Tue, 15 Mar 2011 06:35:10 -0700
7024234: 2/3 jvmti tests fail assert(!_oops_are_stale) failed: oops are stale on Win-AMD64
Summary: Move initialization of the '_instance' field to avoid race with ServiceThread start.
Reviewed-by: dholmes, kamg, never, dsamersoff, ysr, coleenp, acorn
1 /*
2 * Copyright (c) 1997, 2011, 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 #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/methodOop.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
85 # include "adfiles/ad_ppc.hpp"
86 #endif
87 #endif
89 bool DeoptimizationMarker::_is_active = false;
91 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
92 int caller_adjustment,
93 int number_of_frames,
94 intptr_t* frame_sizes,
95 address* frame_pcs,
96 BasicType return_type) {
97 _size_of_deoptimized_frame = size_of_deoptimized_frame;
98 _caller_adjustment = caller_adjustment;
99 _number_of_frames = number_of_frames;
100 _frame_sizes = frame_sizes;
101 _frame_pcs = frame_pcs;
102 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2);
103 _return_type = return_type;
104 _initial_fp = 0;
105 // PD (x86 only)
106 _counter_temp = 0;
107 _unpack_kind = 0;
108 _sender_sp_temp = 0;
110 _total_frame_sizes = size_of_frames();
111 }
114 Deoptimization::UnrollBlock::~UnrollBlock() {
115 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
116 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
117 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
118 }
121 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
122 assert(register_number < RegisterMap::reg_count, "checking register number");
123 return &_register_block[register_number * 2];
124 }
128 int Deoptimization::UnrollBlock::size_of_frames() const {
129 // Acount first for the adjustment of the initial frame
130 int result = _caller_adjustment;
131 for (int index = 0; index < number_of_frames(); index++) {
132 result += frame_sizes()[index];
133 }
134 return result;
135 }
138 void Deoptimization::UnrollBlock::print() {
139 ttyLocker ttyl;
140 tty->print_cr("UnrollBlock");
141 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
142 tty->print( " frame_sizes: ");
143 for (int index = 0; index < number_of_frames(); index++) {
144 tty->print("%d ", frame_sizes()[index]);
145 }
146 tty->cr();
147 }
150 // In order to make fetch_unroll_info work properly with escape
151 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
152 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
153 // of previously eliminated objects occurs in realloc_objects, which is
154 // called from the method fetch_unroll_info_helper below.
155 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
156 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
157 // but makes the entry a little slower. There is however a little dance we have to
158 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
160 // fetch_unroll_info() is called at the beginning of the deoptimization
161 // handler. Note this fact before we start generating temporary frames
162 // that can confuse an asynchronous stack walker. This counter is
163 // decremented at the end of unpack_frames().
164 thread->inc_in_deopt_handler();
166 return fetch_unroll_info_helper(thread);
167 JRT_END
170 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
171 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
173 // Note: there is a safepoint safety issue here. No matter whether we enter
174 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
175 // the vframeArray is created.
176 //
178 // Allocate our special deoptimization ResourceMark
179 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
180 assert(thread->deopt_mark() == NULL, "Pending deopt!");
181 thread->set_deopt_mark(dmark);
183 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
184 RegisterMap map(thread, true);
185 RegisterMap dummy_map(thread, false);
186 // Now get the deoptee with a valid map
187 frame deoptee = stub_frame.sender(&map);
188 // Set the deoptee nmethod
189 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
190 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
192 // Create a growable array of VFrames where each VFrame represents an inlined
193 // Java frame. This storage is allocated with the usual system arena.
194 assert(deoptee.is_compiled_frame(), "Wrong frame type");
195 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
196 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
197 while (!vf->is_top()) {
198 assert(vf->is_compiled_frame(), "Wrong frame type");
199 chunk->push(compiledVFrame::cast(vf));
200 vf = vf->sender();
201 }
202 assert(vf->is_compiled_frame(), "Wrong frame type");
203 chunk->push(compiledVFrame::cast(vf));
205 #ifdef COMPILER2
206 // Reallocate the non-escaping objects and restore their fields. Then
207 // relock objects if synchronization on them was eliminated.
208 if (DoEscapeAnalysis) {
209 if (EliminateAllocations) {
210 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
211 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
213 // The flag return_oop() indicates call sites which return oop
214 // in compiled code. Such sites include java method calls,
215 // runtime calls (for example, used to allocate new objects/arrays
216 // on slow code path) and any other calls generated in compiled code.
217 // It is not guaranteed that we can get such information here only
218 // by analyzing bytecode in deoptimized frames. This is why this flag
219 // is set during method compilation (see Compile::Process_OopMap_Node()).
220 bool save_oop_result = chunk->at(0)->scope()->return_oop();
221 Handle return_value;
222 if (save_oop_result) {
223 // Reallocation may trigger GC. If deoptimization happened on return from
224 // call which returns oop we need to save it since it is not in oopmap.
225 oop result = deoptee.saved_oop_result(&map);
226 assert(result == NULL || result->is_oop(), "must be oop");
227 return_value = Handle(thread, result);
228 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
229 if (TraceDeoptimization) {
230 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
231 }
232 }
233 bool reallocated = false;
234 if (objects != NULL) {
235 JRT_BLOCK
236 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
237 JRT_END
238 }
239 if (reallocated) {
240 reassign_fields(&deoptee, &map, objects);
241 #ifndef PRODUCT
242 if (TraceDeoptimization) {
243 ttyLocker ttyl;
244 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
245 print_objects(objects);
246 }
247 #endif
248 }
249 if (save_oop_result) {
250 // Restore result.
251 deoptee.set_saved_oop_result(&map, return_value());
252 }
253 }
254 if (EliminateLocks) {
255 #ifndef PRODUCT
256 bool first = true;
257 #endif
258 for (int i = 0; i < chunk->length(); i++) {
259 compiledVFrame* cvf = chunk->at(i);
260 assert (cvf->scope() != NULL,"expect only compiled java frames");
261 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
262 if (monitors->is_nonempty()) {
263 relock_objects(monitors, thread);
264 #ifndef PRODUCT
265 if (TraceDeoptimization) {
266 ttyLocker ttyl;
267 for (int j = 0; j < monitors->length(); j++) {
268 MonitorInfo* mi = monitors->at(j);
269 if (mi->eliminated()) {
270 if (first) {
271 first = false;
272 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
273 }
274 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
275 }
276 }
277 }
278 #endif
279 }
280 }
281 }
282 }
283 #endif // COMPILER2
284 // Ensure that no safepoint is taken after pointers have been stored
285 // in fields of rematerialized objects. If a safepoint occurs from here on
286 // out the java state residing in the vframeArray will be missed.
287 No_Safepoint_Verifier no_safepoint;
289 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
291 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
292 thread->set_vframe_array_head(array);
294 // Now that the vframeArray has been created if we have any deferred local writes
295 // added by jvmti then we can free up that structure as the data is now in the
296 // vframeArray
298 if (thread->deferred_locals() != NULL) {
299 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
300 int i = 0;
301 do {
302 // Because of inlining we could have multiple vframes for a single frame
303 // and several of the vframes could have deferred writes. Find them all.
304 if (list->at(i)->id() == array->original().id()) {
305 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
306 list->remove_at(i);
307 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
308 delete dlv;
309 } else {
310 i++;
311 }
312 } while ( i < list->length() );
313 if (list->length() == 0) {
314 thread->set_deferred_locals(NULL);
315 // free the list and elements back to C heap.
