Fri, 12 Feb 2010 15:27:36 -0800
Merge
1 /*
2 * Copyright 1997-2010 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_deoptimization.cpp.incl"
28 bool DeoptimizationMarker::_is_active = false;
30 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
31 int caller_adjustment,
32 int number_of_frames,
33 intptr_t* frame_sizes,
34 address* frame_pcs,
35 BasicType return_type) {
36 _size_of_deoptimized_frame = size_of_deoptimized_frame;
37 _caller_adjustment = caller_adjustment;
38 _number_of_frames = number_of_frames;
39 _frame_sizes = frame_sizes;
40 _frame_pcs = frame_pcs;
41 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2);
42 _return_type = return_type;
43 // PD (x86 only)
44 _counter_temp = 0;
45 _initial_fp = 0;
46 _unpack_kind = 0;
47 _sender_sp_temp = 0;
49 _total_frame_sizes = size_of_frames();
50 }
53 Deoptimization::UnrollBlock::~UnrollBlock() {
54 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
55 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
56 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
57 }
60 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
61 assert(register_number < RegisterMap::reg_count, "checking register number");
62 return &_register_block[register_number * 2];
63 }
67 int Deoptimization::UnrollBlock::size_of_frames() const {
68 // Acount first for the adjustment of the initial frame
69 int result = _caller_adjustment;
70 for (int index = 0; index < number_of_frames(); index++) {
71 result += frame_sizes()[index];
72 }
73 return result;
74 }
77 void Deoptimization::UnrollBlock::print() {
78 ttyLocker ttyl;
79 tty->print_cr("UnrollBlock");
80 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
81 tty->print( " frame_sizes: ");
82 for (int index = 0; index < number_of_frames(); index++) {
83 tty->print("%d ", frame_sizes()[index]);
84 }
85 tty->cr();
86 }
89 // In order to make fetch_unroll_info work properly with escape
90 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
91 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
92 // of previously eliminated objects occurs in realloc_objects, which is
93 // called from the method fetch_unroll_info_helper below.
94 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
95 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
96 // but makes the entry a little slower. There is however a little dance we have to
97 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
99 // fetch_unroll_info() is called at the beginning of the deoptimization
100 // handler. Note this fact before we start generating temporary frames
101 // that can confuse an asynchronous stack walker. This counter is
102 // decremented at the end of unpack_frames().
103 thread->inc_in_deopt_handler();
105 return fetch_unroll_info_helper(thread);
106 JRT_END
109 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
110 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
112 // Note: there is a safepoint safety issue here. No matter whether we enter
113 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
114 // the vframeArray is created.
115 //
117 // Allocate our special deoptimization ResourceMark
118 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
119 assert(thread->deopt_mark() == NULL, "Pending deopt!");
120 thread->set_deopt_mark(dmark);
122 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
123 RegisterMap map(thread, true);
124 RegisterMap dummy_map(thread, false);
125 // Now get the deoptee with a valid map
126 frame deoptee = stub_frame.sender(&map);
128 // Create a growable array of VFrames where each VFrame represents an inlined
129 // Java frame. This storage is allocated with the usual system arena.
130 assert(deoptee.is_compiled_frame(), "Wrong frame type");
131 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
132 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
133 while (!vf->is_top()) {
134 assert(vf->is_compiled_frame(), "Wrong frame type");
135 chunk->push(compiledVFrame::cast(vf));
136 vf = vf->sender();
137 }
138 assert(vf->is_compiled_frame(), "Wrong frame type");
139 chunk->push(compiledVFrame::cast(vf));
141 #ifdef COMPILER2
142 // Reallocate the non-escaping objects and restore their fields. Then
143 // relock objects if synchronization on them was eliminated.
144 if (DoEscapeAnalysis) {
145 if (EliminateAllocations) {
146 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
147 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
149 // The flag return_oop() indicates call sites which return oop
150 // in compiled code. Such sites include java method calls,
151 // runtime calls (for example, used to allocate new objects/arrays
152 // on slow code path) and any other calls generated in compiled code.
153 // It is not guaranteed that we can get such information here only
154 // by analyzing bytecode in deoptimized frames. This is why this flag
155 // is set during method compilation (see Compile::Process_OopMap_Node()).
156 bool save_oop_result = chunk->at(0)->scope()->return_oop();
157 Handle return_value;
158 if (save_oop_result) {
159 // Reallocation may trigger GC. If deoptimization happened on return from
160 // call which returns oop we need to save it since it is not in oopmap.
161 oop result = deoptee.saved_oop_result(&map);
162 assert(result == NULL || result->is_oop(), "must be oop");
163 return_value = Handle(thread, result);
164 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
165 if (TraceDeoptimization) {
166 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
167 }
168 }
169 bool reallocated = false;
170 if (objects != NULL) {
171 JRT_BLOCK
172 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
173 JRT_END
174 }
175 if (reallocated) {
176 reassign_fields(&deoptee, &map, objects);
177 #ifndef PRODUCT
178 if (TraceDeoptimization) {
179 ttyLocker ttyl;
180 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
181 print_objects(objects);
182 }
183 #endif
184 }
185 if (save_oop_result) {
186 // Restore result.
187 deoptee.set_saved_oop_result(&map, return_value());
188 }
189 }
190 if (EliminateLocks) {
191 #ifndef PRODUCT
192 bool first = true;
193 #endif
194 for (int i = 0; i < chunk->length(); i++) {
195 compiledVFrame* cvf = chunk->at(i);
196 assert (cvf->scope() != NULL,"expect only compiled java frames");
197 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
198 if (monitors->is_nonempty()) {
199 relock_objects(monitors, thread);
200 #ifndef PRODUCT
201 if (TraceDeoptimization) {
202 ttyLocker ttyl;
203 for (int j = 0; j < monitors->length(); j++) {
204 MonitorInfo* mi = monitors->at(j);
205 if (mi->eliminated()) {
206 if (first) {
207 first = false;
208 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
209 }
210 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
211 }
212 }
213 }
214 #endif
215 }
216 }
217 }
218 }
219 #endif // COMPILER2
220 // Ensure that no safepoint is taken after pointers have been stored
221 // in fields of rematerialized objects. If a safepoint occurs from here on
222 // out the java state residing in the vframeArray will be missed.
223 No_Safepoint_Verifier no_safepoint;
225 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
227 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
228 thread->set_vframe_array_head(array);
230 // Now that the vframeArray has been created if we have any deferred local writes
231 // added by jvmti then we can free up that structure as the data is now in the
232 // vframeArray
234 if (thread->deferred_locals() != NULL) {
235 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
236 int i = 0;
237 do {
238 // Because of inlining we could have multiple vframes for a single frame
239 // and several of the vframes could have deferred writes. Find them all.
240 if (list->at(i)->id() == array->original().id()) {
241 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
242 list->remove_at(i);
243 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
244 delete dlv;
245 } else {
246 i++;
247 }
248 } while ( i < list->length() );
249 if (list->length() == 0) {
250 thread->set_deferred_locals(NULL);
251 // free the list and elements back to C heap.
252 delete list;
253 }
255 }
257 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
258 CodeBlob* cb = stub_frame.cb();
259 // Verify we have the right vframeArray
260 assert(cb->frame_size() >= 0, "Unexpected frame size");
261 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
263 // If the deopt call site is a MethodHandle invoke call site we have
264 // to adjust the unpack_sp.
