1 /*
2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "code/codeCache.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/nmethod.hpp"
30 #include "code/pcDesc.hpp"
31 #include "code/scopeDesc.hpp"
32 #include "interpreter/bytecode.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/oopMapCache.hpp"
35 #include "memory/allocation.inline.hpp"
36 #include "memory/oopFactory.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "oops/method.hpp"
39 #include "oops/oop.inline.hpp"
40 #include "prims/jvmtiThreadState.hpp"
41 #include "runtime/biasedLocking.hpp"
42 #include "runtime/compilationPolicy.hpp"
43 #include "runtime/deoptimization.hpp"
44 #include "runtime/interfaceSupport.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "runtime/signature.hpp"
47 #include "runtime/stubRoutines.hpp"
48 #include "runtime/thread.hpp"
49 #include "runtime/vframe.hpp"
50 #include "runtime/vframeArray.hpp"
51 #include "runtime/vframe_hp.hpp"
52 #include "utilities/events.hpp"
53 #include "utilities/xmlstream.hpp"
54
55 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
56
57 bool DeoptimizationMarker::_is_active = false;
58
59 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
60 int caller_adjustment,
61 int caller_actual_parameters,
62 int number_of_frames,
63 intptr_t* frame_sizes,
64 address* frame_pcs,
65 BasicType return_type) {
66 _size_of_deoptimized_frame = size_of_deoptimized_frame;
67 _caller_adjustment = caller_adjustment;
68 _caller_actual_parameters = caller_actual_parameters;
69 _number_of_frames = number_of_frames;
70 _frame_sizes = frame_sizes;
71 _frame_pcs = frame_pcs;
72 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
73 _return_type = return_type;
74 _initial_info = 0;
75 // PD (x86 only)
76 _counter_temp = 0;
77 _unpack_kind = 0;
78 _sender_sp_temp = 0;
79
80 _total_frame_sizes = size_of_frames();
81 }
82
83
84 Deoptimization::UnrollBlock::~UnrollBlock() {
85 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
86 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
87 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
88 }
89
90
91 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
92 assert(register_number < RegisterMap::reg_count, "checking register number");
93 return &_register_block[register_number * 2];
94 }
95
96
97
98 int Deoptimization::UnrollBlock::size_of_frames() const {
99 // Acount first for the adjustment of the initial frame
100 int result = _caller_adjustment;
101 for (int index = 0; index < number_of_frames(); index++) {
102 result += frame_sizes()[index];
103 }
104 return result;
105 }
106
107
108 void Deoptimization::UnrollBlock::print() {
109 ttyLocker ttyl;
110 tty->print_cr("UnrollBlock");
111 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
112 tty->print( " frame_sizes: ");
113 for (int index = 0; index < number_of_frames(); index++) {
114 tty->print("%d ", frame_sizes()[index]);
115 }
116 tty->cr();
117 }
118
119
120 // In order to make fetch_unroll_info work properly with escape
121 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
122 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
123 // of previously eliminated objects occurs in realloc_objects, which is
124 // called from the method fetch_unroll_info_helper below.
125 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
126 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
127 // but makes the entry a little slower. There is however a little dance we have to
128 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
129
130 // fetch_unroll_info() is called at the beginning of the deoptimization
131 // handler. Note this fact before we start generating temporary frames
132 // that can confuse an asynchronous stack walker. This counter is
133 // decremented at the end of unpack_frames().
134 thread->inc_in_deopt_handler();
135
136 return fetch_unroll_info_helper(thread);
137 JRT_END
138
139
140 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
141 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
142
143 // Note: there is a safepoint safety issue here. No matter whether we enter
144 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
145 // the vframeArray is created.
146 //
147
148 // Allocate our special deoptimization ResourceMark
149 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
150 assert(thread->deopt_mark() == NULL, "Pending deopt!");
151 thread->set_deopt_mark(dmark);
152
153 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
154 RegisterMap map(thread, true);
155 RegisterMap dummy_map(thread, false);
156 // Now get the deoptee with a valid map
157 frame deoptee = stub_frame.sender(&map);
158 // Set the deoptee nmethod
159 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
160 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
161
162 if (VerifyStack) {
163 thread->validate_frame_layout();
164 }
165
166 // Create a growable array of VFrames where each VFrame represents an inlined
167 // Java frame. This storage is allocated with the usual system arena.
168 assert(deoptee.is_compiled_frame(), "Wrong frame type");
169 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
170 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
171 while (!vf->is_top()) {
172 assert(vf->is_compiled_frame(), "Wrong frame type");
173 chunk->push(compiledVFrame::cast(vf));
174 vf = vf->sender();
175 }
176 assert(vf->is_compiled_frame(), "Wrong frame type");
177 chunk->push(compiledVFrame::cast(vf));
178
179 #ifdef COMPILER2
180 // Reallocate the non-escaping objects and restore their fields. Then
181 // relock objects if synchronization on them was eliminated.
182 if (DoEscapeAnalysis || EliminateNestedLocks) {
183 if (EliminateAllocations) {
184 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
185 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
186
187 // The flag return_oop() indicates call sites which return oop
188 // in compiled code. Such sites include java method calls,
189 // runtime calls (for example, used to allocate new objects/arrays
190 // on slow code path) and any other calls generated in compiled code.
191 // It is not guaranteed that we can get such information here only
192 // by analyzing bytecode in deoptimized frames. This is why this flag
193 // is set during method compilation (see Compile::Process_OopMap_Node()).
194 // If we the previous was popped, we don't have a result
195 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution();
196 Handle return_value;
197 if (save_oop_result) {
198 // Reallocation may trigger GC. If deoptimization happened on return from
199 // call which returns oop we need to save it since it is not in oopmap.
200 oop result = deoptee.saved_oop_result(&map);
201 assert(result == NULL || result->is_oop(), "must be oop");
202 return_value = Handle(thread, result);
203 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
204 if (TraceDeoptimization) {
205 ttyLocker ttyl;
206 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread);
207 }
208 }
209 bool reallocated = false;
210 if (objects != NULL) {
211 JRT_BLOCK
212 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
213 JRT_END
214 }
215 if (reallocated) {
216 reassign_fields(&deoptee, &map, objects);
217 #ifndef PRODUCT
218 if (TraceDeoptimization) {
219 ttyLocker ttyl;
220 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
221 print_objects(objects);
222 }
223 #endif
224 }
225 if (save_oop_result) {
226 // Restore result.
227 deoptee.set_saved_oop_result(&map, return_value());
228 }
229 }
230 if (EliminateLocks) {
231 #ifndef PRODUCT
232 bool first = true;
233 #endif
234 for (int i = 0; i < chunk->length(); i++) {
235 compiledVFrame* cvf = chunk->at(i);
236 assert (cvf->scope() != NULL,"expect only compiled java frames");
237 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
238 if (monitors->is_nonempty()) {
239 relock_objects(monitors, thread);
240 #ifndef PRODUCT
241 if (TraceDeoptimization) {
242 ttyLocker ttyl;
243 for (int j = 0; j < monitors->length(); j++) {
244 MonitorInfo* mi = monitors->at(j);
245 if (mi->eliminated()) {
246 if (first) {
247 first = false;
248 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
249 }
250 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner());
251 }
252 }
253 }
254 #endif
255 }
256 }
257 }
258 }
259 #endif // COMPILER2
260 // Ensure that no safepoint is taken after pointers have been stored
261 // in fields of rematerialized objects. If a safepoint occurs from here on
262 // out the java state residing in the vframeArray will be missed.
263 No_Safepoint_Verifier no_safepoint;
264
265 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
266
267 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
268 thread->set_vframe_array_head(array);
269
270 // Now that the vframeArray has been created if we have any deferred local writes
271 // added by jvmti then we can free up that structure as the data is now in the
272 // vframeArray
273
274 if (thread->deferred_locals() != NULL) {
275 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
276 int i = 0;
277 do {
278 // Because of inlining we could have multiple vframes for a single frame
279 // and several of the vframes could have deferred writes. Find them all.