316 delete list;
317 }
319 }
321 #ifndef SHARK
322 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
323 CodeBlob* cb = stub_frame.cb();
324 // Verify we have the right vframeArray
325 assert(cb->frame_size() >= 0, "Unexpected frame size");
326 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
328 // If the deopt call site is a MethodHandle invoke call site we have
329 // to adjust the unpack_sp.
330 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
331 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
332 unpack_sp = deoptee.unextended_sp();
334 #ifdef ASSERT
335 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
336 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp);
337 #endif
338 #else
339 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
340 #endif // !SHARK
342 // This is a guarantee instead of an assert because if vframe doesn't match
343 // we will unpack the wrong deoptimized frame and wind up in strange places
344 // where it will be very difficult to figure out what went wrong. Better
345 // to die an early death here than some very obscure death later when the
346 // trail is cold.
347 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
348 // in that it will fail to detect a problem when there is one. This needs
349 // more work in tiger timeframe.
350 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
352 int number_of_frames = array->frames();
354 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
355 // virtual activation, which is the reverse of the elements in the vframes array.
356 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames);
357 // +1 because we always have an interpreter return address for the final slot.
358 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1);
359 int callee_parameters = 0;
360 int callee_locals = 0;
361 int popframe_extra_args = 0;
362 // Create an interpreter return address for the stub to use as its return
363 // address so the skeletal frames are perfectly walkable
364 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
366 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
367 // activation be put back on the expression stack of the caller for reexecution
368 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
369 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
370 }
372 //
373 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
374 // frame_sizes/frame_pcs[1] next oldest frame (int)
375 // frame_sizes/frame_pcs[n] youngest frame (int)
376 //
377 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
378 // owns the space for the return address to it's caller). Confusing ain't it.
379 //
380 // The vframe array can address vframes with indices running from
381 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
382 // When we create the skeletal frames we need the oldest frame to be in the zero slot
383 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
384 // so things look a little strange in this loop.
385 //
386 for (int index = 0; index < array->frames(); index++ ) {
387 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
388 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
389 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
390 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
391 callee_locals,
392 index == 0,
393 popframe_extra_args);
394 // This pc doesn't have to be perfect just good enough to identify the frame
395 // as interpreted so the skeleton frame will be walkable
396 // The correct pc will be set when the skeleton frame is completely filled out
397 // The final pc we store in the loop is wrong and will be overwritten below
398 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
400 callee_parameters = array->element(index)->method()->size_of_parameters();
401 callee_locals = array->element(index)->method()->max_locals();
402 popframe_extra_args = 0;
403 }
405 // Compute whether the root vframe returns a float or double value.
406 BasicType return_type;
407 {
408 HandleMark hm;
409 methodHandle method(thread, array->element(0)->method());
410 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
411 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
412 }
414 // Compute information for handling adapters and adjusting the frame size of the caller.
415 int caller_adjustment = 0;
417 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
418 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
419 // than simply use array->sender.pc(). This requires us to walk the current set of frames
420 //
421 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
422 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
424 // Compute the amount the oldest interpreter frame will have to adjust
425 // its caller's stack by. If the caller is a compiled frame then
426 // we pretend that the callee has no parameters so that the
427 // extension counts for the full amount of locals and not just
428 // locals-parms. This is because without a c2i adapter the parm
429 // area as created by the compiled frame will not be usable by
430 // the interpreter. (Depending on the calling convention there
431 // may not even be enough space).
433 // QQQ I'd rather see this pushed down into last_frame_adjust
434 // and have it take the sender (aka caller).
436 if (deopt_sender.is_compiled_frame()) {
437 caller_adjustment = last_frame_adjust(0, callee_locals);
438 } else if (callee_locals > callee_parameters) {
439 // The caller frame may need extending to accommodate
440 // non-parameter locals of the first unpacked interpreted frame.
441 // Compute that adjustment.
442 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
443 }
446 // If the sender is deoptimized the we must retrieve the address of the handler
447 // since the frame will "magically" show the original pc before the deopt
448 // and we'd undo the deopt.
450 frame_pcs[0] = deopt_sender.raw_pc();
452 #ifndef SHARK
453 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
454 #endif // SHARK
456 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
457 caller_adjustment * BytesPerWord,
458 number_of_frames,
459 frame_sizes,
460 frame_pcs,
461 return_type);
462 // On some platforms, we need a way to pass fp to the unpacking code
463 // so the skeletal frames come out correct.
464 info->set_initial_fp((intptr_t) array->sender().fp());
466 if (array->frames() > 1) {
467 if (VerifyStack && TraceDeoptimization) {
468 tty->print_cr("Deoptimizing method containing inlining");
469 }
470 }
472 array->set_unroll_block(info);
473 return info;
474 }
476 // Called to cleanup deoptimization data structures in normal case
477 // after unpacking to stack and when stack overflow error occurs
478 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
479 vframeArray *array) {
481 // Get array if coming from exception
482 if (array == NULL) {
483 array = thread->vframe_array_head();
484 }
485 thread->set_vframe_array_head(NULL);
487 // Free the previous UnrollBlock
488 vframeArray* old_array = thread->vframe_array_last();
489 thread->set_vframe_array_last(array);
491 if (old_array != NULL) {
492 UnrollBlock* old_info = old_array->unroll_block();
493 old_array->set_unroll_block(NULL);
494 delete old_info;
495 delete old_array;
496 }
498 // Deallocate any resource creating in this routine and any ResourceObjs allocated
499 // inside the vframeArray (StackValueCollections)
501 delete thread->deopt_mark();
502 thread->set_deopt_mark(NULL);
503 thread->set_deopt_nmethod(NULL);
506 if (JvmtiExport::can_pop_frame()) {
507 #ifndef CC_INTERP
508 // Regardless of whether we entered this routine with the pending
509 // popframe condition bit set, we should always clear it now
510 thread->clear_popframe_condition();
511 #else
512 // C++ interpeter will clear has_pending_popframe when it enters
513 // with method_resume. For deopt_resume2 we clear it now.
514 if (thread->popframe_forcing_deopt_reexecution())
515 thread->clear_popframe_condition();
516 #endif /* CC_INTERP */
517 }
519 // unpack_frames() is called at the end of the deoptimization handler
520 // and (in C2) at the end of the uncommon trap handler. Note this fact
521 // so that an asynchronous stack walker can work again. This counter is
522 // incremented at the beginning of fetch_unroll_info() and (in C2) at
523 // the beginning of uncommon_trap().
524 thread->dec_in_deopt_handler();
525 }
528 // Return BasicType of value being returned
529 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
531 // We are already active int he special DeoptResourceMark any ResourceObj's we
532 // allocate will be freed at the end of the routine.
534 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
535 // but makes the entry a little slower. There is however a little dance we have to
536 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
537 ResetNoHandleMark rnhm; // No-op in release/product versions
538 HandleMark hm;
540 frame stub_frame = thread->last_frame();
542 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
543 // must point to the vframeArray for the unpack frame.
544 vframeArray* array = thread->vframe_array_head();
546 #ifndef PRODUCT
547 if (TraceDeoptimization) {
548 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
549 }
550 #endif
552 UnrollBlock* info = array->unroll_block();
554 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
555 array->unpack_to_stack(stub_frame, exec_mode);
557 BasicType bt = info->return_type();
559 // If we have an exception pending, claim that the return type is an oop
560 // so the deopt_blob does not overwrite the exception_oop.
