Wed, 13 Mar 2013 09:44:45 +0100
8009761: Deoptimization on sparc doesn't set Llast_SP correctly in the interpreter frames it creates
Summary: deoptimization doesn't set up callee frames so that they restore caller frames correctly.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2012, 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/method.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "prims/jvmtiThreadState.hpp"
40 #include "runtime/biasedLocking.hpp"
41 #include "runtime/compilationPolicy.hpp"
42 #include "runtime/deoptimization.hpp"
43 #include "runtime/interfaceSupport.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/thread.hpp"
48 #include "runtime/vframe.hpp"
49 #include "runtime/vframeArray.hpp"
50 #include "runtime/vframe_hp.hpp"
51 #include "utilities/events.hpp"
52 #include "utilities/xmlstream.hpp"
53 #ifdef TARGET_ARCH_x86
54 # include "vmreg_x86.inline.hpp"
55 #endif
56 #ifdef TARGET_ARCH_sparc
57 # include "vmreg_sparc.inline.hpp"
58 #endif
59 #ifdef TARGET_ARCH_zero
60 # include "vmreg_zero.inline.hpp"
61 #endif
62 #ifdef TARGET_ARCH_arm
63 # include "vmreg_arm.inline.hpp"
64 #endif
65 #ifdef TARGET_ARCH_ppc
66 # include "vmreg_ppc.inline.hpp"
67 #endif
68 #ifdef COMPILER2
69 #ifdef TARGET_ARCH_MODEL_x86_32
70 # include "adfiles/ad_x86_32.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_x86_64
73 # include "adfiles/ad_x86_64.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_sparc
76 # include "adfiles/ad_sparc.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_arm
82 # include "adfiles/ad_arm.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc
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 caller_actual_parameters,
94 int number_of_frames,
95 intptr_t* frame_sizes,
96 address* frame_pcs,
97 BasicType return_type) {
98 _size_of_deoptimized_frame = size_of_deoptimized_frame;
99 _caller_adjustment = caller_adjustment;
100 _caller_actual_parameters = caller_actual_parameters;
101 _number_of_frames = number_of_frames;
102 _frame_sizes = frame_sizes;
103 _frame_pcs = frame_pcs;
104 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
105 _return_type = return_type;
106 _initial_info = 0;
107 // PD (x86 only)
108 _counter_temp = 0;
109 _unpack_kind = 0;
110 _sender_sp_temp = 0;
112 _total_frame_sizes = size_of_frames();
113 }
116 Deoptimization::UnrollBlock::~UnrollBlock() {
117 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
118 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
119 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
120 }
123 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
124 assert(register_number < RegisterMap::reg_count, "checking register number");
125 return &_register_block[register_number * 2];
126 }
130 int Deoptimization::UnrollBlock::size_of_frames() const {
131 // Acount first for the adjustment of the initial frame
132 int result = _caller_adjustment;
133 for (int index = 0; index < number_of_frames(); index++) {
134 result += frame_sizes()[index];
135 }
136 return result;
137 }
140 void Deoptimization::UnrollBlock::print() {
141 ttyLocker ttyl;
142 tty->print_cr("UnrollBlock");
143 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
144 tty->print( " frame_sizes: ");
145 for (int index = 0; index < number_of_frames(); index++) {
146 tty->print("%d ", frame_sizes()[index]);
147 }
148 tty->cr();
149 }
152 // In order to make fetch_unroll_info work properly with escape
153 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
154 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
155 // of previously eliminated objects occurs in realloc_objects, which is
156 // called from the method fetch_unroll_info_helper below.
157 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
158 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
159 // but makes the entry a little slower. There is however a little dance we have to
160 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
162 // fetch_unroll_info() is called at the beginning of the deoptimization
163 // handler. Note this fact before we start generating temporary frames
164 // that can confuse an asynchronous stack walker. This counter is
165 // decremented at the end of unpack_frames().
166 thread->inc_in_deopt_handler();
168 return fetch_unroll_info_helper(thread);
169 JRT_END
172 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
173 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
175 // Note: there is a safepoint safety issue here. No matter whether we enter
176 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
177 // the vframeArray is created.
178 //
180 // Allocate our special deoptimization ResourceMark
181 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
182 assert(thread->deopt_mark() == NULL, "Pending deopt!");
183 thread->set_deopt_mark(dmark);
185 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
186 RegisterMap map(thread, true);
187 RegisterMap dummy_map(thread, false);
188 // Now get the deoptee with a valid map
189 frame deoptee = stub_frame.sender(&map);
190 // Set the deoptee nmethod
191 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
192 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
194 if (VerifyStack) {
195 thread->validate_frame_layout();
196 }
198 // Create a growable array of VFrames where each VFrame represents an inlined
199 // Java frame. This storage is allocated with the usual system arena.
200 assert(deoptee.is_compiled_frame(), "Wrong frame type");
201 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
202 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
203 while (!vf->is_top()) {
204 assert(vf->is_compiled_frame(), "Wrong frame type");
205 chunk->push(compiledVFrame::cast(vf));
206 vf = vf->sender();
207 }
208 assert(vf->is_compiled_frame(), "Wrong frame type");
209 chunk->push(compiledVFrame::cast(vf));
211 #ifdef COMPILER2
212 // Reallocate the non-escaping objects and restore their fields. Then
213 // relock objects if synchronization on them was eliminated.
