='%d'", sd->bci());
1507 xtty->method(sd->method());
1508 xtty->end_elem();
1509 if (sd->is_top()) break;
1510 }
1511 xtty->tail("deoptimized");
1512 }
1513
1514 // Patch the compiled method so that when execution returns to it we will
1515 // deopt the execution state and return to the interpreter.
1516 fr.deoptimize(thread);
1517 }
1518
1519 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) {
1520 // Deoptimize only if the frame comes from compile code.
1521 // Do not deoptimize the frame which is already patched
1522 // during the execution of the loops below.
1523 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1524 return;
1525 }
1526 ResourceMark rm;
1527 DeoptimizationMarker dm;
1528 deoptimize_single_frame(thread, fr, reason);
1529 }
1530
1531 #if INCLUDE_JVMCI
1532 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1533 // there is no exception handler for this pc => deoptimize
1534 cm->make_not_entrant();
1535
1536 // Use Deoptimization::deoptimize for all of its side-effects:
1537 // revoking biases of monitors, gathering traps statistics, logging...
1538 // it also patches the return pc but we do not care about that
1539 // since we return a continuation to the deopt_blob below.
1540 JavaThread* thread = JavaThread::current();
1541 RegisterMap reg_map(thread, UseBiasedLocking);
1542 frame runtime_frame = thread->last_frame();
1543 frame caller_frame = runtime_frame.sender(®_map);
1544 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1545 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler);
1546
1547 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1548 if (trap_mdo != NULL) {
1549 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1550 }
1551
1552 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1553 }
1554 #endif
1555
1556 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1557 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1558 "can only deoptimize other thread at a safepoint");
1559 // Compute frame and register map based on thread and sp.
1560 RegisterMap reg_map(thread, UseBiasedLocking);
1561 frame fr = thread->last_frame();
1562 while (fr.id() != id) {
1563 fr = fr.sender(®_map);
1564 }
1565 deoptimize(thread, fr, ®_map, reason);
1566 }
1567
1568
1569 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1570 if (thread == Thread::current()) {
1571 Deoptimization::deoptimize_frame_internal(thread, id, reason);
1572 } else {
1573 VM_DeoptimizeFrame deopt(thread, id, reason);
1574 VMThread::execute(&deopt);
1575 }
1576 }
1577
1578 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1579 deoptimize_frame(thread, id, Reason_constraint);
1580 }
1581
1582 // JVMTI PopFrame support
1583 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1584 {
1585 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1586 }
1587 JRT_END
1588
1589 MethodData*
1590 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1591 bool create_if_missing) {
1592 Thread* THREAD = thread;
1593 MethodData* mdo = m()->method_data();
1594 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1595 // Build an MDO. Ignore errors like OutOfMemory;
1596 // that simply means we won't have an MDO to update.
1597 Method::build_interpreter_method_data(m, THREAD);
1598 if (HAS_PENDING_EXCEPTION) {
1599 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1600 CLEAR_PENDING_EXCEPTION;
1601 }
1602 mdo = m()->method_data();
1603 }
1604 return mdo;
1605 }
1606
1607 #if COMPILER2_OR_JVMCI
1608 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1609 // in case of an unresolved klass entry, load the class.
1610 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1611 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1612 return;
1613 }
1614
1615 if (!constant_pool->tag_at(index).is_symbol()) return;
1616
1617 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1618 Symbol* symbol = constant_pool->symbol_at(index);
1619
1620 // class name?
1621 if (symbol->char_at(0) != '(') {
1622 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1623 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1624 return;
1625 }
1626
1627 // then it must be a signature!
1628 ResourceMark rm(THREAD);
1629 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1630 if (ss.is_object()) {
1631 Symbol* class_name = ss.as_symbol();
1632 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1633 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1634 }
1635 }
1636 }
1637
1638
1639 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1640 EXCEPTION_MARK;
1641 load_class_by_index(constant_pool, index, THREAD);
1642 if (HAS_PENDING_EXCEPTION) {
1643 // Exception happened during classloading. We ignore the exception here, since it
1644 // is going to be rethrown since the current activation is going to be deoptimized and
1645 // the interpreter will re-execute the bytecode.
1646 CLEAR_PENDING_EXCEPTION;
1647 // Class loading called java code which may have caused a stack
1648 // overflow. If the exception was thrown right before the return
1649 // to the runtime the stack is no longer guarded. Reguard the
1650 // stack otherwise if we return to the uncommon trap blob and the
1651 // stack bang causes a stack overflow we crash.
