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/debugInfoRec.hpp"
28 #include "code/nmethod.hpp"
29 #include "code/pcDesc.hpp"
30 #include "code/scopeDesc.hpp"
31 #include "interpreter/bytecode.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/oopMapCache.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "memory/oopFactory.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/method.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "prims/jvmtiThreadState.hpp"
40 #include "runtime/biasedLocking.hpp"
41 #include "runtime/compilationPolicy.hpp"
42 #include "runtime/deoptimization.hpp"
43 #include "runtime/interfaceSupport.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/thread.hpp"
48 #include "runtime/vframe.hpp"
49 #include "runtime/vframeArray.hpp"
50 #include "runtime/vframe_hp.hpp"
51 #include "utilities/events.hpp"
52 #include "utilities/xmlstream.hpp"
53 #ifdef TARGET_ARCH_x86
54 # include "vmreg_x86.inline.hpp"
55 #endif
56 #ifdef TARGET_ARCH_sparc
57 # include "vmreg_sparc.inline.hpp"
58 #endif
59 #ifdef TARGET_ARCH_zero
60 # include "vmreg_zero.inline.hpp"
61 #endif
62 #ifdef TARGET_ARCH_arm
63 # include "vmreg_arm.inline.hpp"
64 #endif
65 #ifdef TARGET_ARCH_ppc
66 # include "vmreg_ppc.inline.hpp"
67 #endif
68 #ifdef COMPILER2
69 #ifdef TARGET_ARCH_MODEL_x86_32
70 # include "adfiles/ad_x86_32.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_x86_64
73 # include "adfiles/ad_x86_64.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_sparc
76 # include "adfiles/ad_sparc.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_arm
82 # include "adfiles/ad_arm.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc_32
85 # include "adfiles/ad_ppc_32.hpp"
86 #endif
87 #ifdef TARGET_ARCH_MODEL_ppc_64
88 # include "adfiles/ad_ppc_64.hpp"
89 #endif
90 #endif // COMPILER2
91
92 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
93
94 bool DeoptimizationMarker::_is_active = false;
95
96 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
97 int caller_adjustment,
98 int caller_actual_parameters,
99 int number_of_frames,
100 intptr_t* frame_sizes,
101 address* frame_pcs,
102 BasicType return_type) {
103 _size_of_deoptimized_frame = size_of_deoptimized_frame;
104 _caller_adjustment = caller_adjustment;
105 _caller_actual_parameters = caller_actual_parameters;
106 _number_of_frames = number_of_frames;
107 _frame_sizes = frame_sizes;
108 _frame_pcs = frame_pcs;
109 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
110 _return_type = return_type;
111 _initial_info = 0;
112 // PD (x86 only)
113 _counter_temp = 0;
114 _unpack_kind = 0;
115 _sender_sp_temp = 0;
116
117 _total_frame_sizes = size_of_frames();
118 }
119
120
121 Deoptimization::UnrollBlock::~UnrollBlock() {
122 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
123 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
124 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
125 }
126
127
128 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
129 assert(register_number < RegisterMap::reg_count, "checking register number");
130 return &_register_block[register_number * 2];
131 }
132
133
134
135 int Deoptimization::UnrollBlock::size_of_frames() const {
136 // Acount first for the adjustment of the initial frame
137 int result = _caller_adjustment;
138 for (int index = 0; index < number_of_frames(); index++) {
139 result += frame_sizes()[index];
140 }
141 return result;
142 }
143
144
145 void Deoptimization::UnrollBlock::print() {
146 ttyLocker ttyl;
147 tty->print_cr("UnrollBlock");
148 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
149 tty->print( " frame_sizes: ");
150 for (int index = 0; index < number_of_frames(); index++) {
151 tty->print("%d ", frame_sizes()[index]);
152 }
153 tty->cr();
154 }
155
156
157 // In order to make fetch_unroll_info work properly with escape
158 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
159 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
160 // of previously eliminated objects occurs in realloc_objects, which is
161 // called from the method fetch_unroll_info_helper below.
162 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
163 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
164 // but makes the entry a little slower. There is however a little dance we have to
165 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
166
167 // fetch_unroll_info() is called at the beginning of the deoptimization
168 // handler. Note this fact before we start generating temporary frames
169 // that can confuse an asynchronous stack walker. This counter is
170 // decremented at the end of unpack_frames().
171 thread->inc_in_deopt_handler();
172
173 return fetch_unroll_info_helper(thread);
174 JRT_END
175
176
177 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
178 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
179
180 // Note: there is a safepoint safety issue here. No matter whether we enter
181 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
182 // the vframeArray is created.
183 //
184
185 // Allocate our special deoptimization ResourceMark
186 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
187 assert(thread->deopt_mark() == NULL, "Pending deopt!");
188 thread->set_deopt_mark(dmark);
189
190 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
191 RegisterMap map(thread, true);
192 RegisterMap dummy_map(thread, false);
193 // Now get the deoptee with a valid map
194 frame deoptee = stub_frame.sender(&map);
195 // Set the deoptee nmethod
196 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
197 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
198
199 if (VerifyStack) {
200 thread->validate_frame_layout();
201 }
202
203 // Create a growable array of VFrames where each VFrame represents an inlined
204 // Java frame. This storage is allocated with the usual system arena.
205 assert(deoptee.is_compiled_frame(), "Wrong frame type");
206 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
207 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
208 while (!vf->is_top()) {
209 assert(vf->is_compiled_frame(), "Wrong frame type");
210 chunk->push(compiledVFrame::cast(vf));
211 vf = vf->sender();
212 }
213 assert(vf->is_compiled_frame(), "Wrong frame type");
214 chunk->push(compiledVFrame::cast(vf));
215
216 #ifdef COMPILER2
217 // Reallocate the non-escaping objects and restore their fields. Then
218 // relock objects if synchronization on them was eliminated.
219 if (DoEscapeAnalysis || EliminateNestedLocks) {
220 if (EliminateAllocations) {
221 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
222 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
223
224 // The flag return_oop() indicates call sites which return oop
225 // in compiled code. Such sites include java method calls,
226 // runtime calls (for example, used to allocate new objects/arrays
227 // on slow code path) and any other calls generated in compiled code.
228 // It is not guaranteed that we can get such information here only
229 // by analyzing bytecode in deoptimized frames. This is why this flag
230 // is set during method compilation (see Compile::Process_OopMap_Node()).
231 bool save_oop_result = chunk->at(0)->scope()->return_oop();
232 Handle return_value;
233 if (save_oop_result) {
234 // Reallocation may trigger GC. If deoptimization happened on return from
235 // call which returns oop we need to save it since it is not in oopmap.
236 oop result = deoptee.saved_oop_result(&map);
237 assert(result == NULL || result->is_oop(), "must be oop");
238 return_value = Handle(thread, result);
239 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
240 if (TraceDeoptimization) {
241 ttyLocker ttyl;
242 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread);
243 }
244 }
245 bool reallocated = false;
246 if (objects != NULL) {
247 JRT_BLOCK
248 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
249 JRT_END
250 }
251 if (reallocated) {
252 reassign_fields(&deoptee, &map, objects);
253 #ifndef PRODUCT
254 if (TraceDeoptimization) {
255 ttyLocker ttyl;
256 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
257 print_objects(objects);
258 }
259 #endif
260 }
261 if (save_oop_result) {
262 // Restore result.
263 deoptee.set_saved_oop_result(&map, return_value());
264 }
265 }
266 if (EliminateLocks) {
267 #ifndef PRODUCT
268 bool first = true;
269 #endif
270 for (int i = 0; i < chunk->length(); i++) {
271 compiledVFrame* cvf = chunk->at(i);
272 assert (cvf->scope() != NULL,"expect only compiled java frames");
273 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
274 if (monitors->is_nonempty()) {
275 relock_objects(monitors, thread);
276 #ifndef PRODUCT
277 if (TraceDeoptimization) {
278 ttyLocker ttyl;
279 for (int j = 0; j < monitors->length(); j++) {
280 MonitorInfo* mi = monitors->at(j);
281 if (mi->eliminated()) {
282 if (first) {
283 first = false;
284 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
285 }
286 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner());
287 }
288 }
289 }
290 #endif
291 }
292 }
293 }
294 }
295 #endif // COMPILER2
296 // Ensure that no safepoint is taken after pointers have been stored
297 // in fields of rematerialized objects. If a safepoint occurs from here on
298 // out the java state residing in the vframeArray will be missed.