562 if (exec_mode == Unpack_exception)
563 bt = T_OBJECT;
565 // Cleanup thread deopt data
566 cleanup_deopt_info(thread, array);
568 #ifndef PRODUCT
569 if (VerifyStack) {
570 ResourceMark res_mark;
572 // Verify that the just-unpacked frames match the interpreter's
573 // notions of expression stack and locals
574 vframeArray* cur_array = thread->vframe_array_last();
575 RegisterMap rm(thread, false);
576 rm.set_include_argument_oops(false);
577 bool is_top_frame = true;
578 int callee_size_of_parameters = 0;
579 int callee_max_locals = 0;
580 for (int i = 0; i < cur_array->frames(); i++) {
581 vframeArrayElement* el = cur_array->element(i);
582 frame* iframe = el->iframe();
583 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
585 // Get the oop map for this bci
586 InterpreterOopMap mask;
587 int cur_invoke_parameter_size = 0;
588 bool try_next_mask = false;
589 int next_mask_expression_stack_size = -1;
590 int top_frame_expression_stack_adjustment = 0;
591 methodHandle mh(thread, iframe->interpreter_frame_method());
592 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
593 BytecodeStream str(mh);
594 str.set_start(iframe->interpreter_frame_bci());
595 int max_bci = mh->code_size();
596 // Get to the next bytecode if possible
597 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
598 // Check to see if we can grab the number of outgoing arguments
599 // at an uncommon trap for an invoke (where the compiler
600 // generates debug info before the invoke has executed)
601 Bytecodes::Code cur_code = str.next();
602 if (cur_code == Bytecodes::_invokevirtual ||
603 cur_code == Bytecodes::_invokespecial ||
604 cur_code == Bytecodes::_invokestatic ||
605 cur_code == Bytecodes::_invokeinterface) {
606 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
607 Symbol* signature = invoke.signature();
608 ArgumentSizeComputer asc(signature);
609 cur_invoke_parameter_size = asc.size();
610 if (cur_code != Bytecodes::_invokestatic) {
611 // Add in receiver
612 ++cur_invoke_parameter_size;
613 }
614 }
615 if (str.bci() < max_bci) {
616 Bytecodes::Code bc = str.next();
617 if (bc >= 0) {
618 // The interpreter oop map generator reports results before
619 // the current bytecode has executed except in the case of
620 // calls. It seems to be hard to tell whether the compiler
621 // has emitted debug information matching the "state before"
622 // a given bytecode or the state after, so we try both
623 switch (cur_code) {
624 case Bytecodes::_invokevirtual:
625 case Bytecodes::_invokespecial:
626 case Bytecodes::_invokestatic:
627 case Bytecodes::_invokeinterface:
628 case Bytecodes::_athrow:
629 break;
630 default: {
631 InterpreterOopMap next_mask;
632 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
633 next_mask_expression_stack_size = next_mask.expression_stack_size();
634 // Need to subtract off the size of the result type of
635 // the bytecode because this is not described in the
636 // debug info but returned to the interpreter in the TOS
637 // caching register
638 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
639 if (bytecode_result_type != T_ILLEGAL) {
640 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
641 }
642 assert(top_frame_expression_stack_adjustment >= 0, "");
643 try_next_mask = true;
644 break;
645 }
646 }
647 }
648 }
650 // Verify stack depth and oops in frame
651 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
652 if (!(
653 /* SPARC */
654 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
655 /* x86 */
656 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
657 (try_next_mask &&
658 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
659 top_frame_expression_stack_adjustment))) ||
660 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
661 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
662 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
663 )) {
664 ttyLocker ttyl;
666 // Print out some information that will help us debug the problem
667 tty->print_cr("Wrong number of expression stack elements during deoptimization");
668 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
669 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
670 iframe->interpreter_frame_expression_stack_size());
671 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
672 tty->print_cr(" try_next_mask = %d", try_next_mask);
673 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
674 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
675 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
676 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
677 tty->print_cr(" exec_mode = %d", exec_mode);
678 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
679 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
680 tty->print_cr(" Interpreted frames:");
681 for (int k = 0; k < cur_array->frames(); k++) {
682 vframeArrayElement* el = cur_array->element(k);
683 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
684 }
685 cur_array->print_on_2(tty);
686 guarantee(false, "wrong number of expression stack elements during deopt");
687 }
688 VerifyOopClosure verify;
689 iframe->oops_interpreted_do(&verify, &rm, false);
690 callee_size_of_parameters = mh->size_of_parameters();
691 callee_max_locals = mh->max_locals();
692 is_top_frame = false;
693 }
694 }
695 #endif /* !PRODUCT */
698 return bt;
699 JRT_END
702 int Deoptimization::deoptimize_dependents() {
703 Threads::deoptimized_wrt_marked_nmethods();
704 return 0;
705 }
708 #ifdef COMPILER2
709 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
710 Handle pending_exception(thread->pending_exception());
711 const char* exception_file = thread->exception_file();
712 int exception_line = thread->exception_line();
713 thread->clear_pending_exception();
715 for (int i = 0; i < objects->length(); i++) {
716 assert(objects->at(i)->is_object(), "invalid debug information");
717 ObjectValue* sv = (ObjectValue*) objects->at(i);
719 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
720 oop obj = NULL;
722 if (k->oop_is_instance()) {
723 instanceKlass* ik = instanceKlass::cast(k());
724 obj = ik->allocate_instance(CHECK_(false));
725 } else if (k->oop_is_typeArray()) {
726 typeArrayKlass* ak = typeArrayKlass::cast(k());
727 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
728 int len = sv->field_size() / type2size[ak->element_type()];
729 obj = ak->allocate(len, CHECK_(false));
730 } else if (k->oop_is_objArray()) {
731 objArrayKlass* ak = objArrayKlass::cast(k());
732 obj = ak->allocate(sv->field_size(), CHECK_(false));
733 }
735 assert(obj != NULL, "allocation failed");
736 assert(sv->value().is_null(), "redundant reallocation");
737 sv->set_value(obj);
738 }
740 if (pending_exception.not_null()) {
741 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
742 }
744 return true;
745 }
747 // This assumes that the fields are stored in ObjectValue in the same order
748 // they are yielded by do_nonstatic_fields.
749 class FieldReassigner: public FieldClosure {
750 frame* _fr;
751 RegisterMap* _reg_map;
752 ObjectValue* _sv;
753 instanceKlass* _ik;
754 oop _obj;
756 int _i;
757 public:
758 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
759 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
761 int i() const { return _i; }
764 void do_field(fieldDescriptor* fd) {
765 intptr_t val;
766 StackValue* value =
767 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
768 int offset = fd->offset();
769 switch (fd->field_type()) {
770 case T_OBJECT: case T_ARRAY:
771 assert(value->type() == T_OBJECT, "Agreement.");
772 _obj->obj_field_put(offset, value->get_obj()());
773 break;
775 case T_LONG: case T_DOUBLE: {
776 assert(value->type() == T_INT, "Agreement.");
777 StackValue* low =
778 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
779 #ifdef _LP64
780 jlong res = (jlong)low->get_int();
781 #else
782 #ifdef SPARC
783 // For SPARC we have to swap high and low words.