280 if (list->at(i)->id() == array->original().id()) {
281 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
282 list->remove_at(i);
283 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
284 delete dlv;
285 } else {
286 i++;
287 }
288 } while ( i < list->length() );
289 if (list->length() == 0) {
290 thread->set_deferred_locals(NULL);
291 // free the list and elements back to C heap.
292 delete list;
293 }
294
295 }
296
297 #ifndef SHARK
298 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
299 CodeBlob* cb = stub_frame.cb();
300 // Verify we have the right vframeArray
301 assert(cb->frame_size() >= 0, "Unexpected frame size");
302 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
303
304 // If the deopt call site is a MethodHandle invoke call site we have
305 // to adjust the unpack_sp.
306 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
307 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
308 unpack_sp = deoptee.unextended_sp();
309
310 #ifdef ASSERT
311 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
312 #endif
313 #else
314 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
315 #endif // !SHARK
316
317 // This is a guarantee instead of an assert because if vframe doesn't match
318 // we will unpack the wrong deoptimized frame and wind up in strange places
319 // where it will be very difficult to figure out what went wrong. Better
320 // to die an early death here than some very obscure death later when the
321 // trail is cold.
322 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
323 // in that it will fail to detect a problem when there is one. This needs
324 // more work in tiger timeframe.
325 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
326
327 int number_of_frames = array->frames();
328
329 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
330 // virtual activation, which is the reverse of the elements in the vframes array.
331 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
332 // +1 because we always have an interpreter return address for the final slot.
333 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
334 int popframe_extra_args = 0;
335 // Create an interpreter return address for the stub to use as its return
336 // address so the skeletal frames are perfectly walkable
337 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
338
339 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
340 // activation be put back on the expression stack of the caller for reexecution
341 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
342 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
343 }
344
345 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
346 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
347 // than simply use array->sender.pc(). This requires us to walk the current set of frames
348 //
349 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
350 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
351
352 // It's possible that the number of parameters at the call site is
353 // different than number of arguments in the callee when method
354 // handles are used. If the caller is interpreted get the real
355 // value so that the proper amount of space can be added to it's
356 // frame.
357 bool caller_was_method_handle = false;
358 if (deopt_sender.is_interpreted_frame()) {
359 methodHandle method = deopt_sender.interpreter_frame_method();
360 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
361 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
362 // Method handle invokes may involve fairly arbitrary chains of
363 // calls so it's impossible to know how much actual space the
364 // caller has for locals.
365 caller_was_method_handle = true;
366 }
367 }
368
369 //
370 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
371 // frame_sizes/frame_pcs[1] next oldest frame (int)
372 // frame_sizes/frame_pcs[n] youngest frame (int)
373 //
374 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
375 // owns the space for the return address to it's caller). Confusing ain't it.
376 //
377 // The vframe array can address vframes with indices running from
378 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
379 // When we create the skeletal frames we need the oldest frame to be in the zero slot
380 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
381 // so things look a little strange in this loop.
382 //
383 int callee_parameters = 0;
384 int callee_locals = 0;
385 for (int index = 0; index < array->frames(); index++ ) {
386 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
387 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
388 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
389 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
390 callee_locals,
391 index == 0,
392 popframe_extra_args);
393 // This pc doesn't have to be perfect just good enough to identify the frame
394 // as interpreted so the skeleton frame will be walkable
395 // The correct pc will be set when the skeleton frame is completely filled out
396 // The final pc we store in the loop is wrong and will be overwritten below
397 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
398
399 callee_parameters = array->element(index)->method()->size_of_parameters();
400 callee_locals = array->element(index)->method()->max_locals();
401 popframe_extra_args = 0;
402 }
403
404 // Compute whether the root vframe returns a float or double value.
405 BasicType return_type;
406 {
407 HandleMark hm;
408 methodHandle method(thread, array->element(0)->method());
409 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
410 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
411 }
412
413 // Compute information for handling adapters and adjusting the frame size of the caller.
414 int caller_adjustment = 0;
415
416 // Compute the amount the oldest interpreter frame will have to adjust
417 // its caller's stack by. If the caller is a compiled frame then
418 // we pretend that the callee has no parameters so that the
419 // extension counts for the full amount of locals and not just
420 // locals-parms. This is because without a c2i adapter the parm
421 // area as created by the compiled frame will not be usable by
422 // the interpreter. (Depending on the calling convention there
423 // may not even be enough space).
424
425 // QQQ I'd rather see this pushed down into last_frame_adjust
426 // and have it take the sender (aka caller).
427
428 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
429 caller_adjustment = last_frame_adjust(0, callee_locals);
430 } else if (callee_locals > callee_parameters) {
431 // The caller frame may need extending to accommodate
432 // non-parameter locals of the first unpacked interpreted frame.
433 // Compute that adjustment.
434 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
435 }
436
437 // If the sender is deoptimized the we must retrieve the address of the handler
438 // since the frame will "magically" show the original pc before the deopt
439 // and we'd undo the deopt.
440
441 frame_pcs[0] = deopt_sender.raw_pc();
442
443 #ifndef SHARK
444 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
445 #endif // SHARK
446
447 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
448 caller_adjustment * BytesPerWord,
449 caller_was_method_handle ? 0 : callee_parameters,
450 number_of_frames,
451 frame_sizes,
452 frame_pcs,
453 return_type);
454 // On some platforms, we need a way to pass some platform dependent
455 // information to the unpacking code so the skeletal frames come out
456 // correct (initial fp value, unextended sp, ...)
457 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
458
459 if (array->frames() > 1) {
460 if (VerifyStack && TraceDeoptimization) {
461 ttyLocker ttyl;
462 tty->print_cr("Deoptimizing method containing inlining");
463 }
464 }
465
466 array->set_unroll_block(info);
467 return info;
468 }
469
470 // Called to cleanup deoptimization data structures in normal case
471 // after unpacking to stack and when stack overflow error occurs
472 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
473 vframeArray *array) {
474
475 // Get array if coming from exception
476 if (array == NULL) {
477 array = thread->vframe_array_head();
478 }
479 thread->set_vframe_array_head(NULL);
480
481 // Free the previous UnrollBlock
482 vframeArray* old_array = thread->vframe_array_last();
483 thread->set_vframe_array_last(array);
484
485 if (old_array != NULL) {
486 UnrollBlock* old_info = old_array->unroll_block();
487 old_array->set_unroll_block(NULL);
488 delete old_info;
489 delete old_array;
490 }
491
492 // Deallocate any resource creating in this routine and any ResourceObjs allocated
493 // inside the vframeArray (StackValueCollections)
494
495 delete thread->deopt_mark();
496 thread->set_deopt_mark(NULL);
497 thread->set_deopt_nmethod(NULL);
498
499
500 if (JvmtiExport::can_pop_frame()) {
501 #ifndef CC_INTERP
502 // Regardless of whether we entered this routine with the pending
503 // popframe condition bit set, we should always clear it now
504 thread->clear_popframe_condition();
505 #else
506 // C++ interpreter will clear has_pending_popframe when it enters
507 // with method_resume. For deopt_resume2 we clear it now.
508 if (thread->popframe_forcing_deopt_reexecution())
509 thread->clear_popframe_condition();
510 #endif /* CC_INTERP */
511 }
512
513 // unpack_frames() is called at the end of the deoptimization handler
514 // and (in C2) at the end of the uncommon trap handler. Note this fact
515 // so that an asynchronous stack walker can work again. This counter is
516 // incremented at the beginning of fetch_unroll_info() and (in C2) at
517 // the beginning of uncommon_trap().
518 thread->dec_in_deopt_handler();
519 }
520
521
522 // Return BasicType of value being returned
523 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
524
525 // We are already active int he special DeoptResourceMark any ResourceObj's we
526 // allocate will be freed at the end of the routine.