299 No_Safepoint_Verifier no_safepoint;
300
301 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
302
303 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
304 thread->set_vframe_array_head(array);
305
306 // Now that the vframeArray has been created if we have any deferred local writes
307 // added by jvmti then we can free up that structure as the data is now in the
308 // vframeArray
309
310 if (thread->deferred_locals() != NULL) {
311 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
312 int i = 0;
313 do {
314 // Because of inlining we could have multiple vframes for a single frame
315 // and several of the vframes could have deferred writes. Find them all.
316 if (list->at(i)->id() == array->original().id()) {
317 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
318 list->remove_at(i);
319 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
320 delete dlv;
321 } else {
322 i++;
323 }
324 } while ( i < list->length() );
325 if (list->length() == 0) {
326 thread->set_deferred_locals(NULL);
327 // free the list and elements back to C heap.
328 delete list;
329 }
330
331 }
332
333 #ifndef SHARK
334 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
335 CodeBlob* cb = stub_frame.cb();
336 // Verify we have the right vframeArray
337 assert(cb->frame_size() >= 0, "Unexpected frame size");
338 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
339
340 // If the deopt call site is a MethodHandle invoke call site we have
341 // to adjust the unpack_sp.
342 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
343 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
344 unpack_sp = deoptee.unextended_sp();
345
346 #ifdef ASSERT
347 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
348 #endif
349 #else
350 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
351 #endif // !SHARK
352
353 // This is a guarantee instead of an assert because if vframe doesn't match
354 // we will unpack the wrong deoptimized frame and wind up in strange places
355 // where it will be very difficult to figure out what went wrong. Better
356 // to die an early death here than some very obscure death later when the
357 // trail is cold.
358 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
359 // in that it will fail to detect a problem when there is one. This needs
360 // more work in tiger timeframe.
361 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
362
363 int number_of_frames = array->frames();
364
365 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
366 // virtual activation, which is the reverse of the elements in the vframes array.
367 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
368 // +1 because we always have an interpreter return address for the final slot.
369 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
370 int popframe_extra_args = 0;
371 // Create an interpreter return address for the stub to use as its return
372 // address so the skeletal frames are perfectly walkable
373 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
374
375 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
376 // activation be put back on the expression stack of the caller for reexecution
377 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
378 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
379 }
380
381 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
382 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
383 // than simply use array->sender.pc(). This requires us to walk the current set of frames
384 //
385 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
386 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
387
388 // It's possible that the number of parameters at the call site is
389 // different than number of arguments in the callee when method
390 // handles are used. If the caller is interpreted get the real
391 // value so that the proper amount of space can be added to it's
392 // frame.
393 bool caller_was_method_handle = false;
394 if (deopt_sender.is_interpreted_frame()) {
395 methodHandle method = deopt_sender.interpreter_frame_method();
396 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
397 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
398 // Method handle invokes may involve fairly arbitrary chains of
399 // calls so it's impossible to know how much actual space the
400 // caller has for locals.
401 caller_was_method_handle = true;
402 }
403 }
404
405 //
406 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
407 // frame_sizes/frame_pcs[1] next oldest frame (int)
408 // frame_sizes/frame_pcs[n] youngest frame (int)
409 //
410 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
411 // owns the space for the return address to it's caller). Confusing ain't it.
412 //
413 // The vframe array can address vframes with indices running from
414 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
415 // When we create the skeletal frames we need the oldest frame to be in the zero slot
416 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
417 // so things look a little strange in this loop.
418 //
419 int callee_parameters = 0;
420 int callee_locals = 0;
421 for (int index = 0; index < array->frames(); index++ ) {
422 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
423 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
424 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
425 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
426 callee_locals,
427 index == 0,
428 popframe_extra_args);
429 // This pc doesn't have to be perfect just good enough to identify the frame
430 // as interpreted so the skeleton frame will be walkable
431 // The correct pc will be set when the skeleton frame is completely filled out
432 // The final pc we store in the loop is wrong and will be overwritten below
433 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
434
435 callee_parameters = array->element(index)->method()->size_of_parameters();
436 callee_locals = array->element(index)->method()->max_locals();
437 popframe_extra_args = 0;
438 }
439
440 // Compute whether the root vframe returns a float or double value.
441 BasicType return_type;
442 {
443 HandleMark hm;
444 methodHandle method(thread, array->element(0)->method());
445 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
446 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
447 }
448
449 // Compute information for handling adapters and adjusting the frame size of the caller.
450 int caller_adjustment = 0;
451
452 // Compute the amount the oldest interpreter frame will have to adjust
453 // its caller's stack by. If the caller is a compiled frame then
454 // we pretend that the callee has no parameters so that the
455 // extension counts for the full amount of locals and not just
456 // locals-parms. This is because without a c2i adapter the parm
457 // area as created by the compiled frame will not be usable by
458 // the interpreter. (Depending on the calling convention there
459 // may not even be enough space).
460
461 // QQQ I'd rather see this pushed down into last_frame_adjust
462 // and have it take the sender (aka caller).
463
464 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
465 caller_adjustment = last_frame_adjust(0, callee_locals);
466 } else if (callee_locals > callee_parameters) {
467 // The caller frame may need extending to accommodate
468 // non-parameter locals of the first unpacked interpreted frame.
469 // Compute that adjustment.
470 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
471 }
472
473 // If the sender is deoptimized the we must retrieve the address of the handler
474 // since the frame will "magically" show the original pc before the deopt
475 // and we'd undo the deopt.
476
477 frame_pcs[0] = deopt_sender.raw_pc();
478
479 #ifndef SHARK
480 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
481 #endif // SHARK
482
483 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
484 caller_adjustment * BytesPerWord,
485 caller_was_method_handle ? 0 : callee_parameters,
486 number_of_frames,
487 frame_sizes,
488 frame_pcs,
489 return_type);
490 // On some platforms, we need a way to pass some platform dependent
491 // information to the unpacking code so the skeletal frames come out
492 // correct (initial fp value, unextended sp, ...)
493 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
494
495 if (array->frames() > 1) {
496 if (VerifyStack && TraceDeoptimization) {
497 ttyLocker ttyl;
498 tty->print_cr("Deoptimizing method containing inlining");
499 }
500 }
501
502 array->set_unroll_block(info);
503 return info;
504 }
505
506 // Called to cleanup deoptimization data structures in normal case
507 // after unpacking to stack and when stack overflow error occurs
508 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
509 vframeArray *array) {
510
511 // Get array if coming from exception
512 if (array == NULL) {
513 array = thread->vframe_array_head();
514 }
515 thread->set_vframe_array_head(NULL);
516
517 // Free the previous UnrollBlock
518 vframeArray* old_array = thread->vframe_array_last();
519 thread->set_vframe_array_last(array);
520
521 if (old_array != NULL) {
522 UnrollBlock* old_info = old_array->unroll_block();
523 old_array->set_unroll_block(NULL);
524 delete old_info;
525 delete old_array;
526 }
527
528 // Deallocate any resource creating in this routine and any ResourceObjs allocated
529 // inside the vframeArray (StackValueCollections)
530
531 delete thread->deopt_mark();
532 thread->set_deopt_mark(NULL);
533 thread->set_deopt_nmethod(NULL);
534
535
536 if (JvmtiExport::can_pop_frame()) {
537 #ifndef CC_INTERP
538 // Regardless of whether we entered this routine with the pending
539 // popframe condition bit set, we should always clear it now
540 thread->clear_popframe_condition();
541 #else
542 // C++ interpreter will clear has_pending_popframe when it enters
543 // with method_resume. For deopt_resume2 we clear it now.