1652 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1653 JavaThread* thread = (JavaThread*)THREAD;
1654 bool guard_pages_enabled = thread->stack_guards_enabled();
1655 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1656 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1657 }
1658 }
1659
1660 #if INCLUDE_JFR
1661
1662 class DeoptReasonSerializer : public JfrSerializer {
1663 public:
1664 void serialize(JfrCheckpointWriter& writer) {
1665 writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1666 for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1667 writer.write_key((u8)i);
1668 writer.write(Deoptimization::trap_reason_name(i));
1669 }
1670 }
1671 };
1672
1673 class DeoptActionSerializer : public JfrSerializer {
1674 public:
1675 void serialize(JfrCheckpointWriter& writer) {
1676 static const u4 nof_actions = Deoptimization::Action_LIMIT;
1677 writer.write_count(nof_actions);
1678 for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1679 writer.write_key(i);
1680 writer.write(Deoptimization::trap_action_name((int)i));
1681 }
1682 }
1683 };
1684
1685 static void register_serializers() {
1686 static int critical_section = 0;
1687 if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1688 return;
1689 }
1690 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1691 JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1692 }
1693
1694 static void post_deoptimization_event(CompiledMethod* nm,
1695 const Method* method,
1696 int trap_bci,
1697 int instruction,
1698 Deoptimization::DeoptReason reason,
1699 Deoptimization::DeoptAction action) {
1700 assert(nm != NULL, "invariant");
1701 assert(method != NULL, "invariant");
1702 if (EventDeoptimization::is_enabled()) {
1703 static bool serializers_registered = false;
1704 if (!serializers_registered) {
1705 register_serializers();
1706 serializers_registered = true;
1707 }
1708 EventDeoptimization event;
1709 event.set_compileId(nm->compile_id());
1710 event.set_compiler(nm->compiler_type());
1711 event.set_method(method);
1712 event.set_lineNumber(method->line_number_from_bci(trap_bci));
1713 event.set_bci(trap_bci);
1714 event.set_instruction(instruction);
1715 event.set_reason(reason);
1716 event.set_action(action);
1717 event.commit();
1718 }
1719 }
1720
1721 #endif // INCLUDE_JFR
1722
1723 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1724 HandleMark hm;
1725
1726 // uncommon_trap() is called at the beginning of the uncommon trap
1727 // handler. Note this fact before we start generating temporary frames
1728 // that can confuse an asynchronous stack walker. This counter is
1729 // decremented at the end of unpack_frames().
1730 thread->inc_in_deopt_handler();
1731
1732 // We need to update the map if we have biased locking.
1733 #if INCLUDE_JVMCI
1734 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1735 RegisterMap reg_map(thread, true);
1736 #else
1737 RegisterMap reg_map(thread, UseBiasedLocking);
1738 #endif
1739 frame stub_frame = thread->last_frame();
1740 frame fr = stub_frame.sender(®_map);
1741 // Make sure the calling nmethod is not getting deoptimized and removed
1742 // before we are done with it.
1743 nmethodLocker nl(fr.pc());
1744
1745 // Log a message
1746 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1747 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1748
1749 {
1750 ResourceMark rm;
1751
1752 DeoptReason reason = trap_request_reason(trap_request);
1753 DeoptAction action = trap_request_action(trap_request);
1754 #if INCLUDE_JVMCI
1755 int debug_id = trap_request_debug_id(trap_request);
1756 #endif
1757 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1758
1759 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1760 compiledVFrame* cvf = compiledVFrame::cast(vf);
1761
1762 CompiledMethod* nm = cvf->code();
1763
1764 ScopeDesc* trap_scope = cvf->scope();
1765
1766 if (TraceDeoptimization) {
1767 ttyLocker ttyl;
1768 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()
1769 #if INCLUDE_JVMCI
1770 , debug_id
1771 #endif
1772 );
1773 }
1774
1775 methodHandle trap_method(THREAD, trap_scope->method());
1776 int trap_bci = trap_scope->bci();
1777 #if INCLUDE_JVMCI
1778 jlong speculation = thread->pending_failed_speculation();
1779 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1780 nm->as_nmethod()->update_speculation(thread);
1781 } else {
1782 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1783 }
1784
1785 if (trap_bci == SynchronizationEntryBCI) {
1786 trap_bci = 0;
1787 thread->set_pending_monitorenter(true);
1788 }
1789
1790 if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1791 thread->set_pending_transfer_to_interpreter(true);
1792 }
1793 #endif
1794
1795 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1796 // Record this event in the histogram.
1797 gather_statistics(reason, action, trap_bc);
1798
1799 // Ensure that we can record deopt. history:
1800 // Need MDO to record RTM code generation state.