214 if (DoEscapeAnalysis || EliminateNestedLocks) {
215 if (EliminateAllocations) {
216 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
217 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
219 // The flag return_oop() indicates call sites which return oop
220 // in compiled code. Such sites include java method calls,
221 // runtime calls (for example, used to allocate new objects/arrays
222 // on slow code path) and any other calls generated in compiled code.
223 // It is not guaranteed that we can get such information here only
224 // by analyzing bytecode in deoptimized frames. This is why this flag
225 // is set during method compilation (see Compile::Process_OopMap_Node()).
226 bool save_oop_result = chunk->at(0)->scope()->return_oop();
227 Handle return_value;
228 if (save_oop_result) {
229 // Reallocation may trigger GC. If deoptimization happened on return from
230 // call which returns oop we need to save it since it is not in oopmap.
231 oop result = deoptee.saved_oop_result(&map);
232 assert(result == NULL || result->is_oop(), "must be oop");
233 return_value = Handle(thread, result);
234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
235 if (TraceDeoptimization) {
236 ttyLocker ttyl;
237 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
238 }
239 }
240 bool reallocated = false;
241 if (objects != NULL) {
242 JRT_BLOCK
243 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
244 JRT_END
245 }
246 if (reallocated) {
247 reassign_fields(&deoptee, &map, objects);
248 #ifndef PRODUCT
249 if (TraceDeoptimization) {
250 ttyLocker ttyl;
251 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
252 print_objects(objects);
253 }
254 #endif
255 }
256 if (save_oop_result) {
257 // Restore result.
258 deoptee.set_saved_oop_result(&map, return_value());
259 }
260 }
261 if (EliminateLocks) {
262 #ifndef PRODUCT
263 bool first = true;
264 #endif
265 for (int i = 0; i < chunk->length(); i++) {
266 compiledVFrame* cvf = chunk->at(i);
267 assert (cvf->scope() != NULL,"expect only compiled java frames");
268 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
269 if (monitors->is_nonempty()) {
270 relock_objects(monitors, thread);
271 #ifndef PRODUCT
272 if (TraceDeoptimization) {
273 ttyLocker ttyl;
274 for (int j = 0; j < monitors->length(); j++) {
275 MonitorInfo* mi = monitors->at(j);
276 if (mi->eliminated()) {
277 if (first) {
278 first = false;
279 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
280 }
281 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
282 }
283 }
284 }
285 #endif
286 }
287 }
288 }
289 }
290 #endif // COMPILER2
291 // Ensure that no safepoint is taken after pointers have been stored
292 // in fields of rematerialized objects. If a safepoint occurs from here on
293 // out the java state residing in the vframeArray will be missed.
294 No_Safepoint_Verifier no_safepoint;
296 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
298 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
299 thread->set_vframe_array_head(array);
301 // Now that the vframeArray has been created if we have any deferred local writes
302 // added by jvmti then we can free up that structure as the data is now in the
303 // vframeArray
305 if (thread->deferred_locals() != NULL) {
306 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
307 int i = 0;
308 do {
309 // Because of inlining we could have multiple vframes for a single frame
310 // and several of the vframes could have deferred writes. Find them all.
311 if (list->at(i)->id() == array->original().id()) {
312 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
313 list->remove_at(i);
314 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
315 delete dlv;
316 } else {
317 i++;
318 }
319 } while ( i < list->length() );
320 if (list->length() == 0) {
321 thread->set_deferred_locals(NULL);
322 // free the list and elements back to C heap.
323 delete list;
324 }
326 }
328 #ifndef SHARK
329 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
330 CodeBlob* cb = stub_frame.cb();
331 // Verify we have the right vframeArray
332 assert(cb->frame_size() >= 0, "Unexpected frame size");
333 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
335 // If the deopt call site is a MethodHandle invoke call site we have
336 // to adjust the unpack_sp.
337 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
338 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
339 unpack_sp = deoptee.unextended_sp();
341 #ifdef ASSERT
342 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
343 #endif
344 #else
345 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
346 #endif // !SHARK
348 // This is a guarantee instead of an assert because if vframe doesn't match
349 // we will unpack the wrong deoptimized frame and wind up in strange places
350 // where it will be very difficult to figure out what went wrong. Better
351 // to die an early death here than some very obscure death later when the
352 // trail is cold.
353 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
354 // in that it will fail to detect a problem when there is one. This needs
355 // more work in tiger timeframe.
356 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
358 int number_of_frames = array->frames();
360 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
361 // virtual activation, which is the reverse of the elements in the vframes array.
362 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
363 // +1 because we always have an interpreter return address for the final slot.
364 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
365 int popframe_extra_args = 0;
366 // Create an interpreter return address for the stub to use as its return
367 // address so the skeletal frames are perfectly walkable
368 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
370 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
371 // activation be put back on the expression stack of the caller for reexecution
372 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
373 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
374 }
376 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
377 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
378 // than simply use array->sender.pc(). This requires us to walk the current set of frames
379 //
380 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
381 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
383 // It's possible that the number of paramters at the call site is
384 // different than number of arguments in the callee when method
385 // handles are used. If the caller is interpreted get the real
386 // value so that the proper amount of space can be added to it's
387 // frame.
388 bool caller_was_method_handle = false;
389 if (deopt_sender.is_interpreted_frame()) {
390 methodHandle method = deopt_sender.interpreter_frame_method();
391 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
392 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
393 // Method handle invokes may involve fairly arbitrary chains of
394 // calls so it's impossible to know how much actual space the
395 // caller has for locals.