265 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
266 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
267 unpack_sp = deoptee.unextended_sp();
269 #ifdef ASSERT
270 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
271 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp);
272 #endif
273 // This is a guarantee instead of an assert because if vframe doesn't match
274 // we will unpack the wrong deoptimized frame and wind up in strange places
275 // where it will be very difficult to figure out what went wrong. Better
276 // to die an early death here than some very obscure death later when the
277 // trail is cold.
278 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
279 // in that it will fail to detect a problem when there is one. This needs
280 // more work in tiger timeframe.
281 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
283 int number_of_frames = array->frames();
285 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
286 // virtual activation, which is the reverse of the elements in the vframes array.
287 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames);
288 // +1 because we always have an interpreter return address for the final slot.
289 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1);
290 int callee_parameters = 0;
291 int callee_locals = 0;
292 int popframe_extra_args = 0;
293 // Create an interpreter return address for the stub to use as its return
294 // address so the skeletal frames are perfectly walkable
295 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
297 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
298 // activation be put back on the expression stack of the caller for reexecution
299 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
300 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
301 }
303 //
304 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
305 // frame_sizes/frame_pcs[1] next oldest frame (int)
306 // frame_sizes/frame_pcs[n] youngest frame (int)
307 //
308 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
309 // owns the space for the return address to it's caller). Confusing ain't it.
310 //
311 // The vframe array can address vframes with indices running from
312 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
313 // When we create the skeletal frames we need the oldest frame to be in the zero slot
314 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
315 // so things look a little strange in this loop.
316 //
317 for (int index = 0; index < array->frames(); index++ ) {
318 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
319 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
320 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
321 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
322 callee_locals,
323 index == 0,
324 popframe_extra_args);
325 // This pc doesn't have to be perfect just good enough to identify the frame
326 // as interpreted so the skeleton frame will be walkable
327 // The correct pc will be set when the skeleton frame is completely filled out
328 // The final pc we store in the loop is wrong and will be overwritten below
329 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
331 callee_parameters = array->element(index)->method()->size_of_parameters();
332 callee_locals = array->element(index)->method()->max_locals();
333 popframe_extra_args = 0;
334 }
336 // Compute whether the root vframe returns a float or double value.
337 BasicType return_type;
338 {
339 HandleMark hm;
340 methodHandle method(thread, array->element(0)->method());
341 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci());
342 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL;
343 }
345 // Compute information for handling adapters and adjusting the frame size of the caller.
346 int caller_adjustment = 0;
348 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
349 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
350 // than simply use array->sender.pc(). This requires us to walk the current set of frames
351 //
352 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
353 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
355 // Compute the amount the oldest interpreter frame will have to adjust
356 // its caller's stack by. If the caller is a compiled frame then
357 // we pretend that the callee has no parameters so that the
358 // extension counts for the full amount of locals and not just
359 // locals-parms. This is because without a c2i adapter the parm
360 // area as created by the compiled frame will not be usable by
361 // the interpreter. (Depending on the calling convention there
362 // may not even be enough space).
364 // QQQ I'd rather see this pushed down into last_frame_adjust
365 // and have it take the sender (aka caller).
367 if (deopt_sender.is_compiled_frame()) {
368 caller_adjustment = last_frame_adjust(0, callee_locals);
369 } else if (callee_locals > callee_parameters) {
370 // The caller frame may need extending to accommodate
371 // non-parameter locals of the first unpacked interpreted frame.
372 // Compute that adjustment.
373 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
374 }
377 // If the sender is deoptimized the we must retrieve the address of the handler
378 // since the frame will "magically" show the original pc before the deopt
379 // and we'd undo the deopt.
381 frame_pcs[0] = deopt_sender.raw_pc();
383 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
385 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
386 caller_adjustment * BytesPerWord,
387 number_of_frames,
388 frame_sizes,
389 frame_pcs,
390 return_type);
391 #if defined(IA32) || defined(AMD64)
392 // We need a way to pass fp to the unpacking code so the skeletal frames
393 // come out correct. This is only needed for x86 because of c2 using ebp
394 // as an allocatable register. So this update is useless (and harmless)
395 // on the other platforms. It would be nice to do this in a different
396 // way but even the old style deoptimization had a problem with deriving
397 // this value. NEEDS_CLEANUP
398 // Note: now that c1 is using c2's deopt blob we must do this on all
399 // x86 based platforms
400 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes());
401 *fp_addr = array->sender().fp(); // was adapter_caller
402 #endif /* IA32 || AMD64 */
404 if (array->frames() > 1) {
405 if (VerifyStack && TraceDeoptimization) {
406 tty->print_cr("Deoptimizing method containing inlining");
407 }
408 }
410 array->set_unroll_block(info);
411 return info;
412 }
414 // Called to cleanup deoptimization data structures in normal case
415 // after unpacking to stack and when stack overflow error occurs
416 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
417 vframeArray *array) {
419 // Get array if coming from exception
420 if (array == NULL) {
421 array = thread->vframe_array_head();
422 }
423 thread->set_vframe_array_head(NULL);
425 // Free the previous UnrollBlock
426 vframeArray* old_array = thread->vframe_array_last();
427 thread->set_vframe_array_last(array);
429 if (old_array != NULL) {
430 UnrollBlock* old_info = old_array->unroll_block();
431 old_array->set_unroll_block(NULL);
432 delete old_info;
433 delete old_array;
434 }
436 // Deallocate any resource creating in this routine and any ResourceObjs allocated
437 // inside the vframeArray (StackValueCollections)
439 delete thread->deopt_mark();
440 thread->set_deopt_mark(NULL);
443 if (JvmtiExport::can_pop_frame()) {
444 #ifndef CC_INTERP
445 // Regardless of whether we entered this routine with the pending
446 // popframe condition bit set, we should always clear it now
447 thread->clear_popframe_condition();
448 #else
449 // C++ interpeter will clear has_pending_popframe when it enters
450 // with method_resume. For deopt_resume2 we clear it now.
451 if (thread->popframe_forcing_deopt_reexecution())
452 thread->clear_popframe_condition();
453 #endif /* CC_INTERP */
454 }
456 // unpack_frames() is called at the end of the deoptimization handler
457 // and (in C2) at the end of the uncommon trap handler. Note this fact
458 // so that an asynchronous stack walker can work again. This counter is
459 // incremented at the beginning of fetch_unroll_info() and (in C2) at
460 // the beginning of uncommon_trap().
461 thread->dec_in_deopt_handler();
462 }
465 // Return BasicType of value being returned
466 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
468 // We are already active int he special DeoptResourceMark any ResourceObj's we
469 // allocate will be freed at the end of the routine.
471 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
472 // but makes the entry a little slower. There is however a little dance we have to
473 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
474 ResetNoHandleMark rnhm; // No-op in release/product versions
475 HandleMark hm;
477 frame stub_frame = thread->last_frame();
479 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
480 // must point to the vframeArray for the unpack frame.
481 vframeArray* array = thread->vframe_array_head();
483 #ifndef PRODUCT
484 if (TraceDeoptimization) {
485 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
486 }
487 #endif
489 UnrollBlock* info = array->unroll_block();
491 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
492 array->unpack_to_stack(stub_frame, exec_mode);
494 BasicType bt = info->return_type();
496 // If we have an exception pending, claim that the return type is an oop
497 // so the deopt_blob does not overwrite the exception_oop.