544 if (thread->popframe_forcing_deopt_reexecution())
545 thread->clear_popframe_condition();
546 #endif /* CC_INTERP */
547 }
548
549 // unpack_frames() is called at the end of the deoptimization handler
550 // and (in C2) at the end of the uncommon trap handler. Note this fact
551 // so that an asynchronous stack walker can work again. This counter is
552 // incremented at the beginning of fetch_unroll_info() and (in C2) at
553 // the beginning of uncommon_trap().
554 thread->dec_in_deopt_handler();
555 }
556
557
558 // Return BasicType of value being returned
559 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
560
561 // We are already active int he special DeoptResourceMark any ResourceObj's we
562 // allocate will be freed at the end of the routine.
563
564 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
565 // but makes the entry a little slower. There is however a little dance we have to
566 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
567 ResetNoHandleMark rnhm; // No-op in release/product versions
568 HandleMark hm;
569
570 frame stub_frame = thread->last_frame();
571
572 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
573 // must point to the vframeArray for the unpack frame.
574 vframeArray* array = thread->vframe_array_head();
575
576 #ifndef PRODUCT
577 if (TraceDeoptimization) {
578 ttyLocker ttyl;
579 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
580 }
581 #endif
582 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
583 stub_frame.pc(), stub_frame.sp(), exec_mode);
584
585 UnrollBlock* info = array->unroll_block();
586
587 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
588 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
589
590 BasicType bt = info->return_type();
591
592 // If we have an exception pending, claim that the return type is an oop
593 // so the deopt_blob does not overwrite the exception_oop.
594
595 if (exec_mode == Unpack_exception)
596 bt = T_OBJECT;
597
598 // Cleanup thread deopt data
599 cleanup_deopt_info(thread, array);
600
601 #ifndef PRODUCT
602 if (VerifyStack) {
603 ResourceMark res_mark;
604
605 thread->validate_frame_layout();
606
607 // Verify that the just-unpacked frames match the interpreter's
608 // notions of expression stack and locals
609 vframeArray* cur_array = thread->vframe_array_last();
610 RegisterMap rm(thread, false);
611 rm.set_include_argument_oops(false);
612 bool is_top_frame = true;
613 int callee_size_of_parameters = 0;
614 int callee_max_locals = 0;
615 for (int i = 0; i < cur_array->frames(); i++) {
616 vframeArrayElement* el = cur_array->element(i);
617 frame* iframe = el->iframe();
618 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
619
620 // Get the oop map for this bci
621 InterpreterOopMap mask;
622 int cur_invoke_parameter_size = 0;
623 bool try_next_mask = false;
624 int next_mask_expression_stack_size = -1;
625 int top_frame_expression_stack_adjustment = 0;
626 methodHandle mh(thread, iframe->interpreter_frame_method());
627 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
628 BytecodeStream str(mh);
629 str.set_start(iframe->interpreter_frame_bci());
630 int max_bci = mh->code_size();
631 // Get to the next bytecode if possible
632 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
633 // Check to see if we can grab the number of outgoing arguments
634 // at an uncommon trap for an invoke (where the compiler
635 // generates debug info before the invoke has executed)
636 Bytecodes::Code cur_code = str.next();
637 if (cur_code == Bytecodes::_invokevirtual ||
638 cur_code == Bytecodes::_invokespecial ||
639 cur_code == Bytecodes::_invokestatic ||
640 cur_code == Bytecodes::_invokeinterface ||
641 cur_code == Bytecodes::_invokedynamic) {
642 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
643 Symbol* signature = invoke.signature();
644 ArgumentSizeComputer asc(signature);
645 cur_invoke_parameter_size = asc.size();
646 if (invoke.has_receiver()) {
647 // Add in receiver
648 ++cur_invoke_parameter_size;
649 }
650 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
651 callee_size_of_parameters++;
652 }
653 }
654 if (str.bci() < max_bci) {
655 Bytecodes::Code bc = str.next();
656 if (bc >= 0) {
657 // The interpreter oop map generator reports results before
658 // the current bytecode has executed except in the case of
659 // calls. It seems to be hard to tell whether the compiler
660 // has emitted debug information matching the "state before"
661 // a given bytecode or the state after, so we try both
662 switch (cur_code) {
663 case Bytecodes::_invokevirtual:
664 case Bytecodes::_invokespecial:
665 case Bytecodes::_invokestatic:
666 case Bytecodes::_invokeinterface:
667 case Bytecodes::_invokedynamic:
668 case Bytecodes::_athrow:
669 break;
670 default: {
671 InterpreterOopMap next_mask;
672 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
673 next_mask_expression_stack_size = next_mask.expression_stack_size();
674 // Need to subtract off the size of the result type of
675 // the bytecode because this is not described in the
676 // debug info but returned to the interpreter in the TOS
677 // caching register
678 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
679 if (bytecode_result_type != T_ILLEGAL) {
680 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
681 }
682 assert(top_frame_expression_stack_adjustment >= 0, "");
683 try_next_mask = true;
684 break;
685 }
686 }
687 }
688 }
689
690 // Verify stack depth and oops in frame
691 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
692 if (!(
693 /* SPARC */
694 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
695 /* x86 */
696 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
697 (try_next_mask &&
698 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
699 top_frame_expression_stack_adjustment))) ||
700 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
701 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
702 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
703 )) {
704 ttyLocker ttyl;
705
706 // Print out some information that will help us debug the problem
707 tty->print_cr("Wrong number of expression stack elements during deoptimization");
708 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
709 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
710 iframe->interpreter_frame_expression_stack_size());
711 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
712 tty->print_cr(" try_next_mask = %d", try_next_mask);
713 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
714 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
715 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
716 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
717 tty->print_cr(" exec_mode = %d", exec_mode);
718 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
719 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
720 tty->print_cr(" Interpreted frames:");
721 for (int k = 0; k < cur_array->frames(); k++) {
722 vframeArrayElement* el = cur_array->element(k);
723 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
724 }
725 cur_array->print_on_2(tty);
726 guarantee(false, "wrong number of expression stack elements during deopt");
727 }
728 VerifyOopClosure verify;
729 iframe->oops_interpreted_do(&verify, NULL, &rm, false);
730 callee_size_of_parameters = mh->size_of_parameters();
731 callee_max_locals = mh->max_locals();
732 is_top_frame = false;
733 }
734 }
735 #endif /* !PRODUCT */
736
737
738 return bt;
739 JRT_END
740
741
742 int Deoptimization::deoptimize_dependents() {
743 Threads::deoptimized_wrt_marked_nmethods();
744 return 0;
745 }
746
747 Deoptimization::DeoptAction Deoptimization::_unloaded_action
748 = Deoptimization::Action_reinterpret;
749
750 #ifdef COMPILER2
751 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
752 Handle pending_exception(thread->pending_exception());
753 const char* exception_file = thread->exception_file();
754 int exception_line = thread->exception_line();
755 thread->clear_pending_exception();
756
757 for (int i = 0; i < objects->length(); i++) {
758 assert(objects->at(i)->is_object(), "invalid debug information");
759 ObjectValue* sv = (ObjectValue*) objects->at(i);
760
761 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
762 oop obj = NULL;
763
764 if (k->oop_is_instance()) {
765 InstanceKlass* ik = InstanceKlass::cast(k());
766 obj = ik->allocate_instance(CHECK_(false));
767 } else if (k->oop_is_typeArray()) {
768 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
769 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
770 int len = sv->field_size() / type2size[ak->element_type()];
771 obj = ak->allocate(len, CHECK_(false));
772 } else if (k->oop_is_objArray()) {
773 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
774 obj = ak->allocate(sv->field_size(), CHECK_(false));
775 }
776
777 assert(obj != NULL, "allocation failed");
778 assert(sv->value().is_null(), "redundant reallocation");
779 sv->set_value(obj);
780 }
781
782 if (pending_exception.not_null()) {
783 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
784 }
785
786 return true;
787 }
788
789 // This assumes that the fields are stored in ObjectValue in the same order
790 // they are yielded by do_nonstatic_fields.