396 caller_was_method_handle = true;
397 }
398 }
400 //
401 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
402 // frame_sizes/frame_pcs[1] next oldest frame (int)
403 // frame_sizes/frame_pcs[n] youngest frame (int)
404 //
405 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
406 // owns the space for the return address to it's caller). Confusing ain't it.
407 //
408 // The vframe array can address vframes with indices running from
409 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
410 // When we create the skeletal frames we need the oldest frame to be in the zero slot
411 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
412 // so things look a little strange in this loop.
413 //
414 int callee_parameters = 0;
415 int callee_locals = 0;
416 for (int index = 0; index < array->frames(); index++ ) {
417 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
418 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
419 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
420 int caller_parms = callee_parameters;
421 if ((index == array->frames() - 1) && caller_was_method_handle) {
422 caller_parms = 0;
423 }
424 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms,
425 callee_parameters,
426 callee_locals,
427 index == 0,
428 index == array->frames() - 1,
429 popframe_extra_args);
430 // This pc doesn't have to be perfect just good enough to identify the frame
431 // as interpreted so the skeleton frame will be walkable
432 // The correct pc will be set when the skeleton frame is completely filled out
433 // The final pc we store in the loop is wrong and will be overwritten below
434 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
436 callee_parameters = array->element(index)->method()->size_of_parameters();
437 callee_locals = array->element(index)->method()->max_locals();
438 popframe_extra_args = 0;
439 }
441 // Compute whether the root vframe returns a float or double value.
442 BasicType return_type;
443 {
444 HandleMark hm;
445 methodHandle method(thread, array->element(0)->method());
446 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
447 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
448 }
450 // Compute information for handling adapters and adjusting the frame size of the caller.
451 int caller_adjustment = 0;
453 // Compute the amount the oldest interpreter frame will have to adjust
454 // its caller's stack by. If the caller is a compiled frame then
455 // we pretend that the callee has no parameters so that the
456 // extension counts for the full amount of locals and not just
457 // locals-parms. This is because without a c2i adapter the parm
458 // area as created by the compiled frame will not be usable by
459 // the interpreter. (Depending on the calling convention there
460 // may not even be enough space).
462 // QQQ I'd rather see this pushed down into last_frame_adjust
463 // and have it take the sender (aka caller).
465 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
466 caller_adjustment = last_frame_adjust(0, callee_locals);
467 } else if (callee_locals > callee_parameters) {
468 // The caller frame may need extending to accommodate
469 // non-parameter locals of the first unpacked interpreted frame.
470 // Compute that adjustment.
471 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
472 }
474 // If the sender is deoptimized the we must retrieve the address of the handler
475 // since the frame will "magically" show the original pc before the deopt
476 // and we'd undo the deopt.
478 frame_pcs[0] = deopt_sender.raw_pc();
480 #ifndef SHARK
481 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
482 #endif // SHARK
484 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
485 caller_adjustment * BytesPerWord,
486 caller_was_method_handle ? 0 : callee_parameters,
487 number_of_frames,
488 frame_sizes,
489 frame_pcs,
490 return_type);
491 // On some platforms, we need a way to pass some platform dependent
492 // information to the unpacking code so the skeletal frames come out
493 // correct (initial fp value, unextended sp, ...)
494 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
496 if (array->frames() > 1) {
497 if (VerifyStack && TraceDeoptimization) {
498 ttyLocker ttyl;
499 tty->print_cr("Deoptimizing method containing inlining");
500 }
501 }
503 array->set_unroll_block(info);
504 return info;
505 }
507 // Called to cleanup deoptimization data structures in normal case
508 // after unpacking to stack and when stack overflow error occurs
509 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
510 vframeArray *array) {
512 // Get array if coming from exception
513 if (array == NULL) {
514 array = thread->vframe_array_head();
515 }
516 thread->set_vframe_array_head(NULL);
518 // Free the previous UnrollBlock
519 vframeArray* old_array = thread->vframe_array_last();
520 thread->set_vframe_array_last(array);
522 if (old_array != NULL) {
523 UnrollBlock* old_info = old_array->unroll_block();
524 old_array->set_unroll_block(NULL);
525 delete old_info;
526 delete old_array;
527 }
529 // Deallocate any resource creating in this routine and any ResourceObjs allocated
530 // inside the vframeArray (StackValueCollections)
532 delete thread->deopt_mark();
533 thread->set_deopt_mark(NULL);
534 thread->set_deopt_nmethod(NULL);
537 if (JvmtiExport::can_pop_frame()) {
538 #ifndef CC_INTERP
539 // Regardless of whether we entered this routine with the pending
540 // popframe condition bit set, we should always clear it now
541 thread->clear_popframe_condition();
542 #else
543 // C++ interpeter will clear has_pending_popframe when it enters
544 // with method_resume. For deopt_resume2 we clear it now.
545 if (thread->popframe_forcing_deopt_reexecution())
546 thread->clear_popframe_condition();
547 #endif /* CC_INTERP */
548 }
550 // unpack_frames() is called at the end of the deoptimization handler
551 // and (in C2) at the end of the uncommon trap handler. Note this fact
552 // so that an asynchronous stack walker can work again. This counter is
553 // incremented at the beginning of fetch_unroll_info() and (in C2) at
554 // the beginning of uncommon_trap().