499 if (exec_mode == Unpack_exception)
500 bt = T_OBJECT;
502 // Cleanup thread deopt data
503 cleanup_deopt_info(thread, array);
505 #ifndef PRODUCT
506 if (VerifyStack) {
507 ResourceMark res_mark;
509 // Verify that the just-unpacked frames match the interpreter's
510 // notions of expression stack and locals
511 vframeArray* cur_array = thread->vframe_array_last();
512 RegisterMap rm(thread, false);
513 rm.set_include_argument_oops(false);
514 bool is_top_frame = true;
515 int callee_size_of_parameters = 0;
516 int callee_max_locals = 0;
517 for (int i = 0; i < cur_array->frames(); i++) {
518 vframeArrayElement* el = cur_array->element(i);
519 frame* iframe = el->iframe();
520 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
522 // Get the oop map for this bci
523 InterpreterOopMap mask;
524 int cur_invoke_parameter_size = 0;
525 bool try_next_mask = false;
526 int next_mask_expression_stack_size = -1;
527 int top_frame_expression_stack_adjustment = 0;
528 methodHandle mh(thread, iframe->interpreter_frame_method());
529 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
530 BytecodeStream str(mh);
531 str.set_start(iframe->interpreter_frame_bci());
532 int max_bci = mh->code_size();
533 // Get to the next bytecode if possible
534 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
535 // Check to see if we can grab the number of outgoing arguments
536 // at an uncommon trap for an invoke (where the compiler
537 // generates debug info before the invoke has executed)
538 Bytecodes::Code cur_code = str.next();
539 if (cur_code == Bytecodes::_invokevirtual ||
540 cur_code == Bytecodes::_invokespecial ||
541 cur_code == Bytecodes::_invokestatic ||
542 cur_code == Bytecodes::_invokeinterface) {
543 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci());
544 symbolHandle signature(thread, invoke->signature());
545 ArgumentSizeComputer asc(signature);
546 cur_invoke_parameter_size = asc.size();
547 if (cur_code != Bytecodes::_invokestatic) {
548 // Add in receiver
549 ++cur_invoke_parameter_size;
550 }
551 }
552 if (str.bci() < max_bci) {
553 Bytecodes::Code bc = str.next();
554 if (bc >= 0) {
555 // The interpreter oop map generator reports results before
556 // the current bytecode has executed except in the case of
557 // calls. It seems to be hard to tell whether the compiler
558 // has emitted debug information matching the "state before"
559 // a given bytecode or the state after, so we try both
560 switch (cur_code) {
561 case Bytecodes::_invokevirtual:
562 case Bytecodes::_invokespecial:
563 case Bytecodes::_invokestatic:
564 case Bytecodes::_invokeinterface:
565 case Bytecodes::_athrow:
566 break;
567 default: {
568 InterpreterOopMap next_mask;
569 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
570 next_mask_expression_stack_size = next_mask.expression_stack_size();
571 // Need to subtract off the size of the result type of
572 // the bytecode because this is not described in the
573 // debug info but returned to the interpreter in the TOS
574 // caching register
575 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
576 if (bytecode_result_type != T_ILLEGAL) {
577 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
578 }
579 assert(top_frame_expression_stack_adjustment >= 0, "");
580 try_next_mask = true;
581 break;
582 }
583 }
584 }
585 }
587 // Verify stack depth and oops in frame
588 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
589 if (!(
590 /* SPARC */
591 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
592 /* x86 */
593 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
594 (try_next_mask &&
595 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
596 top_frame_expression_stack_adjustment))) ||
597 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
598 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
599 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
600 )) {
601 ttyLocker ttyl;
603 // Print out some information that will help us debug the problem
604 tty->print_cr("Wrong number of expression stack elements during deoptimization");
605 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
606 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
607 iframe->interpreter_frame_expression_stack_size());
608 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
609 tty->print_cr(" try_next_mask = %d", try_next_mask);
610 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
611 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
612 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
613 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
614 tty->print_cr(" exec_mode = %d", exec_mode);
615 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
616 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
617 tty->print_cr(" Interpreted frames:");
618 for (int k = 0; k < cur_array->frames(); k++) {
619 vframeArrayElement* el = cur_array->element(k);
620 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
621 }
622 cur_array->print_on_2(tty);
623 guarantee(false, "wrong number of expression stack elements during deopt");
624 }
625 VerifyOopClosure verify;
626 iframe->oops_interpreted_do(&verify, &rm, false);
627 callee_size_of_parameters = mh->size_of_parameters();
628 callee_max_locals = mh->max_locals();
629 is_top_frame = false;
630 }
631 }
632 #endif /* !PRODUCT */
635 return bt;
636 JRT_END
639 int Deoptimization::deoptimize_dependents() {
640 Threads::deoptimized_wrt_marked_nmethods();
641 return 0;
642 }
645 #ifdef COMPILER2
646 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
647 Handle pending_exception(thread->pending_exception());
648 const char* exception_file = thread->exception_file();
649 int exception_line = thread->exception_line();
650 thread->clear_pending_exception();
652 for (int i = 0; i < objects->length(); i++) {
653 assert(objects->at(i)->is_object(), "invalid debug information");
654 ObjectValue* sv = (ObjectValue*) objects->at(i);
656 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
657 oop obj = NULL;
659 if (k->oop_is_instance()) {
660 instanceKlass* ik = instanceKlass::cast(k());
661 obj = ik->allocate_instance(CHECK_(false));
662 } else if (k->oop_is_typeArray()) {
663 typeArrayKlass* ak = typeArrayKlass::cast(k());
664 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
665 int len = sv->field_size() / type2size[ak->element_type()];
666 obj = ak->allocate(len, CHECK_(false));
667 } else if (k->oop_is_objArray()) {
668 objArrayKlass* ak = objArrayKlass::cast(k());
669 obj = ak->allocate(sv->field_size(), CHECK_(false));
670 }
672 assert(obj != NULL, "allocation failed");
673 assert(sv->value().is_null(), "redundant reallocation");
674 sv->set_value(obj);
675 }
677 if (pending_exception.not_null()) {
678 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
679 }
681 return true;
682 }
684 // This assumes that the fields are stored in ObjectValue in the same order
685 // they are yielded by do_nonstatic_fields.
686 class FieldReassigner: public FieldClosure {
687 frame* _fr;
688 RegisterMap* _reg_map;
689 ObjectValue* _sv;
690 instanceKlass* _ik;
691 oop _obj;
693 int _i;
694 public:
695 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
696 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
698 int i() const { return _i; }
701 void do_field(fieldDescriptor* fd) {
702 intptr_t val;
703 StackValue* value =
704 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
705 int offset = fd->offset();
706 switch (fd->field_type()) {
707 case T_OBJECT: case T_ARRAY:
708 assert(value->type() == T_OBJECT, "Agreement.");
709 _obj->obj_field_put(offset, value->get_obj()());
710 break;
712 case T_LONG: case T_DOUBLE: {
713 assert(value->type() == T_INT, "Agreement.");
714 StackValue* low =
715 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
716 #ifdef _LP64
717 jlong res = (jlong)low->get_int();
718 #else
719 #ifdef SPARC
720 // For SPARC we have to swap high and low words.
721 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
722 #else
723 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
724 #endif //SPARC
725 #endif
726 _obj->long_field_put(offset, res);
727 break;
728 }
729 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
730 case T_INT: case T_FLOAT: // 4 bytes.