791 class FieldReassigner: public FieldClosure {
792 frame* _fr;
793 RegisterMap* _reg_map;
794 ObjectValue* _sv;
795 InstanceKlass* _ik;
796 oop _obj;
797
798 int _i;
799 public:
800 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
801 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
802
803 int i() const { return _i; }
804
805
806 void do_field(fieldDescriptor* fd) {
807 intptr_t val;
808 StackValue* value =
809 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
810 int offset = fd->offset();
811 switch (fd->field_type()) {
812 case T_OBJECT: case T_ARRAY:
813 assert(value->type() == T_OBJECT, "Agreement.");
814 _obj->obj_field_put(offset, value->get_obj()());
815 break;
816
817 case T_LONG: case T_DOUBLE: {
818 assert(value->type() == T_INT, "Agreement.");
819 StackValue* low =
820 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
821 #ifdef _LP64
822 jlong res = (jlong)low->get_int();
823 #else
824 #ifdef SPARC
825 // For SPARC we have to swap high and low words.
826 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
827 #else
828 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
829 #endif //SPARC
830 #endif
831 _obj->long_field_put(offset, res);
832 break;
833 }
834 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
835 case T_INT: case T_FLOAT: // 4 bytes.
836 assert(value->type() == T_INT, "Agreement.");
837 val = value->get_int();
838 _obj->int_field_put(offset, (jint)*((jint*)&val));
839 break;
840
841 case T_SHORT: case T_CHAR: // 2 bytes
842 assert(value->type() == T_INT, "Agreement.");
843 val = value->get_int();
844 _obj->short_field_put(offset, (jshort)*((jint*)&val));
845 break;
846
847 case T_BOOLEAN: case T_BYTE: // 1 byte
848 assert(value->type() == T_INT, "Agreement.");
849 val = value->get_int();
850 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
851 break;
852
853 default:
854 ShouldNotReachHere();
855 }
856 _i++;
857 }
858 };
859
860 // restore elements of an eliminated type array
861 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
862 int index = 0;
863 intptr_t val;
864
865 for (int i = 0; i < sv->field_size(); i++) {
866 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
867 switch(type) {
868 case T_LONG: case T_DOUBLE: {
869 assert(value->type() == T_INT, "Agreement.");
870 StackValue* low =
871 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
872 #ifdef _LP64
873 jlong res = (jlong)low->get_int();
874 #else
875 #ifdef SPARC
876 // For SPARC we have to swap high and low words.
877 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
878 #else
879 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
880 #endif //SPARC
881 #endif
882 obj->long_at_put(index, res);
883 break;
884 }
885
886 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
887 case T_INT: case T_FLOAT: // 4 bytes.
888 assert(value->type() == T_INT, "Agreement.");
889 val = value->get_int();
890 obj->int_at_put(index, (jint)*((jint*)&val));
891 break;
892
893 case T_SHORT: case T_CHAR: // 2 bytes
894 assert(value->type() == T_INT, "Agreement.");
895 val = value->get_int();
896 obj->short_at_put(index, (jshort)*((jint*)&val));
897 break;
898
899 case T_BOOLEAN: case T_BYTE: // 1 byte
900 assert(value->type() == T_INT, "Agreement.");
901 val = value->get_int();
902 obj->bool_at_put(index, (jboolean)*((jint*)&val));
903 break;
904
905 default:
906 ShouldNotReachHere();
907 }
908 index++;
909 }
910 }
911
912
913 // restore fields of an eliminated object array
914 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
915 for (int i = 0; i < sv->field_size(); i++) {
916 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
917 assert(value->type() == T_OBJECT, "object element expected");
918 obj->obj_at_put(i, value->get_obj()());
919 }
920 }
921
922
923 // restore fields of all eliminated objects and arrays
924 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
925 for (int i = 0; i < objects->length(); i++) {
926 ObjectValue* sv = (ObjectValue*) objects->at(i);
927 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
928 Handle obj = sv->value();
929 assert(obj.not_null(), "reallocation was missed");
930
931 if (k->oop_is_instance()) {
932 InstanceKlass* ik = InstanceKlass::cast(k());
933 FieldReassigner reassign(fr, reg_map, sv, obj());
934 ik->do_nonstatic_fields(&reassign);
935 } else if (k->oop_is_typeArray()) {
936 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
937 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
938 } else if (k->oop_is_objArray()) {
939 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
940 }
941 }
942 }
943
944
945 // relock objects for which synchronization was eliminated
946 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
947 for (int i = 0; i < monitors->length(); i++) {
948 MonitorInfo* mon_info = monitors->at(i);
949 if (mon_info->eliminated()) {
950 assert(mon_info->owner() != NULL, "reallocation was missed");
951 Handle obj = Handle(mon_info->owner());
952 markOop mark = obj->mark();
953 if (UseBiasedLocking && mark->has_bias_pattern()) {
954 // New allocated objects may have the mark set to anonymously biased.
955 // Also the deoptimized method may called methods with synchronization
956 // where the thread-local object is bias locked to the current thread.
957 assert(mark->is_biased_anonymously() ||
958 mark->biased_locker() == thread, "should be locked to current thread");
959 // Reset mark word to unbiased prototype.
960 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
961 obj->set_mark(unbiased_prototype);
962 }
963 BasicLock* lock = mon_info->lock();
964 ObjectSynchronizer::slow_enter(obj, lock, thread);
965 }
966 assert(mon_info->owner()->is_locked(), "object must be locked now");
967 }
968 }
969
970
971 #ifndef PRODUCT
972 // print information about reallocated objects
973 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
974 fieldDescriptor fd;
975
976 for (int i = 0; i < objects->length(); i++) {
977 ObjectValue* sv = (ObjectValue*) objects->at(i);
978 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
979 Handle obj = sv->value();
980
981 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()());
982 k->print_value();
983 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
984 tty->cr();
985
986 if (Verbose) {
987 k->oop_print_on(obj(), tty);
988 }
989 }
990 }
991 #endif
992 #endif // COMPILER2
993
994 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
995 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
996
997 #ifndef PRODUCT
998 if (TraceDeoptimization) {
999 ttyLocker ttyl;
1000 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
1001 fr.print_on(tty);
1002 tty->print_cr(" Virtual frames (innermost first):");
1003 for (int index = 0; index < chunk->length(); index++) {
1004 compiledVFrame* vf = chunk->at(index);
1005 tty->print(" %2d - ", index);
1006 vf->print_value();
1007 int bci = chunk->at(index)->raw_bci();
1008 const char* code_name;
1009 if (bci == SynchronizationEntryBCI) {
1010 code_name = "sync entry";
1011 } else {
1012 Bytecodes::Code code = vf->method()->code_at(bci);
1013 code_name = Bytecodes::name(code);
1014 }
1015 tty->print(" - %s", code_name);
1016 tty->print_cr(" @ bci %d ", bci);
1017 if (Verbose) {
1018 vf->print();
1019 tty->cr();
1020 }
1021 }
1022 }
1023 #endif
1024
1025 // Register map for next frame (used for stack crawl). We capture
1026 // the state of the deopt'ing frame's caller. Thus if we need to
1027 // stuff a C2I adapter we can properly fill in the callee-save
1028 // register locations.
1029 frame caller = fr.sender(reg_map);
1030 int frame_size = caller.sp() - fr.sp();
1031
1032 frame sender = caller;
1033
1034 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1035 // the vframeArray containing the unpacking information is allocated in the C heap.
1036 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1037 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
1038
1039 // Compare the vframeArray to the collected vframes
1040 assert(array->structural_compare(thread, chunk), "just checking");
1041
1042 #ifndef PRODUCT
1043 if (TraceDeoptimization) {
1044 ttyLocker ttyl;
1045 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1046 }
1047 #endif // PRODUCT
1048
1049 return array;
1050 }
1051
1052
1053 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1054 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1055 for (int i = 0; i < monitors->length(); i++) {
1056 MonitorInfo* mon_info = monitors->at(i);
1057 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1058 objects_to_revoke->append(Handle(mon_info->owner()));
1059 }
1060 }
1061 }
1062
1063
1064 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1065 if (!UseBiasedLocking) {
1066 return;
1067 }
1068
1069 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1070
1071 // Unfortunately we don't have a RegisterMap available in most of
1072 // the places we want to call this routine so we need to walk the
1073 // stack again to update the register map.