784 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
785 #else
786 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
787 #endif //SPARC
788 #endif
789 _obj->long_field_put(offset, res);
790 break;
791 }
792 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
793 case T_INT: case T_FLOAT: // 4 bytes.
794 assert(value->type() == T_INT, "Agreement.");
795 val = value->get_int();
796 _obj->int_field_put(offset, (jint)*((jint*)&val));
797 break;
799 case T_SHORT: case T_CHAR: // 2 bytes
800 assert(value->type() == T_INT, "Agreement.");
801 val = value->get_int();
802 _obj->short_field_put(offset, (jshort)*((jint*)&val));
803 break;
805 case T_BOOLEAN: case T_BYTE: // 1 byte
806 assert(value->type() == T_INT, "Agreement.");
807 val = value->get_int();
808 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
809 break;
811 default:
812 ShouldNotReachHere();
813 }
814 _i++;
815 }
816 };
818 // restore elements of an eliminated type array
819 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
820 int index = 0;
821 intptr_t val;
823 for (int i = 0; i < sv->field_size(); i++) {
824 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
825 switch(type) {
826 case T_LONG: case T_DOUBLE: {
827 assert(value->type() == T_INT, "Agreement.");
828 StackValue* low =
829 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
830 #ifdef _LP64
831 jlong res = (jlong)low->get_int();
832 #else
833 #ifdef SPARC
834 // For SPARC we have to swap high and low words.
835 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
836 #else
837 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
838 #endif //SPARC
839 #endif
840 obj->long_at_put(index, res);
841 break;
842 }
844 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
845 case T_INT: case T_FLOAT: // 4 bytes.
846 assert(value->type() == T_INT, "Agreement.");
847 val = value->get_int();
848 obj->int_at_put(index, (jint)*((jint*)&val));
849 break;
851 case T_SHORT: case T_CHAR: // 2 bytes
852 assert(value->type() == T_INT, "Agreement.");
853 val = value->get_int();
854 obj->short_at_put(index, (jshort)*((jint*)&val));
855 break;
857 case T_BOOLEAN: case T_BYTE: // 1 byte
858 assert(value->type() == T_INT, "Agreement.");
859 val = value->get_int();
860 obj->bool_at_put(index, (jboolean)*((jint*)&val));
861 break;
863 default:
864 ShouldNotReachHere();
865 }
866 index++;
867 }
868 }
871 // restore fields of an eliminated object array
872 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
873 for (int i = 0; i < sv->field_size(); i++) {
874 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
875 assert(value->type() == T_OBJECT, "object element expected");
876 obj->obj_at_put(i, value->get_obj()());
877 }
878 }
881 // restore fields of all eliminated objects and arrays
882 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
883 for (int i = 0; i < objects->length(); i++) {
884 ObjectValue* sv = (ObjectValue*) objects->at(i);
885 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
886 Handle obj = sv->value();
887 assert(obj.not_null(), "reallocation was missed");
889 if (k->oop_is_instance()) {
890 instanceKlass* ik = instanceKlass::cast(k());
891 FieldReassigner reassign(fr, reg_map, sv, obj());
892 ik->do_nonstatic_fields(&reassign);
893 } else if (k->oop_is_typeArray()) {
894 typeArrayKlass* ak = typeArrayKlass::cast(k());
895 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
896 } else if (k->oop_is_objArray()) {
897 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
898 }
899 }
900 }
903 // relock objects for which synchronization was eliminated
904 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
905 for (int i = 0; i < monitors->length(); i++) {
906 MonitorInfo* mon_info = monitors->at(i);
907 if (mon_info->eliminated()) {
908 assert(mon_info->owner() != NULL, "reallocation was missed");
909 Handle obj = Handle(mon_info->owner());
910 markOop mark = obj->mark();
911 if (UseBiasedLocking && mark->has_bias_pattern()) {
912 // New allocated objects may have the mark set to anonymously biased.
913 // Also the deoptimized method may called methods with synchronization
914 // where the thread-local object is bias locked to the current thread.
915 assert(mark->is_biased_anonymously() ||
916 mark->biased_locker() == thread, "should be locked to current thread");
917 // Reset mark word to unbiased prototype.
918 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
919 obj->set_mark(unbiased_prototype);
920 }
921 BasicLock* lock = mon_info->lock();
922 ObjectSynchronizer::slow_enter(obj, lock, thread);
923 }
924 assert(mon_info->owner()->is_locked(), "object must be locked now");
925 }
926 }
929 #ifndef PRODUCT
930 // print information about reallocated objects
931 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
932 fieldDescriptor fd;
934 for (int i = 0; i < objects->length(); i++) {
935 ObjectValue* sv = (ObjectValue*) objects->at(i);
936 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
937 Handle obj = sv->value();
939 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
940 k->as_klassOop()->print_value();
941 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
942 tty->cr();
944 if (Verbose) {
945 k->oop_print_on(obj(), tty);
946 }
947 }
948 }
949 #endif
950 #endif // COMPILER2
952 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
954 #ifndef PRODUCT
955 if (TraceDeoptimization) {
956 ttyLocker ttyl;
957 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
958 fr.print_on(tty);
959 tty->print_cr(" Virtual frames (innermost first):");
960 for (int index = 0; index < chunk->length(); index++) {
961 compiledVFrame* vf = chunk->at(index);
962 tty->print(" %2d - ", index);
963 vf->print_value();
964 int bci = chunk->at(index)->raw_bci();
965 const char* code_name;
966 if (bci == SynchronizationEntryBCI) {
967 code_name = "sync entry";
968 } else {
969 Bytecodes::Code code = vf->method()->code_at(bci);
970 code_name = Bytecodes::name(code);
971 }
972 tty->print(" - %s", code_name);
973 tty->print_cr(" @ bci %d ", bci);
974 if (Verbose) {
975 vf->print();
976 tty->cr();
977 }
978 }
979 }
980 #endif
982 // Register map for next frame (used for stack crawl). We capture
983 // the state of the deopt'ing frame's caller. Thus if we need to
984 // stuff a C2I adapter we can properly fill in the callee-save
985 // register locations.
986 frame caller = fr.sender(reg_map);
987 int frame_size = caller.sp() - fr.sp();
989 frame sender = caller;
991 // Since the Java thread being deoptimized will eventually adjust it's own stack,
992 // the vframeArray containing the unpacking information is allocated in the C heap.
993 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
994 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
996 // Compare the vframeArray to the collected vframes
997 assert(array->structural_compare(thread, chunk), "just checking");
998 Events::log("# vframes = %d", (intptr_t)chunk->length());
1000 #ifndef PRODUCT
1001 if (TraceDeoptimization) {
1002 ttyLocker ttyl;
1003 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1004 }
1005 #endif // PRODUCT
1007 return array;
1008 }
1011 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1012 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1013 for (int i = 0; i < monitors->length(); i++) {
1014 MonitorInfo* mon_info = monitors->at(i);
1015 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1016 objects_to_revoke->append(Handle(mon_info->owner()));
1017 }
1018 }
1019 }
1022 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1023 if (!UseBiasedLocking) {
1024 return;
1025 }
1027 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1029 // Unfortunately we don't have a RegisterMap available in most of
1030 // the places we want to call this routine so we need to walk the
1031 // stack again to update the register map.