731 assert(value->type() == T_INT, "Agreement.");
732 val = value->get_int();
733 _obj->int_field_put(offset, (jint)*((jint*)&val));
734 break;
736 case T_SHORT: case T_CHAR: // 2 bytes
737 assert(value->type() == T_INT, "Agreement.");
738 val = value->get_int();
739 _obj->short_field_put(offset, (jshort)*((jint*)&val));
740 break;
742 case T_BOOLEAN: case T_BYTE: // 1 byte
743 assert(value->type() == T_INT, "Agreement.");
744 val = value->get_int();
745 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
746 break;
748 default:
749 ShouldNotReachHere();
750 }
751 _i++;
752 }
753 };
755 // restore elements of an eliminated type array
756 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
757 int index = 0;
758 intptr_t val;
760 for (int i = 0; i < sv->field_size(); i++) {
761 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
762 switch(type) {
763 case T_LONG: case T_DOUBLE: {
764 assert(value->type() == T_INT, "Agreement.");
765 StackValue* low =
766 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
767 #ifdef _LP64
768 jlong res = (jlong)low->get_int();
769 #else
770 #ifdef SPARC
771 // For SPARC we have to swap high and low words.
772 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
773 #else
774 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
775 #endif //SPARC
776 #endif
777 obj->long_at_put(index, res);
778 break;
779 }
781 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
782 case T_INT: case T_FLOAT: // 4 bytes.
783 assert(value->type() == T_INT, "Agreement.");
784 val = value->get_int();
785 obj->int_at_put(index, (jint)*((jint*)&val));
786 break;
788 case T_SHORT: case T_CHAR: // 2 bytes
789 assert(value->type() == T_INT, "Agreement.");
790 val = value->get_int();
791 obj->short_at_put(index, (jshort)*((jint*)&val));
792 break;
794 case T_BOOLEAN: case T_BYTE: // 1 byte
795 assert(value->type() == T_INT, "Agreement.");
796 val = value->get_int();
797 obj->bool_at_put(index, (jboolean)*((jint*)&val));
798 break;
800 default:
801 ShouldNotReachHere();
802 }
803 index++;
804 }
805 }
808 // restore fields of an eliminated object array
809 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
810 for (int i = 0; i < sv->field_size(); i++) {
811 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
812 assert(value->type() == T_OBJECT, "object element expected");
813 obj->obj_at_put(i, value->get_obj()());
814 }
815 }
818 // restore fields of all eliminated objects and arrays
819 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
820 for (int i = 0; i < objects->length(); i++) {
821 ObjectValue* sv = (ObjectValue*) objects->at(i);
822 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
823 Handle obj = sv->value();
824 assert(obj.not_null(), "reallocation was missed");
826 if (k->oop_is_instance()) {
827 instanceKlass* ik = instanceKlass::cast(k());
828 FieldReassigner reassign(fr, reg_map, sv, obj());
829 ik->do_nonstatic_fields(&reassign);
830 } else if (k->oop_is_typeArray()) {
831 typeArrayKlass* ak = typeArrayKlass::cast(k());
832 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
833 } else if (k->oop_is_objArray()) {
834 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
835 }
836 }
837 }
840 // relock objects for which synchronization was eliminated
841 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
842 for (int i = 0; i < monitors->length(); i++) {
843 MonitorInfo* mon_info = monitors->at(i);
844 if (mon_info->eliminated()) {
845 assert(mon_info->owner() != NULL, "reallocation was missed");
846 Handle obj = Handle(mon_info->owner());
847 markOop mark = obj->mark();
848 if (UseBiasedLocking && mark->has_bias_pattern()) {
849 // New allocated objects may have the mark set to anonymously biased.
850 // Also the deoptimized method may called methods with synchronization
851 // where the thread-local object is bias locked to the current thread.
852 assert(mark->is_biased_anonymously() ||
853 mark->biased_locker() == thread, "should be locked to current thread");
854 // Reset mark word to unbiased prototype.
855 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
856 obj->set_mark(unbiased_prototype);
857 }
858 BasicLock* lock = mon_info->lock();
859 ObjectSynchronizer::slow_enter(obj, lock, thread);
860 }
861 assert(mon_info->owner()->is_locked(), "object must be locked now");
862 }
863 }
866 #ifndef PRODUCT
867 // print information about reallocated objects
868 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
869 fieldDescriptor fd;
871 for (int i = 0; i < objects->length(); i++) {
872 ObjectValue* sv = (ObjectValue*) objects->at(i);
873 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
874 Handle obj = sv->value();
876 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
877 k->as_klassOop()->print_value();
878 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
879 tty->cr();
881 if (Verbose) {
882 k->oop_print_on(obj(), tty);
883 }
884 }
885 }
886 #endif
887 #endif // COMPILER2
889 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
891 #ifndef PRODUCT
892 if (TraceDeoptimization) {
893 ttyLocker ttyl;
894 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
895 fr.print_on(tty);
896 tty->print_cr(" Virtual frames (innermost first):");
897 for (int index = 0; index < chunk->length(); index++) {
898 compiledVFrame* vf = chunk->at(index);
899 tty->print(" %2d - ", index);
900 vf->print_value();
901 int bci = chunk->at(index)->raw_bci();
902 const char* code_name;
903 if (bci == SynchronizationEntryBCI) {
904 code_name = "sync entry";
905 } else {
906 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci);
907 code_name = Bytecodes::name(code);
908 }
909 tty->print(" - %s", code_name);
910 tty->print_cr(" @ bci %d ", bci);
911 if (Verbose) {
912 vf->print();
913 tty->cr();
914 }
915 }
916 }
917 #endif
919 // Register map for next frame (used for stack crawl). We capture
920 // the state of the deopt'ing frame's caller. Thus if we need to
921 // stuff a C2I adapter we can properly fill in the callee-save
922 // register locations.
923 frame caller = fr.sender(reg_map);
924 int frame_size = caller.sp() - fr.sp();
926 frame sender = caller;
928 // Since the Java thread being deoptimized will eventually adjust it's own stack,
929 // the vframeArray containing the unpacking information is allocated in the C heap.
930 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
931 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
933 // Compare the vframeArray to the collected vframes
934 assert(array->structural_compare(thread, chunk), "just checking");
935 Events::log("# vframes = %d", (intptr_t)chunk->length());
937 #ifndef PRODUCT
938 if (TraceDeoptimization) {
939 ttyLocker ttyl;
940 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
941 }
942 #endif // PRODUCT
944 return array;
945 }
948 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
949 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
950 for (int i = 0; i < monitors->length(); i++) {
951 MonitorInfo* mon_info = monitors->at(i);
952 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
953 objects_to_revoke->append(Handle(mon_info->owner()));
954 }
955 }
956 }
959 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
960 if (!UseBiasedLocking) {
961 return;
962 }
964 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
966 // Unfortunately we don't have a RegisterMap available in most of
967 // the places we want to call this routine so we need to walk the
968 // stack again to update the register map.