1074 if (map == NULL || !map->update_map()) {
1075 StackFrameStream sfs(thread, true);
1076 bool found = false;
1077 while (!found && !sfs.is_done()) {
1078 frame* cur = sfs.current();
1079 sfs.next();
1080 found = cur->id() == fr.id();
1081 }
1082 assert(found, "frame to be deoptimized not found on target thread's stack");
1083 map = sfs.register_map();
1084 }
1085
1086 vframe* vf = vframe::new_vframe(&fr, map, thread);
1087 compiledVFrame* cvf = compiledVFrame::cast(vf);
1088 // Revoke monitors' biases in all scopes
1089 while (!cvf->is_top()) {
1090 collect_monitors(cvf, objects_to_revoke);
1091 cvf = compiledVFrame::cast(cvf->sender());
1092 }
1093 collect_monitors(cvf, objects_to_revoke);
1094
1095 if (SafepointSynchronize::is_at_safepoint()) {
1096 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1097 } else {
1098 BiasedLocking::revoke(objects_to_revoke);
1099 }
1100 }
1101
1102
1103 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1104 if (!UseBiasedLocking) {
1105 return;
1106 }
1107
1108 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1109 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1110 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1111 if (jt->has_last_Java_frame()) {
1112 StackFrameStream sfs(jt, true);
1113 while (!sfs.is_done()) {
1114 frame* cur = sfs.current();
1115 if (cb->contains(cur->pc())) {
1116 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1117 compiledVFrame* cvf = compiledVFrame::cast(vf);
1118 // Revoke monitors' biases in all scopes
1119 while (!cvf->is_top()) {
1120 collect_monitors(cvf, objects_to_revoke);
1121 cvf = compiledVFrame::cast(cvf->sender());
1122 }
1123 collect_monitors(cvf, objects_to_revoke);
1124 }
1125 sfs.next();
1126 }
1127 }
1128 }
1129 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1130 }
1131
1132
1133 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1134 assert(fr.can_be_deoptimized(), "checking frame type");
1135
1136 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1137
1138 // Patch the nmethod so that when execution returns to it we will
1139 // deopt the execution state and return to the interpreter.
1140 fr.deoptimize(thread);
1141 }
1142
1143 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1144 // Deoptimize only if the frame comes from compile code.
1145 // Do not deoptimize the frame which is already patched
1146 // during the execution of the loops below.
1147 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1148 return;
1149 }
1150 ResourceMark rm;
1151 DeoptimizationMarker dm;
1152 if (UseBiasedLocking) {
1153 revoke_biases_of_monitors(thread, fr, map);
1154 }
1155 deoptimize_single_frame(thread, fr);
1156
1157 }
1158
1159
1160 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1161 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1162 "can only deoptimize other thread at a safepoint");
1163 // Compute frame and register map based on thread and sp.
1164 RegisterMap reg_map(thread, UseBiasedLocking);
1165 frame fr = thread->last_frame();
1166 while (fr.id() != id) {
1167 fr = fr.sender(®_map);
1168 }
1169 deoptimize(thread, fr, ®_map);
1170 }
1171
1172
1173 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1174 if (thread == Thread::current()) {
1175 Deoptimization::deoptimize_frame_internal(thread, id);
1176 } else {
1177 VM_DeoptimizeFrame deopt(thread, id);
1178 VMThread::execute(&deopt);
1179 }
1180 }
1181
1182
1183 // JVMTI PopFrame support
1184 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1185 {
1186 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1187 }
1188 JRT_END
1189
1190 MethodData*
1191 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1192 bool create_if_missing) {
1193 Thread* THREAD = thread;
1194 MethodData* mdo = m()->method_data();
1195 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1196 // Build an MDO. Ignore errors like OutOfMemory;
1197 // that simply means we won't have an MDO to update.
1198 Method::build_interpreter_method_data(m, THREAD);
1199 if (HAS_PENDING_EXCEPTION) {
1200 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1201 CLEAR_PENDING_EXCEPTION;
1202 }
1203 mdo = m()->method_data();
1204 }
1205 return mdo;
1206 }
1207
1208 #if defined(COMPILER2) || defined(SHARK)
1209 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1210 // in case of an unresolved klass entry, load the class.
1211 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1212 Klass* tk = constant_pool->klass_at(index, CHECK);
1213 return;
1214 }
1215
1216 if (!constant_pool->tag_at(index).is_symbol()) return;
1217
1218 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1219 Symbol* symbol = constant_pool->symbol_at(index);
1220
1221 // class name?
1222 if (symbol->byte_at(0) != '(') {
1223 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1224 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1225 return;
1226 }
1227
1228 // then it must be a signature!
1229 ResourceMark rm(THREAD);
1230 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1231 if (ss.is_object()) {
1232 Symbol* class_name = ss.as_symbol(CHECK);
1233 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1234 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1235 }
1236 }
1237 }
1238
1239
1240 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1241 EXCEPTION_MARK;
1242 load_class_by_index(constant_pool, index, THREAD);
1243 if (HAS_PENDING_EXCEPTION) {
1244 // Exception happened during classloading. We ignore the exception here, since it
1245 // is going to be rethrown since the current activation is going to be deoptimized and
1246 // the interpreter will re-execute the bytecode.
1247 CLEAR_PENDING_EXCEPTION;
1248 // Class loading called java code which may have caused a stack
1249 // overflow. If the exception was thrown right before the return
1250 // to the runtime the stack is no longer guarded. Reguard the
1251 // stack otherwise if we return to the uncommon trap blob and the
1252 // stack bang causes a stack overflow we crash.
1253 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1254 JavaThread* thread = (JavaThread*)THREAD;
1255 bool guard_pages_enabled = thread->stack_yellow_zone_enabled();
1256 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1257 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1258 }
1259 }
1260
1261 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1262 HandleMark hm;
1263
1264 // uncommon_trap() is called at the beginning of the uncommon trap
1265 // handler. Note this fact before we start generating temporary frames
1266 // that can confuse an asynchronous stack walker. This counter is
1267 // decremented at the end of unpack_frames().
1268 thread->inc_in_deopt_handler();
1269
1270 // We need to update the map if we have biased locking.
1271 RegisterMap reg_map(thread, UseBiasedLocking);
1272 frame stub_frame = thread->last_frame();
1273 frame fr = stub_frame.sender(®_map);
1274 // Make sure the calling nmethod is not getting deoptimized and removed
1275 // before we are done with it.
1276 nmethodLocker nl(fr.pc());
1277
1278 // Log a message
1279 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1280 trap_request, fr.pc());
1281
1282 {
1283 ResourceMark rm;
1284
1285 // Revoke biases of any monitors in the frame to ensure we can migrate them
1286 revoke_biases_of_monitors(thread, fr, ®_map);
1287
1288 DeoptReason reason = trap_request_reason(trap_request);
1289 DeoptAction action = trap_request_action(trap_request);
1290 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1291
1292 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1293 compiledVFrame* cvf = compiledVFrame::cast(vf);
1294
1295 nmethod* nm = cvf->code();
1296
1297 ScopeDesc* trap_scope = cvf->scope();
1298 methodHandle trap_method = trap_scope->method();
1299 int trap_bci = trap_scope->bci();
1300 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1301
1302 // Record this event in the histogram.
1303 gather_statistics(reason, action, trap_bc);
1304
1305 // Ensure that we can record deopt. history:
1306 // Need MDO to record RTM code generation state.
1307 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1308
1309 MethodData* trap_mdo =
1310 get_method_data(thread, trap_method, create_if_missing);
1311
1312 // Log a message
1313 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1314 trap_reason_name(reason), trap_action_name(action), fr.pc(),
1315 trap_method->name_and_sig_as_C_string(), trap_bci);
1316
1317 // Print a bunch of diagnostics, if requested.