1032 if (map == NULL || !map->update_map()) {
1033 StackFrameStream sfs(thread, true);
1034 bool found = false;
1035 while (!found && !sfs.is_done()) {
1036 frame* cur = sfs.current();
1037 sfs.next();
1038 found = cur->id() == fr.id();
1039 }
1040 assert(found, "frame to be deoptimized not found on target thread's stack");
1041 map = sfs.register_map();
1042 }
1044 vframe* vf = vframe::new_vframe(&fr, map, thread);
1045 compiledVFrame* cvf = compiledVFrame::cast(vf);
1046 // Revoke monitors' biases in all scopes
1047 while (!cvf->is_top()) {
1048 collect_monitors(cvf, objects_to_revoke);
1049 cvf = compiledVFrame::cast(cvf->sender());
1050 }
1051 collect_monitors(cvf, objects_to_revoke);
1053 if (SafepointSynchronize::is_at_safepoint()) {
1054 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1055 } else {
1056 BiasedLocking::revoke(objects_to_revoke);
1057 }
1058 }
1061 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1062 if (!UseBiasedLocking) {
1063 return;
1064 }
1066 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1067 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1068 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1069 if (jt->has_last_Java_frame()) {
1070 StackFrameStream sfs(jt, true);
1071 while (!sfs.is_done()) {
1072 frame* cur = sfs.current();
1073 if (cb->contains(cur->pc())) {
1074 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1075 compiledVFrame* cvf = compiledVFrame::cast(vf);
1076 // Revoke monitors' biases in all scopes
1077 while (!cvf->is_top()) {
1078 collect_monitors(cvf, objects_to_revoke);
1079 cvf = compiledVFrame::cast(cvf->sender());
1080 }
1081 collect_monitors(cvf, objects_to_revoke);
1082 }
1083 sfs.next();
1084 }
1085 }
1086 }
1087 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1088 }
1091 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1092 assert(fr.can_be_deoptimized(), "checking frame type");
1094 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1096 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id());
1098 // Patch the nmethod so that when execution returns to it we will
1099 // deopt the execution state and return to the interpreter.
1100 fr.deoptimize(thread);
1101 }
1103 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1104 // Deoptimize only if the frame comes from compile code.
1105 // Do not deoptimize the frame which is already patched
1106 // during the execution of the loops below.
1107 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1108 return;
1109 }
1110 ResourceMark rm;
1111 DeoptimizationMarker dm;
1112 if (UseBiasedLocking) {
1113 revoke_biases_of_monitors(thread, fr, map);
1114 }
1115 deoptimize_single_frame(thread, fr);
1117 }
1120 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1121 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1122 "can only deoptimize other thread at a safepoint");
1123 // Compute frame and register map based on thread and sp.
1124 RegisterMap reg_map(thread, UseBiasedLocking);
1125 frame fr = thread->last_frame();
1126 while (fr.id() != id) {
1127 fr = fr.sender(®_map);
1128 }
1129 deoptimize(thread, fr, ®_map);
1130 }
1133 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1134 if (thread == Thread::current()) {
1135 Deoptimization::deoptimize_frame_internal(thread, id);
1136 } else {
1137 VM_DeoptimizeFrame deopt(thread, id);
1138 VMThread::execute(&deopt);
1139 }
1140 }
1143 // JVMTI PopFrame support
1144 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1145 {
1146 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1147 }
1148 JRT_END
1151 #if defined(COMPILER2) || defined(SHARK)
1152 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1153 // in case of an unresolved klass entry, load the class.
1154 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1155 klassOop tk = constant_pool->klass_at(index, CHECK);
1156 return;
1157 }
1159 if (!constant_pool->tag_at(index).is_symbol()) return;
1161 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
1162 Symbol* symbol = constant_pool->symbol_at(index);
1164 // class name?
1165 if (symbol->byte_at(0) != '(') {
1166 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1167 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1168 return;
1169 }
1171 // then it must be a signature!
1172 ResourceMark rm(THREAD);
1173 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1174 if (ss.is_object()) {
1175 Symbol* class_name = ss.as_symbol(CHECK);
1176 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1177 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1178 }
1179 }
1180 }
1183 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1184 EXCEPTION_MARK;
1185 load_class_by_index(constant_pool, index, THREAD);
1186 if (HAS_PENDING_EXCEPTION) {
1187 // Exception happened during classloading. We ignore the exception here, since it
1188 // is going to be rethrown since the current activation is going to be deoptimzied and
1189 // the interpreter will re-execute the bytecode.
1190 CLEAR_PENDING_EXCEPTION;
1191 }
1192 }
1194 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1195 HandleMark hm;
1197 // uncommon_trap() is called at the beginning of the uncommon trap
1198 // handler. Note this fact before we start generating temporary frames
1199 // that can confuse an asynchronous stack walker. This counter is
1200 // decremented at the end of unpack_frames().
1201 thread->inc_in_deopt_handler();
1203 // We need to update the map if we have biased locking.
1204 RegisterMap reg_map(thread, UseBiasedLocking);
1205 frame stub_frame = thread->last_frame();
1206 frame fr = stub_frame.sender(®_map);
1207 // Make sure the calling nmethod is not getting deoptimized and removed
1208 // before we are done with it.
1209 nmethodLocker nl(fr.pc());
1211 {
1212 ResourceMark rm;
1214 // Revoke biases of any monitors in the frame to ensure we can migrate them
1215 revoke_biases_of_monitors(thread, fr, ®_map);
1217 DeoptReason reason = trap_request_reason(trap_request);
1218 DeoptAction action = trap_request_action(trap_request);
1219 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1221 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request);
1222 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1223 compiledVFrame* cvf = compiledVFrame::cast(vf);
1225 nmethod* nm = cvf->code();
1227 ScopeDesc* trap_scope = cvf->scope();
1228 methodHandle trap_method = trap_scope->method();
1229 int trap_bci = trap_scope->bci();
1230 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1232 // Record this event in the histogram.
1233 gather_statistics(reason, action, trap_bc);
1235 // Ensure that we can record deopt. history:
1236 bool create_if_missing = ProfileTraps;
1238 methodDataHandle trap_mdo
1239 (THREAD, get_method_data(thread, trap_method, create_if_missing));
1241 // Print a bunch of diagnostics, if requested.
1242 if (TraceDeoptimization || LogCompilation) {
1243 ResourceMark rm;
1244 ttyLocker ttyl;
1245 char buf[100];
1246 if (xtty != NULL) {
1247 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1248 os::current_thread_id(),
1249 format_trap_request(buf, sizeof(buf), trap_request));
1250 nm->log_identity(xtty);
1251 }
1252 Symbol* class_name = NULL;
1253 bool unresolved = false;
1254 if (unloaded_class_index >= 0) {
1255 constantPoolHandle constants (THREAD, trap_method->constants());
1256 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1257 class_name = constants->klass_name_at(unloaded_class_index);
1258 unresolved = true;
1259 if (xtty != NULL)
1260 xtty->print(" unresolved='1'");
1261 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1262 class_name = constants->symbol_at(unloaded_class_index);
1263 }
1264 if (xtty != NULL)
1265 xtty->name(class_name);
1266 }
1267 if (xtty != NULL && trap_mdo.not_null()) {
1268 // Dump the relevant MDO state.
1269 // This is the deopt count for the current reason, any previous
1270 // reasons or recompiles seen at this point.