527
528 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
529 // but makes the entry a little slower. There is however a little dance we have to
530 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
531 ResetNoHandleMark rnhm; // No-op in release/product versions
532 HandleMark hm;
533
534 frame stub_frame = thread->last_frame();
535
536 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
537 // must point to the vframeArray for the unpack frame.
538 vframeArray* array = thread->vframe_array_head();
539
540 #ifndef PRODUCT
541 if (TraceDeoptimization) {
542 ttyLocker ttyl;
543 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
544 }
545 #endif
546 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
547 stub_frame.pc(), stub_frame.sp(), exec_mode);
548
549 UnrollBlock* info = array->unroll_block();
550
551 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
552 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
553
554 BasicType bt = info->return_type();
555
556 // If we have an exception pending, claim that the return type is an oop
557 // so the deopt_blob does not overwrite the exception_oop.
558
559 if (exec_mode == Unpack_exception)
560 bt = T_OBJECT;
561
562 // Cleanup thread deopt data
563 cleanup_deopt_info(thread, array);
564
565 #ifndef PRODUCT
566 if (VerifyStack) {
567 ResourceMark res_mark;
568
569 thread->validate_frame_layout();
570
571 // Verify that the just-unpacked frames match the interpreter's
572 // notions of expression stack and locals
573 vframeArray* cur_array = thread->vframe_array_last();
574 RegisterMap rm(thread, false);
575 rm.set_include_argument_oops(false);
576 bool is_top_frame = true;
577 int callee_size_of_parameters = 0;
578 int callee_max_locals = 0;
579 for (int i = 0; i < cur_array->frames(); i++) {
580 vframeArrayElement* el = cur_array->element(i);
581 frame* iframe = el->iframe();
582 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
583
584 // Get the oop map for this bci
585 InterpreterOopMap mask;
586 int cur_invoke_parameter_size = 0;
587 bool try_next_mask = false;
588 int next_mask_expression_stack_size = -1;
589 int top_frame_expression_stack_adjustment = 0;
590 methodHandle mh(thread, iframe->interpreter_frame_method());
591 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
592 BytecodeStream str(mh);
593 str.set_start(iframe->interpreter_frame_bci());
594 int max_bci = mh->code_size();
595 // Get to the next bytecode if possible
596 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
597 // Check to see if we can grab the number of outgoing arguments
598 // at an uncommon trap for an invoke (where the compiler
599 // generates debug info before the invoke has executed)
600 Bytecodes::Code cur_code = str.next();
601 if (cur_code == Bytecodes::_invokevirtual ||
602 cur_code == Bytecodes::_invokespecial ||
603 cur_code == Bytecodes::_invokestatic ||
604 cur_code == Bytecodes::_invokeinterface ||
605 cur_code == Bytecodes::_invokedynamic) {
606 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
607 Symbol* signature = invoke.signature();
608 ArgumentSizeComputer asc(signature);
609 cur_invoke_parameter_size = asc.size();
610 if (invoke.has_receiver()) {
611 // Add in receiver
612 ++cur_invoke_parameter_size;
613 }
614 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
615 callee_size_of_parameters++;
616 }
617 }
618 if (str.bci() < max_bci) {
619 Bytecodes::Code bc = str.next();
620 if (bc >= 0) {
621 // The interpreter oop map generator reports results before
622 // the current bytecode has executed except in the case of
623 // calls. It seems to be hard to tell whether the compiler
624 // has emitted debug information matching the "state before"
625 // a given bytecode or the state after, so we try both
626 switch (cur_code) {
627 case Bytecodes::_invokevirtual:
628 case Bytecodes::_invokespecial:
629 case Bytecodes::_invokestatic:
630 case Bytecodes::_invokeinterface:
631 case Bytecodes::_invokedynamic:
632 case Bytecodes::_athrow:
633 break;
634 default: {
635 InterpreterOopMap next_mask;
636 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
637 next_mask_expression_stack_size = next_mask.expression_stack_size();
638 // Need to subtract off the size of the result type of
639 // the bytecode because this is not described in the
640 // debug info but returned to the interpreter in the TOS
641 // caching register
642 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
643 if (bytecode_result_type != T_ILLEGAL) {
644 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
645 }
646 assert(top_frame_expression_stack_adjustment >= 0, "");
647 try_next_mask = true;
648 break;
649 }
650 }
651 }
652 }
653
654 // Verify stack depth and oops in frame
655 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
656 if (!(
657 /* SPARC */
658 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
659 /* x86 */
660 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
661 (try_next_mask &&
662 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
663 top_frame_expression_stack_adjustment))) ||
664 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
665 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
666 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
667 )) {
668 ttyLocker ttyl;
669
670 // Print out some information that will help us debug the problem
671 tty->print_cr("Wrong number of expression stack elements during deoptimization");
672 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
673 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
674 iframe->interpreter_frame_expression_stack_size());
675 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
676 tty->print_cr(" try_next_mask = %d", try_next_mask);
677 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
678 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
679 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
680 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
681 tty->print_cr(" exec_mode = %d", exec_mode);
682 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
683 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
684 tty->print_cr(" Interpreted frames:");
685 for (int k = 0; k < cur_array->frames(); k++) {
686 vframeArrayElement* el = cur_array->element(k);
687 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
688 }
689 cur_array->print_on_2(tty);
690 guarantee(false, "wrong number of expression stack elements during deopt");
691 }
692 VerifyOopClosure verify;
693 iframe->oops_interpreted_do(&verify, NULL, &rm, false);
694 callee_size_of_parameters = mh->size_of_parameters();
695 callee_max_locals = mh->max_locals();
696 is_top_frame = false;
697 }
698 }
699 #endif /* !PRODUCT */
700
701
702 return bt;
703 JRT_END
704
705
706 int Deoptimization::deoptimize_dependents() {
707 Threads::deoptimized_wrt_marked_nmethods();
708 return 0;
709 }
710
711 Deoptimization::DeoptAction Deoptimization::_unloaded_action
712 = Deoptimization::Action_reinterpret;
713
714 #ifdef COMPILER2
715 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
716 Handle pending_exception(thread->pending_exception());
717 const char* exception_file = thread->exception_file();
718 int exception_line = thread->exception_line();
719 thread->clear_pending_exception();
720
721 for (int i = 0; i < objects->length(); i++) {
722 assert(objects->at(i)->is_object(), "invalid debug information");
723 ObjectValue* sv = (ObjectValue*) objects->at(i);
724
725 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
726 oop obj = NULL;
727
728 if (k->oop_is_instance()) {
729 InstanceKlass* ik = InstanceKlass::cast(k());
730 obj = ik->allocate_instance(CHECK_(false));
731 } else if (k->oop_is_typeArray()) {
732 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
733 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
734 int len = sv->field_size() / type2size[ak->element_type()];
735 obj = ak->allocate(len, CHECK_(false));
736 } else if (k->oop_is_objArray()) {
737 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
738 obj = ak->allocate(sv->field_size(), CHECK_(false));
739 }
740
741 assert(obj != NULL, "allocation failed");
742 assert(sv->value().is_null(), "redundant reallocation");
743 sv->set_value(obj);
744 }
745
746 if (pending_exception.not_null()) {
747 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
748 }
749
750 return true;
751 }
752
753 // This assumes that the fields are stored in ObjectValue in the same order
754 // they are yielded by do_nonstatic_fields.
755 class FieldReassigner: public FieldClosure {
756 frame* _fr;
757 RegisterMap* _reg_map;
758 ObjectValue* _sv;
759 InstanceKlass* _ik;
760 oop _obj;
761
762 int _i;
763 public:
764 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
765 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
766
767 int i() const { return _i; }
768
769
770 void do_field(fieldDescriptor* fd) {
771 intptr_t val;
772 StackValue* value =
773 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
774 int offset = fd->offset();
775 switch (fd->field_type()) {
776 case T_OBJECT: case T_ARRAY:
777 assert(value->type() == T_OBJECT, "Agreement.");
778 _obj->obj_field_put(offset, value->get_obj()());
779 break;
780
781 case T_LONG: case T_DOUBLE: {
782 assert(value->type() == T_INT, "Agreement.");
783 StackValue* low =
784 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
785 #ifdef _LP64
786 jlong res = (jlong)low->get_int();
787 #else
788 #ifdef SPARC
789 // For SPARC we have to swap high and low words.