1318 if (TraceDeoptimization || LogCompilation) {
1319 ResourceMark rm;
1320 ttyLocker ttyl;
1321 char buf[100];
1322 if (xtty != NULL) {
1323 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1324 os::current_thread_id(),
1325 format_trap_request(buf, sizeof(buf), trap_request));
1326 nm->log_identity(xtty);
1327 }
1328 Symbol* class_name = NULL;
1329 bool unresolved = false;
1330 if (unloaded_class_index >= 0) {
1331 constantPoolHandle constants (THREAD, trap_method->constants());
1332 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1333 class_name = constants->klass_name_at(unloaded_class_index);
1334 unresolved = true;
1335 if (xtty != NULL)
1336 xtty->print(" unresolved='1'");
1337 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1338 class_name = constants->symbol_at(unloaded_class_index);
1339 }
1340 if (xtty != NULL)
1341 xtty->name(class_name);
1342 }
1343 if (xtty != NULL && trap_mdo != NULL && reason < MethodData::_trap_hist_limit) {
1344 // Dump the relevant MDO state.
1345 // This is the deopt count for the current reason, any previous
1346 // reasons or recompiles seen at this point.
1347 int dcnt = trap_mdo->trap_count(reason);
1348 if (dcnt != 0)
1349 xtty->print(" count='%d'", dcnt);
1350 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1351 int dos = (pdata == NULL)? 0: pdata->trap_state();
1352 if (dos != 0) {
1353 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1354 if (trap_state_is_recompiled(dos)) {
1355 int recnt2 = trap_mdo->overflow_recompile_count();
1356 if (recnt2 != 0)
1357 xtty->print(" recompiles2='%d'", recnt2);
1358 }
1359 }
1360 }
1361 if (xtty != NULL) {
1362 xtty->stamp();
1363 xtty->end_head();
1364 }
1365 if (TraceDeoptimization) { // make noise on the tty
1366 tty->print("Uncommon trap occurred in");
1367 nm->method()->print_short_name(tty);
1368 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1369 fr.pc(),
1370 os::current_thread_id(),
1371 trap_reason_name(reason),
1372 trap_action_name(action),
1373 unloaded_class_index);
1374 if (class_name != NULL) {
1375 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1376 class_name->print_symbol_on(tty);
1377 }
1378 tty->cr();
1379 }
1380 if (xtty != NULL) {
1381 // Log the precise location of the trap.
1382 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1383 xtty->begin_elem("jvms bci='%d'", sd->bci());
1384 xtty->method(sd->method());
1385 xtty->end_elem();
1386 if (sd->is_top()) break;
1387 }
1388 xtty->tail("uncommon_trap");
1389 }
1390 }
1391 // (End diagnostic printout.)
1392
1393 // Load class if necessary
1394 if (unloaded_class_index >= 0) {
1395 constantPoolHandle constants(THREAD, trap_method->constants());
1396 load_class_by_index(constants, unloaded_class_index);
1397 }
1398
1399 // Flush the nmethod if necessary and desirable.
1400 //
1401 // We need to avoid situations where we are re-flushing the nmethod
1402 // because of a hot deoptimization site. Repeated flushes at the same
1403 // point need to be detected by the compiler and avoided. If the compiler
1404 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1405 // module must take measures to avoid an infinite cycle of recompilation
1406 // and deoptimization. There are several such measures:
1407 //
1408 // 1. If a recompilation is ordered a second time at some site X
1409 // and for the same reason R, the action is adjusted to 'reinterpret',
1410 // to give the interpreter time to exercise the method more thoroughly.
1411 // If this happens, the method's overflow_recompile_count is incremented.
1412 //
1413 // 2. If the compiler fails to reduce the deoptimization rate, then
1414 // the method's overflow_recompile_count will begin to exceed the set
1415 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1416 // is adjusted to 'make_not_compilable', and the method is abandoned
1417 // to the interpreter. This is a performance hit for hot methods,
1418 // but is better than a disastrous infinite cycle of recompilations.
1419 // (Actually, only the method containing the site X is abandoned.)
1420 //
1421 // 3. In parallel with the previous measures, if the total number of
1422 // recompilations of a method exceeds the much larger set limit
1423 // PerMethodRecompilationCutoff, the method is abandoned.
1424 // This should only happen if the method is very large and has
1425 // many "lukewarm" deoptimizations. The code which enforces this
1426 // limit is elsewhere (class nmethod, class Method).
1427 //
1428 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1429 // to recompile at each bytecode independently of the per-BCI cutoff.
1430 //
1431 // The decision to update code is up to the compiler, and is encoded
1432 // in the Action_xxx code. If the compiler requests Action_none
1433 // no trap state is changed, no compiled code is changed, and the
1434 // computation suffers along in the interpreter.
1435 //
1436 // The other action codes specify various tactics for decompilation
1437 // and recompilation. Action_maybe_recompile is the loosest, and
1438 // allows the compiled code to stay around until enough traps are seen,
1439 // and until the compiler gets around to recompiling the trapping method.
1440 //
1441 // The other actions cause immediate removal of the present code.
1442
1443 bool update_trap_state = (reason != Reason_tenured);
1444 bool make_not_entrant = false;
1445 bool make_not_compilable = false;
1446 bool reprofile = false;
1447 switch (action) {
1448 case Action_none:
1449 // Keep the old code.
1450 update_trap_state = false;
1451 break;
1452 case Action_maybe_recompile:
1453 // Do not need to invalidate the present code, but we can
1454 // initiate another
1455 // Start compiler without (necessarily) invalidating the nmethod.
1456 // The system will tolerate the old code, but new code should be
1457 // generated when possible.
1458 break;
1459 case Action_reinterpret:
1460 // Go back into the interpreter for a while, and then consider
1461 // recompiling form scratch.
1462 make_not_entrant = true;
1463 // Reset invocation counter for outer most method.
1464 // This will allow the interpreter to exercise the bytecodes
1465 // for a while before recompiling.
1466 // By contrast, Action_make_not_entrant is immediate.
1467 //
1468 // Note that the compiler will track null_check, null_assert,
1469 // range_check, and class_check events and log them as if they
1470 // had been traps taken from compiled code. This will update
1471 // the MDO trap history so that the next compilation will
1472 // properly detect hot trap sites.
1473 reprofile = true;
1474 break;
1475 case Action_make_not_entrant:
1476 // Request immediate recompilation, and get rid of the old code.
1477 // Make them not entrant, so next time they are called they get
1478 // recompiled. Unloaded classes are loaded now so recompile before next
1479 // time they are called. Same for uninitialized. The interpreter will
1480 // link the missing class, if any.
1481 make_not_entrant = true;
1482 break;
1483 case Action_make_not_compilable:
1484 // Give up on compiling this method at all.
1485 make_not_entrant = true;
1486 make_not_compilable = true;
1487 break;
1488 default:
1489 ShouldNotReachHere();
1490 }
1491
1492 // Setting +ProfileTraps fixes the following, on all platforms:
1493 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1494 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1495 // recompile relies on a MethodData* to record heroic opt failures.
1496
1497 // Whether the interpreter is producing MDO data or not, we also need
1498 // to use the MDO to detect hot deoptimization points and control
1499 // aggressive optimization.
1500 bool inc_recompile_count = false;
1501 ProfileData* pdata = NULL;
1502 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1503 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1504 uint this_trap_count = 0;
1505 bool maybe_prior_trap = false;
1506 bool maybe_prior_recompile = false;
1507 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1508 nm->method(),
1509 //outputs:
1510 this_trap_count,
1511 maybe_prior_trap,
1512 maybe_prior_recompile);
1513 // Because the interpreter also counts null, div0, range, and class
1514 // checks, these traps from compiled code are double-counted.
1515 // This is harmless; it just means that the PerXTrapLimit values
1516 // are in effect a little smaller than they look.
1517
1518 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1519 if (per_bc_reason != Reason_none) {
1520 // Now take action based on the partially known per-BCI history.
1521 if (maybe_prior_trap
1522 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1523 // If there are too many traps at this BCI, force a recompile.
1524 // This will allow the compiler to see the limit overflow, and
1525 // take corrective action, if possible. The compiler generally
1526 // does not use the exact PerBytecodeTrapLimit value, but instead
1527 // changes its tactics if it sees any traps at all. This provides
1528 // a little hysteresis, delaying a recompile until a trap happens
1529 // several times.