1271 int dcnt = trap_mdo->trap_count(reason);
1272 if (dcnt != 0)
1273 xtty->print(" count='%d'", dcnt);
1274 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1275 int dos = (pdata == NULL)? 0: pdata->trap_state();
1276 if (dos != 0) {
1277 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1278 if (trap_state_is_recompiled(dos)) {
1279 int recnt2 = trap_mdo->overflow_recompile_count();
1280 if (recnt2 != 0)
1281 xtty->print(" recompiles2='%d'", recnt2);
1282 }
1283 }
1284 }
1285 if (xtty != NULL) {
1286 xtty->stamp();
1287 xtty->end_head();
1288 }
1289 if (TraceDeoptimization) { // make noise on the tty
1290 tty->print("Uncommon trap occurred in");
1291 nm->method()->print_short_name(tty);
1292 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
1293 fr.pc(),
1294 (int) os::current_thread_id(),
1295 trap_reason_name(reason),
1296 trap_action_name(action),
1297 unloaded_class_index);
1298 if (class_name != NULL) {
1299 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1300 class_name->print_symbol_on(tty);
1301 }
1302 tty->cr();
1303 }
1304 if (xtty != NULL) {
1305 // Log the precise location of the trap.
1306 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1307 xtty->begin_elem("jvms bci='%d'", sd->bci());
1308 xtty->method(sd->method());
1309 xtty->end_elem();
1310 if (sd->is_top()) break;
1311 }
1312 xtty->tail("uncommon_trap");
1313 }
1314 }
1315 // (End diagnostic printout.)
1317 // Load class if necessary
1318 if (unloaded_class_index >= 0) {
1319 constantPoolHandle constants(THREAD, trap_method->constants());
1320 load_class_by_index(constants, unloaded_class_index);
1321 }
1323 // Flush the nmethod if necessary and desirable.
1324 //
1325 // We need to avoid situations where we are re-flushing the nmethod
1326 // because of a hot deoptimization site. Repeated flushes at the same
1327 // point need to be detected by the compiler and avoided. If the compiler
1328 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1329 // module must take measures to avoid an infinite cycle of recompilation
1330 // and deoptimization. There are several such measures:
1331 //
1332 // 1. If a recompilation is ordered a second time at some site X
1333 // and for the same reason R, the action is adjusted to 'reinterpret',
1334 // to give the interpreter time to exercise the method more thoroughly.
1335 // If this happens, the method's overflow_recompile_count is incremented.
1336 //
1337 // 2. If the compiler fails to reduce the deoptimization rate, then
1338 // the method's overflow_recompile_count will begin to exceed the set
1339 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1340 // is adjusted to 'make_not_compilable', and the method is abandoned
1341 // to the interpreter. This is a performance hit for hot methods,
1342 // but is better than a disastrous infinite cycle of recompilations.
1343 // (Actually, only the method containing the site X is abandoned.)
1344 //
1345 // 3. In parallel with the previous measures, if the total number of
1346 // recompilations of a method exceeds the much larger set limit
1347 // PerMethodRecompilationCutoff, the method is abandoned.
1348 // This should only happen if the method is very large and has
1349 // many "lukewarm" deoptimizations. The code which enforces this
1350 // limit is elsewhere (class nmethod, class methodOopDesc).
1351 //
1352 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1353 // to recompile at each bytecode independently of the per-BCI cutoff.
1354 //
1355 // The decision to update code is up to the compiler, and is encoded
1356 // in the Action_xxx code. If the compiler requests Action_none
1357 // no trap state is changed, no compiled code is changed, and the
1358 // computation suffers along in the interpreter.
1359 //
1360 // The other action codes specify various tactics for decompilation
1361 // and recompilation. Action_maybe_recompile is the loosest, and
1362 // allows the compiled code to stay around until enough traps are seen,
1363 // and until the compiler gets around to recompiling the trapping method.
1364 //
1365 // The other actions cause immediate removal of the present code.
1367 bool update_trap_state = true;
1368 bool make_not_entrant = false;
1369 bool make_not_compilable = false;
1370 bool reprofile = false;
1371 switch (action) {
1372 case Action_none:
1373 // Keep the old code.
1374 update_trap_state = false;
1375 break;
1376 case Action_maybe_recompile:
1377 // Do not need to invalidate the present code, but we can
1378 // initiate another
1379 // Start compiler without (necessarily) invalidating the nmethod.
1380 // The system will tolerate the old code, but new code should be
1381 // generated when possible.
1382 break;
1383 case Action_reinterpret:
1384 // Go back into the interpreter for a while, and then consider
1385 // recompiling form scratch.
1386 make_not_entrant = true;
1387 // Reset invocation counter for outer most method.
1388 // This will allow the interpreter to exercise the bytecodes
1389 // for a while before recompiling.
1390 // By contrast, Action_make_not_entrant is immediate.
1391 //
1392 // Note that the compiler will track null_check, null_assert,
1393 // range_check, and class_check events and log them as if they
1394 // had been traps taken from compiled code. This will update
1395 // the MDO trap history so that the next compilation will
1396 // properly detect hot trap sites.
1397 reprofile = true;
1398 break;
1399 case Action_make_not_entrant:
1400 // Request immediate recompilation, and get rid of the old code.
1401 // Make them not entrant, so next time they are called they get
1402 // recompiled. Unloaded classes are loaded now so recompile before next
1403 // time they are called. Same for uninitialized. The interpreter will
1404 // link the missing class, if any.
1405 make_not_entrant = true;
1406 break;
1407 case Action_make_not_compilable:
1408 // Give up on compiling this method at all.
1409 make_not_entrant = true;
1410 make_not_compilable = true;
1411 break;
1412 default:
1413 ShouldNotReachHere();
1414 }
1416 // Setting +ProfileTraps fixes the following, on all platforms:
1417 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1418 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1419 // recompile relies on a methodDataOop to record heroic opt failures.
1421 // Whether the interpreter is producing MDO data or not, we also need
1422 // to use the MDO to detect hot deoptimization points and control
1423 // aggressive optimization.
1424 bool inc_recompile_count = false;
1425 ProfileData* pdata = NULL;
1426 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
1427 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
1428 uint this_trap_count = 0;
1429 bool maybe_prior_trap = false;
1430 bool maybe_prior_recompile = false;
1431 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1432 //outputs:
1433 this_trap_count,
1434 maybe_prior_trap,
1435 maybe_prior_recompile);
1436 // Because the interpreter also counts null, div0, range, and class
1437 // checks, these traps from compiled code are double-counted.
1438 // This is harmless; it just means that the PerXTrapLimit values
1439 // are in effect a little smaller than they look.
1441 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1442 if (per_bc_reason != Reason_none) {
1443 // Now take action based on the partially known per-BCI history.
1444 if (maybe_prior_trap
1445 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1446 // If there are too many traps at this BCI, force a recompile.
1447 // This will allow the compiler to see the limit overflow, and
1448 // take corrective action, if possible. The compiler generally
1449 // does not use the exact PerBytecodeTrapLimit value, but instead
1450 // changes its tactics if it sees any traps at all. This provides
1451 // a little hysteresis, delaying a recompile until a trap happens
1452 // several times.
1453 //
1454 // Actually, since there is only one bit of counter per BCI,
1455 // the possible per-BCI counts are {0,1,(per-method count)}.
1456 // This produces accurate results if in fact there is only
1457 // one hot trap site, but begins to get fuzzy if there are
1458 // many sites. For example, if there are ten sites each
1459 // trapping two or more times, they each get the blame for
1460 // all of their traps.
1461 make_not_entrant = true;
1462 }
1464 // Detect repeated recompilation at the same BCI, and enforce a limit.
