Wed, 03 Jul 2019 20:42:37 +0800
Merge
1 /*
2 * Copyright (c) 1997, 2016, 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 */
25 /*
26 * This file has been modified by Loongson Technology in 2015. These
27 * modifications are Copyright (c) 2015 Loongson Technology, and are made
28 * available on the same license terms set forth above.
29 */
31 #include "precompiled.hpp"
32 #include "classfile/systemDictionary.hpp"
33 #include "code/debugInfoRec.hpp"
34 #include "code/nmethod.hpp"
35 #include "code/pcDesc.hpp"
36 #include "code/scopeDesc.hpp"
37 #include "interpreter/bytecode.hpp"
38 #include "interpreter/interpreter.hpp"
39 #include "interpreter/oopMapCache.hpp"
40 #include "memory/allocation.inline.hpp"
41 #include "memory/oopFactory.hpp"
42 #include "memory/resourceArea.hpp"
43 #include "oops/method.hpp"
44 #include "oops/oop.inline.hpp"
45 #include "prims/jvmtiThreadState.hpp"
46 #include "runtime/biasedLocking.hpp"
47 #include "runtime/compilationPolicy.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/interfaceSupport.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "runtime/signature.hpp"
52 #include "runtime/stubRoutines.hpp"
53 #include "runtime/thread.hpp"
54 #include "runtime/vframe.hpp"
55 #include "runtime/vframeArray.hpp"
56 #include "runtime/vframe_hp.hpp"
57 #include "utilities/events.hpp"
58 #include "utilities/xmlstream.hpp"
59 #ifdef TARGET_ARCH_x86
60 # include "vmreg_x86.inline.hpp"
61 #endif
62 #ifdef TARGET_ARCH_sparc
63 # include "vmreg_sparc.inline.hpp"
64 #endif
65 #ifdef TARGET_ARCH_zero
66 # include "vmreg_zero.inline.hpp"
67 #endif
68 #ifdef TARGET_ARCH_arm
69 # include "vmreg_arm.inline.hpp"
70 #endif
71 #ifdef TARGET_ARCH_ppc
72 # include "vmreg_ppc.inline.hpp"
73 #endif
74 #ifdef TARGET_ARCH_mips
75 # include "vmreg_mips.inline.hpp"
76 #endif
77 #ifdef COMPILER2
78 #if defined AD_MD_HPP
79 # include AD_MD_HPP
80 #elif defined TARGET_ARCH_MODEL_x86_32
81 # include "adfiles/ad_x86_32.hpp"
82 #elif defined TARGET_ARCH_MODEL_x86_64
83 # include "adfiles/ad_x86_64.hpp"
84 #elif defined TARGET_ARCH_MODEL_sparc
85 # include "adfiles/ad_sparc.hpp"
86 #elif defined TARGET_ARCH_MODEL_zero
87 # include "adfiles/ad_zero.hpp"
88 #elif defined TARGET_ARCH_MODEL_ppc_64
89 # include "adfiles/ad_ppc_64.hpp"
90 #endif
91 #ifdef TARGET_ARCH_MODEL_mips_64
92 # include "adfiles/ad_mips_64.hpp"
93 #endif
94 #endif // COMPILER2
96 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
98 bool DeoptimizationMarker::_is_active = false;
100 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
101 int caller_adjustment,
102 int caller_actual_parameters,
103 int number_of_frames,
104 intptr_t* frame_sizes,
105 address* frame_pcs,
106 BasicType return_type) {
107 _size_of_deoptimized_frame = size_of_deoptimized_frame;
108 _caller_adjustment = caller_adjustment;
109 _caller_actual_parameters = caller_actual_parameters;
110 _number_of_frames = number_of_frames;
111 _frame_sizes = frame_sizes;
112 _frame_pcs = frame_pcs;
113 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
114 _return_type = return_type;
115 _initial_info = 0;
116 // PD (x86 only)
117 _counter_temp = 0;
118 _unpack_kind = 0;
119 _sender_sp_temp = 0;
121 _total_frame_sizes = size_of_frames();
122 }
125 Deoptimization::UnrollBlock::~UnrollBlock() {
126 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
127 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
128 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
129 }
132 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
133 assert(register_number < RegisterMap::reg_count, "checking register number");
134 return &_register_block[register_number * 2];
135 }
139 int Deoptimization::UnrollBlock::size_of_frames() const {
140 // Acount first for the adjustment of the initial frame
141 int result = _caller_adjustment;
142 for (int index = 0; index < number_of_frames(); index++) {
143 result += frame_sizes()[index];
144 }
145 return result;
146 }
149 void Deoptimization::UnrollBlock::print() {
150 ttyLocker ttyl;
151 tty->print_cr("UnrollBlock");
152 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
153 tty->print( " frame_sizes: ");
154 for (int index = 0; index < number_of_frames(); index++) {
155 tty->print("%d ", frame_sizes()[index]);
156 }
157 tty->cr();
158 }
161 // In order to make fetch_unroll_info work properly with escape
162 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
163 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
164 // of previously eliminated objects occurs in realloc_objects, which is
165 // called from the method fetch_unroll_info_helper below.
166 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
167 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
168 // but makes the entry a little slower. There is however a little dance we have to
169 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
171 // fetch_unroll_info() is called at the beginning of the deoptimization
172 // handler. Note this fact before we start generating temporary frames
173 // that can confuse an asynchronous stack walker. This counter is
174 // decremented at the end of unpack_frames().
175 thread->inc_in_deopt_handler();
177 return fetch_unroll_info_helper(thread);
178 JRT_END
181 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
182 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
184 // Note: there is a safepoint safety issue here. No matter whether we enter
185 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
186 // the vframeArray is created.
187 //
189 // Allocate our special deoptimization ResourceMark
190 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
191 assert(thread->deopt_mark() == NULL, "Pending deopt!");
192 thread->set_deopt_mark(dmark);
194 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
195 RegisterMap map(thread, true);
196 RegisterMap dummy_map(thread, false);
197 // Now get the deoptee with a valid map
198 frame deoptee = stub_frame.sender(&map);
199 // Set the deoptee nmethod
200 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
201 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
203 if (VerifyStack) {
204 thread->validate_frame_layout();
205 }
207 // Create a growable array of VFrames where each VFrame represents an inlined
208 // Java frame. This storage is allocated with the usual system arena.
209 assert(deoptee.is_compiled_frame(), "Wrong frame type");
210 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
211 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
212 while (!vf->is_top()) {
213 assert(vf->is_compiled_frame(), "Wrong frame type");
214 chunk->push(compiledVFrame::cast(vf));
215 vf = vf->sender();
216 }
217 assert(vf->is_compiled_frame(), "Wrong frame type");
218 chunk->push(compiledVFrame::cast(vf));
220 bool realloc_failures = false;
222 #ifdef COMPILER2
223 // Reallocate the non-escaping objects and restore their fields. Then
224 // relock objects if synchronization on them was eliminated.
225 if (DoEscapeAnalysis || EliminateNestedLocks) {
226 if (EliminateAllocations) {
227 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
228 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
230 // The flag return_oop() indicates call sites which return oop
231 // in compiled code. Such sites include java method calls,
232 // runtime calls (for example, used to allocate new objects/arrays
233 // on slow code path) and any other calls generated in compiled code.
234 // It is not guaranteed that we can get such information here only
235 // by analyzing bytecode in deoptimized frames. This is why this flag
236 // is set during method compilation (see Compile::Process_OopMap_Node()).
237 // If the previous frame was popped, we don't have a result.
238 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution();
239 Handle return_value;
240 if (save_oop_result) {
241 // Reallocation may trigger GC. If deoptimization happened on return from
242 // call which returns oop we need to save it since it is not in oopmap.
243 oop result = deoptee.saved_oop_result(&map);
244 assert(result == NULL || result->is_oop(), "must be oop");
245 return_value = Handle(thread, result);
246 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
247 if (TraceDeoptimization) {
248 ttyLocker ttyl;
249 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread);
250 }
251 }
252 if (objects != NULL) {
253 JRT_BLOCK
254 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
255 JRT_END
256 reassign_fields(&deoptee, &map, objects, realloc_failures);
257 #ifndef PRODUCT
258 if (TraceDeoptimization) {
259 ttyLocker ttyl;
260 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
261 print_objects(objects, realloc_failures);
262 }
263 #endif
264 }
265 if (save_oop_result) {
266 // Restore result.