969 if (map == NULL || !map->update_map()) {
970 StackFrameStream sfs(thread, true);
971 bool found = false;
972 while (!found && !sfs.is_done()) {
973 frame* cur = sfs.current();
974 sfs.next();
975 found = cur->id() == fr.id();
976 }
977 assert(found, "frame to be deoptimized not found on target thread's stack");
978 map = sfs.register_map();
979 }
981 vframe* vf = vframe::new_vframe(&fr, map, thread);
982 compiledVFrame* cvf = compiledVFrame::cast(vf);
983 // Revoke monitors' biases in all scopes
984 while (!cvf->is_top()) {
985 collect_monitors(cvf, objects_to_revoke);
986 cvf = compiledVFrame::cast(cvf->sender());
987 }
988 collect_monitors(cvf, objects_to_revoke);
990 if (SafepointSynchronize::is_at_safepoint()) {
991 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
992 } else {
993 BiasedLocking::revoke(objects_to_revoke);
994 }
995 }
998 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
999 if (!UseBiasedLocking) {
1000 return;
1001 }
1003 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1004 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1005 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1006 if (jt->has_last_Java_frame()) {
1007 StackFrameStream sfs(jt, true);
1008 while (!sfs.is_done()) {
1009 frame* cur = sfs.current();
1010 if (cb->contains(cur->pc())) {
1011 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1012 compiledVFrame* cvf = compiledVFrame::cast(vf);
1013 // Revoke monitors' biases in all scopes
1014 while (!cvf->is_top()) {
1015 collect_monitors(cvf, objects_to_revoke);
1016 cvf = compiledVFrame::cast(cvf->sender());
1017 }
1018 collect_monitors(cvf, objects_to_revoke);
1019 }
1020 sfs.next();
1021 }
1022 }
1023 }
1024 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1025 }
1028 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1029 assert(fr.can_be_deoptimized(), "checking frame type");
1031 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1033 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id());
1035 // Patch the nmethod so that when execution returns to it we will
1036 // deopt the execution state and return to the interpreter.
1037 fr.deoptimize(thread);
1038 }
1040 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1041 // Deoptimize only if the frame comes from compile code.
1042 // Do not deoptimize the frame which is already patched
1043 // during the execution of the loops below.
1044 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1045 return;
1046 }
1047 ResourceMark rm;
1048 DeoptimizationMarker dm;
1049 if (UseBiasedLocking) {
1050 revoke_biases_of_monitors(thread, fr, map);
1051 }
1052 deoptimize_single_frame(thread, fr);
1054 }
1057 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1058 // Compute frame and register map based on thread and sp.
1059 RegisterMap reg_map(thread, UseBiasedLocking);
1060 frame fr = thread->last_frame();
1061 while (fr.id() != id) {
1062 fr = fr.sender(®_map);
1063 }
1064 deoptimize(thread, fr, ®_map);
1065 }
1068 // JVMTI PopFrame support
1069 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1070 {
1071 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1072 }
1073 JRT_END
1076 #ifdef COMPILER2
1077 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1078 // in case of an unresolved klass entry, load the class.
1079 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1080 klassOop tk = constant_pool->klass_at(index, CHECK);
1081 return;
1082 }
1084 if (!constant_pool->tag_at(index).is_symbol()) return;
1086 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
1087 symbolHandle symbol (THREAD, constant_pool->symbol_at(index));
1089 // class name?
1090 if (symbol->byte_at(0) != '(') {
1091 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1092 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1093 return;
1094 }
1096 // then it must be a signature!
1097 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1098 if (ss.is_object()) {
1099 symbolOop s = ss.as_symbol(CHECK);
1100 symbolHandle class_name (THREAD, s);
1101 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1102 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1103 }
1104 }
1105 }
1108 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1109 EXCEPTION_MARK;
1110 load_class_by_index(constant_pool, index, THREAD);
1111 if (HAS_PENDING_EXCEPTION) {
1112 // Exception happened during classloading. We ignore the exception here, since it
1113 // is going to be rethrown since the current activation is going to be deoptimzied and
1114 // the interpreter will re-execute the bytecode.
1115 CLEAR_PENDING_EXCEPTION;
1116 }
1117 }
1119 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1120 HandleMark hm;
1122 // uncommon_trap() is called at the beginning of the uncommon trap
1123 // handler. Note this fact before we start generating temporary frames
1124 // that can confuse an asynchronous stack walker. This counter is
1125 // decremented at the end of unpack_frames().
1126 thread->inc_in_deopt_handler();
1128 // We need to update the map if we have biased locking.
1129 RegisterMap reg_map(thread, UseBiasedLocking);
1130 frame stub_frame = thread->last_frame();
1131 frame fr = stub_frame.sender(®_map);
1132 // Make sure the calling nmethod is not getting deoptimized and removed
1133 // before we are done with it.
1134 nmethodLocker nl(fr.pc());
1136 {
1137 ResourceMark rm;
1139 // Revoke biases of any monitors in the frame to ensure we can migrate them
1140 revoke_biases_of_monitors(thread, fr, ®_map);
1142 DeoptReason reason = trap_request_reason(trap_request);
1143 DeoptAction action = trap_request_action(trap_request);
1144 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1146 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request);
1147 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1148 compiledVFrame* cvf = compiledVFrame::cast(vf);
1150 nmethod* nm = cvf->code();
1152 ScopeDesc* trap_scope = cvf->scope();
1153 methodHandle trap_method = trap_scope->method();
1154 int trap_bci = trap_scope->bci();
1155 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code();
1157 // Record this event in the histogram.
1158 gather_statistics(reason, action, trap_bc);
1160 // Ensure that we can record deopt. history:
1161 bool create_if_missing = ProfileTraps;
1163 methodDataHandle trap_mdo
1164 (THREAD, get_method_data(thread, trap_method, create_if_missing));
1166 // Print a bunch of diagnostics, if requested.
1167 if (TraceDeoptimization || LogCompilation) {
1168 ResourceMark rm;
1169 ttyLocker ttyl;
1170 char buf[100];
1171 if (xtty != NULL) {
1172 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1173 os::current_thread_id(),
1174 format_trap_request(buf, sizeof(buf), trap_request));
1175 nm->log_identity(xtty);
1176 }
1177 symbolHandle class_name;
1178 bool unresolved = false;
1179 if (unloaded_class_index >= 0) {
1180 constantPoolHandle constants (THREAD, trap_method->constants());
1181 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1182 class_name = symbolHandle(THREAD,
1183 constants->klass_name_at(unloaded_class_index));
1184 unresolved = true;
1185 if (xtty != NULL)
1186 xtty->print(" unresolved='1'");
1187 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1188 class_name = symbolHandle(THREAD,
1189 constants->symbol_at(unloaded_class_index));
1190 }
1191 if (xtty != NULL)
1192 xtty->name(class_name);
1193 }
1194 if (xtty != NULL && trap_mdo.not_null()) {
1195 // Dump the relevant MDO state.
1196 // This is the deopt count for the current reason, any previous
1197 // reasons or recompiles seen at this point.
1198 int dcnt = trap_mdo->trap_count(reason);
1199 if (dcnt != 0)
1200 xtty->print(" count='%d'", dcnt);
1201 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1202 int dos = (pdata == NULL)? 0: pdata->trap_state();
1203 if (dos != 0) {
1204 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1205 if (trap_state_is_recompiled(dos)) {
1206 int recnt2 = trap_mdo->overflow_recompile_count();
1207 if (recnt2 != 0)
1208 xtty->print(" recompiles2='%d'", recnt2);
1209 }
1210 }
1211 }
1212 if (xtty != NULL) {
1213 xtty->stamp();
1214 xtty->end_head();
1215 }
1216 if (TraceDeoptimization) { // make noise on the tty
1217 tty->print("Uncommon trap occurred in");
1218 nm->method()->print_short_name(tty);
1219 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
1220 fr.pc(),
1221 (int) os::current_thread_id(),
1222 trap_reason_name(reason),
1223 trap_action_name(action),
1224 unloaded_class_index);
1225 if (class_name.not_null()) {
1226 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1227 class_name->print_symbol_on(tty);
1228 }
1229 tty->cr();
1230 }
1231 if (xtty != NULL) {
1232 // Log the precise location of the trap.