1465 if (make_not_entrant && maybe_prior_recompile) {
1466 // More than one recompile at this point.
1467 inc_recompile_count = maybe_prior_trap;
1468 }
1469 } else {
1470 // For reasons which are not recorded per-bytecode, we simply
1471 // force recompiles unconditionally.
1472 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1473 make_not_entrant = true;
1474 }
1476 // Go back to the compiler if there are too many traps in this method.
1477 if (this_trap_count >= (uint)PerMethodTrapLimit) {
1478 // If there are too many traps in this method, force a recompile.
1479 // This will allow the compiler to see the limit overflow, and
1480 // take corrective action, if possible.
1481 // (This condition is an unlikely backstop only, because the
1482 // PerBytecodeTrapLimit is more likely to take effect first,
1483 // if it is applicable.)
1484 make_not_entrant = true;
1485 }
1487 // Here's more hysteresis: If there has been a recompile at
1488 // this trap point already, run the method in the interpreter
1489 // for a while to exercise it more thoroughly.
1490 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1491 reprofile = true;
1492 }
1494 }
1496 // Take requested actions on the method:
1498 // Recompile
1499 if (make_not_entrant) {
1500 if (!nm->make_not_entrant()) {
1501 return; // the call did not change nmethod's state
1502 }
1504 if (pdata != NULL) {
1505 // Record the recompilation event, if any.
1506 int tstate0 = pdata->trap_state();
1507 int tstate1 = trap_state_set_recompiled(tstate0, true);
1508 if (tstate1 != tstate0)
1509 pdata->set_trap_state(tstate1);
1510 }
1511 }
1513 if (inc_recompile_count) {
1514 trap_mdo->inc_overflow_recompile_count();
1515 if ((uint)trap_mdo->overflow_recompile_count() >
1516 (uint)PerBytecodeRecompilationCutoff) {
1517 // Give up on the method containing the bad BCI.
1518 if (trap_method() == nm->method()) {
1519 make_not_compilable = true;
1520 } else {
1521 trap_method->set_not_compilable(CompLevel_full_optimization);
1522 // But give grace to the enclosing nm->method().
1523 }
1524 }
1525 }
1527 // Reprofile
1528 if (reprofile) {
1529 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1530 }
1532 // Give up compiling
1533 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1534 assert(make_not_entrant, "consistent");
1535 nm->method()->set_not_compilable(CompLevel_full_optimization);
1536 }
1538 } // Free marked resources
1540 }
1541 JRT_END
1543 methodDataOop
1544 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1545 bool create_if_missing) {
1546 Thread* THREAD = thread;
1547 methodDataOop mdo = m()->method_data();
1548 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1549 // Build an MDO. Ignore errors like OutOfMemory;
1550 // that simply means we won't have an MDO to update.
1551 methodOopDesc::build_interpreter_method_data(m, THREAD);
1552 if (HAS_PENDING_EXCEPTION) {
1553 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1554 CLEAR_PENDING_EXCEPTION;
1555 }
1556 mdo = m()->method_data();
1557 }
1558 return mdo;
1559 }
1561 ProfileData*
1562 Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
1563 int trap_bci,
1564 Deoptimization::DeoptReason reason,
1565 //outputs:
1566 uint& ret_this_trap_count,
1567 bool& ret_maybe_prior_trap,
1568 bool& ret_maybe_prior_recompile) {
1569 uint prior_trap_count = trap_mdo->trap_count(reason);
1570 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1572 // If the runtime cannot find a place to store trap history,
1573 // it is estimated based on the general condition of the method.
1574 // If the method has ever been recompiled, or has ever incurred
1575 // a trap with the present reason , then this BCI is assumed
1576 // (pessimistically) to be the culprit.
1577 bool maybe_prior_trap = (prior_trap_count != 0);
1578 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1579 ProfileData* pdata = NULL;
1582 // For reasons which are recorded per bytecode, we check per-BCI data.
1583 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1584 if (per_bc_reason != Reason_none) {
1585 // Find the profile data for this BCI. If there isn't one,
1586 // try to allocate one from the MDO's set of spares.
1587 // This will let us detect a repeated trap at this point.
1588 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
1590 if (pdata != NULL) {
1591 // Query the trap state of this profile datum.
1592 int tstate0 = pdata->trap_state();
1593 if (!trap_state_has_reason(tstate0, per_bc_reason))
1594 maybe_prior_trap = false;
1595 if (!trap_state_is_recompiled(tstate0))
1596 maybe_prior_recompile = false;
1598 // Update the trap state of this profile datum.
1599 int tstate1 = tstate0;
1600 // Record the reason.
1601 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1602 // Store the updated state on the MDO, for next time.
1603 if (tstate1 != tstate0)
1604 pdata->set_trap_state(tstate1);
1605 } else {
1606 if (LogCompilation && xtty != NULL) {
1607 ttyLocker ttyl;
1608 // Missing MDP? Leave a small complaint in the log.
1609 xtty->elem("missing_mdp bci='%d'", trap_bci);
1610 }
1611 }
1612 }
1614 // Return results:
1615 ret_this_trap_count = this_trap_count;
1616 ret_maybe_prior_trap = maybe_prior_trap;
1617 ret_maybe_prior_recompile = maybe_prior_recompile;
1618 return pdata;
1619 }
1621 void
1622 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1623 ResourceMark rm;
1624 // Ignored outputs:
1625 uint ignore_this_trap_count;
1626 bool ignore_maybe_prior_trap;
1627 bool ignore_maybe_prior_recompile;
1628 query_update_method_data(trap_mdo, trap_bci,
1629 (DeoptReason)reason,
1630 ignore_this_trap_count,
1631 ignore_maybe_prior_trap,
1632 ignore_maybe_prior_recompile);
1633 }
1635 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1637 // Still in Java no safepoints
1638 {
1639 // This enters VM and may safepoint
1640 uncommon_trap_inner(thread, trap_request);
1641 }
1642 return fetch_unroll_info_helper(thread);
1643 }
1645 // Local derived constants.
1646 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1647 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1648 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1650 //---------------------------trap_state_reason---------------------------------
1651 Deoptimization::DeoptReason
1652 Deoptimization::trap_state_reason(int trap_state) {
1653 // This assert provides the link between the width of DataLayout::trap_bits
1654 // and the encoding of "recorded" reasons. It ensures there are enough
1655 // bits to store all needed reasons in the per-BCI MDO profile.
1656 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1657 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1658 trap_state -= recompile_bit;
1659 if (trap_state == DS_REASON_MASK) {
1660 return Reason_many;
1661 } else {
1662 assert((int)Reason_none == 0, "state=0 => Reason_none");
1663 return (DeoptReason)trap_state;
1664 }
1665 }
1666 //-------------------------trap_state_has_reason-------------------------------
1667 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1668 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1669 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1670 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1671 trap_state -= recompile_bit;
1672 if (trap_state == DS_REASON_MASK) {
1673 return -1; // true, unspecifically (bottom of state lattice)
1674 } else if (trap_state == reason) {
1675 return 1; // true, definitely
1676 } else if (trap_state == 0) {
1677 return 0; // false, definitely (top of state lattice)
1678 } else {
1679 return 0; // false, definitely
1680 }
1681 }
1682 //-------------------------trap_state_add_reason-------------------------------
1683 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1684 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1685 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1686 trap_state -= recompile_bit;
1687 if (trap_state == DS_REASON_MASK) {
1688 return trap_state + recompile_bit; // already at state lattice bottom
1689 } else if (trap_state == reason) {
1690 return trap_state + recompile_bit; // the condition is already true
1691 } else if (trap_state == 0) {
1692 return reason + recompile_bit; // no condition has yet been true
1693 } else {
1694 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1695 }
1696 }
1697 //-----------------------trap_state_is_recompiled------------------------------
1698 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1699 return (trap_state & DS_RECOMPILE_BIT) != 0;
1700 }
1701 //-----------------------trap_state_set_recompiled-----------------------------
1702 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1703 if (z) return trap_state | DS_RECOMPILE_BIT;
1704 else return trap_state & ~DS_RECOMPILE_BIT;
1705 }
1706 //---------------------------format_trap_state---------------------------------
1707 // This is used for debugging and diagnostics, including hotspot.log output.