555 thread->dec_in_deopt_handler();
556 }
559 // Return BasicType of value being returned
560 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
562 // We are already active int he special DeoptResourceMark any ResourceObj's we
563 // allocate will be freed at the end of the routine.
565 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
566 // but makes the entry a little slower. There is however a little dance we have to
567 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
568 ResetNoHandleMark rnhm; // No-op in release/product versions
569 HandleMark hm;
571 frame stub_frame = thread->last_frame();
573 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
574 // must point to the vframeArray for the unpack frame.
575 vframeArray* array = thread->vframe_array_head();
577 #ifndef PRODUCT
578 if (TraceDeoptimization) {
579 ttyLocker ttyl;
580 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
581 }
582 #endif
583 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
584 stub_frame.pc(), stub_frame.sp(), exec_mode);
586 UnrollBlock* info = array->unroll_block();
588 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
589 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
591 BasicType bt = info->return_type();
593 // If we have an exception pending, claim that the return type is an oop
594 // so the deopt_blob does not overwrite the exception_oop.
596 if (exec_mode == Unpack_exception)
597 bt = T_OBJECT;
599 // Cleanup thread deopt data
600 cleanup_deopt_info(thread, array);
602 #ifndef PRODUCT
603 if (VerifyStack) {
604 ResourceMark res_mark;
606 thread->validate_frame_layout();
608 // Verify that the just-unpacked frames match the interpreter's
609 // notions of expression stack and locals
610 vframeArray* cur_array = thread->vframe_array_last();
611 RegisterMap rm(thread, false);
612 rm.set_include_argument_oops(false);
613 bool is_top_frame = true;
614 int callee_size_of_parameters = 0;
615 int callee_max_locals = 0;
616 for (int i = 0; i < cur_array->frames(); i++) {
617 vframeArrayElement* el = cur_array->element(i);
618 frame* iframe = el->iframe();
619 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
621 // Get the oop map for this bci
622 InterpreterOopMap mask;
623 int cur_invoke_parameter_size = 0;
624 bool try_next_mask = false;
625 int next_mask_expression_stack_size = -1;
626 int top_frame_expression_stack_adjustment = 0;
627 methodHandle mh(thread, iframe->interpreter_frame_method());
628 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
629 BytecodeStream str(mh);
630 str.set_start(iframe->interpreter_frame_bci());
631 int max_bci = mh->code_size();
632 // Get to the next bytecode if possible
633 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
634 // Check to see if we can grab the number of outgoing arguments
635 // at an uncommon trap for an invoke (where the compiler
636 // generates debug info before the invoke has executed)
637 Bytecodes::Code cur_code = str.next();
638 if (cur_code == Bytecodes::_invokevirtual ||
639 cur_code == Bytecodes::_invokespecial ||
640 cur_code == Bytecodes::_invokestatic ||
641 cur_code == Bytecodes::_invokeinterface) {
642 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
643 Symbol* signature = invoke.signature();
644 ArgumentSizeComputer asc(signature);
645 cur_invoke_parameter_size = asc.size();
646 if (cur_code != Bytecodes::_invokestatic) {
647 // Add in receiver
648 ++cur_invoke_parameter_size;
649 }
650 }
651 if (str.bci() < max_bci) {
652 Bytecodes::Code bc = str.next();
653 if (bc >= 0) {
654 // The interpreter oop map generator reports results before
655 // the current bytecode has executed except in the case of
656 // calls. It seems to be hard to tell whether the compiler
657 // has emitted debug information matching the "state before"
658 // a given bytecode or the state after, so we try both
659 switch (cur_code) {
660 case Bytecodes::_invokevirtual:
661 case Bytecodes::_invokespecial:
662 case Bytecodes::_invokestatic:
663 case Bytecodes::_invokeinterface:
664 case Bytecodes::_athrow:
665 break;
666 default: {
667 InterpreterOopMap next_mask;
668 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
669 next_mask_expression_stack_size = next_mask.expression_stack_size();
670 // Need to subtract off the size of the result type of
671 // the bytecode because this is not described in the
672 // debug info but returned to the interpreter in the TOS
673 // caching register
674 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
675 if (bytecode_result_type != T_ILLEGAL) {
676 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
677 }
678 assert(top_frame_expression_stack_adjustment >= 0, "");
679 try_next_mask = true;
680 break;
681 }
682 }
683 }
684 }
686 // Verify stack depth and oops in frame
687 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
688 if (!(
689 /* SPARC */
690 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
691 /* x86 */
692 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
693 (try_next_mask &&
694 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
695 top_frame_expression_stack_adjustment))) ||
696 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
697 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
698 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
699 )) {
700 ttyLocker ttyl;
702 // Print out some information that will help us debug the problem
703 tty->print_cr("Wrong number of expression stack elements during deoptimization");
704 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