1233 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1234 xtty->begin_elem("jvms bci='%d'", sd->bci());
1235 xtty->method(sd->method());
1236 xtty->end_elem();
1237 if (sd->is_top()) break;
1238 }
1239 xtty->tail("uncommon_trap");
1240 }
1241 }
1242 // (End diagnostic printout.)
1244 // Load class if necessary
1245 if (unloaded_class_index >= 0) {
1246 constantPoolHandle constants(THREAD, trap_method->constants());
1247 load_class_by_index(constants, unloaded_class_index);
1248 }
1250 // Flush the nmethod if necessary and desirable.
1251 //
1252 // We need to avoid situations where we are re-flushing the nmethod
1253 // because of a hot deoptimization site. Repeated flushes at the same
1254 // point need to be detected by the compiler and avoided. If the compiler
1255 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1256 // module must take measures to avoid an infinite cycle of recompilation
1257 // and deoptimization. There are several such measures:
1258 //
1259 // 1. If a recompilation is ordered a second time at some site X
1260 // and for the same reason R, the action is adjusted to 'reinterpret',
1261 // to give the interpreter time to exercise the method more thoroughly.
1262 // If this happens, the method's overflow_recompile_count is incremented.
1263 //
1264 // 2. If the compiler fails to reduce the deoptimization rate, then
1265 // the method's overflow_recompile_count will begin to exceed the set
1266 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1267 // is adjusted to 'make_not_compilable', and the method is abandoned
1268 // to the interpreter. This is a performance hit for hot methods,
1269 // but is better than a disastrous infinite cycle of recompilations.
1270 // (Actually, only the method containing the site X is abandoned.)
1271 //
1272 // 3. In parallel with the previous measures, if the total number of
1273 // recompilations of a method exceeds the much larger set limit
1274 // PerMethodRecompilationCutoff, the method is abandoned.
1275 // This should only happen if the method is very large and has
1276 // many "lukewarm" deoptimizations. The code which enforces this
1277 // limit is elsewhere (class nmethod, class methodOopDesc).
1278 //
1279 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1280 // to recompile at each bytecode independently of the per-BCI cutoff.
1281 //
1282 // The decision to update code is up to the compiler, and is encoded
1283 // in the Action_xxx code. If the compiler requests Action_none
1284 // no trap state is changed, no compiled code is changed, and the
1285 // computation suffers along in the interpreter.
1286 //
1287 // The other action codes specify various tactics for decompilation
1288 // and recompilation. Action_maybe_recompile is the loosest, and
1289 // allows the compiled code to stay around until enough traps are seen,
1290 // and until the compiler gets around to recompiling the trapping method.
1291 //
1292 // The other actions cause immediate removal of the present code.
1294 bool update_trap_state = true;
1295 bool make_not_entrant = false;
1296 bool make_not_compilable = false;
1297 bool reset_counters = false;
1298 switch (action) {
1299 case Action_none:
1300 // Keep the old code.
1301 update_trap_state = false;
1302 break;
1303 case Action_maybe_recompile:
1304 // Do not need to invalidate the present code, but we can
1305 // initiate another
1306 // Start compiler without (necessarily) invalidating the nmethod.
1307 // The system will tolerate the old code, but new code should be
1308 // generated when possible.
1309 break;
1310 case Action_reinterpret:
1311 // Go back into the interpreter for a while, and then consider
1312 // recompiling form scratch.
1313 make_not_entrant = true;
1314 // Reset invocation counter for outer most method.
1315 // This will allow the interpreter to exercise the bytecodes
1316 // for a while before recompiling.
1317 // By contrast, Action_make_not_entrant is immediate.
1318 //
1319 // Note that the compiler will track null_check, null_assert,
1320 // range_check, and class_check events and log them as if they
1321 // had been traps taken from compiled code. This will update
1322 // the MDO trap history so that the next compilation will
1323 // properly detect hot trap sites.
1324 reset_counters = true;
1325 break;
1326 case Action_make_not_entrant:
1327 // Request immediate recompilation, and get rid of the old code.
1328 // Make them not entrant, so next time they are called they get
1329 // recompiled. Unloaded classes are loaded now so recompile before next
1330 // time they are called. Same for uninitialized. The interpreter will
1331 // link the missing class, if any.
1332 make_not_entrant = true;
1333 break;
1334 case Action_make_not_compilable:
1335 // Give up on compiling this method at all.
1336 make_not_entrant = true;
1337 make_not_compilable = true;
1338 break;
1339 default:
1340 ShouldNotReachHere();
1341 }
1343 // Setting +ProfileTraps fixes the following, on all platforms:
1344 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1345 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1346 // recompile relies on a methodDataOop to record heroic opt failures.
1348 // Whether the interpreter is producing MDO data or not, we also need
1349 // to use the MDO to detect hot deoptimization points and control
1350 // aggressive optimization.
1351 bool inc_recompile_count = false;
1352 ProfileData* pdata = NULL;
1353 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
1354 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
1355 uint this_trap_count = 0;
1356 bool maybe_prior_trap = false;
1357 bool maybe_prior_recompile = false;
1358 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1359 //outputs:
1360 this_trap_count,
1361 maybe_prior_trap,
1362 maybe_prior_recompile);
1363 // Because the interpreter also counts null, div0, range, and class
1364 // checks, these traps from compiled code are double-counted.
1365 // This is harmless; it just means that the PerXTrapLimit values
1366 // are in effect a little smaller than they look.
1368 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1369 if (per_bc_reason != Reason_none) {
1370 // Now take action based on the partially known per-BCI history.
1371 if (maybe_prior_trap
1372 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1373 // If there are too many traps at this BCI, force a recompile.
1374 // This will allow the compiler to see the limit overflow, and
1375 // take corrective action, if possible. The compiler generally
1376 // does not use the exact PerBytecodeTrapLimit value, but instead
1377 // changes its tactics if it sees any traps at all. This provides
1378 // a little hysteresis, delaying a recompile until a trap happens
1379 // several times.
1380 //
1381 // Actually, since there is only one bit of counter per BCI,
1382 // the possible per-BCI counts are {0,1,(per-method count)}.
1383 // This produces accurate results if in fact there is only
1384 // one hot trap site, but begins to get fuzzy if there are
1385 // many sites. For example, if there are ten sites each
1386 // trapping two or more times, they each get the blame for
1387 // all of their traps.
1388 make_not_entrant = true;
1389 }
1391 // Detect repeated recompilation at the same BCI, and enforce a limit.
1392 if (make_not_entrant && maybe_prior_recompile) {
1393 // More than one recompile at this point.
1394 inc_recompile_count = maybe_prior_trap;
1395 }
1396 } else {
1397 // For reasons which are not recorded per-bytecode, we simply
1398 // force recompiles unconditionally.
1399 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1400 make_not_entrant = true;
1401 }
1403 // Go back to the compiler if there are too many traps in this method.
1404 if (this_trap_count >= (uint)PerMethodTrapLimit) {
1405 // If there are too many traps in this method, force a recompile.
1406 // This will allow the compiler to see the limit overflow, and
1407 // take corrective action, if possible.
1408 // (This condition is an unlikely backstop only, because the
1409 // PerBytecodeTrapLimit is more likely to take effect first,
1410 // if it is applicable.)