1801 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1802
1803 methodHandle profiled_method;
1804 #if INCLUDE_JVMCI
1805 if (nm->is_compiled_by_jvmci()) {
1806 profiled_method = methodHandle(THREAD, nm->method());
1807 } else {
1808 profiled_method = trap_method;
1809 }
1810 #else
1811 profiled_method = trap_method;
1812 #endif
1813
1814 MethodData* trap_mdo =
1815 get_method_data(thread, profiled_method, create_if_missing);
1816
1817 JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);)
1818
1819 // Log a message
1820 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1821 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1822 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1823
1824 // Print a bunch of diagnostics, if requested.
1825 if (TraceDeoptimization || LogCompilation) {
1826 ResourceMark rm;
1827 ttyLocker ttyl;
1828 char buf[100];
1829 if (xtty != NULL) {
1830 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1831 os::current_thread_id(),
1832 format_trap_request(buf, sizeof(buf), trap_request));
1833 #if INCLUDE_JVMCI
1834 if (speculation != 0) {
1835 xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1836 }
1837 #endif
1838 nm->log_identity(xtty);
1839 }
1840 Symbol* class_name = NULL;
1841 bool unresolved = false;
1842 if (unloaded_class_index >= 0) {
1843 constantPoolHandle constants (THREAD, trap_method->constants());
1844 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1845 class_name = constants->klass_name_at(unloaded_class_index);
1846 unresolved = true;
1847 if (xtty != NULL)
1848 xtty->print(" unresolved='1'");
1849 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1850 class_name = constants->symbol_at(unloaded_class_index);
1851 }
1852 if (xtty != NULL)
1853 xtty->name(class_name);
1854 }
1855 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1856 // Dump the relevant MDO state.
1857 // This is the deopt count for the current reason, any previous
1858 // reasons or recompiles seen at this point.
1859 int dcnt = trap_mdo->trap_count(reason);
1860 if (dcnt != 0)
1861 xtty->print(" count='%d'", dcnt);
1862 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1863 int dos = (pdata == NULL)? 0: pdata->trap_state();
1864 if (dos != 0) {
1865 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1866 if (trap_state_is_recompiled(dos)) {
1867 int recnt2 = trap_mdo->overflow_recompile_count();
1868 if (recnt2 != 0)
1869 xtty->print(" recompiles2='%d'", recnt2);
1870 }
1871 }
1872 }
1873 if (xtty != NULL) {
1874 xtty->stamp();
1875 xtty->end_head();
1876 }
1877 if (TraceDeoptimization) { // make noise on the tty
1878 tty->print("Uncommon trap occurred in");
1879 nm->method()->print_short_name(tty);
1880 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
1881 #if INCLUDE_JVMCI
1882 if (nm->is_nmethod()) {
1883 const char* installed_code_name = nm->as_nmethod()->jvmci_name();
1884 if (installed_code_name != NULL) {
1885 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
1886 }
1887 }
1888 #endif
1889 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
1890 p2i(fr.pc()),
1891 os::current_thread_id(),
1892 trap_reason_name(reason),
1893 trap_action_name(action),
1894 unloaded_class_index
1895 #if INCLUDE_JVMCI
1896 , debug_id
1897 #endif
1898 );
1899 if (class_name != NULL) {
1900 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1901 class_name->print_symbol_on(tty);
1902 }
1903 tty->cr();
1904 }
1905 if (xtty != NULL) {
1906 // Log the precise location of the trap.
1907 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1908 xtty->begin_elem("jvms bci='%d'", sd->bci());
1909 xtty->method(sd->method());
1910 xtty->end_elem();
1911 if (sd->is_top()) break;
1912 }
1913 xtty->tail("uncommon_trap");
1914 }
1915 }
1916 // (End diagnostic printout.)
1917
1918 // Load class if necessary
1919 if (unloaded_class_index >= 0) {
1920 constantPoolHandle constants(THREAD, trap_method->constants());
1921 load_class_by_index(constants, unloaded_class_index);
1922 }
1923
1924 // Flush the nmethod if necessary and desirable.
1925 //
1926 // We need to avoid situations where we are re-flushing the nmethod
1927 // because of a hot deoptimization site. Repeated flushes at the same
1928 // point need to be detected by the compiler and avoided. If the compiler
1929 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1930 // module must take measures to avoid an infinite cycle of recompilation
1931 // and deoptimization. There are several such measures:
1932 //
1933 // 1. If a recompilation is ordered a second time at some site X
1934 // and for the same reason R, the action is adjusted to 'reinterpret',
1935 // to give the interpreter time to exercise the method more thoroughly.