1530 //
1531 // Actually, since there is only one bit of counter per BCI,
1532 // the possible per-BCI counts are {0,1,(per-method count)}.
1533 // This produces accurate results if in fact there is only
1534 // one hot trap site, but begins to get fuzzy if there are
1535 // many sites. For example, if there are ten sites each
1536 // trapping two or more times, they each get the blame for
1537 // all of their traps.
1538 make_not_entrant = true;
1539 }
1540
1541 // Detect repeated recompilation at the same BCI, and enforce a limit.
1542 if (make_not_entrant && maybe_prior_recompile) {
1543 // More than one recompile at this point.
1544 inc_recompile_count = maybe_prior_trap;
1545 }
1546 } else {
1547 // For reasons which are not recorded per-bytecode, we simply
1548 // force recompiles unconditionally.
1549 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1550 make_not_entrant = true;
1551 }
1552
1553 // Go back to the compiler if there are too many traps in this method.
1554 if (this_trap_count >= per_method_trap_limit(reason)) {
1555 // If there are too many traps in this method, force a recompile.
1556 // This will allow the compiler to see the limit overflow, and
1557 // take corrective action, if possible.
1558 // (This condition is an unlikely backstop only, because the
1559 // PerBytecodeTrapLimit is more likely to take effect first,
1560 // if it is applicable.)
1561 make_not_entrant = true;
1562 }
1563
1564 // Here's more hysteresis: If there has been a recompile at
1565 // this trap point already, run the method in the interpreter
1566 // for a while to exercise it more thoroughly.
1567 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1568 reprofile = true;
1569 }
1570 }
1571
1572 // Take requested actions on the method:
1573
1574 // Recompile
1575 if (make_not_entrant) {
1576 if (!nm->make_not_entrant()) {
1577 return; // the call did not change nmethod's state
1578 }
1579
1580 if (pdata != NULL) {
1581 // Record the recompilation event, if any.
1582 int tstate0 = pdata->trap_state();
1583 int tstate1 = trap_state_set_recompiled(tstate0, true);
1584 if (tstate1 != tstate0)
1585 pdata->set_trap_state(tstate1);
1586 }
1587
1588 #if INCLUDE_RTM_OPT
1589 // Restart collecting RTM locking abort statistic if the method
1590 // is recompiled for a reason other than RTM state change.
1591 // Assume that in new recompiled code the statistic could be different,
1592 // for example, due to different inlining.
1593 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1594 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) {
1595 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1596 }
1597 #endif
1598 // For code aging we count traps separately here, using make_not_entrant()
1599 // as a guard against simultaneous deopts in multiple threads.
1600 if (reason == Reason_tenured && trap_mdo != NULL) {
1601 trap_mdo->inc_tenure_traps();
1602 }
1603 }
1604
1605 if (inc_recompile_count) {
1606 trap_mdo->inc_overflow_recompile_count();
1607 if ((uint)trap_mdo->overflow_recompile_count() >
1608 (uint)PerBytecodeRecompilationCutoff) {
1609 // Give up on the method containing the bad BCI.
1610 if (trap_method() == nm->method()) {
1611 make_not_compilable = true;
1612 } else {
1613 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1614 // But give grace to the enclosing nm->method().
1615 }
1616 }
1617 }
1618
1619 // Reprofile
1620 if (reprofile) {
1621 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1622 }
1623
1624 // Give up compiling
1625 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1626 assert(make_not_entrant, "consistent");
1627 nm->method()->set_not_compilable(CompLevel_full_optimization);
1628 }
1629
1630 } // Free marked resources
1631
1632 }
1633 JRT_END
1634
1635 ProfileData*
1636 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1637 int trap_bci,
1638 Deoptimization::DeoptReason reason,
1639 Method* compiled_method,
1640 //outputs:
1641 uint& ret_this_trap_count,
1642 bool& ret_maybe_prior_trap,
1643 bool& ret_maybe_prior_recompile) {
1644 uint prior_trap_count = trap_mdo->trap_count(reason);
1645 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1646
1647 // If the runtime cannot find a place to store trap history,
1648 // it is estimated based on the general condition of the method.
1649 // If the method has ever been recompiled, or has ever incurred
1650 // a trap with the present reason , then this BCI is assumed
1651 // (pessimistically) to be the culprit.
1652 bool maybe_prior_trap = (prior_trap_count != 0);
1653 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1654 ProfileData* pdata = NULL;
1655
1656
1657 // For reasons which are recorded per bytecode, we check per-BCI data.
1658 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1659 if (per_bc_reason != Reason_none) {
1660 // Find the profile data for this BCI. If there isn't one,
1661 // try to allocate one from the MDO's set of spares.
1662 // This will let us detect a repeated trap at this point.
1663 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1664
1665 if (pdata != NULL) {
1666 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
1667 if (LogCompilation && xtty != NULL) {
1668 ttyLocker ttyl;
1669 // no more room for speculative traps in this MDO
1670 xtty->elem("speculative_traps_oom");
1671 }
1672 }
1673 // Query the trap state of this profile datum.
1674 int tstate0 = pdata->trap_state();
1675 if (!trap_state_has_reason(tstate0, per_bc_reason))
1676 maybe_prior_trap = false;
1677 if (!trap_state_is_recompiled(tstate0))
1678 maybe_prior_recompile = false;
1679
1680 // Update the trap state of this profile datum.
1681 int tstate1 = tstate0;
1682 // Record the reason.
1683 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1684 // Store the updated state on the MDO, for next time.
1685 if (tstate1 != tstate0)
1686 pdata->set_trap_state(tstate1);
1687 } else {
1688 if (LogCompilation && xtty != NULL) {
1689 ttyLocker ttyl;
1690 // Missing MDP? Leave a small complaint in the log.
1691 xtty->elem("missing_mdp bci='%d'", trap_bci);
1692 }
1693 }
1694 }
1695
1696 // Return results:
1697 ret_this_trap_count = this_trap_count;
1698 ret_maybe_prior_trap = maybe_prior_trap;
1699 ret_maybe_prior_recompile = maybe_prior_recompile;
1700 return pdata;
1701 }
1702
1703 void
1704 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1705 ResourceMark rm;
1706 // Ignored outputs:
1707 uint ignore_this_trap_count;
1708 bool ignore_maybe_prior_trap;
1709 bool ignore_maybe_prior_recompile;
1710 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
1711 query_update_method_data(trap_mdo, trap_bci,
1712 (DeoptReason)reason,
1713 NULL,
1714 ignore_this_trap_count,
1715 ignore_maybe_prior_trap,
1716 ignore_maybe_prior_recompile);
1717 }
1718
1719 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1720
1721 // Still in Java no safepoints
1722 {
1723 // This enters VM and may safepoint
1724 uncommon_trap_inner(thread, trap_request);
1725 }
1726 return fetch_unroll_info_helper(thread);
1727 }
1728
1729 // Local derived constants.
1730 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1731 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1732 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1733
1734 //---------------------------trap_state_reason---------------------------------
1735 Deoptimization::DeoptReason
1736 Deoptimization::trap_state_reason(int trap_state) {
1737 // This assert provides the link between the width of DataLayout::trap_bits
1738 // and the encoding of "recorded" reasons. It ensures there are enough
1739 // bits to store all needed reasons in the per-BCI MDO profile.