267 deoptee.set_saved_oop_result(&map, return_value());
268 }
269 }
270 if (EliminateLocks) {
271 #ifndef PRODUCT
272 bool first = true;
273 #endif
274 for (int i = 0; i < chunk->length(); i++) {
275 compiledVFrame* cvf = chunk->at(i);
276 assert (cvf->scope() != NULL,"expect only compiled java frames");
277 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
278 if (monitors->is_nonempty()) {
279 relock_objects(monitors, thread, realloc_failures);
280 #ifndef PRODUCT
281 if (TraceDeoptimization) {
282 ttyLocker ttyl;
283 for (int j = 0; j < monitors->length(); j++) {
284 MonitorInfo* mi = monitors->at(j);
285 if (mi->eliminated()) {
286 if (first) {
287 first = false;
288 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
289 }
290 if (mi->owner_is_scalar_replaced()) {
291 Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
292 tty->print_cr(" failed reallocation for klass %s", k->external_name());
293 } else {
294 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner());
295 }
296 }
297 }
298 }
299 #endif
300 }
301 }
302 }
303 }
304 #endif // COMPILER2
305 // Ensure that no safepoint is taken after pointers have been stored
306 // in fields of rematerialized objects. If a safepoint occurs from here on
307 // out the java state residing in the vframeArray will be missed.
308 No_Safepoint_Verifier no_safepoint;
310 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
311 #ifdef COMPILER2
312 if (realloc_failures) {
313 pop_frames_failed_reallocs(thread, array);
314 }
315 #endif
317 assert(thread->vframe_array_head() == NULL, "Pending deopt!");
318 thread->set_vframe_array_head(array);
320 // Now that the vframeArray has been created if we have any deferred local writes
321 // added by jvmti then we can free up that structure as the data is now in the
322 // vframeArray
324 if (thread->deferred_locals() != NULL) {
325 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
326 int i = 0;
327 do {
328 // Because of inlining we could have multiple vframes for a single frame
329 // and several of the vframes could have deferred writes. Find them all.
330 if (list->at(i)->id() == array->original().id()) {
331 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
332 list->remove_at(i);
333 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
334 delete dlv;
335 } else {
336 i++;
337 }
338 } while ( i < list->length() );
339 if (list->length() == 0) {
340 thread->set_deferred_locals(NULL);
341 // free the list and elements back to C heap.
342 delete list;
343 }
345 }
347 #ifndef SHARK
348 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
349 CodeBlob* cb = stub_frame.cb();
350 // Verify we have the right vframeArray
351 assert(cb->frame_size() >= 0, "Unexpected frame size");
352 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
354 // If the deopt call site is a MethodHandle invoke call site we have
355 // to adjust the unpack_sp.
356 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
357 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
358 unpack_sp = deoptee.unextended_sp();
360 #ifdef ASSERT
361 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
362 #endif
363 #else
364 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
365 #endif // !SHARK
367 // This is a guarantee instead of an assert because if vframe doesn't match
368 // we will unpack the wrong deoptimized frame and wind up in strange places
369 // where it will be very difficult to figure out what went wrong. Better
370 // to die an early death here than some very obscure death later when the
371 // trail is cold.
372 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
373 // in that it will fail to detect a problem when there is one. This needs
374 // more work in tiger timeframe.
375 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
377 int number_of_frames = array->frames();
379 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
380 // virtual activation, which is the reverse of the elements in the vframes array.
381 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
382 // +1 because we always have an interpreter return address for the final slot.
383 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
384 int popframe_extra_args = 0;
385 // Create an interpreter return address for the stub to use as its return
386 // address so the skeletal frames are perfectly walkable
387 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
389 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
390 // activation be put back on the expression stack of the caller for reexecution
391 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
392 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
393 }
395 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
396 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
397 // than simply use array->sender.pc(). This requires us to walk the current set of frames
398 //
399 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
400 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
402 // It's possible that the number of paramters at the call site is
403 // different than number of arguments in the callee when method
404 // handles are used. If the caller is interpreted get the real
405 // value so that the proper amount of space can be added to it's
406 // frame.
407 bool caller_was_method_handle = false;
408 if (deopt_sender.is_interpreted_frame()) {
409 methodHandle method = deopt_sender.interpreter_frame_method();
410 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
411 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
412 // Method handle invokes may involve fairly arbitrary chains of
413 // calls so it's impossible to know how much actual space the
414 // caller has for locals.
415 caller_was_method_handle = true;
416 }
417 }
419 //
420 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
421 // frame_sizes/frame_pcs[1] next oldest frame (int)
422 // frame_sizes/frame_pcs[n] youngest frame (int)
423 //
424 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
425 // owns the space for the return address to it's caller). Confusing ain't it.
426 //
427 // The vframe array can address vframes with indices running from
428 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
429 // When we create the skeletal frames we need the oldest frame to be in the zero slot
430 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
431 // so things look a little strange in this loop.
432 //
433 int callee_parameters = 0;
434 int callee_locals = 0;
435 for (int index = 0; index < array->frames(); index++ ) {
436 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
437 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
438 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
439 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
440 callee_locals,
441 index == 0,
442 popframe_extra_args);
443 // This pc doesn't have to be perfect just good enough to identify the frame
444 // as interpreted so the skeleton frame will be walkable
445 // The correct pc will be set when the skeleton frame is completely filled out
446 // The final pc we store in the loop is wrong and will be overwritten below
447 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
449 callee_parameters = array->element(index)->method()->size_of_parameters();
450 callee_locals = array->element(index)->method()->max_locals();
451 popframe_extra_args = 0;
452 }
454 // Compute whether the root vframe returns a float or double value.
455 BasicType return_type;
456 {
457 HandleMark hm;
458 methodHandle method(thread, array->element(0)->method());
459 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
460 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
461 }
463 // Compute information for handling adapters and adjusting the frame size of the caller.
464 int caller_adjustment = 0;
466 // Compute the amount the oldest interpreter frame will have to adjust
467 // its caller's stack by. If the caller is a compiled frame then
468 // we pretend that the callee has no parameters so that the
469 // extension counts for the full amount of locals and not just
470 // locals-parms. This is because without a c2i adapter the parm
471 // area as created by the compiled frame will not be usable by
472 // the interpreter. (Depending on the calling convention there
473 // may not even be enough space).
475 // QQQ I'd rather see this pushed down into last_frame_adjust
476 // and have it take the sender (aka caller).
478 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
479 caller_adjustment = last_frame_adjust(0, callee_locals);
480 } else if (callee_locals > callee_parameters) {
481 // The caller frame may need extending to accommodate
482 // non-parameter locals of the first unpacked interpreted frame.
483 // Compute that adjustment.
484 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
485 }
487 // If the sender is deoptimized the we must retrieve the address of the handler
488 // since the frame will "magically" show the original pc before the deopt
489 // and we'd undo the deopt.
491 frame_pcs[0] = deopt_sender.raw_pc();
493 #ifndef SHARK
494 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
495 #endif // SHARK
497 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
498 caller_adjustment * BytesPerWord,
499 caller_was_method_handle ? 0 : callee_parameters,
500 number_of_frames,
501 frame_sizes,
502 frame_pcs,
503 return_type);
504 // On some platforms, we need a way to pass some platform dependent
505 // information to the unpacking code so the skeletal frames come out
506 // correct (initial fp value, unextended sp, ...)
507 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
509 if (array->frames() > 1) {
510 if (VerifyStack && TraceDeoptimization) {
511 ttyLocker ttyl;
512 tty->print_cr("Deoptimizing method containing inlining");
513 }
514 }
516 array->set_unroll_block(info);
517 return info;
518 }
520 // Called to cleanup deoptimization data structures in normal case
521 // after unpacking to stack and when stack overflow error occurs
522 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
523 vframeArray *array) {
525 // Get array if coming from exception
526 if (array == NULL) {
527 array = thread->vframe_array_head();
528 }
529 thread->set_vframe_array_head(NULL);
531 // Free the previous UnrollBlock
532 vframeArray* old_array = thread->vframe_array_last();
533 thread->set_vframe_array_last(array);
535 if (old_array != NULL) {
536 UnrollBlock* old_info = old_array->unroll_block();
537 old_array->set_unroll_block(NULL);
538 delete old_info;
539 delete old_array;
540 }
542 // Deallocate any resource creating in this routine and any ResourceObjs allocated
543 // inside the vframeArray (StackValueCollections)
545 delete thread->deopt_mark();
546 thread->set_deopt_mark(NULL);
547 thread->set_deopt_nmethod(NULL);
550 if (JvmtiExport::can_pop_frame()) {
551 #ifndef CC_INTERP
552 // Regardless of whether we entered this routine with the pending
553 // popframe condition bit set, we should always clear it now
554 thread->clear_popframe_condition();
555 #else
556 // C++ interpeter will clear has_pending_popframe when it enters
557 // with method_resume. For deopt_resume2 we clear it now.