705 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
706 iframe->interpreter_frame_expression_stack_size());
707 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
708 tty->print_cr(" try_next_mask = %d", try_next_mask);
709 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
710 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
711 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
712 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
713 tty->print_cr(" exec_mode = %d", exec_mode);
714 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
715 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
716 tty->print_cr(" Interpreted frames:");
717 for (int k = 0; k < cur_array->frames(); k++) {
718 vframeArrayElement* el = cur_array->element(k);
719 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
720 }
721 cur_array->print_on_2(tty);
722 guarantee(false, "wrong number of expression stack elements during deopt");
723 }
724 VerifyOopClosure verify;
725 iframe->oops_interpreted_do(&verify, NULL, &rm, false);
726 callee_size_of_parameters = mh->size_of_parameters();
727 callee_max_locals = mh->max_locals();
728 is_top_frame = false;
729 }
730 }
731 #endif /* !PRODUCT */
734 return bt;
735 JRT_END
738 int Deoptimization::deoptimize_dependents() {
739 Threads::deoptimized_wrt_marked_nmethods();
740 return 0;
741 }
744 #ifdef COMPILER2
745 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
746 Handle pending_exception(thread->pending_exception());
747 const char* exception_file = thread->exception_file();
748 int exception_line = thread->exception_line();
749 thread->clear_pending_exception();
751 for (int i = 0; i < objects->length(); i++) {
752 assert(objects->at(i)->is_object(), "invalid debug information");
753 ObjectValue* sv = (ObjectValue*) objects->at(i);
755 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
756 oop obj = NULL;
758 if (k->oop_is_instance()) {
759 InstanceKlass* ik = InstanceKlass::cast(k());
760 obj = ik->allocate_instance(CHECK_(false));
761 } else if (k->oop_is_typeArray()) {
762 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
763 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
764 int len = sv->field_size() / type2size[ak->element_type()];
765 obj = ak->allocate(len, CHECK_(false));
766 } else if (k->oop_is_objArray()) {
767 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
768 obj = ak->allocate(sv->field_size(), CHECK_(false));
769 }
771 assert(obj != NULL, "allocation failed");
772 assert(sv->value().is_null(), "redundant reallocation");
773 sv->set_value(obj);
774 }
776 if (pending_exception.not_null()) {
777 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
778 }
780 return true;
781 }
783 // This assumes that the fields are stored in ObjectValue in the same order
784 // they are yielded by do_nonstatic_fields.
785 class FieldReassigner: public FieldClosure {
786 frame* _fr;
787 RegisterMap* _reg_map;
788 ObjectValue* _sv;
789 InstanceKlass* _ik;
790 oop _obj;
792 int _i;
793 public:
794 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
795 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
797 int i() const { return _i; }
800 void do_field(fieldDescriptor* fd) {
801 intptr_t val;
802 StackValue* value =
803 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
804 int offset = fd->offset();
805 switch (fd->field_type()) {
806 case T_OBJECT: case T_ARRAY:
807 assert(value->type() == T_OBJECT, "Agreement.");
808 _obj->obj_field_put(offset, value->get_obj()());
809 break;
811 case T_LONG: case T_DOUBLE: {
812 assert(value->type() == T_INT, "Agreement.");
813 StackValue* low =
814 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
815 #ifdef _LP64
816 jlong res = (jlong)low->get_int();
817 #else
818 #ifdef SPARC
819 // For SPARC we have to swap high and low words.
820 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
821 #else
822 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
823 #endif //SPARC
824 #endif
825 _obj->long_field_put(offset, res);
826 break;
827 }
828 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
829 case T_INT: case T_FLOAT: // 4 bytes.
830 assert(value->type() == T_INT, "Agreement.");
831 val = value->get_int();
832 _obj->int_field_put(offset, (jint)*((jint*)&val));
833 break;
835 case T_SHORT: case T_CHAR: // 2 bytes
836 assert(value->type() == T_INT, "Agreement.");
837 val = value->get_int();
838 _obj->short_field_put(offset, (jshort)*((jint*)&val));
839 break;
841 case T_BOOLEAN: case T_BYTE: // 1 byte
842 assert(value->type() == T_INT, "Agreement.");
843 val = value->get_int();
844 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
845 break;
847 default:
848 ShouldNotReachHere();
849 }
850 _i++;
851 }
852 };
854 // restore elements of an eliminated type array
855 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
856 int index = 0;
857 intptr_t val;
859 for (int i = 0; i < sv->field_size(); i++) {
860 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
861 switch(type) {
862 case T_LONG: case T_DOUBLE: {
863 assert(value->type() == T_INT, "Agreement.");
864 StackValue* low =
865 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
866 #ifdef _LP64
867 jlong res = (jlong)low->get_int();
868 #else
869 #ifdef SPARC
870 // For SPARC we have to swap high and low words.
871 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
872 #else
873 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
874 #endif //SPARC
875 #endif
876 obj->long_at_put(index, res);
877 break;
878 }
880 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
881 case T_INT: case T_FLOAT: // 4 bytes.