1936 // If this happens, the method's overflow_recompile_count is incremented.
1937 //
1938 // 2. If the compiler fails to reduce the deoptimization rate, then
1939 // the method's overflow_recompile_count will begin to exceed the set
1940 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1941 // is adjusted to 'make_not_compilable', and the method is abandoned
1942 // to the interpreter. This is a performance hit for hot methods,
1943 // but is better than a disastrous infinite cycle of recompilations.
1944 // (Actually, only the method containing the site X is abandoned.)
1945 //
1946 // 3. In parallel with the previous measures, if the total number of
1947 // recompilations of a method exceeds the much larger set limit
1948 // PerMethodRecompilationCutoff, the method is abandoned.
1949 // This should only happen if the method is very large and has
1950 // many "lukewarm" deoptimizations. The code which enforces this
1951 // limit is elsewhere (class nmethod, class Method).
1952 //
1953 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1954 // to recompile at each bytecode independently of the per-BCI cutoff.
1955 //
1956 // The decision to update code is up to the compiler, and is encoded
1957 // in the Action_xxx code. If the compiler requests Action_none
1958 // no trap state is changed, no compiled code is changed, and the
1959 // computation suffers along in the interpreter.
1960 //
1961 // The other action codes specify various tactics for decompilation
1962 // and recompilation. Action_maybe_recompile is the loosest, and
1963 // allows the compiled code to stay around until enough traps are seen,
1964 // and until the compiler gets around to recompiling the trapping method.
1965 //
1966 // The other actions cause immediate removal of the present code.
1967
1968 // Traps caused by injected profile shouldn't pollute trap counts.
1969 bool injected_profile_trap = trap_method->has_injected_profile() &&
1970 (reason == Reason_intrinsic || reason == Reason_unreached);
1971
1972 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
1973 bool make_not_entrant = false;
1974 bool make_not_compilable = false;
1975 bool reprofile = false;
1976 switch (action) {
1977 case Action_none:
1978 // Keep the old code.
1979 update_trap_state = false;
1980 break;
1981 case Action_maybe_recompile:
1982 // Do not need to invalidate the present code, but we can
1983 // initiate another
1984 // Start compiler without (necessarily) invalidating the nmethod.
1985 // The system will tolerate the old code, but new code should be
1986 // generated when possible.
1987 break;
1988 case Action_reinterpret:
1989 // Go back into the interpreter for a while, and then consider
1990 // recompiling form scratch.
1991 make_not_entrant = true;
1992 // Reset invocation counter for outer most method.
1993 // This will allow the interpreter to exercise the bytecodes
1994 // for a while before recompiling.
1995 // By contrast, Action_make_not_entrant is immediate.
1996 //
1997 // Note that the compiler will track null_check, null_assert,
1998 // range_check, and class_check events and log them as if they
1999 // had been traps taken from compiled code. This will update
2000 // the MDO trap history so that the next compilation will
2001 // properly detect hot trap sites.
2002 reprofile = true;
2003 break;
2004 case Action_make_not_entrant:
2005 // Request immediate recompilation, and get rid of the old code.
2006 // Make them not entrant, so next time they are called they get
2007 // recompiled. Unloaded classes are loaded now so recompile before next
2008 // time they are called. Same for uninitialized. The interpreter will
2009 // link the missing class, if any.
2010 make_not_entrant = true;
2011 break;
2012 case Action_make_not_compilable:
2013 // Give up on compiling this method at all.
2014 make_not_entrant = true;
2015 make_not_compilable = true;
2016 break;
2017 default:
2018 ShouldNotReachHere();
2019 }
2020
2021 // Setting +ProfileTraps fixes the following, on all platforms:
2022 // 4852688: ProfileInterpreter is off by default for ia64. The result is
2023 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2024 // recompile relies on a MethodData* to record heroic opt failures.
2025
2026 // Whether the interpreter is producing MDO data or not, we also need
2027 // to use the MDO to detect hot deoptimization points and control
2028 // aggressive optimization.
2029 bool inc_recompile_count = false;
2030 ProfileData* pdata = NULL;
2031 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
2032 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
2033 uint this_trap_count = 0;
2034 bool maybe_prior_trap = false;
2035 bool maybe_prior_recompile = false;
2036 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2037 #if INCLUDE_JVMCI
2038 nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2039 #endif
2040 nm->method(),
2041 //outputs:
2042 this_trap_count,
2043 maybe_prior_trap,
2044 maybe_prior_recompile);
2045 // Because the interpreter also counts null, div0, range, and class
2046 // checks, these traps from compiled code are double-counted.