558 if (thread->popframe_forcing_deopt_reexecution())
559 thread->clear_popframe_condition();
560 #endif /* CC_INTERP */
561 }
563 // unpack_frames() is called at the end of the deoptimization handler
564 // and (in C2) at the end of the uncommon trap handler. Note this fact
565 // so that an asynchronous stack walker can work again. This counter is
566 // incremented at the beginning of fetch_unroll_info() and (in C2) at
567 // the beginning of uncommon_trap().
568 thread->dec_in_deopt_handler();
569 }
572 // Return BasicType of value being returned
573 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
575 // We are already active int he special DeoptResourceMark any ResourceObj's we
576 // allocate will be freed at the end of the routine.
578 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
579 // but makes the entry a little slower. There is however a little dance we have to
580 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
581 ResetNoHandleMark rnhm; // No-op in release/product versions
582 HandleMark hm;
584 frame stub_frame = thread->last_frame();
586 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
587 // must point to the vframeArray for the unpack frame.
588 vframeArray* array = thread->vframe_array_head();
590 #ifndef PRODUCT
591 if (TraceDeoptimization) {
592 ttyLocker ttyl;
593 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
594 }
595 #endif
596 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
597 stub_frame.pc(), stub_frame.sp(), exec_mode);
599 UnrollBlock* info = array->unroll_block();
601 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
602 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
604 BasicType bt = info->return_type();
606 // If we have an exception pending, claim that the return type is an oop
607 // so the deopt_blob does not overwrite the exception_oop.
609 if (exec_mode == Unpack_exception)
610 bt = T_OBJECT;
612 // Cleanup thread deopt data
613 cleanup_deopt_info(thread, array);
615 #ifndef PRODUCT
616 if (VerifyStack) {
617 ResourceMark res_mark;
619 thread->validate_frame_layout();
621 // Verify that the just-unpacked frames match the interpreter's
622 // notions of expression stack and locals
623 vframeArray* cur_array = thread->vframe_array_last();
624 RegisterMap rm(thread, false);
625 rm.set_include_argument_oops(false);
626 bool is_top_frame = true;
627 int callee_size_of_parameters = 0;
628 int callee_max_locals = 0;
629 for (int i = 0; i < cur_array->frames(); i++) {
630 vframeArrayElement* el = cur_array->element(i);
631 frame* iframe = el->iframe();
632 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
634 // Get the oop map for this bci
635 InterpreterOopMap mask;
636 int cur_invoke_parameter_size = 0;
637 bool try_next_mask = false;
638 int next_mask_expression_stack_size = -1;
639 int top_frame_expression_stack_adjustment = 0;
640 methodHandle mh(thread, iframe->interpreter_frame_method());
641 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
642 BytecodeStream str(mh);
643 str.set_start(iframe->interpreter_frame_bci());
644 int max_bci = mh->code_size();
645 // Get to the next bytecode if possible
646 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
647 // Check to see if we can grab the number of outgoing arguments
648 // at an uncommon trap for an invoke (where the compiler
649 // generates debug info before the invoke has executed)
650 Bytecodes::Code cur_code = str.next();
651 if (cur_code == Bytecodes::_invokevirtual ||
652 cur_code == Bytecodes::_invokespecial ||
653 cur_code == Bytecodes::_invokestatic ||
654 cur_code == Bytecodes::_invokeinterface ||
655 cur_code == Bytecodes::_invokedynamic) {
656 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
657 Symbol* signature = invoke.signature();
658 ArgumentSizeComputer asc(signature);
659 cur_invoke_parameter_size = asc.size();
660 if (invoke.has_receiver()) {
661 // Add in receiver
662 ++cur_invoke_parameter_size;
663 }
664 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
665 callee_size_of_parameters++;
666 }
667 }
668 if (str.bci() < max_bci) {
669 Bytecodes::Code bc = str.next();
670 if (bc >= 0) {
671 // The interpreter oop map generator reports results before
672 // the current bytecode has executed except in the case of
673 // calls. It seems to be hard to tell whether the compiler
674 // has emitted debug information matching the "state before"
675 // a given bytecode or the state after, so we try both
676 switch (cur_code) {
677 case Bytecodes::_invokevirtual:
678 case Bytecodes::_invokespecial:
679 case Bytecodes::_invokestatic:
680 case Bytecodes::_invokeinterface:
681 case Bytecodes::_invokedynamic:
682 case Bytecodes::_athrow:
683 break;
684 default: {
685 InterpreterOopMap next_mask;
686 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
687 next_mask_expression_stack_size = next_mask.expression_stack_size();
688 // Need to subtract off the size of the result type of
689 // the bytecode because this is not described in the
690 // debug info but returned to the interpreter in the TOS
691 // caching register
692 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
693 if (bytecode_result_type != T_ILLEGAL) {
694 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
695 }
696 assert(top_frame_expression_stack_adjustment >= 0, "");
697 try_next_mask = true;
698 break;
699 }
700 }
701 }
702 }
704 // Verify stack depth and oops in frame
705 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
706 if (!(
707 /* SPARC */
708 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
709 /* x86 */
710 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
711 (try_next_mask &&
712 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
713 top_frame_expression_stack_adjustment))) ||
714 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
715 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
716 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
717 )) {
718 ttyLocker ttyl;
720 // Print out some information that will help us debug the problem
721 tty->print_cr("Wrong number of expression stack elements during deoptimization");
722 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
723 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
724 iframe->interpreter_frame_expression_stack_size());
725 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
726 tty->print_cr(" try_next_mask = %d", try_next_mask);
727 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
728 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
729 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
730 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
731 tty->print_cr(" exec_mode = %d", exec_mode);
732 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
733 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
734 tty->print_cr(" Interpreted frames:");
735 for (int k = 0; k < cur_array->frames(); k++) {
736 vframeArrayElement* el = cur_array->element(k);
737 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
738 }
739 cur_array->print_on_2(tty);
740 guarantee(false, "wrong number of expression stack elements during deopt");
741 }
742 VerifyOopClosure verify;
743 iframe->oops_interpreted_do(&verify, NULL, &rm, false);
744 callee_size_of_parameters = mh->size_of_parameters();
745 callee_max_locals = mh->max_locals();
746 is_top_frame = false;
747 }
748 }
749 #endif /* !PRODUCT */
752 return bt;
753 JRT_END
756 int Deoptimization::deoptimize_dependents() {
757 Threads::deoptimized_wrt_marked_nmethods();
758 return 0;
759 }
762 #ifdef COMPILER2
763 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
764 Handle pending_exception(thread->pending_exception());
765 const char* exception_file = thread->exception_file();
766 int exception_line = thread->exception_line();
767 thread->clear_pending_exception();
769 bool failures = false;
771 for (int i = 0; i < objects->length(); i++) {
772 assert(objects->at(i)->is_object(), "invalid debug information");
773 ObjectValue* sv = (ObjectValue*) objects->at(i);
775 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
776 oop obj = NULL;
778 if (k->oop_is_instance()) {
779 InstanceKlass* ik = InstanceKlass::cast(k());
780 obj = ik->allocate_instance(THREAD);
781 } else if (k->oop_is_typeArray()) {
782 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
783 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
784 int len = sv->field_size() / type2size[ak->element_type()];
785 obj = ak->allocate(len, THREAD);
786 } else if (k->oop_is_objArray()) {
787 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
788 obj = ak->allocate(sv->field_size(), THREAD);
789 }
791 if (obj == NULL) {
792 failures = true;
793 }
795 assert(sv->value().is_null(), "redundant reallocation");
796 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
797 CLEAR_PENDING_EXCEPTION;
798 sv->set_value(obj);
799 }
801 if (failures) {
802 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
803 } else if (pending_exception.not_null()) {
804 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
805 }
807 return failures;
808 }
810 // This assumes that the fields are stored in ObjectValue in the same order
811 // they are yielded by do_nonstatic_fields.