1411 make_not_entrant = true;
1412 }
1414 // Here's more hysteresis: If there has been a recompile at
1415 // this trap point already, run the method in the interpreter
1416 // for a while to exercise it more thoroughly.
1417 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1418 reset_counters = true;
1419 }
1421 }
1423 // Take requested actions on the method:
1425 // Recompile
1426 if (make_not_entrant) {
1427 if (!nm->make_not_entrant()) {
1428 return; // the call did not change nmethod's state
1429 }
1431 if (pdata != NULL) {
1432 // Record the recompilation event, if any.
1433 int tstate0 = pdata->trap_state();
1434 int tstate1 = trap_state_set_recompiled(tstate0, true);
1435 if (tstate1 != tstate0)
1436 pdata->set_trap_state(tstate1);
1437 }
1438 }
1440 if (inc_recompile_count) {
1441 trap_mdo->inc_overflow_recompile_count();
1442 if ((uint)trap_mdo->overflow_recompile_count() >
1443 (uint)PerBytecodeRecompilationCutoff) {
1444 // Give up on the method containing the bad BCI.
1445 if (trap_method() == nm->method()) {
1446 make_not_compilable = true;
1447 } else {
1448 trap_method->set_not_compilable();
1449 // But give grace to the enclosing nm->method().
1450 }
1451 }
1452 }
1454 // Reset invocation counters
1455 if (reset_counters) {
1456 if (nm->is_osr_method())
1457 reset_invocation_counter(trap_scope, CompileThreshold);
1458 else
1459 reset_invocation_counter(trap_scope);
1460 }
1462 // Give up compiling
1463 if (make_not_compilable && !nm->method()->is_not_compilable()) {
1464 assert(make_not_entrant, "consistent");
1465 nm->method()->set_not_compilable();
1466 }
1468 } // Free marked resources
1470 }
1471 JRT_END
1473 methodDataOop
1474 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1475 bool create_if_missing) {
1476 Thread* THREAD = thread;
1477 methodDataOop mdo = m()->method_data();
1478 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1479 // Build an MDO. Ignore errors like OutOfMemory;
1480 // that simply means we won't have an MDO to update.
1481 methodOopDesc::build_interpreter_method_data(m, THREAD);
1482 if (HAS_PENDING_EXCEPTION) {
1483 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1484 CLEAR_PENDING_EXCEPTION;
1485 }
1486 mdo = m()->method_data();
1487 }
1488 return mdo;
1489 }
1491 ProfileData*
1492 Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
1493 int trap_bci,
1494 Deoptimization::DeoptReason reason,
1495 //outputs:
1496 uint& ret_this_trap_count,
1497 bool& ret_maybe_prior_trap,
1498 bool& ret_maybe_prior_recompile) {
1499 uint prior_trap_count = trap_mdo->trap_count(reason);
1500 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1502 // If the runtime cannot find a place to store trap history,
1503 // it is estimated based on the general condition of the method.
1504 // If the method has ever been recompiled, or has ever incurred
1505 // a trap with the present reason , then this BCI is assumed
1506 // (pessimistically) to be the culprit.
1507 bool maybe_prior_trap = (prior_trap_count != 0);
1508 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1509 ProfileData* pdata = NULL;
1512 // For reasons which are recorded per bytecode, we check per-BCI data.
1513 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1514 if (per_bc_reason != Reason_none) {
1515 // Find the profile data for this BCI. If there isn't one,
1516 // try to allocate one from the MDO's set of spares.
1517 // This will let us detect a repeated trap at this point.
1518 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
1520 if (pdata != NULL) {
1521 // Query the trap state of this profile datum.
1522 int tstate0 = pdata->trap_state();
1523 if (!trap_state_has_reason(tstate0, per_bc_reason))
1524 maybe_prior_trap = false;
1525 if (!trap_state_is_recompiled(tstate0))
1526 maybe_prior_recompile = false;
1528 // Update the trap state of this profile datum.
1529 int tstate1 = tstate0;
1530 // Record the reason.
1531 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1532 // Store the updated state on the MDO, for next time.
1533 if (tstate1 != tstate0)
1534 pdata->set_trap_state(tstate1);
1535 } else {
1536 if (LogCompilation && xtty != NULL) {
1537 ttyLocker ttyl;
1538 // Missing MDP? Leave a small complaint in the log.
1539 xtty->elem("missing_mdp bci='%d'", trap_bci);
1540 }
1541 }
1542 }
1544 // Return results:
1545 ret_this_trap_count = this_trap_count;
1546 ret_maybe_prior_trap = maybe_prior_trap;
1547 ret_maybe_prior_recompile = maybe_prior_recompile;
1548 return pdata;
1549 }
1551 void
1552 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1553 ResourceMark rm;
1554 // Ignored outputs:
1555 uint ignore_this_trap_count;
1556 bool ignore_maybe_prior_trap;
1557 bool ignore_maybe_prior_recompile;
1558 query_update_method_data(trap_mdo, trap_bci,
1559 (DeoptReason)reason,
1560 ignore_this_trap_count,
1561 ignore_maybe_prior_trap,
1562 ignore_maybe_prior_recompile);
1563 }
1565 void Deoptimization::reset_invocation_counter(ScopeDesc* trap_scope, jint top_count) {
1566 ScopeDesc* sd = trap_scope;
1567 for (; !sd->is_top(); sd = sd->sender()) {
1568 // Reset ICs of inlined methods, since they can trigger compilations also.
1569 sd->method()->invocation_counter()->reset();
1570 }
1571 InvocationCounter* c = sd->method()->invocation_counter();
1572 if (top_count != _no_count) {
1573 // It was an OSR method, so bump the count higher.
1574 c->set(c->state(), top_count);
1575 } else {
1576 c->reset();
1577 }
1578 sd->method()->backedge_counter()->reset();
1579 }
1581 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1583 // Still in Java no safepoints
1584 {
1585 // This enters VM and may safepoint
1586 uncommon_trap_inner(thread, trap_request);
1587 }
1588 return fetch_unroll_info_helper(thread);
1589 }
1591 // Local derived constants.
1592 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1593 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1594 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1596 //---------------------------trap_state_reason---------------------------------
1597 Deoptimization::DeoptReason
1598 Deoptimization::trap_state_reason(int trap_state) {
1599 // This assert provides the link between the width of DataLayout::trap_bits
1600 // and the encoding of "recorded" reasons. It ensures there are enough
1601 // bits to store all needed reasons in the per-BCI MDO profile.