2047 // This is harmless; it just means that the PerXTrapLimit values
2048 // are in effect a little smaller than they look.
2049
2050 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2051 if (per_bc_reason != Reason_none) {
2052 // Now take action based on the partially known per-BCI history.
2053 if (maybe_prior_trap
2054 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2055 // If there are too many traps at this BCI, force a recompile.
2056 // This will allow the compiler to see the limit overflow, and
2057 // take corrective action, if possible. The compiler generally
2058 // does not use the exact PerBytecodeTrapLimit value, but instead
2059 // changes its tactics if it sees any traps at all. This provides
2060 // a little hysteresis, delaying a recompile until a trap happens
2061 // several times.
2062 //
2063 // Actually, since there is only one bit of counter per BCI,
2064 // the possible per-BCI counts are {0,1,(per-method count)}.
2065 // This produces accurate results if in fact there is only
2066 // one hot trap site, but begins to get fuzzy if there are
2067 // many sites. For example, if there are ten sites each
2068 // trapping two or more times, they each get the blame for
2069 // all of their traps.
2070 make_not_entrant = true;
2071 }
2072
2073 // Detect repeated recompilation at the same BCI, and enforce a limit.
2074 if (make_not_entrant && maybe_prior_recompile) {
2075 // More than one recompile at this point.
2076 inc_recompile_count = maybe_prior_trap;
2077 }
2078 } else {
2079 // For reasons which are not recorded per-bytecode, we simply
2080 // force recompiles unconditionally.
2081 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2082 make_not_entrant = true;
2083 }
2084
2085 // Go back to the compiler if there are too many traps in this method.
2086 if (this_trap_count >= per_method_trap_limit(reason)) {
2087 // If there are too many traps in this method, force a recompile.
2088 // This will allow the compiler to see the limit overflow, and
2089 // take corrective action, if possible.
2090 // (This condition is an unlikely backstop only, because the
2091 // PerBytecodeTrapLimit is more likely to take effect first,
2092 // if it is applicable.)
2093 make_not_entrant = true;
2094 }
2095
2096 // Here's more hysteresis: If there has been a recompile at
2097 // this trap point already, run the method in the interpreter
2098 // for a while to exercise it more thoroughly.
2099 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2100 reprofile = true;
2101 }
2102 }
2103
2104 // Take requested actions on the method:
2105
2106 // Recompile
2107 if (make_not_entrant) {
2108 if (!nm->make_not_entrant()) {
2109 return; // the call did not change nmethod's state
2110 }
2111
2112 if (pdata != NULL) {
2113 // Record the recompilation event, if any.
2114 int tstate0 = pdata->trap_state();
2115 int tstate1 = trap_state_set_recompiled(tstate0, true);
2116 if (tstate1 != tstate0)
2117 pdata->set_trap_state(tstate1);
2118 }
2119
2120 #if INCLUDE_RTM_OPT
2121 // Restart collecting RTM locking abort statistic if the method
2122 // is recompiled for a reason other than RTM state change.
2123 // Assume that in new recompiled code the statistic could be different,
2124 // for example, due to different inlining.
2125 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2126 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2127 trap_mdo->atomic_set_rtm_state(ProfileRTM);
2128 }
2129 #endif
2130 // For code aging we count traps separately here, using make_not_entrant()
2131 // as a guard against simultaneous deopts in multiple threads.
2132 if (reason == Reason_tenured && trap_mdo != NULL) {
2133 trap_mdo->inc_tenure_traps();
2134 }
2135 }
2136
2137 if (inc_recompile_count) {
2138 trap_mdo->inc_overflow_recompile_count();
2139 if ((uint)trap_mdo->overflow_recompile_count() >
2140 (uint)PerBytecodeRecompilationCutoff) {
2141 // Give up on the method containing the bad BCI.
2142 if (trap_method() == nm->method()) {
2143 make_not_compilable = true;
2144 } else {
2145 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2146 // But give grace to the enclosing nm->method().
2147 }
2148 }
2149 }
2150
2151 // Reprofile
2152 if (reprofile) {
2153 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2154 }
2155
2156 // Give up compiling
2157 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2158 assert(make_not_entrant, "consistent");
2159 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2160 }
2161
2162 } // Free marked resources
2163
2164 }
2165 JRT_END
2166
2167 ProfileData*
2168 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2169 int trap_bci,
2170 Deoptimization::DeoptReason reason,
2171 bool update_total_trap_count,
2172 #if INCLUDE_JVMCI
2173 bool is_osr,
2174 #endif
2175 Method* compiled_method,
2176 //outputs:
2177 uint& ret_this_trap_count,
2178 bool& ret_maybe_prior_trap,
2179 bool& ret_maybe_prior_recompile) {
2180 bool maybe_prior_trap = false;
2181 bool maybe_prior_recompile = false;
2182 uint this_trap_count = 0;
2183 if (update_total_trap_count) {
2184 uint idx = reason;
2185 #if INCLUDE_JVMCI
2186 if (is_osr) {
2187 idx += Reason_LIMIT;
2188 }
2189 #endif
2190 uint prior_trap_count = trap_mdo->trap_count(idx);
2191 this_trap_count = trap_mdo->inc_trap_count(idx);
2192
2193 // If the runtime cannot find a place to store trap history,
2194 // it is estimated based on the general condition of the method.