790 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
791 #else
792 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
793 #endif //SPARC
794 #endif
795 _obj->long_field_put(offset, res);
796 break;
797 }
798 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
799 case T_INT: case T_FLOAT: // 4 bytes.
800 assert(value->type() == T_INT, "Agreement.");
801 val = value->get_int();
802 _obj->int_field_put(offset, (jint)*((jint*)&val));
803 break;
804
805 case T_SHORT: case T_CHAR: // 2 bytes
806 assert(value->type() == T_INT, "Agreement.");
807 val = value->get_int();
808 _obj->short_field_put(offset, (jshort)*((jint*)&val));
809 break;
810
811 case T_BOOLEAN: case T_BYTE: // 1 byte
812 assert(value->type() == T_INT, "Agreement.");
813 val = value->get_int();
814 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
815 break;
816
817 default:
818 ShouldNotReachHere();
819 }
820 _i++;
821 }
822 };
823
824 // restore elements of an eliminated type array
825 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
826 int index = 0;
827 intptr_t val;
828
829 for (int i = 0; i < sv->field_size(); i++) {
830 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
831 switch(type) {
832 case T_LONG: case T_DOUBLE: {
833 assert(value->type() == T_INT, "Agreement.");
834 StackValue* low =
835 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
836 #ifdef _LP64
837 jlong res = (jlong)low->get_int();
838 #else
839 #ifdef SPARC
840 // For SPARC we have to swap high and low words.
841 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
842 #else
843 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
844 #endif //SPARC
845 #endif
846 obj->long_at_put(index, res);
847 break;
848 }
849
850 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
851 case T_INT: case T_FLOAT: // 4 bytes.
852 assert(value->type() == T_INT, "Agreement.");
853 val = value->get_int();
854 obj->int_at_put(index, (jint)*((jint*)&val));
855 break;
856
857 case T_SHORT: case T_CHAR: // 2 bytes
858 assert(value->type() == T_INT, "Agreement.");
859 val = value->get_int();
860 obj->short_at_put(index, (jshort)*((jint*)&val));
861 break;
862
863 case T_BOOLEAN: case T_BYTE: // 1 byte
864 assert(value->type() == T_INT, "Agreement.");
865 val = value->get_int();
866 obj->bool_at_put(index, (jboolean)*((jint*)&val));
867 break;
868
869 default:
870 ShouldNotReachHere();
871 }
872 index++;
873 }
874 }
875
876
877 // restore fields of an eliminated object array
878 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
879 for (int i = 0; i < sv->field_size(); i++) {
880 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
881 assert(value->type() == T_OBJECT, "object element expected");
882 obj->obj_at_put(i, value->get_obj()());
883 }
884 }
885
886
887 // restore fields of all eliminated objects and arrays
888 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
889 for (int i = 0; i < objects->length(); i++) {
890 ObjectValue* sv = (ObjectValue*) objects->at(i);
891 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
892 Handle obj = sv->value();
893 assert(obj.not_null(), "reallocation was missed");
894
895 if (k->oop_is_instance()) {
896 InstanceKlass* ik = InstanceKlass::cast(k());
897 FieldReassigner reassign(fr, reg_map, sv, obj());
898 ik->do_nonstatic_fields(&reassign);
899 } else if (k->oop_is_typeArray()) {
900 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
901 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
902 } else if (k->oop_is_objArray()) {
903 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
904 }
905 }
906 }
907
908
909 // relock objects for which synchronization was eliminated
910 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
911 for (int i = 0; i < monitors->length(); i++) {
912 MonitorInfo* mon_info = monitors->at(i);
913 if (mon_info->eliminated()) {
914 assert(mon_info->owner() != NULL, "reallocation was missed");
915 Handle obj = Handle(mon_info->owner());
916 markOop mark = obj->mark();
917 if (UseBiasedLocking && mark->has_bias_pattern()) {
918 // New allocated objects may have the mark set to anonymously biased.
919 // Also the deoptimized method may called methods with synchronization
920 // where the thread-local object is bias locked to the current thread.
921 assert(mark->is_biased_anonymously() ||
922 mark->biased_locker() == thread, "should be locked to current thread");
923 // Reset mark word to unbiased prototype.
924 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
925 obj->set_mark(unbiased_prototype);
926 }
927 BasicLock* lock = mon_info->lock();
928 ObjectSynchronizer::slow_enter(obj, lock, thread);
929 }
930 assert(mon_info->owner()->is_locked(), "object must be locked now");
931 }
932 }
933
934
935 #ifndef PRODUCT
936 // print information about reallocated objects
937 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
938 fieldDescriptor fd;
939
940 for (int i = 0; i < objects->length(); i++) {
941 ObjectValue* sv = (ObjectValue*) objects->at(i);
942 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
943 Handle obj = sv->value();
944
945 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()());
946 k->print_value();
947 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
948 tty->cr();
949
950 if (Verbose) {
951 k->oop_print_on(obj(), tty);
952 }
953 }
954 }
955 #endif
956 #endif // COMPILER2
957
958 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
959 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
960
961 #ifndef PRODUCT
962 if (TraceDeoptimization) {
963 ttyLocker ttyl;
964 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
965 fr.print_on(tty);
966 tty->print_cr(" Virtual frames (innermost first):");
967 for (int index = 0; index < chunk->length(); index++) {
968 compiledVFrame* vf = chunk->at(index);
969 tty->print(" %2d - ", index);
970 vf->print_value();
971 int bci = chunk->at(index)->raw_bci();
972 const char* code_name;
973 if (bci == SynchronizationEntryBCI) {
974 code_name = "sync entry";
975 } else {
976 Bytecodes::Code code = vf->method()->code_at(bci);
977 code_name = Bytecodes::name(code);
978 }
979 tty->print(" - %s", code_name);
980 tty->print_cr(" @ bci %d ", bci);
981 if (Verbose) {
982 vf->print();
983 tty->cr();
984 }
985 }
986 }
987 #endif
988
989 // Register map for next frame (used for stack crawl). We capture
990 // the state of the deopt'ing frame's caller. Thus if we need to
991 // stuff a C2I adapter we can properly fill in the callee-save
992 // register locations.
993 frame caller = fr.sender(reg_map);
994 int frame_size = caller.sp() - fr.sp();
995
996 frame sender = caller;
997
998 // Since the Java thread being deoptimized will eventually adjust it's own stack,
999 // the vframeArray containing the unpacking information is allocated in the C heap.
1000 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1001 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
1002
1003 // Compare the vframeArray to the collected vframes
1004 assert(array->structural_compare(thread, chunk), "just checking");
1005
1006 #ifndef PRODUCT
1007 if (TraceDeoptimization) {
1008 ttyLocker ttyl;
1009 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1010 }
1011 #endif // PRODUCT
1012
1013 return array;
1014 }
1015
1016
1017 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1018 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1019 for (int i = 0; i < monitors->length(); i++) {
1020 MonitorInfo* mon_info = monitors->at(i);
1021 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1022 objects_to_revoke->append(Handle(mon_info->owner()));
1023 }
1024 }
1025 }
1026
1027
1028 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1029 if (!UseBiasedLocking) {
1030 return;
1031 }
1032
1033 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1034
1035 // Unfortunately we don't have a RegisterMap available in most of
1036 // the places we want to call this routine so we need to walk the
1037 // stack again to update the register map.