1602 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1603 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1604 trap_state -= recompile_bit;
1605 if (trap_state == DS_REASON_MASK) {
1606 return Reason_many;
1607 } else {
1608 assert((int)Reason_none == 0, "state=0 => Reason_none");
1609 return (DeoptReason)trap_state;
1610 }
1611 }
1612 //-------------------------trap_state_has_reason-------------------------------
1613 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1614 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1615 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1616 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1617 trap_state -= recompile_bit;
1618 if (trap_state == DS_REASON_MASK) {
1619 return -1; // true, unspecifically (bottom of state lattice)
1620 } else if (trap_state == reason) {
1621 return 1; // true, definitely
1622 } else if (trap_state == 0) {
1623 return 0; // false, definitely (top of state lattice)
1624 } else {
1625 return 0; // false, definitely
1626 }
1627 }
1628 //-------------------------trap_state_add_reason-------------------------------
1629 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1630 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1631 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1632 trap_state -= recompile_bit;
1633 if (trap_state == DS_REASON_MASK) {
1634 return trap_state + recompile_bit; // already at state lattice bottom
1635 } else if (trap_state == reason) {
1636 return trap_state + recompile_bit; // the condition is already true
1637 } else if (trap_state == 0) {
1638 return reason + recompile_bit; // no condition has yet been true
1639 } else {
1640 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1641 }
1642 }
1643 //-----------------------trap_state_is_recompiled------------------------------
1644 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1645 return (trap_state & DS_RECOMPILE_BIT) != 0;
1646 }
1647 //-----------------------trap_state_set_recompiled-----------------------------
1648 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1649 if (z) return trap_state | DS_RECOMPILE_BIT;
1650 else return trap_state & ~DS_RECOMPILE_BIT;
1651 }
1652 //---------------------------format_trap_state---------------------------------
1653 // This is used for debugging and diagnostics, including hotspot.log output.
1654 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1655 int trap_state) {
1656 DeoptReason reason = trap_state_reason(trap_state);
1657 bool recomp_flag = trap_state_is_recompiled(trap_state);
1658 // Re-encode the state from its decoded components.
1659 int decoded_state = 0;
1660 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1661 decoded_state = trap_state_add_reason(decoded_state, reason);
1662 if (recomp_flag)
1663 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1664 // If the state re-encodes properly, format it symbolically.
1665 // Because this routine is used for debugging and diagnostics,
1666 // be robust even if the state is a strange value.
1667 size_t len;
1668 if (decoded_state != trap_state) {
1669 // Random buggy state that doesn't decode??
1670 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1671 } else {
1672 len = jio_snprintf(buf, buflen, "%s%s",
1673 trap_reason_name(reason),
1674 recomp_flag ? " recompiled" : "");
1675 }
1676 if (len >= buflen)
1677 buf[buflen-1] = '\0';
1678 return buf;
1679 }
1682 //--------------------------------statics--------------------------------------
1683 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1684 = Deoptimization::Action_reinterpret;
1685 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1686 // Note: Keep this in sync. with enum DeoptReason.
1687 "none",
1688 "null_check",
1689 "null_assert",
1690 "range_check",
1691 "class_check",
1692 "array_check",
1693 "intrinsic",
1694 "bimorphic",
1695 "unloaded",
1696 "uninitialized",
1697 "unreached",
1698 "unhandled",
1699 "constraint",
1700 "div0_check",
1701 "age",
1702 "predicate"
1703 };
1704 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1705 // Note: Keep this in sync. with enum DeoptAction.
1706 "none",
1707 "maybe_recompile",
1708 "reinterpret",
1709 "make_not_entrant",
1710 "make_not_compilable"
1711 };
1713 const char* Deoptimization::trap_reason_name(int reason) {
1714 if (reason == Reason_many) return "many";
1715 if ((uint)reason < Reason_LIMIT)
1716 return _trap_reason_name[reason];
1717 static char buf[20];
1718 sprintf(buf, "reason%d", reason);
1719 return buf;
1720 }
1721 const char* Deoptimization::trap_action_name(int action) {
1722 if ((uint)action < Action_LIMIT)
1723 return _trap_action_name[action];
1724 static char buf[20];
1725 sprintf(buf, "action%d", action);
1726 return buf;
1727 }
1729 // This is used for debugging and diagnostics, including hotspot.log output.
1730 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1731 int trap_request) {
1732 jint unloaded_class_index = trap_request_index(trap_request);
1733 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1734 const char* action = trap_action_name(trap_request_action(trap_request));
1735 size_t len;
1736 if (unloaded_class_index < 0) {
1737 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1738 reason, action);
1739 } else {
1740 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1741 reason, action, unloaded_class_index);
1742 }
1743 if (len >= buflen)
1744 buf[buflen-1] = '\0';
1745 return buf;
1746 }
1748 juint Deoptimization::_deoptimization_hist
1749 [Deoptimization::Reason_LIMIT]
1750 [1 + Deoptimization::Action_LIMIT]
1751 [Deoptimization::BC_CASE_LIMIT]
1752 = {0};
1754 enum {
1755 LSB_BITS = 8,
1756 LSB_MASK = right_n_bits(LSB_BITS)
1757 };
1759 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1760 Bytecodes::Code bc) {
1761 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1762 assert(action >= 0 && action < Action_LIMIT, "oob");
1763 _deoptimization_hist[Reason_none][0][0] += 1; // total
1764 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1765 juint* cases = _deoptimization_hist[reason][1+action];
1766 juint* bc_counter_addr = NULL;
1767 juint bc_counter = 0;
1768 // Look for an unused counter, or an exact match to this BC.
1769 if (bc != Bytecodes::_illegal) {
1770 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1771 juint* counter_addr = &cases[bc_case];
1772 juint counter = *counter_addr;
1773 if ((counter == 0 && bc_counter_addr == NULL)
1774 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1775 // this counter is either free or is already devoted to this BC
1776 bc_counter_addr = counter_addr;
1777 bc_counter = counter | bc;
1778 }
1779 }
1780 }
1781 if (bc_counter_addr == NULL) {
1782 // Overflow, or no given bytecode.
1783 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1784 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1785 }
1786 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1787 }
1789 jint Deoptimization::total_deoptimization_count() {
1790 return _deoptimization_hist[Reason_none][0][0];
1791 }
1793 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1794 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1795 return _deoptimization_hist[reason][0][0];
1796 }
1798 void Deoptimization::print_statistics() {
1799 juint total = total_deoptimization_count();
1800 juint account = total;
1801 if (total != 0) {
1802 ttyLocker ttyl;
1803 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1804 tty->print_cr("Deoptimization traps recorded:");
1805 #define PRINT_STAT_LINE(name, r) \
1806 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1807 PRINT_STAT_LINE("total", total);
1808 // For each non-zero entry in the histogram, print the reason,
1809 // the action, and (if specifically known) the type of bytecode.
1810 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1811 for (int action = 0; action < Action_LIMIT; action++) {
1812 juint* cases = _deoptimization_hist[reason][1+action];
1813 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1814 juint counter = cases[bc_case];
1815 if (counter != 0) {
1816 char name[1*K];
1817 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1818 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1819 bc = Bytecodes::_illegal;
1820 sprintf(name, "%s/%s/%s",
1821 trap_reason_name(reason),
1822 trap_action_name(action),
1823 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1824 juint r = counter >> LSB_BITS;
1825 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1826 account -= r;
1827 }
1828 }
1829 }
1830 }
1831 if (account != 0) {
1832 PRINT_STAT_LINE("unaccounted", account);
1833 }
1834 #undef PRINT_STAT_LINE
1835 if (xtty != NULL) xtty->tail("statistics");
1836 }
1837 }
1838 #else // COMPILER2
1841 // Stubs for C1 only system.
1842 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1843 return false;
1844 }
1846 const char* Deoptimization::trap_reason_name(int reason) {
1847 return "unknown";
1848 }
1850 void Deoptimization::print_statistics() {
1851 // no output
1852 }
1854 void
1855 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1856 // no udpate
1857 }
1859 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1860 return 0;
1861 }
1863 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1864 Bytecodes::Code bc) {
1865 // no update
1866 }
1868 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1869 int trap_state) {
1870 jio_snprintf(buf, buflen, "#%d", trap_state);
1871 return buf;
1872 }
1874 #endif // COMPILER2