2195 // If the method has ever been recompiled, or has ever incurred
2196 // a trap with the present reason , then this BCI is assumed
2197 // (pessimistically) to be the culprit.
2198 maybe_prior_trap = (prior_trap_count != 0);
2199 maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2200 }
2201 ProfileData* pdata = NULL;
2202
2203
2204 // For reasons which are recorded per bytecode, we check per-BCI data.
2205 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2206 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2207 if (per_bc_reason != Reason_none) {
2208 // Find the profile data for this BCI. If there isn't one,
2209 // try to allocate one from the MDO's set of spares.
2210 // This will let us detect a repeated trap at this point.
2211 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2212
2213 if (pdata != NULL) {
2214 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2215 if (LogCompilation && xtty != NULL) {
2216 ttyLocker ttyl;
2217 // no more room for speculative traps in this MDO
2218 xtty->elem("speculative_traps_oom");
2219 }
2220 }
2221 // Query the trap state of this profile datum.
2222 int tstate0 = pdata->trap_state();
2223 if (!trap_state_has_reason(tstate0, per_bc_reason))
2224 maybe_prior_trap = false;
2225 if (!trap_state_is_recompiled(tstate0))
2226 maybe_prior_recompile = false;
2227
2228 // Update the trap state of this profile datum.
2229 int tstate1 = tstate0;
2230 // Record the reason.
2231 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2232 // Store the updated state on the MDO, for next time.
2233 if (tstate1 != tstate0)
2234 pdata->set_trap_state(tstate1);
2235 } else {
2236 if (LogCompilation && xtty != NULL) {
2237 ttyLocker ttyl;
2238 // Missing MDP? Leave a small complaint in the log.
2239 xtty->elem("missing_mdp bci='%d'", trap_bci);
2240 }
2241 }
2242 }
2243
2244 // Return results:
2245 ret_this_trap_count = this_trap_count;
2246 ret_maybe_prior_trap = maybe_prior_trap;
2247 ret_maybe_prior_recompile = maybe_prior_recompile;
2248 return pdata;
2249 }
2250
2251 void
2252 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2253 ResourceMark rm;
2254 // Ignored outputs:
2255 uint ignore_this_trap_count;
2256 bool ignore_maybe_prior_trap;
2257 bool ignore_maybe_prior_recompile;
2258 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2259 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2260 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2261 query_update_method_data(trap_mdo, trap_bci,
2262 (DeoptReason)reason,
2263 update_total_counts,
2264 #if INCLUDE_JVMCI
2265 false,
2266 #endif
2267 NULL,
2268 ignore_this_trap_count,
2269 ignore_maybe_prior_trap,
2270 ignore_maybe_prior_recompile);
2271 }
2272
2273 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2274 if (TraceDeoptimization) {
2275 tty->print("Uncommon trap ");
2276 }
2277 // Still in Java no safepoints
2278 {
2279 // This enters VM and may safepoint
2280 uncommon_trap_inner(thread, trap_request);
2281 }
2282 return fetch_unroll_info_helper(thread, exec_mode);
2283 }
2284
2285 // Local derived constants.
2286 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2287 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1;
2288 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2289
2290 //---------------------------trap_state_reason---------------------------------
2291 Deoptimization::DeoptReason
2292 Deoptimization::trap_state_reason(int trap_state) {
2293 // This assert provides the link between the width of DataLayout::trap_bits
2294 // and the encoding of "recorded" reasons. It ensures there are enough
2295 // bits to store all needed reasons in the per-BCI MDO profile.