1038 if (map == NULL || !map->update_map()) {
1039 StackFrameStream sfs(thread, true);
1040 bool found = false;
1041 while (!found && !sfs.is_done()) {
1042 frame* cur = sfs.current();
1043 sfs.next();
1044 found = cur->id() == fr.id();
1045 }
1046 assert(found, "frame to be deoptimized not found on target thread's stack");
1047 map = sfs.register_map();
1048 }
1049
1050 vframe* vf = vframe::new_vframe(&fr, map, thread);
1051 compiledVFrame* cvf = compiledVFrame::cast(vf);
1052 // Revoke monitors' biases in all scopes
1053 while (!cvf->is_top()) {
1054 collect_monitors(cvf, objects_to_revoke);
1055 cvf = compiledVFrame::cast(cvf->sender());
1056 }
1057 collect_monitors(cvf, objects_to_revoke);
1058
1059 if (SafepointSynchronize::is_at_safepoint()) {
1060 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1061 } else {
1062 BiasedLocking::revoke(objects_to_revoke);
1063 }
1064 }
1065
1066
1067 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1068 if (!UseBiasedLocking) {
1069 return;
1070 }
1071
1072 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1073 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1074 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1075 if (jt->has_last_Java_frame()) {
1076 StackFrameStream sfs(jt, true);
1077 while (!sfs.is_done()) {
1078 frame* cur = sfs.current();
1079 if (cb->contains(cur->pc())) {
1080 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1081 compiledVFrame* cvf = compiledVFrame::cast(vf);
1082 // Revoke monitors' biases in all scopes
1083 while (!cvf->is_top()) {
1084 collect_monitors(cvf, objects_to_revoke);
1085 cvf = compiledVFrame::cast(cvf->sender());
1086 }
1087 collect_monitors(cvf, objects_to_revoke);
1088 }
1089 sfs.next();
1090 }
1091 }
1092 }
1093 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1094 }
1095
1096
1097 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1098 assert(fr.can_be_deoptimized(), "checking frame type");
1099
1100 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1101
1102 // Patch the nmethod so that when execution returns to it we will
1103 // deopt the execution state and return to the interpreter.
1104 fr.deoptimize(thread);
1105 }
1106
1107 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1108 // Deoptimize only if the frame comes from compile code.
1109 // Do not deoptimize the frame which is already patched
1110 // during the execution of the loops below.
1111 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1112 return;
1113 }
1114 ResourceMark rm;
1115 DeoptimizationMarker dm;
1116 if (UseBiasedLocking) {
1117 revoke_biases_of_monitors(thread, fr, map);
1118 }
1119 deoptimize_single_frame(thread, fr);
1120
1121 }
1122
1123
1124 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1125 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1126 "can only deoptimize other thread at a safepoint");
1127 // Compute frame and register map based on thread and sp.
1128 RegisterMap reg_map(thread, UseBiasedLocking);
1129 frame fr = thread->last_frame();
1130 while (fr.id() != id) {
1131 fr = fr.sender(®_map);
1132 }
1133 deoptimize(thread, fr, ®_map);
1134 }
1135
1136
1137 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1138 if (thread == Thread::current()) {
1139 Deoptimization::deoptimize_frame_internal(thread, id);
1140 } else {
1141 VM_DeoptimizeFrame deopt(thread, id);
1142 VMThread::execute(&deopt);
1143 }
1144 }
1145
1146
1147 // JVMTI PopFrame support
1148 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1149 {
1150 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1151 }
1152 JRT_END
1153
1154 MethodData*
1155 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1156 bool create_if_missing) {
1157 Thread* THREAD = thread;
1158 MethodData* mdo = m()->method_data();
1159 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1160 // Build an MDO. Ignore errors like OutOfMemory;
1161 // that simply means we won't have an MDO to update.
1162 Method::build_interpreter_method_data(m, THREAD);
1163 if (HAS_PENDING_EXCEPTION) {
1164 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1165 CLEAR_PENDING_EXCEPTION;
1166 }
1167 mdo = m()->method_data();
1168 }
1169 return mdo;
1170 }
1171
1172 #if defined(COMPILER2) || defined(SHARK)
1173 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1174 // in case of an unresolved klass entry, load the class.
1175 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1176 Klass* tk = constant_pool->klass_at(index, CHECK);
1177 return;
1178 }
1179
1180 if (!constant_pool->tag_at(index).is_symbol()) return;
1181
1182 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1183 Symbol* symbol = constant_pool->symbol_at(index);
1184
1185 // class name?
1186 if (symbol->byte_at(0) != '(') {
1187 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1188 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1189 return;
1190 }
1191
1192 // then it must be a signature!
1193 ResourceMark rm(THREAD);
1194 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1195 if (ss.is_object()) {
1196 Symbol* class_name = ss.as_symbol(CHECK);
1197 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1198 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1199 }
1200 }
1201 }
1202
1203
1204 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1205 EXCEPTION_MARK;
1206 load_class_by_index(constant_pool, index, THREAD);
1207 if (HAS_PENDING_EXCEPTION) {
1208 // Exception happened during classloading. We ignore the exception here, since it
1209 // is going to be rethrown since the current activation is going to be deoptimized and
1210 // the interpreter will re-execute the bytecode.
1211 CLEAR_PENDING_EXCEPTION;
1212 // Class loading called java code which may have caused a stack
1213 // overflow. If the exception was thrown right before the return
1214 // to the runtime the stack is no longer guarded. Reguard the
1215 // stack otherwise if we return to the uncommon trap blob and the
1216 // stack bang causes a stack overflow we crash.
1217 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1218 JavaThread* thread = (JavaThread*)THREAD;
1219 bool guard_pages_enabled = thread->stack_yellow_zone_enabled();
1220 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1221 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1222 }
1223 }
1224
1225 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1226 HandleMark hm;
1227
1228 // uncommon_trap() is called at the beginning of the uncommon trap
1229 // handler. Note this fact before we start generating temporary frames
1230 // that can confuse an asynchronous stack walker. This counter is
1231 // decremented at the end of unpack_frames().
1232 thread->inc_in_deopt_handler();
1233
1234 // We need to update the map if we have biased locking.
1235 RegisterMap reg_map(thread, UseBiasedLocking);
1236 frame stub_frame = thread->last_frame();
1237 frame fr = stub_frame.sender(®_map);
1238 // Make sure the calling nmethod is not getting deoptimized and removed
1239 // before we are done with it.
1240 nmethodLocker nl(fr.pc());
1241
1242 // Log a message
1243 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1244 trap_request, fr.pc());
1245
1246 {
1247 ResourceMark rm;
1248
1249 // Revoke biases of any monitors in the frame to ensure we can migrate them
1250 revoke_biases_of_monitors(thread, fr, ®_map);
1251
1252 DeoptReason reason = trap_request_reason(trap_request);
1253 DeoptAction action = trap_request_action(trap_request);
1254 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1255
1256 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1257 compiledVFrame* cvf = compiledVFrame::cast(vf);
1258
1259 nmethod* nm = cvf->code();
1260
1261 ScopeDesc* trap_scope = cvf->scope();
1262 methodHandle trap_method = trap_scope->method();
1263 int trap_bci = trap_scope->bci();
1264 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1265
1266 // Record this event in the histogram.
1267 gather_statistics(reason, action, trap_bc);
1268
1269 // Ensure that we can record deopt. history:
1270 // Need MDO to record RTM code generation state.
1271 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1272
1273 MethodData* trap_mdo =
1274 get_method_data(thread, trap_method, create_if_missing);
1275
1276 // Log a message
1277 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1278 trap_reason_name(reason), trap_action_name(action), fr.pc(),
1279 trap_method->name_and_sig_as_C_string(), trap_bci);
1280
1281 // Print a bunch of diagnostics, if requested.
1282 if (TraceDeoptimization || LogCompilation) {
1283 ResourceMark rm;
1284 ttyLocker ttyl;
1285 char buf[100];
1286 if (xtty != NULL) {
1287 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1288 os::current_thread_id(),
1289 format_trap_request(buf, sizeof(buf), trap_request));
1290 nm->log_identity(xtty);
1291 }
1292 Symbol* class_name = NULL;
1293 bool unresolved = false;
1294 if (unloaded_class_index >= 0) {
1295 constantPoolHandle constants (THREAD, trap_method->constants());
1296 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1297 class_name = constants->klass_name_at(unloaded_class_index);
1298 unresolved = true;
1299 if (xtty != NULL)
1300 xtty->print(" unresolved='1'");
1301 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1302 class_name = constants->symbol_at(unloaded_class_index);
1303 }
1304 if (xtty != NULL)
1305 xtty->name(class_name);
1306 }
1307 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
1308 // Dump the relevant MDO state.