812 class FieldReassigner: public FieldClosure {
813 frame* _fr;
814 RegisterMap* _reg_map;
815 ObjectValue* _sv;
816 InstanceKlass* _ik;
817 oop _obj;
819 int _i;
820 public:
821 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
822 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
824 int i() const { return _i; }
827 void do_field(fieldDescriptor* fd) {
828 intptr_t val;
829 StackValue* value =
830 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
831 int offset = fd->offset();
832 switch (fd->field_type()) {
833 case T_OBJECT: case T_ARRAY:
834 assert(value->type() == T_OBJECT, "Agreement.");
835 _obj->obj_field_put(offset, value->get_obj()());
836 break;
838 case T_LONG: case T_DOUBLE: {
839 assert(value->type() == T_INT, "Agreement.");
840 StackValue* low =
841 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
842 #ifdef _LP64
843 jlong res = (jlong)low->get_int();
844 #else
845 #ifdef SPARC
846 // For SPARC we have to swap high and low words.
847 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
848 #else
849 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
850 #endif //SPARC
851 #endif
852 _obj->long_field_put(offset, res);
853 break;
854 }
855 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
856 case T_INT: case T_FLOAT: // 4 bytes.
857 assert(value->type() == T_INT, "Agreement.");
858 val = value->get_int();
859 _obj->int_field_put(offset, (jint)*((jint*)&val));
860 break;
862 case T_SHORT:
863 assert(value->type() == T_INT, "Agreement.");
864 val = value->get_int();
865 _obj->short_field_put(offset, (jshort)*((jint*)&val));
866 break;
868 case T_CHAR:
869 assert(value->type() == T_INT, "Agreement.");
870 val = value->get_int();
871 _obj->char_field_put(offset, (jchar)*((jint*)&val));
872 break;
874 case T_BYTE:
875 assert(value->type() == T_INT, "Agreement.");
876 val = value->get_int();
877 _obj->byte_field_put(offset, (jbyte)*((jint*)&val));
878 break;
880 case T_BOOLEAN:
881 assert(value->type() == T_INT, "Agreement.");
882 val = value->get_int();
883 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
884 break;
886 default:
887 ShouldNotReachHere();
888 }
889 _i++;
890 }
891 };
893 // restore elements of an eliminated type array
894 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
895 int index = 0;
896 intptr_t val;
898 for (int i = 0; i < sv->field_size(); i++) {
899 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
900 switch(type) {
901 case T_LONG: case T_DOUBLE: {
902 assert(value->type() == T_INT, "Agreement.");
903 StackValue* low =
904 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
905 #ifdef _LP64
906 jlong res = (jlong)low->get_int();
907 #else
908 #ifdef SPARC
909 // For SPARC we have to swap high and low words.
910 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
911 #else
912 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
913 #endif //SPARC
914 #endif
915 obj->long_at_put(index, res);
916 break;
917 }
919 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
920 case T_INT: case T_FLOAT: // 4 bytes.
921 assert(value->type() == T_INT, "Agreement.");
922 val = value->get_int();
923 obj->int_at_put(index, (jint)*((jint*)&val));
924 break;
926 case T_SHORT:
927 assert(value->type() == T_INT, "Agreement.");
928 val = value->get_int();
929 obj->short_at_put(index, (jshort)*((jint*)&val));
930 break;
932 case T_CHAR:
933 assert(value->type() == T_INT, "Agreement.");
934 val = value->get_int();
935 obj->char_at_put(index, (jchar)*((jint*)&val));
936 break;
938 case T_BYTE:
939 assert(value->type() == T_INT, "Agreement.");
940 val = value->get_int();
941 obj->byte_at_put(index, (jbyte)*((jint*)&val));
942 break;
944 case T_BOOLEAN:
945 assert(value->type() == T_INT, "Agreement.");
946 val = value->get_int();
947 obj->bool_at_put(index, (jboolean)*((jint*)&val));
948 break;
950 default:
951 ShouldNotReachHere();
952 }
953 index++;
954 }
955 }
958 // restore fields of an eliminated object array
959 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
960 for (int i = 0; i < sv->field_size(); i++) {
961 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
962 assert(value->type() == T_OBJECT, "object element expected");
963 obj->obj_at_put(i, value->get_obj()());
964 }
965 }
968 // restore fields of all eliminated objects and arrays
969 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
970 for (int i = 0; i < objects->length(); i++) {
971 ObjectValue* sv = (ObjectValue*) objects->at(i);
972 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
973 Handle obj = sv->value();
974 assert(obj.not_null() || realloc_failures, "reallocation was missed");
975 if (obj.is_null()) {
976 continue;
977 }
979 if (k->oop_is_instance()) {
980 InstanceKlass* ik = InstanceKlass::cast(k());
981 FieldReassigner reassign(fr, reg_map, sv, obj());
982 ik->do_nonstatic_fields(&reassign);
983 } else if (k->oop_is_typeArray()) {
984 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
985 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
986 } else if (k->oop_is_objArray()) {
987 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
988 }
989 }
990 }
993 // relock objects for which synchronization was eliminated
994 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
995 for (int i = 0; i < monitors->length(); i++) {
996 MonitorInfo* mon_info = monitors->at(i);
997 if (mon_info->eliminated()) {
998 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
999 if (!mon_info->owner_is_scalar_replaced()) {
1000 Handle obj = Handle(mon_info->owner());
1001 markOop mark = obj->mark();
1002 if (UseBiasedLocking && mark->has_bias_pattern()) {
1003 // New allocated objects may have the mark set to anonymously biased.
1004 // Also the deoptimized method may called methods with synchronization
1005 // where the thread-local object is bias locked to the current thread.
1006 assert(mark->is_biased_anonymously() ||
1007 mark->biased_locker() == thread, "should be locked to current thread");
1008 // Reset mark word to unbiased prototype.
1009 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
1010 obj->set_mark(unbiased_prototype);
1011 }
1012 BasicLock* lock = mon_info->lock();
1013 ObjectSynchronizer::slow_enter(obj, lock, thread);
1014 assert(mon_info->owner()->is_locked(), "object must be locked now");
1015 }
1016 }
1017 }
1018 }
1021 #ifndef PRODUCT
1022 // print information about reallocated objects
1023 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1024 fieldDescriptor fd;
1026 for (int i = 0; i < objects->length(); i++) {
1027 ObjectValue* sv = (ObjectValue*) objects->at(i);
1028 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
1029 Handle obj = sv->value();
1031 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()());
1032 k->print_value();
1033 assert(obj.not_null() || realloc_failures, "reallocation was missed");
1034 if (obj.is_null()) {
1035 tty->print(" allocation failed");
1036 } else {
1037 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1038 }
1039 tty->cr();
1041 if (Verbose && !obj.is_null()) {
1042 k->oop_print_on(obj(), tty);
1043 }
1044 }
1045 }
1046 #endif
1047 #endif // COMPILER2
1049 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1050 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
1052 #ifndef PRODUCT
1053 if (TraceDeoptimization) {
1054 ttyLocker ttyl;
1055 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
1056 fr.print_on(tty);
1057 tty->print_cr(" Virtual frames (innermost first):");
1058 for (int index = 0; index < chunk->length(); index++) {
1059 compiledVFrame* vf = chunk->at(index);
1060 tty->print(" %2d - ", index);
1061 vf->print_value();
1062 int bci = chunk->at(index)->raw_bci();
1063 const char* code_name;
1064 if (bci == SynchronizationEntryBCI) {
1065 code_name = "sync entry";
1066 } else {
1067 Bytecodes::Code code = vf->method()->code_at(bci);
1068 code_name = Bytecodes::name(code);
1069 }
1070 tty->print(" - %s", code_name);
1071 tty->print_cr(" @ bci %d ", bci);
1072 if (Verbose) {
1073 vf->print();
1074 tty->cr();
1075 }
1076 }
1077 }
1078 #endif
1080 // Register map for next frame (used for stack crawl). We capture
1081 // the state of the deopt'ing frame's caller. Thus if we need to
1082 // stuff a C2I adapter we can properly fill in the callee-save
1083 // register locations.