882 assert(value->type() == T_INT, "Agreement.");
883 val = value->get_int();
884 obj->int_at_put(index, (jint)*((jint*)&val));
885 break;
887 case T_SHORT: case T_CHAR: // 2 bytes
888 assert(value->type() == T_INT, "Agreement.");
889 val = value->get_int();
890 obj->short_at_put(index, (jshort)*((jint*)&val));
891 break;
893 case T_BOOLEAN: case T_BYTE: // 1 byte
894 assert(value->type() == T_INT, "Agreement.");
895 val = value->get_int();
896 obj->bool_at_put(index, (jboolean)*((jint*)&val));
897 break;
899 default:
900 ShouldNotReachHere();
901 }
902 index++;
903 }
904 }
907 // restore fields of an eliminated object array
908 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
909 for (int i = 0; i < sv->field_size(); i++) {
910 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
911 assert(value->type() == T_OBJECT, "object element expected");
912 obj->obj_at_put(i, value->get_obj()());
913 }
914 }
917 // restore fields of all eliminated objects and arrays
918 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
919 for (int i = 0; i < objects->length(); i++) {
920 ObjectValue* sv = (ObjectValue*) objects->at(i);
921 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
922 Handle obj = sv->value();
923 assert(obj.not_null(), "reallocation was missed");
925 if (k->oop_is_instance()) {
926 InstanceKlass* ik = InstanceKlass::cast(k());
927 FieldReassigner reassign(fr, reg_map, sv, obj());
928 ik->do_nonstatic_fields(&reassign);
929 } else if (k->oop_is_typeArray()) {
930 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
931 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
932 } else if (k->oop_is_objArray()) {
933 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
934 }
935 }
936 }
939 // relock objects for which synchronization was eliminated
940 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
941 for (int i = 0; i < monitors->length(); i++) {
942 MonitorInfo* mon_info = monitors->at(i);
943 if (mon_info->eliminated()) {
944 assert(mon_info->owner() != NULL, "reallocation was missed");
945 Handle obj = Handle(mon_info->owner());
946 markOop mark = obj->mark();
947 if (UseBiasedLocking && mark->has_bias_pattern()) {
948 // New allocated objects may have the mark set to anonymously biased.
949 // Also the deoptimized method may called methods with synchronization
950 // where the thread-local object is bias locked to the current thread.
951 assert(mark->is_biased_anonymously() ||
952 mark->biased_locker() == thread, "should be locked to current thread");
953 // Reset mark word to unbiased prototype.
954 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
955 obj->set_mark(unbiased_prototype);
956 }
957 BasicLock* lock = mon_info->lock();
958 ObjectSynchronizer::slow_enter(obj, lock, thread);
959 }
960 assert(mon_info->owner()->is_locked(), "object must be locked now");
961 }
962 }
965 #ifndef PRODUCT
966 // print information about reallocated objects
967 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
968 fieldDescriptor fd;
970 for (int i = 0; i < objects->length(); i++) {
971 ObjectValue* sv = (ObjectValue*) objects->at(i);
972 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
973 Handle obj = sv->value();
975 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
976 k->print_value();
977 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
978 tty->cr();
980 if (Verbose) {
981 k->oop_print_on(obj(), tty);
982 }
983 }
984 }
985 #endif
986 #endif // COMPILER2
988 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
989 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
991 #ifndef PRODUCT
992 if (TraceDeoptimization) {
993 ttyLocker ttyl;
994 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
995 fr.print_on(tty);
996 tty->print_cr(" Virtual frames (innermost first):");
997 for (int index = 0; index < chunk->length(); index++) {
998 compiledVFrame* vf = chunk->at(index);
999 tty->print(" %2d - ", index);
1000 vf->print_value();
1001 int bci = chunk->at(index)->raw_bci();
1002 const char* code_name;
1003 if (bci == SynchronizationEntryBCI) {
1004 code_name = "sync entry";
1005 } else {
1006 Bytecodes::Code code = vf->method()->code_at(bci);
1007 code_name = Bytecodes::name(code);
1008 }
1009 tty->print(" - %s", code_name);
1010 tty->print_cr(" @ bci %d ", bci);
1011 if (Verbose) {
1012 vf->print();
1013 tty->cr();
1014 }
1015 }
1016 }
1017 #endif
1019 // Register map for next frame (used for stack crawl). We capture
1020 // the state of the deopt'ing frame's caller. Thus if we need to
1021 // stuff a C2I adapter we can properly fill in the callee-save
1022 // register locations.
1023 frame caller = fr.sender(reg_map);
1024 int frame_size = caller.sp() - fr.sp();
1026 frame sender = caller;
1028 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1029 // the vframeArray containing the unpacking information is allocated in the C heap.
1030 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1031 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
1033 // Compare the vframeArray to the collected vframes
1034 assert(array->structural_compare(thread, chunk), "just checking");
1036 #ifndef PRODUCT
1037 if (TraceDeoptimization) {
1038 ttyLocker ttyl;
1039 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1040 }
1041 #endif // PRODUCT
1043 return array;
1044 }
1047 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1048 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1049 for (int i = 0; i < monitors->length(); i++) {
1050 MonitorInfo* mon_info = monitors->at(i);
1051 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1052 objects_to_revoke->append(Handle(mon_info->owner()));
1053 }
1054 }
1055 }
1058 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1059 if (!UseBiasedLocking) {
1060 return;
1061 }
1063 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1065 // Unfortunately we don't have a RegisterMap available in most of
1066 // the places we want to call this routine so we need to walk the
1067 // stack again to update the register map.