1309 // This is the deopt count for the current reason, any previous
1310 // reasons or recompiles seen at this point.
1311 int dcnt = trap_mdo->trap_count(reason);
1312 if (dcnt != 0)
1313 xtty->print(" count='%d'", dcnt);
1314 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1315 int dos = (pdata == NULL)? 0: pdata->trap_state();
1316 if (dos != 0) {
1317 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1318 if (trap_state_is_recompiled(dos)) {
1319 int recnt2 = trap_mdo->overflow_recompile_count();
1320 if (recnt2 != 0)
1321 xtty->print(" recompiles2='%d'", recnt2);
1322 }
1323 }
1324 }
1325 if (xtty != NULL) {
1326 xtty->stamp();
1327 xtty->end_head();
1328 }
1329 if (TraceDeoptimization) { // make noise on the tty
1330 tty->print("Uncommon trap occurred in");
1331 nm->method()->print_short_name(tty);
1332 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1333 fr.pc(),
1334 os::current_thread_id(),
1335 trap_reason_name(reason),
1336 trap_action_name(action),
1337 unloaded_class_index);
1338 if (class_name != NULL) {
1339 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1340 class_name->print_symbol_on(tty);
1341 }
1342 tty->cr();
1343 }
1344 if (xtty != NULL) {
1345 // Log the precise location of the trap.
1346 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1347 xtty->begin_elem("jvms bci='%d'", sd->bci());
1348 xtty->method(sd->method());
1349 xtty->end_elem();
1350 if (sd->is_top()) break;
1351 }
1352 xtty->tail("uncommon_trap");
1353 }
1354 }
1355 // (End diagnostic printout.)
1356
1357 // Load class if necessary
1358 if (unloaded_class_index >= 0) {
1359 constantPoolHandle constants(THREAD, trap_method->constants());
1360 load_class_by_index(constants, unloaded_class_index);
1361 }
1362
1363 // Flush the nmethod if necessary and desirable.
1364 //
1365 // We need to avoid situations where we are re-flushing the nmethod
1366 // because of a hot deoptimization site. Repeated flushes at the same
1367 // point need to be detected by the compiler and avoided. If the compiler
1368 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1369 // module must take measures to avoid an infinite cycle of recompilation
1370 // and deoptimization. There are several such measures:
1371 //
1372 // 1. If a recompilation is ordered a second time at some site X
1373 // and for the same reason R, the action is adjusted to 'reinterpret',
1374 // to give the interpreter time to exercise the method more thoroughly.
1375 // If this happens, the method's overflow_recompile_count is incremented.
1376 //
1377 // 2. If the compiler fails to reduce the deoptimization rate, then
1378 // the method's overflow_recompile_count will begin to exceed the set
1379 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1380 // is adjusted to 'make_not_compilable', and the method is abandoned
1381 // to the interpreter. This is a performance hit for hot methods,
1382 // but is better than a disastrous infinite cycle of recompilations.
1383 // (Actually, only the method containing the site X is abandoned.)
1384 //
1385 // 3. In parallel with the previous measures, if the total number of
1386 // recompilations of a method exceeds the much larger set limit
1387 // PerMethodRecompilationCutoff, the method is abandoned.
1388 // This should only happen if the method is very large and has
1389 // many "lukewarm" deoptimizations. The code which enforces this
1390 // limit is elsewhere (class nmethod, class Method).
1391 //
1392 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1393 // to recompile at each bytecode independently of the per-BCI cutoff.
1394 //
1395 // The decision to update code is up to the compiler, and is encoded
1396 // in the Action_xxx code. If the compiler requests Action_none
1397 // no trap state is changed, no compiled code is changed, and the
1398 // computation suffers along in the interpreter.
1399 //
1400 // The other action codes specify various tactics for decompilation
1401 // and recompilation. Action_maybe_recompile is the loosest, and
1402 // allows the compiled code to stay around until enough traps are seen,
1403 // and until the compiler gets around to recompiling the trapping method.
1404 //
1405 // The other actions cause immediate removal of the present code.
1406
1407 bool update_trap_state = (reason != Reason_tenured);
1408 bool make_not_entrant = false;
1409 bool make_not_compilable = false;
1410 bool reprofile = false;
1411 switch (action) {
1412 case Action_none:
1413 // Keep the old code.
1414 update_trap_state = false;
1415 break;
1416 case Action_maybe_recompile:
1417 // Do not need to invalidate the present code, but we can
1418 // initiate another
1419 // Start compiler without (necessarily) invalidating the nmethod.
1420 // The system will tolerate the old code, but new code should be
1421 // generated when possible.
1422 break;
1423 case Action_reinterpret:
1424 // Go back into the interpreter for a while, and then consider
1425 // recompiling form scratch.
1426 make_not_entrant = true;
1427 // Reset invocation counter for outer most method.
1428 // This will allow the interpreter to exercise the bytecodes
1429 // for a while before recompiling.
1430 // By contrast, Action_make_not_entrant is immediate.
1431 //
1432 // Note that the compiler will track null_check, null_assert,
1433 // range_check, and class_check events and log them as if they
1434 // had been traps taken from compiled code. This will update
1435 // the MDO trap history so that the next compilation will
1436 // properly detect hot trap sites.
1437 reprofile = true;
1438 break;
1439 case Action_make_not_entrant:
1440 // Request immediate recompilation, and get rid of the old code.
1441 // Make them not entrant, so next time they are called they get
1442 // recompiled. Unloaded classes are loaded now so recompile before next
1443 // time they are called. Same for uninitialized. The interpreter will
1444 // link the missing class, if any.
1445 make_not_entrant = true;
1446 break;
1447 case Action_make_not_compilable:
1448 // Give up on compiling this method at all.
1449 make_not_entrant = true;
1450 make_not_compilable = true;
1451 break;
1452 default:
1453 ShouldNotReachHere();
1454 }
1455
1456 // Setting +ProfileTraps fixes the following, on all platforms:
1457 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1458 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1459 // recompile relies on a MethodData* to record heroic opt failures.
1460
1461 // Whether the interpreter is producing MDO data or not, we also need
1462 // to use the MDO to detect hot deoptimization points and control
1463 // aggressive optimization.
1464 bool inc_recompile_count = false;
1465 ProfileData* pdata = NULL;
1466 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1467 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1468 uint this_trap_count = 0;
1469 bool maybe_prior_trap = false;
1470 bool maybe_prior_recompile = false;
1471 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1472 nm->method(),
1473 //outputs:
1474 this_trap_count,
1475 maybe_prior_trap,
1476 maybe_prior_recompile);
1477 // Because the interpreter also counts null, div0, range, and class
1478 // checks, these traps from compiled code are double-counted.
1479 // This is harmless; it just means that the PerXTrapLimit values
1480 // are in effect a little smaller than they look.
1481
1482 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1483 if (per_bc_reason != Reason_none) {
1484 // Now take action based on the partially known per-BCI history.
1485 if (maybe_prior_trap
1486 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1487 // If there are too many traps at this BCI, force a recompile.
1488 // This will allow the compiler to see the limit overflow, and
1489 // take corrective action, if possible. The compiler generally
1490 // does not use the exact PerBytecodeTrapLimit value, but instead
1491 // changes its tactics if it sees any traps at all. This provides
1492 // a little hysteresis, delaying a recompile until a trap happens
1493 // several times.
1494 //
1495 // Actually, since there is only one bit of counter per BCI,
1496 // the possible per-BCI counts are {0,1,(per-method count)}.