1084 frame caller = fr.sender(reg_map);
1085 int frame_size = caller.sp() - fr.sp();
1087 frame sender = caller;
1089 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1090 // the vframeArray containing the unpacking information is allocated in the C heap.
1091 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1092 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1094 // Compare the vframeArray to the collected vframes
1095 assert(array->structural_compare(thread, chunk), "just checking");
1097 #ifndef PRODUCT
1098 if (TraceDeoptimization) {
1099 ttyLocker ttyl;
1100 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1101 }
1102 #endif // PRODUCT
1104 return array;
1105 }
1107 #ifdef COMPILER2
1108 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1109 // Reallocation of some scalar replaced objects failed. Record
1110 // that we need to pop all the interpreter frames for the
1111 // deoptimized compiled frame.
1112 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1113 thread->set_frames_to_pop_failed_realloc(array->frames());
1114 // Unlock all monitors here otherwise the interpreter will see a
1115 // mix of locked and unlocked monitors (because of failed
1116 // reallocations of synchronized objects) and be confused.
1117 for (int i = 0; i < array->frames(); i++) {
1118 MonitorChunk* monitors = array->element(i)->monitors();
1119 if (monitors != NULL) {
1120 for (int j = 0; j < monitors->number_of_monitors(); j++) {
1121 BasicObjectLock* src = monitors->at(j);
1122 if (src->obj() != NULL) {
1123 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1124 }
1125 }
1126 array->element(i)->free_monitors(thread);
1127 #ifdef ASSERT
1128 array->element(i)->set_removed_monitors();
1129 #endif
1130 }
1131 }
1132 }
1133 #endif
1135 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1136 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1137 for (int i = 0; i < monitors->length(); i++) {
1138 MonitorInfo* mon_info = monitors->at(i);
1139 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1140 objects_to_revoke->append(Handle(mon_info->owner()));
1141 }
1142 }
1143 }
1146 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1147 if (!UseBiasedLocking) {
1148 return;
1149 }
1151 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1153 // Unfortunately we don't have a RegisterMap available in most of
1154 // the places we want to call this routine so we need to walk the
1155 // stack again to update the register map.
1156 if (map == NULL || !map->update_map()) {
1157 StackFrameStream sfs(thread, true);
1158 bool found = false;
1159 while (!found && !sfs.is_done()) {
1160 frame* cur = sfs.current();
1161 sfs.next();
1162 found = cur->id() == fr.id();
1163 }
1164 assert(found, "frame to be deoptimized not found on target thread's stack");
1165 map = sfs.register_map();
1166 }
1168 vframe* vf = vframe::new_vframe(&fr, map, thread);
1169 compiledVFrame* cvf = compiledVFrame::cast(vf);
1170 // Revoke monitors' biases in all scopes
1171 while (!cvf->is_top()) {
1172 collect_monitors(cvf, objects_to_revoke);
1173 cvf = compiledVFrame::cast(cvf->sender());
1174 }
1175 collect_monitors(cvf, objects_to_revoke);
1177 if (SafepointSynchronize::is_at_safepoint()) {
1178 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1179 } else {
1180 BiasedLocking::revoke(objects_to_revoke);
1181 }
1182 }
1185 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1186 if (!UseBiasedLocking) {
1187 return;
1188 }
1190 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1191 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1192 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1193 if (jt->has_last_Java_frame()) {
1194 StackFrameStream sfs(jt, true);
1195 while (!sfs.is_done()) {
1196 frame* cur = sfs.current();
1197 if (cb->contains(cur->pc())) {
1198 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1199 compiledVFrame* cvf = compiledVFrame::cast(vf);
1200 // Revoke monitors' biases in all scopes
1201 while (!cvf->is_top()) {
1202 collect_monitors(cvf, objects_to_revoke);
1203 cvf = compiledVFrame::cast(cvf->sender());
1204 }
1205 collect_monitors(cvf, objects_to_revoke);
1206 }
1207 sfs.next();
1208 }
1209 }
1210 }
1211 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1212 }
1215 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1216 assert(fr.can_be_deoptimized(), "checking frame type");
1218 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1220 // Patch the nmethod so that when execution returns to it we will
1221 // deopt the execution state and return to the interpreter.
1222 fr.deoptimize(thread);
1223 }
1225 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1226 // Deoptimize only if the frame comes from compile code.
1227 // Do not deoptimize the frame which is already patched
1228 // during the execution of the loops below.
1229 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1230 return;
1231 }
1232 ResourceMark rm;
1233 DeoptimizationMarker dm;
1234 if (UseBiasedLocking) {
1235 revoke_biases_of_monitors(thread, fr, map);
1236 }
1237 deoptimize_single_frame(thread, fr);
1239 }
1242 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1243 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1244 "can only deoptimize other thread at a safepoint");
1245 // Compute frame and register map based on thread and sp.
1246 RegisterMap reg_map(thread, UseBiasedLocking);
1247 frame fr = thread->last_frame();
1248 while (fr.id() != id) {
1249 fr = fr.sender(®_map);
1250 }
1251 deoptimize(thread, fr, ®_map);
1252 }
1255 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1256 if (thread == Thread::current()) {
1257 Deoptimization::deoptimize_frame_internal(thread, id);
1258 } else {
1259 VM_DeoptimizeFrame deopt(thread, id);
1260 VMThread::execute(&deopt);
1261 }
1262 }
1265 // JVMTI PopFrame support
1266 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1267 {
1268 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1269 }
1270 JRT_END
1273 #if defined(COMPILER2) || defined(SHARK)
1274 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1275 // in case of an unresolved klass entry, load the class.
1276 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1277 Klass* tk = constant_pool->klass_at(index, CHECK);
1278 return;
1279 }
1281 if (!constant_pool->tag_at(index).is_symbol()) return;
1283 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1284 Symbol* symbol = constant_pool->symbol_at(index);
1286 // class name?
1287 if (symbol->byte_at(0) != '(') {
1288 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1289 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1290 return;
1291 }
1293 // then it must be a signature!
1294 ResourceMark rm(THREAD);
1295 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1296 if (ss.is_object()) {
1297 Symbol* class_name = ss.as_symbol(CHECK);
1298 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1299 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1300 }
1301 }
1302 }
1305 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1306 EXCEPTION_MARK;
1307 load_class_by_index(constant_pool, index, THREAD);
1308 if (HAS_PENDING_EXCEPTION) {
1309 // Exception happened during classloading. We ignore the exception here, since it
1310 // is going to be rethrown since the current activation is going to be deoptimized and
1311 // the interpreter will re-execute the bytecode.
1312 CLEAR_PENDING_EXCEPTION;
1313 // Class loading called java code which may have caused a stack
1314 // overflow. If the exception was thrown right before the return
1315 // to the runtime the stack is no longer guarded. Reguard the
1316 // stack otherwise if we return to the uncommon trap blob and the
1317 // stack bang causes a stack overflow we crash.
1318 assert(THREAD->is_Java_thread(), "only a java thread can be here");
1319 JavaThread* thread = (JavaThread*)THREAD;
1320 bool guard_pages_enabled = thread->stack_yellow_zone_enabled();
1321 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1322 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1323 }
1324 }
1326 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1327 HandleMark hm;
1329 // uncommon_trap() is called at the beginning of the uncommon trap
1330 // handler. Note this fact before we start generating temporary frames
1331 // that can confuse an asynchronous stack walker. This counter is
1332 // decremented at the end of unpack_frames().
1333 thread->inc_in_deopt_handler();
1335 // We need to update the map if we have biased locking.
1336 RegisterMap reg_map(thread, UseBiasedLocking);
1337 frame stub_frame = thread->last_frame();
1338 frame fr = stub_frame.sender(®_map);
1339 // Make sure the calling nmethod is not getting deoptimized and removed
1340 // before we are done with it.
1341 nmethodLocker nl(fr.pc());
1343 // Log a message
1344 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1345 trap_request, fr.pc());
1347 {
1348 ResourceMark rm;
1350 // Revoke biases of any monitors in the frame to ensure we can migrate them
1351 revoke_biases_of_monitors(thread, fr, ®_map);
1353 DeoptReason reason = trap_request_reason(trap_request);
1354 DeoptAction action = trap_request_action(trap_request);
1355 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1357 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1358 compiledVFrame* cvf = compiledVFrame::cast(vf);
1360 nmethod* nm = cvf->code();
1362 ScopeDesc* trap_scope = cvf->scope();
1363 methodHandle trap_method = trap_scope->method();
1364 int trap_bci = trap_scope->bci();
1365 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1367 // Record this event in the histogram.