2296 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2297 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2298 trap_state -= recompile_bit;
2299 if (trap_state == DS_REASON_MASK) {
2300 return Reason_many;
2301 } else {
2302 assert((int)Reason_none == 0, "state=0 => Reason_none");
2303 return (DeoptReason)trap_state;
2304 }
2305 }
2306 //-------------------------trap_state_has_reason-------------------------------
2307 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2308 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2309 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2310 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2311 trap_state -= recompile_bit;
2312 if (trap_state == DS_REASON_MASK) {
2313 return -1; // true, unspecifically (bottom of state lattice)
2314 } else if (trap_state == reason) {
2315 return 1; // true, definitely
2316 } else if (trap_state == 0) {
2317 return 0; // false, definitely (top of state lattice)
2318 } else {
2319 return 0; // false, definitely
2320 }
2321 }
2322 //-------------------------trap_state_add_reason-------------------------------
2323 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2324 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2325 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2326 trap_state -= recompile_bit;
2327 if (trap_state == DS_REASON_MASK) {
2328 return trap_state + recompile_bit; // already at state lattice bottom
2329 } else if (trap_state == reason) {
2330 return trap_state + recompile_bit; // the condition is already true
2331 } else if (trap_state == 0) {
2332 return reason + recompile_bit; // no condition has yet been true
2333 } else {
2334 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
2335 }
2336 }
2337 //-----------------------trap_state_is_recompiled------------------------------
2338 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2339 return (trap_state & DS_RECOMPILE_BIT) != 0;
2340 }
2341 //-----------------------trap_state_set_recompiled-----------------------------
2342 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2343 if (z) return trap_state | DS_RECOMPILE_BIT;
2344 else return trap_state & ~DS_RECOMPILE_BIT;
2345 }
2346 //---------------------------format_trap_state---------------------------------
2347 // This is used for debugging and diagnostics, including LogFile output.
2348 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2349 int trap_state) {
2350 assert(buflen > 0, "sanity");
2351 DeoptReason reason = trap_state_reason(trap_state);
2352 bool recomp_flag = trap_state_is_recompiled(trap_state);
2353 // Re-encode the state from its decoded components.
2354 int decoded_state = 0;
2355 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2356 decoded_state = trap_state_add_reason(decoded_state, reason);
2357 if (recomp_flag)
2358 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2359 // If the state re-encodes properly, format it symbolically.
2360 // Because this routine is used for debugging and diagnostics,
2361 // be robust even if the state is a strange value.
2362 size_t len;
2363 if (decoded_state != trap_state) {
2364 // Random buggy state that doesn't decode??
2365 len = jio_snprintf(buf, buflen, "#%d", trap_state);
2366 } else {
2367 len = jio_snprintf(buf, buflen, "%s%s",
2368 trap_reason_name(reason),
2369 recomp_flag ? " recompiled" : "");
2370 }
2371 return buf;
2372 }
2373
2374
2375 //--------------------------------statics--------------------------------------
2376 const char* Deoptimization::_trap_reason_name[] = {
2377 // Note: Keep this in sync. with enum DeoptReason.
2378 "none",
2379 "null_check",
2380 "null_assert" JVMCI_ONLY("_or_unreached0"),
2381 "range_check",
2382 "class_check",
2383 "array_check",
2384 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2385 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2386 "profile_predicate",
2387 "unloaded",
2388 "uninitialized",
2389 "initialized",
2390 "unreached",
2391 "unhandled",
2392 "constraint",
2393 "div0_check",
2394 "age",
2395 "predicate",
2396 "loop_limit_check",
2397 "speculate_class_check",
2398 "speculate_null_check",
2399 "speculate_null_assert",
2400 "rtm_state_change",
2401 "unstable_if",
2402 "unstable_fused_if",
2403 #if INCLUDE_JVMCI
2404 "aliasing",
2405 "transfer_to_interpreter",
2406 "not_compiled_exception_handler",
2407 "unresolved",
2408 "jsr_mismatch",
2409 #endif
2410 "tenured"
2411 };
2412 const char* Deoptimization::_trap_action_name[] = {
2413 // Note: Keep this in sync. with enum DeoptAction.
2414 "none",
2415 "maybe_recompile",
2416 "reinterpret",
2417 "make_not_entrant",
2418 "make_not_compilable"
2419 };
2420
2421 const char* Deoptimization::trap_reason_name(int reason) {
2422 // Check that every reason has a name
2423 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2424
2425 if (reason == Reason_many) return "many";
2426 if ((uint)reason < Reason_LIMIT)
2427 return _trap_reason_name[reason];
2428 static char buf[20];
2429 sprintf(buf, "reason%d", reason);
2430 return buf;
2431 }
2432 const char* Deoptimization::trap_action_name(int action) {
2433 // Check that every action has a name
2434 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2435
2436 if ((uint)action < Action_LIMIT)
2437 return _trap_action_name[action];
2438 static char buf[20];
2439 sprintf(buf, "action%d", action);
2440 return buf;
2441 }
2442
2443 // This is used for debugging and diagnostics, including LogFile output.