1740 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1741 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1742 trap_state -= recompile_bit;
1743 if (trap_state == DS_REASON_MASK) {
1744 return Reason_many;
1745 } else {
1746 assert((int)Reason_none == 0, "state=0 => Reason_none");
1747 return (DeoptReason)trap_state;
1748 }
1749 }
1750 //-------------------------trap_state_has_reason-------------------------------
1751 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1752 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1753 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1754 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1755 trap_state -= recompile_bit;
1756 if (trap_state == DS_REASON_MASK) {
1757 return -1; // true, unspecifically (bottom of state lattice)
1758 } else if (trap_state == reason) {
1759 return 1; // true, definitely
1760 } else if (trap_state == 0) {
1761 return 0; // false, definitely (top of state lattice)
1762 } else {
1763 return 0; // false, definitely
1764 }
1765 }
1766 //-------------------------trap_state_add_reason-------------------------------
1767 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1768 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1769 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1770 trap_state -= recompile_bit;
1771 if (trap_state == DS_REASON_MASK) {
1772 return trap_state + recompile_bit; // already at state lattice bottom
1773 } else if (trap_state == reason) {
1774 return trap_state + recompile_bit; // the condition is already true
1775 } else if (trap_state == 0) {
1776 return reason + recompile_bit; // no condition has yet been true
1777 } else {
1778 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1779 }
1780 }
1781 //-----------------------trap_state_is_recompiled------------------------------
1782 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1783 return (trap_state & DS_RECOMPILE_BIT) != 0;
1784 }
1785 //-----------------------trap_state_set_recompiled-----------------------------
1786 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1787 if (z) return trap_state | DS_RECOMPILE_BIT;
1788 else return trap_state & ~DS_RECOMPILE_BIT;
1789 }
1790 //---------------------------format_trap_state---------------------------------
1791 // This is used for debugging and diagnostics, including LogFile output.
1792 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1793 int trap_state) {
1794 DeoptReason reason = trap_state_reason(trap_state);
1795 bool recomp_flag = trap_state_is_recompiled(trap_state);
1796 // Re-encode the state from its decoded components.
1797 int decoded_state = 0;
1798 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1799 decoded_state = trap_state_add_reason(decoded_state, reason);
1800 if (recomp_flag)
1801 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1802 // If the state re-encodes properly, format it symbolically.
1803 // Because this routine is used for debugging and diagnostics,
1804 // be robust even if the state is a strange value.
1805 size_t len;
1806 if (decoded_state != trap_state) {
1807 // Random buggy state that doesn't decode??
1808 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1809 } else {
1810 len = jio_snprintf(buf, buflen, "%s%s",
1811 trap_reason_name(reason),
1812 recomp_flag ? " recompiled" : "");
1813 }
1814 if (len >= buflen)
1815 buf[buflen-1] = '\0';
1816 return buf;
1817 }
1818
1819
1820 //--------------------------------statics--------------------------------------
1821 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1822 // Note: Keep this in sync. with enum DeoptReason.
1823 "none",
1824 "null_check",
1825 "null_assert",
1826 "range_check",
1827 "class_check",
1828 "array_check",
1829 "intrinsic",
1830 "bimorphic",
1831 "unloaded",
1832 "uninitialized",
1833 "unreached",
1834 "unhandled",
1835 "constraint",
1836 "div0_check",
1837 "age",
1838 "predicate",
1839 "loop_limit_check",
1840 "speculate_class_check",
1841 "speculate_null_check",
1842 "rtm_state_change",
1843 "tenured"
1844 };
1845 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1846 // Note: Keep this in sync. with enum DeoptAction.
1847 "none",
1848 "maybe_recompile",
1849 "reinterpret",
1850 "make_not_entrant",
1851 "make_not_compilable"
1852 };
1853
1854 const char* Deoptimization::trap_reason_name(int reason) {
1855 if (reason == Reason_many) return "many";
1856 if ((uint)reason < Reason_LIMIT) {
1857 const char* name = _trap_reason_name[reason];
1858 assert(name != NULL, "trap reason name is undefined");
1859 return name;
1860 }
1861 static char buf[20];
1862 sprintf(buf, "reason%d", reason);
1863 return buf;
1864 }
1865 const char* Deoptimization::trap_action_name(int action) {
1866 if ((uint)action < Action_LIMIT) {
1867 const char* name = _trap_action_name[action];
1868 assert(name != NULL, "trap action name is undefined");
1869 return name;
1870 }
1871 static char buf[20];
1872 sprintf(buf, "action%d", action);
1873 return buf;
1874 }
1875
1876 // This is used for debugging and diagnostics, including LogFile output.
1877 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1878 int trap_request) {
1879 jint unloaded_class_index = trap_request_index(trap_request);
1880 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1881 const char* action = trap_action_name(trap_request_action(trap_request));
1882 size_t len;
1883 if (unloaded_class_index < 0) {
1884 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1885 reason, action);
1886 } else {
1887 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1888 reason, action, unloaded_class_index);
1889 }
1890 if (len >= buflen)
1891 buf[buflen-1] = '\0';
1892 return buf;
1893 }
1894
1895 juint Deoptimization::_deoptimization_hist
1896 [Deoptimization::Reason_LIMIT]
1897 [1 + Deoptimization::Action_LIMIT]
1898 [Deoptimization::BC_CASE_LIMIT]
1899 = {0};
1900
1901 enum {
1902 LSB_BITS = 8,
1903 LSB_MASK = right_n_bits(LSB_BITS)
1904 };
1905
1906 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1907 Bytecodes::Code bc) {
1908 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1909 assert(action >= 0 && action < Action_LIMIT, "oob");
1910 _deoptimization_hist[Reason_none][0][0] += 1; // total
1911 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1912 juint* cases = _deoptimization_hist[reason][1+action];
1913 juint* bc_counter_addr = NULL;
1914 juint bc_counter = 0;
1915 // Look for an unused counter, or an exact match to this BC.
1916 if (bc != Bytecodes::_illegal) {
1917 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1918 juint* counter_addr = &cases[bc_case];
1919 juint counter = *counter_addr;
1920 if ((counter == 0 && bc_counter_addr == NULL)
1921 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1922 // this counter is either free or is already devoted to this BC
1923 bc_counter_addr = counter_addr;
1924 bc_counter = counter | bc;
1925 }
1926 }
1927 }
1928 if (bc_counter_addr == NULL) {
1929 // Overflow, or no given bytecode.
1930 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1931 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1932 }
1933 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1934 }
1935
1936 jint Deoptimization::total_deoptimization_count() {
1937 return _deoptimization_hist[Reason_none][0][0];
1938 }
1939
1940 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1941 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1942 return _deoptimization_hist[reason][0][0];
1943 }
1944
1945 void Deoptimization::print_statistics() {
1946 juint total = total_deoptimization_count();
1947 juint account = total;
1948 if (total != 0) {
1949 ttyLocker ttyl;
1950 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1951 tty->print_cr("Deoptimization traps recorded:");
1952 #define PRINT_STAT_LINE(name, r) \
1953 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1954 PRINT_STAT_LINE("total", total);
1955 // For each non-zero entry in the histogram, print the reason,
1956 // the action, and (if specifically known) the type of bytecode.
1957 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1958 for (int action = 0; action < Action_LIMIT; action++) {
1959 juint* cases = _deoptimization_hist[reason][1+action];
1960 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1961 juint counter = cases[bc_case];
1962 if (counter != 0) {
1963 char name[1*K];
1964 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1965 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1966 bc = Bytecodes::_illegal;
1967 sprintf(name, "%s/%s/%s",
1968 trap_reason_name(reason),
1969 trap_action_name(action),
1970 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1971 juint r = counter >> LSB_BITS;
1972 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1973 account -= r;
1974 }
1975 }
1976 }
1977 }
1978 if (account != 0) {
1979 PRINT_STAT_LINE("unaccounted", account);
1980 }
1981 #undef PRINT_STAT_LINE
1982 if (xtty != NULL) xtty->tail("statistics");
1983 }
1984 }
1985 #else // COMPILER2 || SHARK
1986
1987
1988 // Stubs for C1 only system.
1989 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1990 return false;
1991 }
1992
1993 const char* Deoptimization::trap_reason_name(int reason) {
1994 return "unknown";
1995 }
1996
1997 void Deoptimization::print_statistics() {
1998 // no output
1999 }
2000
2001 void
2002 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2003 // no udpate
2004 }
2005
2006 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2007 return 0;
2008 }
2009
2010 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2011 Bytecodes::Code bc) {
2012 // no update
2013 }
2014
2015 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2016 int trap_state) {
2017 jio_snprintf(buf, buflen, "#%d", trap_state);
2018 return buf;
2019 }
2020
2021 #endif // COMPILER2 || SHARK