1368 gather_statistics(reason, action, trap_bc);
1370 // Ensure that we can record deopt. history:
1371 // Need MDO to record RTM code generation state.
1372 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking );
1374 MethodData* trap_mdo =
1375 get_method_data(thread, trap_method, create_if_missing);
1377 // Log a message
1378 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1379 trap_reason_name(reason), trap_action_name(action), fr.pc(),
1380 trap_method->name_and_sig_as_C_string(), trap_bci);
1382 // Print a bunch of diagnostics, if requested.
1383 if (TraceDeoptimization || LogCompilation) {
1384 ResourceMark rm;
1385 ttyLocker ttyl;
1386 char buf[100];
1387 if (xtty != NULL) {
1388 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1389 os::current_thread_id(),
1390 format_trap_request(buf, sizeof(buf), trap_request));
1391 nm->log_identity(xtty);
1392 }
1393 Symbol* class_name = NULL;
1394 bool unresolved = false;
1395 if (unloaded_class_index >= 0) {
1396 constantPoolHandle constants (THREAD, trap_method->constants());
1397 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1398 class_name = constants->klass_name_at(unloaded_class_index);
1399 unresolved = true;
1400 if (xtty != NULL)
1401 xtty->print(" unresolved='1'");
1402 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1403 class_name = constants->symbol_at(unloaded_class_index);
1404 }
1405 if (xtty != NULL)
1406 xtty->name(class_name);
1407 }
1408 if (xtty != NULL && trap_mdo != NULL) {
1409 // Dump the relevant MDO state.
1410 // This is the deopt count for the current reason, any previous
1411 // reasons or recompiles seen at this point.
1412 int dcnt = trap_mdo->trap_count(reason);
1413 if (dcnt != 0)
1414 xtty->print(" count='%d'", dcnt);
1415 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1416 int dos = (pdata == NULL)? 0: pdata->trap_state();
1417 if (dos != 0) {
1418 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1419 if (trap_state_is_recompiled(dos)) {
1420 int recnt2 = trap_mdo->overflow_recompile_count();
1421 if (recnt2 != 0)
1422 xtty->print(" recompiles2='%d'", recnt2);
1423 }
1424 }
1425 }
1426 if (xtty != NULL) {
1427 xtty->stamp();
1428 xtty->end_head();
1429 }
1430 if (TraceDeoptimization) { // make noise on the tty
1431 tty->print("Uncommon trap occurred in");
1432 nm->method()->print_short_name(tty);
1433 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1434 fr.pc(),
1435 os::current_thread_id(),
1436 trap_reason_name(reason),
1437 trap_action_name(action),
1438 unloaded_class_index);
1439 if (class_name != NULL) {
1440 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1441 class_name->print_symbol_on(tty);
1442 }
1443 tty->cr();
1444 }
1445 if (xtty != NULL) {
1446 // Log the precise location of the trap.
1447 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1448 xtty->begin_elem("jvms bci='%d'", sd->bci());
1449 xtty->method(sd->method());
1450 xtty->end_elem();
1451 if (sd->is_top()) break;
1452 }
1453 xtty->tail("uncommon_trap");
1454 }
1455 }
1456 // (End diagnostic printout.)
1458 // Load class if necessary
1459 if (unloaded_class_index >= 0) {
1460 constantPoolHandle constants(THREAD, trap_method->constants());
1461 load_class_by_index(constants, unloaded_class_index);
1462 }
1464 // Flush the nmethod if necessary and desirable.
1465 //
1466 // We need to avoid situations where we are re-flushing the nmethod
1467 // because of a hot deoptimization site. Repeated flushes at the same
1468 // point need to be detected by the compiler and avoided. If the compiler
1469 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1470 // module must take measures to avoid an infinite cycle of recompilation
1471 // and deoptimization. There are several such measures:
1472 //
1473 // 1. If a recompilation is ordered a second time at some site X
1474 // and for the same reason R, the action is adjusted to 'reinterpret',
1475 // to give the interpreter time to exercise the method more thoroughly.
1476 // If this happens, the method's overflow_recompile_count is incremented.
1477 //
1478 // 2. If the compiler fails to reduce the deoptimization rate, then
1479 // the method's overflow_recompile_count will begin to exceed the set
1480 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1481 // is adjusted to 'make_not_compilable', and the method is abandoned
1482 // to the interpreter. This is a performance hit for hot methods,
1483 // but is better than a disastrous infinite cycle of recompilations.
1484 // (Actually, only the method containing the site X is abandoned.)
1485 //
1486 // 3. In parallel with the previous measures, if the total number of
1487 // recompilations of a method exceeds the much larger set limit
1488 // PerMethodRecompilationCutoff, the method is abandoned.
1489 // This should only happen if the method is very large and has
1490 // many "lukewarm" deoptimizations. The code which enforces this
1491 // limit is elsewhere (class nmethod, class Method).
1492 //
1493 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1494 // to recompile at each bytecode independently of the per-BCI cutoff.
1495 //
1496 // The decision to update code is up to the compiler, and is encoded
1497 // in the Action_xxx code. If the compiler requests Action_none
1498 // no trap state is changed, no compiled code is changed, and the
1499 // computation suffers along in the interpreter.
1500 //
1501 // The other action codes specify various tactics for decompilation
1502 // and recompilation. Action_maybe_recompile is the loosest, and
1503 // allows the compiled code to stay around until enough traps are seen,
1504 // and until the compiler gets around to recompiling the trapping method.
1505 //
1506 // The other actions cause immediate removal of the present code.
1508 // Traps caused by injected profile shouldn't pollute trap counts.
1509 bool injected_profile_trap = trap_method->has_injected_profile() &&
1510 (reason == Reason_intrinsic || reason == Reason_unreached);
1511 bool update_trap_state = !injected_profile_trap;
1512 bool make_not_entrant = false;
1513 bool make_not_compilable = false;
1514 bool reprofile = false;
1515 switch (action) {
1516 case Action_none:
1517 // Keep the old code.
1518 update_trap_state = false;
1519 break;
1520 case Action_maybe_recompile:
1521 // Do not need to invalidate the present code, but we can
1522 // initiate another
1523 // Start compiler without (necessarily) invalidating the nmethod.
1524 // The system will tolerate the old code, but new code should be
1525 // generated when possible.
1526 break;
1527 case Action_reinterpret:
1528 // Go back into the interpreter for a while, and then consider
1529 // recompiling form scratch.
1530 make_not_entrant = true;
1531 // Reset invocation counter for outer most method.
1532 // This will allow the interpreter to exercise the bytecodes
1533 // for a while before recompiling.
1534 // By contrast, Action_make_not_entrant is immediate.
1535 //
1536 // Note that the compiler will track null_check, null_assert,
1537 // range_check, and class_check events and log them as if they
1538 // had been traps taken from compiled code. This will update
1539 // the MDO trap history so that the next compilation will
1540 // properly detect hot trap sites.
1541 reprofile = true;
1542 break;
1543 case Action_make_not_entrant:
1544 // Request immediate recompilation, and get rid of the old code.
1545 // Make them not entrant, so next time they are called they get
1546 // recompiled. Unloaded classes are loaded now so recompile before next
1547 // time they are called. Same for uninitialized. The interpreter will
1548 // link the missing class, if any.
1549 make_not_entrant = true;
1550 break;
1551 case Action_make_not_compilable:
1552 // Give up on compiling this method at all.
1553 make_not_entrant = true;
1554 make_not_compilable = true;
1555 break;
1556 default:
1557 ShouldNotReachHere();
1558 }
1560 // Setting +ProfileTraps fixes the following, on all platforms:
1561 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1562 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1563 // recompile relies on a MethodData* to record heroic opt failures.
1565 // Whether the interpreter is producing MDO data or not, we also need
1566 // to use the MDO to detect hot deoptimization points and control
1567 // aggressive optimization.
1568 bool inc_recompile_count = false;
1569 ProfileData* pdata = NULL;
1570 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1571 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1572 uint this_trap_count = 0;
1573 bool maybe_prior_trap = false;
1574 bool maybe_prior_recompile = false;
1575 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1576 nm->method(),
1577 //outputs:
1578 this_trap_count,
1579 maybe_prior_trap,
1580 maybe_prior_recompile);
1581 // Because the interpreter also counts null, div0, range, and class
1582 // checks, these traps from compiled code are double-counted.