1068 if (map == NULL || !map->update_map()) {
1069 StackFrameStream sfs(thread, true);
1070 bool found = false;
1071 while (!found && !sfs.is_done()) {
1072 frame* cur = sfs.current();
1073 sfs.next();
1074 found = cur->id() == fr.id();
1075 }
1076 assert(found, "frame to be deoptimized not found on target thread's stack");
1077 map = sfs.register_map();
1078 }
1080 vframe* vf = vframe::new_vframe(&fr, map, thread);
1081 compiledVFrame* cvf = compiledVFrame::cast(vf);
1082 // Revoke monitors' biases in all scopes
1083 while (!cvf->is_top()) {
1084 collect_monitors(cvf, objects_to_revoke);
1085 cvf = compiledVFrame::cast(cvf->sender());
1086 }
1087 collect_monitors(cvf, objects_to_revoke);
1089 if (SafepointSynchronize::is_at_safepoint()) {
1090 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1091 } else {
1092 BiasedLocking::revoke(objects_to_revoke);
1093 }
1094 }
1097 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1098 if (!UseBiasedLocking) {
1099 return;
1100 }
1102 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1103 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1104 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1105 if (jt->has_last_Java_frame()) {
1106 StackFrameStream sfs(jt, true);
1107 while (!sfs.is_done()) {
1108 frame* cur = sfs.current();
1109 if (cb->contains(cur->pc())) {
1110 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1111 compiledVFrame* cvf = compiledVFrame::cast(vf);
1112 // Revoke monitors' biases in all scopes
1113 while (!cvf->is_top()) {
1114 collect_monitors(cvf, objects_to_revoke);
1115 cvf = compiledVFrame::cast(cvf->sender());
1116 }
1117 collect_monitors(cvf, objects_to_revoke);
1118 }
1119 sfs.next();
1120 }
1121 }
1122 }
1123 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1124 }
1127 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1128 assert(fr.can_be_deoptimized(), "checking frame type");
1130 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1132 // Patch the nmethod so that when execution returns to it we will
1133 // deopt the execution state and return to the interpreter.
1134 fr.deoptimize(thread);
1135 }
1137 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1138 // Deoptimize only if the frame comes from compile code.
1139 // Do not deoptimize the frame which is already patched
1140 // during the execution of the loops below.
1141 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1142 return;
1143 }
1144 ResourceMark rm;
1145 DeoptimizationMarker dm;
1146 if (UseBiasedLocking) {
1147 revoke_biases_of_monitors(thread, fr, map);
1148 }
1149 deoptimize_single_frame(thread, fr);
1151 }
1154 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1155 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1156 "can only deoptimize other thread at a safepoint");
1157 // Compute frame and register map based on thread and sp.
1158 RegisterMap reg_map(thread, UseBiasedLocking);
1159 frame fr = thread->last_frame();
1160 while (fr.id() != id) {
1161 fr = fr.sender(®_map);
1162 }
1163 deoptimize(thread, fr, ®_map);
1164 }
1167 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1168 if (thread == Thread::current()) {
1169 Deoptimization::deoptimize_frame_internal(thread, id);
1170 } else {
1171 VM_DeoptimizeFrame deopt(thread, id);
1172 VMThread::execute(&deopt);
1173 }
1174 }
1177 // JVMTI PopFrame support
1178 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1179 {
1180 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1181 }
1182 JRT_END
1185 #if defined(COMPILER2) || defined(SHARK)
1186 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1187 // in case of an unresolved klass entry, load the class.
1188 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1189 Klass* tk = constant_pool->klass_at(index, CHECK);
1190 return;
1191 }
1193 if (!constant_pool->tag_at(index).is_symbol()) return;
1195 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1196 Symbol* symbol = constant_pool->symbol_at(index);
1198 // class name?
1199 if (symbol->byte_at(0) != '(') {
1200 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1201 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1202 return;
1203 }
1205 // then it must be a signature!
1206 ResourceMark rm(THREAD);
1207 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1208 if (ss.is_object()) {
1209 Symbol* class_name = ss.as_symbol(CHECK);
1210 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1211 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1212 }
1213 }
1214 }
1217 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1218 EXCEPTION_MARK;
1219 load_class_by_index(constant_pool, index, THREAD);
1220 if (HAS_PENDING_EXCEPTION) {
1221 // Exception happened during classloading. We ignore the exception here, since it
1222 // is going to be rethrown since the current activation is going to be deoptimzied and
1223 // the interpreter will re-execute the bytecode.
1224 CLEAR_PENDING_EXCEPTION;
1225 }
1226 }
1228 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1229 HandleMark hm;
1231 // uncommon_trap() is called at the beginning of the uncommon trap
1232 // handler. Note this fact before we start generating temporary frames
1233 // that can confuse an asynchronous stack walker. This counter is
1234 // decremented at the end of unpack_frames().
1235 thread->inc_in_deopt_handler();
1237 // We need to update the map if we have biased locking.
1238 RegisterMap reg_map(thread, UseBiasedLocking);
1239 frame stub_frame = thread->last_frame();
1240 frame fr = stub_frame.sender(®_map);
1241 // Make sure the calling nmethod is not getting deoptimized and removed
1242 // before we are done with it.
1243 nmethodLocker nl(fr.pc());
1245 // Log a message
1246 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1247 trap_request, fr.pc());
1249 {
1250 ResourceMark rm;
1252 // Revoke biases of any monitors in the frame to ensure we can migrate them
1253 revoke_biases_of_monitors(thread, fr, ®_map);
1255 DeoptReason reason = trap_request_reason(trap_request);
1256 DeoptAction action = trap_request_action(trap_request);
1257 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1259 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1260 compiledVFrame* cvf = compiledVFrame::cast(vf);
1262 nmethod* nm = cvf->code();
1264 ScopeDesc* trap_scope = cvf->scope();
1265 methodHandle trap_method = trap_scope->method();
1266 int trap_bci = trap_scope->bci();
1267 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1269 // Record this event in the histogram.