2444 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2445 int trap_request) {
2446 jint unloaded_class_index = trap_request_index(trap_request);
2447 const char* reason = trap_reason_name(trap_request_reason(trap_request));
2448 const char* action = trap_action_name(trap_request_action(trap_request));
2449 #if INCLUDE_JVMCI
2450 int debug_id = trap_request_debug_id(trap_request);
2451 #endif
2452 size_t len;
2453 if (unloaded_class_index < 0) {
2454 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2455 reason, action
2456 #if INCLUDE_JVMCI
2457 ,debug_id
2458 #endif
2459 );
2460 } else {
2461 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2462 reason, action, unloaded_class_index
2463 #if INCLUDE_JVMCI
2464 ,debug_id
2465 #endif
2466 );
2467 }
2468 return buf;
2469 }
2470
2471 juint Deoptimization::_deoptimization_hist
2472 [Deoptimization::Reason_LIMIT]
2473 [1 + Deoptimization::Action_LIMIT]
2474 [Deoptimization::BC_CASE_LIMIT]
2475 = {0};
2476
2477 enum {
2478 LSB_BITS = 8,
2479 LSB_MASK = right_n_bits(LSB_BITS)
2480 };
2481
2482 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2483 Bytecodes::Code bc) {
2484 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2485 assert(action >= 0 && action < Action_LIMIT, "oob");
2486 _deoptimization_hist[Reason_none][0][0] += 1; // total
2487 _deoptimization_hist[reason][0][0] += 1; // per-reason total
2488 juint* cases = _deoptimization_hist[reason][1+action];
2489 juint* bc_counter_addr = NULL;
2490 juint bc_counter = 0;
2491 // Look for an unused counter, or an exact match to this BC.
2492 if (bc != Bytecodes::_illegal) {
2493 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2494 juint* counter_addr = &cases[bc_case];
2495 juint counter = *counter_addr;
2496 if ((counter == 0 && bc_counter_addr == NULL)
2497 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2498 // this counter is either free or is already devoted to this BC
2499 bc_counter_addr = counter_addr;
2500 bc_counter = counter | bc;
2501 }
2502 }
2503 }
2504 if (bc_counter_addr == NULL) {
2505 // Overflow, or no given bytecode.
2506 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2507 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2508 }
2509 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2510 }
2511
2512 jint Deoptimization::total_deoptimization_count() {
2513 return _deoptimization_hist[Reason_none][0][0];
2514 }
2515
2516 void Deoptimization::print_statistics() {
2517 juint total = total_deoptimization_count();
2518 juint account = total;
2519 if (total != 0) {
2520 ttyLocker ttyl;
2521 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2522 tty->print_cr("Deoptimization traps recorded:");
2523 #define PRINT_STAT_LINE(name, r) \
2524 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2525 PRINT_STAT_LINE("total", total);
2526 // For each non-zero entry in the histogram, print the reason,
2527 // the action, and (if specifically known) the type of bytecode.
2528 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2529 for (int action = 0; action < Action_LIMIT; action++) {
2530 juint* cases = _deoptimization_hist[reason][1+action];
2531 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2532 juint counter = cases[bc_case];
2533 if (counter != 0) {
2534 char name[1*K];
2535 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2536 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2537 bc = Bytecodes::_illegal;
2538 sprintf(name, "%s/%s/%s",
2539 trap_reason_name(reason),
2540 trap_action_name(action),
2541 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2542 juint r = counter >> LSB_BITS;
2543 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2544 account -= r;
2545 }
2546 }
2547 }
2548 }
2549 if (account != 0) {
2550 PRINT_STAT_LINE("unaccounted", account);
2551 }
2552 #undef PRINT_STAT_LINE
2553 if (xtty != NULL) xtty->tail("statistics");
2554 }
2555 }
2556 #else // COMPILER2_OR_JVMCI
2557
2558
2559 // Stubs for C1 only system.
2560 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2561 return false;
2562 }
2563
2564 const char* Deoptimization::trap_reason_name(int reason) {
2565 return "unknown";
2566 }
2567
2568 void Deoptimization::print_statistics() {
2569 // no output
2570 }
2571
2572 void
2573 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2574 // no udpate
2575 }
2576
2577 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2578 return 0;
2579 }
2580
2581 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2582 Bytecodes::Code bc) {
2583 // no update
2584 }
2585
2586 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2587 int trap_state) {
2588 jio_snprintf(buf, buflen, "#%d", trap_state);
2589 return buf;
2590 }
2591
2592 #endif // COMPILER2_OR_JVMCI