1583 // This is harmless; it just means that the PerXTrapLimit values
1584 // are in effect a little smaller than they look.
1586 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1587 if (per_bc_reason != Reason_none) {
1588 // Now take action based on the partially known per-BCI history.
1589 if (maybe_prior_trap
1590 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1591 // If there are too many traps at this BCI, force a recompile.
1592 // This will allow the compiler to see the limit overflow, and
1593 // take corrective action, if possible. The compiler generally
1594 // does not use the exact PerBytecodeTrapLimit value, but instead
1595 // changes its tactics if it sees any traps at all. This provides
1596 // a little hysteresis, delaying a recompile until a trap happens
1597 // several times.
1598 //
1599 // Actually, since there is only one bit of counter per BCI,
1600 // the possible per-BCI counts are {0,1,(per-method count)}.
1601 // This produces accurate results if in fact there is only
1602 // one hot trap site, but begins to get fuzzy if there are
1603 // many sites. For example, if there are ten sites each
1604 // trapping two or more times, they each get the blame for
1605 // all of their traps.
1606 make_not_entrant = true;
1607 }
1609 // Detect repeated recompilation at the same BCI, and enforce a limit.
1610 if (make_not_entrant && maybe_prior_recompile) {
1611 // More than one recompile at this point.
1612 inc_recompile_count = maybe_prior_trap;
1613 }
1614 } else {
1615 // For reasons which are not recorded per-bytecode, we simply
1616 // force recompiles unconditionally.
1617 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1618 make_not_entrant = true;
1619 }
1621 // Go back to the compiler if there are too many traps in this method.
1622 if (this_trap_count >= per_method_trap_limit(reason)) {
1623 // If there are too many traps in this method, force a recompile.
1624 // This will allow the compiler to see the limit overflow, and
1625 // take corrective action, if possible.
1626 // (This condition is an unlikely backstop only, because the
1627 // PerBytecodeTrapLimit is more likely to take effect first,
1628 // if it is applicable.)
1629 make_not_entrant = true;
1630 }
1632 // Here's more hysteresis: If there has been a recompile at
1633 // this trap point already, run the method in the interpreter
1634 // for a while to exercise it more thoroughly.
1635 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1636 reprofile = true;
1637 }
1639 }
1641 // Take requested actions on the method:
1643 // Recompile
1644 if (make_not_entrant) {
1645 if (!nm->make_not_entrant()) {
1646 return; // the call did not change nmethod's state
1647 }
1649 if (pdata != NULL) {
1650 // Record the recompilation event, if any.
1651 int tstate0 = pdata->trap_state();
1652 int tstate1 = trap_state_set_recompiled(tstate0, true);
1653 if (tstate1 != tstate0)
1654 pdata->set_trap_state(tstate1);
1655 }
1657 #if INCLUDE_RTM_OPT
1658 // Restart collecting RTM locking abort statistic if the method
1659 // is recompiled for a reason other than RTM state change.
1660 // Assume that in new recompiled code the statistic could be different,
1661 // for example, due to different inlining.
1662 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1663 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) {
1664 trap_mdo->atomic_set_rtm_state(ProfileRTM);
1665 }
1666 #endif
1667 }
1669 if (inc_recompile_count) {
1670 trap_mdo->inc_overflow_recompile_count();
1671 if ((uint)trap_mdo->overflow_recompile_count() >
1672 (uint)PerBytecodeRecompilationCutoff) {
1673 // Give up on the method containing the bad BCI.
1674 if (trap_method() == nm->method()) {
1675 make_not_compilable = true;
1676 } else {
1677 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1678 // But give grace to the enclosing nm->method().
1679 }
1680 }
1681 }
1683 // Reprofile
1684 if (reprofile) {
1685 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1686 }
1688 // Give up compiling
1689 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1690 assert(make_not_entrant, "consistent");
1691 nm->method()->set_not_compilable(CompLevel_full_optimization);
1692 }
1694 } // Free marked resources
1696 }
1697 JRT_END
1699 MethodData*
1700 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1701 bool create_if_missing) {
1702 Thread* THREAD = thread;
1703 MethodData* mdo = m()->method_data();
1704 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1705 // Build an MDO. Ignore errors like OutOfMemory;
1706 // that simply means we won't have an MDO to update.
1707 Method::build_interpreter_method_data(m, THREAD);
1708 if (HAS_PENDING_EXCEPTION) {
1709 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1710 CLEAR_PENDING_EXCEPTION;
1711 }
1712 mdo = m()->method_data();
1713 }
1714 return mdo;
1715 }
1717 ProfileData*
1718 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1719 int trap_bci,
1720 Deoptimization::DeoptReason reason,
1721 Method* compiled_method,
1722 //outputs:
1723 uint& ret_this_trap_count,
1724 bool& ret_maybe_prior_trap,
1725 bool& ret_maybe_prior_recompile) {
1726 uint prior_trap_count = trap_mdo->trap_count(reason);
1727 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1729 // If the runtime cannot find a place to store trap history,
1730 // it is estimated based on the general condition of the method.
1731 // If the method has ever been recompiled, or has ever incurred
1732 // a trap with the present reason , then this BCI is assumed
1733 // (pessimistically) to be the culprit.
1734 bool maybe_prior_trap = (prior_trap_count != 0);
1735 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1736 ProfileData* pdata = NULL;
1739 // For reasons which are recorded per bytecode, we check per-BCI data.
1740 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1741 if (per_bc_reason != Reason_none) {
1742 // Find the profile data for this BCI. If there isn't one,
1743 // try to allocate one from the MDO's set of spares.
1744 // This will let us detect a repeated trap at this point.
1745 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1747 if (pdata != NULL) {
1748 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
1749 if (LogCompilation && xtty != NULL) {
1750 ttyLocker ttyl;
1751 // no more room for speculative traps in this MDO
1752 xtty->elem("speculative_traps_oom");
1753 }
1754 }
1755 // Query the trap state of this profile datum.
1756 int tstate0 = pdata->trap_state();
1757 if (!trap_state_has_reason(tstate0, per_bc_reason))
1758 maybe_prior_trap = false;
1759 if (!trap_state_is_recompiled(tstate0))
1760 maybe_prior_recompile = false;
1762 // Update the trap state of this profile datum.
1763 int tstate1 = tstate0;
1764 // Record the reason.
1765 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1766 // Store the updated state on the MDO, for next time.
1767 if (tstate1 != tstate0)
1768 pdata->set_trap_state(tstate1);
1769 } else {
1770 if (LogCompilation && xtty != NULL) {
1771 ttyLocker ttyl;
1772 // Missing MDP? Leave a small complaint in the log.
1773 xtty->elem("missing_mdp bci='%d'", trap_bci);
1774 }
1775 }
1776 }
1778 // Return results:
1779 ret_this_trap_count = this_trap_count;
1780 ret_maybe_prior_trap = maybe_prior_trap;
1781 ret_maybe_prior_recompile = maybe_prior_recompile;
1782 return pdata;
1783 }
1785 void
1786 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1787 ResourceMark rm;
1788 // Ignored outputs:
1789 uint ignore_this_trap_count;
1790 bool ignore_maybe_prior_trap;
1791 bool ignore_maybe_prior_recompile;
1792 assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
1793 query_update_method_data(trap_mdo, trap_bci,
1794 (DeoptReason)reason,
1795 NULL,
1796 ignore_this_trap_count,
1797 ignore_maybe_prior_trap,
1798 ignore_maybe_prior_recompile);
1799 }
1801 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1803 // Still in Java no safepoints
1804 {
1805 // This enters VM and may safepoint
1806 uncommon_trap_inner(thread, trap_request);
1807 }
1808 return fetch_unroll_info_helper(thread);
1809 }
1811 // Local derived constants.
1812 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1813 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1814 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1816 //---------------------------trap_state_reason---------------------------------
1817 Deoptimization::DeoptReason
1818 Deoptimization::trap_state_reason(int trap_state) {
1819 // This assert provides the link between the width of DataLayout::trap_bits
1820 // and the encoding of "recorded" reasons. It ensures there are enough
1821 // bits to store all needed reasons in the per-BCI MDO profile.