1708 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1709 int trap_state) {
1710 DeoptReason reason = trap_state_reason(trap_state);
1711 bool recomp_flag = trap_state_is_recompiled(trap_state);
1712 // Re-encode the state from its decoded components.
1713 int decoded_state = 0;
1714 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1715 decoded_state = trap_state_add_reason(decoded_state, reason);
1716 if (recomp_flag)
1717 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1718 // If the state re-encodes properly, format it symbolically.
1719 // Because this routine is used for debugging and diagnostics,
1720 // be robust even if the state is a strange value.
1721 size_t len;
1722 if (decoded_state != trap_state) {
1723 // Random buggy state that doesn't decode??
1724 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1725 } else {
1726 len = jio_snprintf(buf, buflen, "%s%s",
1727 trap_reason_name(reason),
1728 recomp_flag ? " recompiled" : "");
1729 }
1730 if (len >= buflen)
1731 buf[buflen-1] = '\0';
1732 return buf;
1733 }
1736 //--------------------------------statics--------------------------------------
1737 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1738 = Deoptimization::Action_reinterpret;
1739 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1740 // Note: Keep this in sync. with enum DeoptReason.
1741 "none",
1742 "null_check",
1743 "null_assert",
1744 "range_check",
1745 "class_check",
1746 "array_check",
1747 "intrinsic",
1748 "bimorphic",
1749 "unloaded",
1750 "uninitialized",
1751 "unreached",
1752 "unhandled",
1753 "constraint",
1754 "div0_check",
1755 "age",
1756 "predicate"
1757 };
1758 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1759 // Note: Keep this in sync. with enum DeoptAction.
1760 "none",
1761 "maybe_recompile",
1762 "reinterpret",
1763 "make_not_entrant",
1764 "make_not_compilable"
1765 };
1767 const char* Deoptimization::trap_reason_name(int reason) {
1768 if (reason == Reason_many) return "many";
1769 if ((uint)reason < Reason_LIMIT)
1770 return _trap_reason_name[reason];
1771 static char buf[20];
1772 sprintf(buf, "reason%d", reason);
1773 return buf;
1774 }
1775 const char* Deoptimization::trap_action_name(int action) {
1776 if ((uint)action < Action_LIMIT)
1777 return _trap_action_name[action];
1778 static char buf[20];
1779 sprintf(buf, "action%d", action);
1780 return buf;
1781 }
1783 // This is used for debugging and diagnostics, including hotspot.log output.
1784 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1785 int trap_request) {
1786 jint unloaded_class_index = trap_request_index(trap_request);
1787 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1788 const char* action = trap_action_name(trap_request_action(trap_request));
1789 size_t len;
1790 if (unloaded_class_index < 0) {
1791 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1792 reason, action);
1793 } else {
1794 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1795 reason, action, unloaded_class_index);
1796 }
1797 if (len >= buflen)
1798 buf[buflen-1] = '\0';
1799 return buf;
1800 }
1802 juint Deoptimization::_deoptimization_hist
1803 [Deoptimization::Reason_LIMIT]
1804 [1 + Deoptimization::Action_LIMIT]
1805 [Deoptimization::BC_CASE_LIMIT]
1806 = {0};
1808 enum {
1809 LSB_BITS = 8,
1810 LSB_MASK = right_n_bits(LSB_BITS)
1811 };
1813 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1814 Bytecodes::Code bc) {
1815 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1816 assert(action >= 0 && action < Action_LIMIT, "oob");
1817 _deoptimization_hist[Reason_none][0][0] += 1; // total
1818 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1819 juint* cases = _deoptimization_hist[reason][1+action];
1820 juint* bc_counter_addr = NULL;
1821 juint bc_counter = 0;
1822 // Look for an unused counter, or an exact match to this BC.
1823 if (bc != Bytecodes::_illegal) {
1824 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1825 juint* counter_addr = &cases[bc_case];
1826 juint counter = *counter_addr;
1827 if ((counter == 0 && bc_counter_addr == NULL)
1828 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1829 // this counter is either free or is already devoted to this BC
1830 bc_counter_addr = counter_addr;
1831 bc_counter = counter | bc;
1832 }
1833 }
1834 }
1835 if (bc_counter_addr == NULL) {
1836 // Overflow, or no given bytecode.
1837 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1838 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1839 }
1840 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1841 }
1843 jint Deoptimization::total_deoptimization_count() {
1844 return _deoptimization_hist[Reason_none][0][0];
1845 }
1847 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1848 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1849 return _deoptimization_hist[reason][0][0];
1850 }
1852 void Deoptimization::print_statistics() {
1853 juint total = total_deoptimization_count();
1854 juint account = total;
1855 if (total != 0) {
1856 ttyLocker ttyl;
1857 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1858 tty->print_cr("Deoptimization traps recorded:");
1859 #define PRINT_STAT_LINE(name, r) \
1860 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1861 PRINT_STAT_LINE("total", total);
1862 // For each non-zero entry in the histogram, print the reason,
1863 // the action, and (if specifically known) the type of bytecode.
1864 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1865 for (int action = 0; action < Action_LIMIT; action++) {
1866 juint* cases = _deoptimization_hist[reason][1+action];
1867 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1868 juint counter = cases[bc_case];
1869 if (counter != 0) {
1870 char name[1*K];
1871 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1872 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1873 bc = Bytecodes::_illegal;
1874 sprintf(name, "%s/%s/%s",
1875 trap_reason_name(reason),
1876 trap_action_name(action),
1877 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1878 juint r = counter >> LSB_BITS;
1879 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1880 account -= r;
1881 }
1882 }
1883 }
1884 }
1885 if (account != 0) {
1886 PRINT_STAT_LINE("unaccounted", account);
1887 }
1888 #undef PRINT_STAT_LINE
1889 if (xtty != NULL) xtty->tail("statistics");
1890 }
1891 }
1892 #else // COMPILER2 || SHARK
1895 // Stubs for C1 only system.
1896 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1897 return false;
1898 }
1900 const char* Deoptimization::trap_reason_name(int reason) {
1901 return "unknown";
1902 }
1904 void Deoptimization::print_statistics() {
1905 // no output
1906 }
1908 void
1909 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1910 // no udpate
1911 }
1913 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1914 return 0;
1915 }
1917 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1918 Bytecodes::Code bc) {
1919 // no update
1920 }
1922 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1923 int trap_state) {
1924 jio_snprintf(buf, buflen, "#%d", trap_state);
1925 return buf;
1926 }
1928 #endif // COMPILER2 || SHARK