1497 // This produces accurate results if in fact there is only
1498 // one hot trap site, but begins to get fuzzy if there are
1499 // many sites. For example, if there are ten sites each
1500 // trapping two or more times, they each get the blame for
1501 // all of their traps.
1502 make_not_entrant = true;
1503 }
1504
1505 // Detect repeated recompilation at the same BCI, and enforce a limit.
1506 if (make_not_entrant && maybe_prior_recompile) {
1507 // More than one recompile at this point.
1508 inc_recompile_count = maybe_prior_trap;
1509 }
1510 } else {
1511 // For reasons which are not recorded per-bytecode, we simply
1512 // force recompiles unconditionally.
1513 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1514 make_not_entrant = true;
1515 }
1516
1517 // Go back to the compiler if there are too many traps in this method.
1518 if (this_trap_count >= per_method_trap_limit(reason)) {
1519 // If there are too many traps in this method, force a recompile.
1520 // This will allow the compiler to see the limit overflow, and
1521 // take corrective action, if possible.
1522 // (This condition is an unlikely backstop only, because the
1523 // PerBytecodeTrapLimit is more likely to take effect first,
1524 // if it is applicable.)
1525 make_not_entrant = true;
1526 }
1527
1528 // Here's more hysteresis: If there has been a recompile at
1529 // this trap point already, run the method in the interpreter
1530 // for a while to exercise it more thoroughly.
1531 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1532 reprofile = true;
1533 }
1534 }
1535
1536 // Take requested actions on the method:
1537
1538 // Recompile
1539 if (make_not_entrant) {
1540 if (!nm->make_not_entrant()) {
1541 return; // the call did not change nmethod's state
1542 }
1543
1544 if (pdata != NULL) {
1545 // Record the recompilation event, if any.
1546 int tstate0 = pdata->trap_state();
1547 int tstate1 = trap_state_set_recompiled(tstate0, true);
1548 if (tstate1 != tstate0)
1549 pdata->set_trap_state(tstate1);
1550 }
1551
1552 #if INCLUDE_RTM_OPT
1553 // Restart collecting RTM locking abort statistic if the method
1554 // is recompiled for a reason other than RTM state change.
1555 // Assume that in new recompiled code the statistic could be different,
1556 // for example, due to different inlining.
1557 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1558 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) {
1559 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1560 }
1561 #endif
1562 // For code aging we count traps separately here, using make_not_entrant()
1563 // as a guard against simultaneous deopts in multiple threads.
1564 if (reason == Reason_tenured && trap_mdo != NULL) {
1565 trap_mdo->inc_tenure_traps();
1566 }
1567 }
1568
1569 if (inc_recompile_count) {
1570 trap_mdo->inc_overflow_recompile_count();
1571 if ((uint)trap_mdo->overflow_recompile_count() >
1572 (uint)PerBytecodeRecompilationCutoff) {
1573 // Give up on the method containing the bad BCI.
1574 if (trap_method() == nm->method()) {
1575 make_not_compilable = true;
1576 } else {
1577 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1578 // But give grace to the enclosing nm->method().
1579 }
1580 }
1581 }
1582
1583 // Reprofile
1584 if (reprofile) {
1585 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1586 }
1587
1588 // Give up compiling
1589 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1590 assert(make_not_entrant, "consistent");
1591 nm->method()->set_not_compilable(CompLevel_full_optimization);
1592 }
1593
1594 } // Free marked resources
1595
1596 }
1597 JRT_END
1598
1599 ProfileData*
1600 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1601 int trap_bci,
1602 Deoptimization::DeoptReason reason,
1603 Method* compiled_method,
1604 //outputs:
1605 uint& ret_this_trap_count,
1606 bool& ret_maybe_prior_trap,
1607 bool& ret_maybe_prior_recompile) {
1608 uint prior_trap_count = trap_mdo->trap_count(reason);
1609 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1610
1611 // If the runtime cannot find a place to store trap history,
1612 // it is estimated based on the general condition of the method.
1613 // If the method has ever been recompiled, or has ever incurred
1614 // a trap with the present reason , then this BCI is assumed
1615 // (pessimistically) to be the culprit.
1616 bool maybe_prior_trap = (prior_trap_count != 0);
1617 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1618 ProfileData* pdata = NULL;
1619
1620
1621 // For reasons which are recorded per bytecode, we check per-BCI data.
1622 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1623 if (per_bc_reason != Reason_none) {
1624 // Find the profile data for this BCI. If there isn't one,
1625 // try to allocate one from the MDO's set of spares.
1626 // This will let us detect a repeated trap at this point.
1627 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1628
1629 if (pdata != NULL) {
1630 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
1631 if (LogCompilation && xtty != NULL) {
1632 ttyLocker ttyl;
1633 // no more room for speculative traps in this MDO
1634 xtty->elem("speculative_traps_oom");
1635 }
1636 }
1637 // Query the trap state of this profile datum.
1638 int tstate0 = pdata->trap_state();
1639 if (!trap_state_has_reason(tstate0, per_bc_reason))
1640 maybe_prior_trap = false;
1641 if (!trap_state_is_recompiled(tstate0))
1642 maybe_prior_recompile = false;
1643
1644 // Update the trap state of this profile datum.
1645 int tstate1 = tstate0;
1646 // Record the reason.
1647 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1648 // Store the updated state on the MDO, for next time.
1649 if (tstate1 != tstate0)
1650 pdata->set_trap_state(tstate1);
1651 } else {
1652 if (LogCompilation && xtty != NULL) {
1653 ttyLocker ttyl;
1654 // Missing MDP? Leave a small complaint in the log.
1655 xtty->elem("missing_mdp bci='%d'", trap_bci);
1656 }
1657 }
1658 }
1659
1660 // Return results:
1661 ret_this_trap_count = this_trap_count;
1662 ret_maybe_prior_trap = maybe_prior_trap;
1663 ret_maybe_prior_recompile = maybe_prior_recompile;
1664 return pdata;
1665 }
1666
1667 void
1668 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1669 ResourceMark rm;
1670 // Ignored outputs:
1671 uint ignore_this_trap_count;
1672 bool ignore_maybe_prior_trap;
1673 bool ignore_maybe_prior_recompile;
1674 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
1675 query_update_method_data(trap_mdo, trap_bci,
1676 (DeoptReason)reason,
1677 NULL,
1678 ignore_this_trap_count,
1679 ignore_maybe_prior_trap,
1680 ignore_maybe_prior_recompile);
1681 }
1682
1683 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1684
1685 // Still in Java no safepoints
1686 {
1687 // This enters VM and may safepoint
1688 uncommon_trap_inner(thread, trap_request);
1689 }
1690 return fetch_unroll_info_helper(thread);
1691 }
1692
1693 // Local derived constants.
1694 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1695 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1696 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1697
1698 //---------------------------trap_state_reason---------------------------------
1699 Deoptimization::DeoptReason
1700 Deoptimization::trap_state_reason(int trap_state) {
1701 // This assert provides the link between the width of DataLayout::trap_bits
1702 // and the encoding of "recorded" reasons. It ensures there are enough
1703 // bits to store all needed reasons in the per-BCI MDO profile.