1822 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1823 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1824 trap_state -= recompile_bit;
1825 if (trap_state == DS_REASON_MASK) {
1826 return Reason_many;
1827 } else {
1828 assert((int)Reason_none == 0, "state=0 => Reason_none");
1829 return (DeoptReason)trap_state;
1830 }
1831 }
1832 //-------------------------trap_state_has_reason-------------------------------
1833 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1834 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1835 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1836 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1837 trap_state -= recompile_bit;
1838 if (trap_state == DS_REASON_MASK) {
1839 return -1; // true, unspecifically (bottom of state lattice)
1840 } else if (trap_state == reason) {
1841 return 1; // true, definitely
1842 } else if (trap_state == 0) {
1843 return 0; // false, definitely (top of state lattice)
1844 } else {
1845 return 0; // false, definitely
1846 }
1847 }
1848 //-------------------------trap_state_add_reason-------------------------------
1849 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1850 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1851 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1852 trap_state -= recompile_bit;
1853 if (trap_state == DS_REASON_MASK) {
1854 return trap_state + recompile_bit; // already at state lattice bottom
1855 } else if (trap_state == reason) {
1856 return trap_state + recompile_bit; // the condition is already true
1857 } else if (trap_state == 0) {
1858 return reason + recompile_bit; // no condition has yet been true
1859 } else {
1860 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1861 }
1862 }
1863 //-----------------------trap_state_is_recompiled------------------------------
1864 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1865 return (trap_state & DS_RECOMPILE_BIT) != 0;
1866 }
1867 //-----------------------trap_state_set_recompiled-----------------------------
1868 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1869 if (z) return trap_state | DS_RECOMPILE_BIT;
1870 else return trap_state & ~DS_RECOMPILE_BIT;
1871 }
1872 //---------------------------format_trap_state---------------------------------
1873 // This is used for debugging and diagnostics, including LogFile output.
1874 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1875 int trap_state) {
1876 DeoptReason reason = trap_state_reason(trap_state);
1877 bool recomp_flag = trap_state_is_recompiled(trap_state);
1878 // Re-encode the state from its decoded components.
1879 int decoded_state = 0;
1880 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1881 decoded_state = trap_state_add_reason(decoded_state, reason);
1882 if (recomp_flag)
1883 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1884 // If the state re-encodes properly, format it symbolically.
1885 // Because this routine is used for debugging and diagnostics,
1886 // be robust even if the state is a strange value.
1887 size_t len;
1888 if (decoded_state != trap_state) {
1889 // Random buggy state that doesn't decode??
1890 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1891 } else {
1892 len = jio_snprintf(buf, buflen, "%s%s",
1893 trap_reason_name(reason),
1894 recomp_flag ? " recompiled" : "");
1895 }
1896 return buf;
1897 }
1900 //--------------------------------statics--------------------------------------
1901 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1902 = Deoptimization::Action_reinterpret;
1903 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1904 // Note: Keep this in sync. with enum DeoptReason.
1905 "none",
1906 "null_check",
1907 "null_assert",
1908 "range_check",
1909 "class_check",
1910 "array_check",
1911 "intrinsic",
1912 "bimorphic",
1913 "unloaded",
1914 "uninitialized",
1915 "unreached",
1916 "unhandled",
1917 "constraint",
1918 "div0_check",
1919 "age",
1920 "predicate",
1921 "loop_limit_check",
1922 "speculate_class_check",
1923 "rtm_state_change",
1924 "unstable_if"
1925 };
1926 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1927 // Note: Keep this in sync. with enum DeoptAction.
1928 "none",
1929 "maybe_recompile",
1930 "reinterpret",
1931 "make_not_entrant",
1932 "make_not_compilable"
1933 };
1935 const char* Deoptimization::trap_reason_name(int reason) {
1936 if (reason == Reason_many) return "many";
1937 if ((uint)reason < Reason_LIMIT)
1938 return _trap_reason_name[reason];
1939 static char buf[20];
1940 sprintf(buf, "reason%d", reason);
1941 return buf;
1942 }
1943 const char* Deoptimization::trap_action_name(int action) {
1944 if ((uint)action < Action_LIMIT)
1945 return _trap_action_name[action];
1946 static char buf[20];
1947 sprintf(buf, "action%d", action);
1948 return buf;
1949 }
1951 // This is used for debugging and diagnostics, including LogFile output.
1952 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1953 int trap_request) {
1954 jint unloaded_class_index = trap_request_index(trap_request);
1955 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1956 const char* action = trap_action_name(trap_request_action(trap_request));
1957 size_t len;
1958 if (unloaded_class_index < 0) {
1959 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1960 reason, action);
1961 } else {
1962 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1963 reason, action, unloaded_class_index);
1964 }
1965 return buf;
1966 }
1968 juint Deoptimization::_deoptimization_hist
1969 [Deoptimization::Reason_LIMIT]
1970 [1 + Deoptimization::Action_LIMIT]
1971 [Deoptimization::BC_CASE_LIMIT]
1972 = {0};
1974 enum {
1975 LSB_BITS = 8,
1976 LSB_MASK = right_n_bits(LSB_BITS)
1977 };
1979 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1980 Bytecodes::Code bc) {
1981 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1982 assert(action >= 0 && action < Action_LIMIT, "oob");
1983 _deoptimization_hist[Reason_none][0][0] += 1; // total
1984 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1985 juint* cases = _deoptimization_hist[reason][1+action];
1986 juint* bc_counter_addr = NULL;
1987 juint bc_counter = 0;
1988 // Look for an unused counter, or an exact match to this BC.
1989 if (bc != Bytecodes::_illegal) {
1990 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1991 juint* counter_addr = &cases[bc_case];
1992 juint counter = *counter_addr;
1993 if ((counter == 0 && bc_counter_addr == NULL)
1994 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1995 // this counter is either free or is already devoted to this BC
1996 bc_counter_addr = counter_addr;
1997 bc_counter = counter | bc;
1998 }
1999 }
2000 }
2001 if (bc_counter_addr == NULL) {
2002 // Overflow, or no given bytecode.
2003 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2004 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
2005 }
2006 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2007 }
2009 jint Deoptimization::total_deoptimization_count() {
2010 return _deoptimization_hist[Reason_none][0][0];
2011 }
2013 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2014 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2015 return _deoptimization_hist[reason][0][0];
2016 }
2018 void Deoptimization::print_statistics() {
2019 juint total = total_deoptimization_count();
2020 juint account = total;
2021 if (total != 0) {
2022 ttyLocker ttyl;
2023 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
2024 tty->print_cr("Deoptimization traps recorded:");
2025 #define PRINT_STAT_LINE(name, r) \
2026 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2027 PRINT_STAT_LINE("total", total);
2028 // For each non-zero entry in the histogram, print the reason,
2029 // the action, and (if specifically known) the type of bytecode.
2030 for (int reason = 0; reason < Reason_LIMIT; reason++) {
2031 for (int action = 0; action < Action_LIMIT; action++) {
2032 juint* cases = _deoptimization_hist[reason][1+action];
2033 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2034 juint counter = cases[bc_case];
2035 if (counter != 0) {
2036 char name[1*K];
2037 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2038 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2039 bc = Bytecodes::_illegal;
2040 sprintf(name, "%s/%s/%s",
2041 trap_reason_name(reason),
2042 trap_action_name(action),
2043 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2044 juint r = counter >> LSB_BITS;
2045 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2046 account -= r;
2047 }
2048 }
2049 }
2050 }
2051 if (account != 0) {
2052 PRINT_STAT_LINE("unaccounted", account);
2053 }
2054 #undef PRINT_STAT_LINE
2055 if (xtty != NULL) xtty->tail("statistics");
2056 }
2057 }
2058 #else // COMPILER2 || SHARK
2061 // Stubs for C1 only system.
2062 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2063 return false;
2064 }
2066 const char* Deoptimization::trap_reason_name(int reason) {
2067 return "unknown";
2068 }
2070 void Deoptimization::print_statistics() {
2071 // no output
2072 }
2074 void
2075 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2076 // no udpate
2077 }
2079 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2080 return 0;
2081 }
2083 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2084 Bytecodes::Code bc) {
2085 // no update
2086 }
2088 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2089 int trap_state) {
2090 jio_snprintf(buf, buflen, "#%d", trap_state);
2091 return buf;
2092 }
2094 #endif // COMPILER2 || SHARK