1704 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1705 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1706 trap_state -= recompile_bit;
1707 if (trap_state == DS_REASON_MASK) {
1708 return Reason_many;
1709 } else {
1710 assert((int)Reason_none == 0, "state=0 => Reason_none");
1711 return (DeoptReason)trap_state;
1712 }
1713 }
1714 //-------------------------trap_state_has_reason-------------------------------
1715 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1716 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1717 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1718 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1719 trap_state -= recompile_bit;
1720 if (trap_state == DS_REASON_MASK) {
1721 return -1; // true, unspecifically (bottom of state lattice)
1722 } else if (trap_state == reason) {
1723 return 1; // true, definitely
1724 } else if (trap_state == 0) {
1725 return 0; // false, definitely (top of state lattice)
1726 } else {
1727 return 0; // false, definitely
1728 }
1729 }
1730 //-------------------------trap_state_add_reason-------------------------------
1731 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1732 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1733 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1734 trap_state -= recompile_bit;
1735 if (trap_state == DS_REASON_MASK) {
1736 return trap_state + recompile_bit; // already at state lattice bottom
1737 } else if (trap_state == reason) {
1738 return trap_state + recompile_bit; // the condition is already true
1739 } else if (trap_state == 0) {
1740 return reason + recompile_bit; // no condition has yet been true
1741 } else {
1742 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1743 }
1744 }
1745 //-----------------------trap_state_is_recompiled------------------------------
1746 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1747 return (trap_state & DS_RECOMPILE_BIT) != 0;
1748 }
1749 //-----------------------trap_state_set_recompiled-----------------------------
1750 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1751 if (z) return trap_state | DS_RECOMPILE_BIT;
1752 else return trap_state & ~DS_RECOMPILE_BIT;
1753 }
1754 //---------------------------format_trap_state---------------------------------
1755 // This is used for debugging and diagnostics, including LogFile output.
1756 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1757 int trap_state) {
1758 DeoptReason reason = trap_state_reason(trap_state);
1759 bool recomp_flag = trap_state_is_recompiled(trap_state);
1760 // Re-encode the state from its decoded components.
1761 int decoded_state = 0;
1762 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1763 decoded_state = trap_state_add_reason(decoded_state, reason);
1764 if (recomp_flag)
1765 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1766 // If the state re-encodes properly, format it symbolically.
1767 // Because this routine is used for debugging and diagnostics,
1768 // be robust even if the state is a strange value.
1769 size_t len;
1770 if (decoded_state != trap_state) {
1771 // Random buggy state that doesn't decode??
1772 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1773 } else {
1774 len = jio_snprintf(buf, buflen, "%s%s",
1775 trap_reason_name(reason),
1776 recomp_flag ? " recompiled" : "");
1777 }
1778 if (len >= buflen)
1779 buf[buflen-1] = '\0';
1780 return buf;
1781 }
1782
1783
1784 //--------------------------------statics--------------------------------------
1785 const char* Deoptimization::_trap_reason_name[] = {
1786 // Note: Keep this in sync. with enum DeoptReason.
1787 "none",
1788 "null_check",
1789 "null_assert",
1790 "range_check",
1791 "class_check",
1792 "array_check",
1793 "intrinsic",
1794 "bimorphic",
1795 "unloaded",
1796 "uninitialized",
1797 "unreached",
1798 "unhandled",
1799 "constraint",
1800 "div0_check",
1801 "age",
1802 "predicate",
1803 "loop_limit_check",
1804 "speculate_class_check",
1805 "speculate_null_check",
1806 "rtm_state_change",
1807 "unstable_if",
1808 "tenured"
1809 };
1810 const char* Deoptimization::_trap_action_name[] = {
1811 // Note: Keep this in sync. with enum DeoptAction.
1812 "none",
1813 "maybe_recompile",
1814 "reinterpret",
1815 "make_not_entrant",
1816 "make_not_compilable"
1817 };
1818
1819 const char* Deoptimization::trap_reason_name(int reason) {
1820 // Check that every reason has a name
1821 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
1822
1823 if (reason == Reason_many) return "many";
1824 if ((uint)reason < Reason_LIMIT)
1825 return _trap_reason_name[reason];
1826 static char buf[20];
1827 sprintf(buf, "reason%d", reason);
1828 return buf;
1829 }
1830 const char* Deoptimization::trap_action_name(int action) {
1831 // Check that every action has a name
1832 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
1833
1834 if ((uint)action < Action_LIMIT)
1835 return _trap_action_name[action];
1836 static char buf[20];
1837 sprintf(buf, "action%d", action);
1838 return buf;
1839 }
1840
1841 // This is used for debugging and diagnostics, including LogFile output.
1842 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1843 int trap_request) {
1844 jint unloaded_class_index = trap_request_index(trap_request);
1845 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1846 const char* action = trap_action_name(trap_request_action(trap_request));
1847 size_t len;
1848 if (unloaded_class_index < 0) {
1849 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1850 reason, action);
1851 } else {
1852 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1853 reason, action, unloaded_class_index);
1854 }
1855 if (len >= buflen)
1856 buf[buflen-1] = '\0';
1857 return buf;
1858 }
1859
1860 juint Deoptimization::_deoptimization_hist
1861 [Deoptimization::Reason_LIMIT]
1862 [1 + Deoptimization::Action_LIMIT]
1863 [Deoptimization::BC_CASE_LIMIT]
1864 = {0};
1865
1866 enum {
1867 LSB_BITS = 8,
1868 LSB_MASK = right_n_bits(LSB_BITS)
1869 };
1870
1871 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1872 Bytecodes::Code bc) {
1873 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1874 assert(action >= 0 && action < Action_LIMIT, "oob");
1875 _deoptimization_hist[Reason_none][0][0] += 1; // total
1876 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1877 juint* cases = _deoptimization_hist[reason][1+action];
1878 juint* bc_counter_addr = NULL;
1879 juint bc_counter = 0;
1880 // Look for an unused counter, or an exact match to this BC.
1881 if (bc != Bytecodes::_illegal) {
1882 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1883 juint* counter_addr = &cases[bc_case];
1884 juint counter = *counter_addr;
1885 if ((counter == 0 && bc_counter_addr == NULL)
1886 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1887 // this counter is either free or is already devoted to this BC
1888 bc_counter_addr = counter_addr;
1889 bc_counter = counter | bc;
1890 }
1891 }
1892 }
1893 if (bc_counter_addr == NULL) {
1894 // Overflow, or no given bytecode.
1895 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1896 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1897 }
1898 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1899 }
1900
1901 jint Deoptimization::total_deoptimization_count() {
1902 return _deoptimization_hist[Reason_none][0][0];
1903 }
1904
1905 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1906 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1907 return _deoptimization_hist[reason][0][0];
1908 }
1909
1910 void Deoptimization::print_statistics() {
1911 juint total = total_deoptimization_count();
1912 juint account = total;
1913 if (total != 0) {
1914 ttyLocker ttyl;
1915 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1916 tty->print_cr("Deoptimization traps recorded:");
1917 #define PRINT_STAT_LINE(name, r) \
1918 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1919 PRINT_STAT_LINE("total", total);
1920 // For each non-zero entry in the histogram, print the reason,
1921 // the action, and (if specifically known) the type of bytecode.
1922 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1923 for (int action = 0; action < Action_LIMIT; action++) {
1924 juint* cases = _deoptimization_hist[reason][1+action];
1925 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1926 juint counter = cases[bc_case];
1927 if (counter != 0) {
1928 char name[1*K];
1929 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1930 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1931 bc = Bytecodes::_illegal;
1932 sprintf(name, "%s/%s/%s",
1933 trap_reason_name(reason),
1934 trap_action_name(action),
1935 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1936 juint r = counter >> LSB_BITS;
1937 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1938 account -= r;
1939 }
1940 }
1941 }
1942 }
1943 if (account != 0) {
1944 PRINT_STAT_LINE("unaccounted", account);
1945 }
1946 #undef PRINT_STAT_LINE
1947 if (xtty != NULL) xtty->tail("statistics");
1948 }
1949 }
1950 #else // COMPILER2 || SHARK
1951
1952
1953 // Stubs for C1 only system.
1954 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1955 return false;
1956 }
1957
1958 const char* Deoptimization::trap_reason_name(int reason) {
1959 return "unknown";
1960 }
1961
1962 void Deoptimization::print_statistics() {
1963 // no output
1964 }
1965
1966 void
1967 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1968 // no udpate
1969 }
1970
1971 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1972 return 0;
1973 }
1974
1975 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1976 Bytecodes::Code bc) {
1977 // no update
1978 }
1979
1980 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1981 int trap_state) {
1982 jio_snprintf(buf, buflen, "#%d", trap_state);
1983 return buf;
1984 }
1985
1986 #endif // COMPILER2 || SHARK