1270 gather_statistics(reason, action, trap_bc);
1272 // Ensure that we can record deopt. history:
1273 bool create_if_missing = ProfileTraps;
1275 MethodData* trap_mdo =
1276 get_method_data(thread, trap_method, create_if_missing);
1278 // Log a message
1279 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1280 trap_reason_name(reason), trap_action_name(action), fr.pc(),
1281 trap_method->name_and_sig_as_C_string(), trap_bci);
1283 // Print a bunch of diagnostics, if requested.
1284 if (TraceDeoptimization || LogCompilation) {
1285 ResourceMark rm;
1286 ttyLocker ttyl;
1287 char buf[100];
1288 if (xtty != NULL) {
1289 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1290 os::current_thread_id(),
1291 format_trap_request(buf, sizeof(buf), trap_request));
1292 nm->log_identity(xtty);
1293 }
1294 Symbol* class_name = NULL;
1295 bool unresolved = false;
1296 if (unloaded_class_index >= 0) {
1297 constantPoolHandle constants (THREAD, trap_method->constants());
1298 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1299 class_name = constants->klass_name_at(unloaded_class_index);
1300 unresolved = true;
1301 if (xtty != NULL)
1302 xtty->print(" unresolved='1'");
1303 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1304 class_name = constants->symbol_at(unloaded_class_index);
1305 }
1306 if (xtty != NULL)
1307 xtty->name(class_name);
1308 }
1309 if (xtty != NULL && trap_mdo != NULL) {
1310 // Dump the relevant MDO state.
1311 // This is the deopt count for the current reason, any previous
1312 // reasons or recompiles seen at this point.
1313 int dcnt = trap_mdo->trap_count(reason);
1314 if (dcnt != 0)
1315 xtty->print(" count='%d'", dcnt);
1316 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1317 int dos = (pdata == NULL)? 0: pdata->trap_state();
1318 if (dos != 0) {
1319 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1320 if (trap_state_is_recompiled(dos)) {
1321 int recnt2 = trap_mdo->overflow_recompile_count();
1322 if (recnt2 != 0)
1323 xtty->print(" recompiles2='%d'", recnt2);
1324 }
1325 }
1326 }
1327 if (xtty != NULL) {
1328 xtty->stamp();
1329 xtty->end_head();
1330 }
1331 if (TraceDeoptimization) { // make noise on the tty
1332 tty->print("Uncommon trap occurred in");
1333 nm->method()->print_short_name(tty);
1334 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1335 fr.pc(),
1336 os::current_thread_id(),
1337 trap_reason_name(reason),
1338 trap_action_name(action),
1339 unloaded_class_index);
1340 if (class_name != NULL) {
1341 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1342 class_name->print_symbol_on(tty);
1343 }
1344 tty->cr();
1345 }
1346 if (xtty != NULL) {
1347 // Log the precise location of the trap.
1348 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1349 xtty->begin_elem("jvms bci='%d'", sd->bci());
1350 xtty->method(sd->method());
1351 xtty->end_elem();
1352 if (sd->is_top()) break;
1353 }
1354 xtty->tail("uncommon_trap");
1355 }
1356 }
1357 // (End diagnostic printout.)
1359 // Load class if necessary
1360 if (unloaded_class_index >= 0) {
1361 constantPoolHandle constants(THREAD, trap_method->constants());
1362 load_class_by_index(constants, unloaded_class_index);
1363 }
1365 // Flush the nmethod if necessary and desirable.
1366 //
1367 // We need to avoid situations where we are re-flushing the nmethod
1368 // because of a hot deoptimization site. Repeated flushes at the same
1369 // point need to be detected by the compiler and avoided. If the compiler
1370 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1371 // module must take measures to avoid an infinite cycle of recompilation
1372 // and deoptimization. There are several such measures:
1373 //
1374 // 1. If a recompilation is ordered a second time at some site X
1375 // and for the same reason R, the action is adjusted to 'reinterpret',
1376 // to give the interpreter time to exercise the method more thoroughly.
1377 // If this happens, the method's overflow_recompile_count is incremented.
1378 //
1379 // 2. If the compiler fails to reduce the deoptimization rate, then
1380 // the method's overflow_recompile_count will begin to exceed the set
1381 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1382 // is adjusted to 'make_not_compilable', and the method is abandoned
1383 // to the interpreter. This is a performance hit for hot methods,
1384 // but is better than a disastrous infinite cycle of recompilations.
1385 // (Actually, only the method containing the site X is abandoned.)
1386 //
1387 // 3. In parallel with the previous measures, if the total number of
1388 // recompilations of a method exceeds the much larger set limit
1389 // PerMethodRecompilationCutoff, the method is abandoned.
1390 // This should only happen if the method is very large and has
1391 // many "lukewarm" deoptimizations. The code which enforces this
1392 // limit is elsewhere (class nmethod, class Method).
1393 //
1394 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1395 // to recompile at each bytecode independently of the per-BCI cutoff.
1396 //
1397 // The decision to update code is up to the compiler, and is encoded
1398 // in the Action_xxx code. If the compiler requests Action_none
1399 // no trap state is changed, no compiled code is changed, and the
1400 // computation suffers along in the interpreter.
1401 //
1402 // The other action codes specify various tactics for decompilation
1403 // and recompilation. Action_maybe_recompile is the loosest, and
1404 // allows the compiled code to stay around until enough traps are seen,
1405 // and until the compiler gets around to recompiling the trapping method.
1406 //
1407 // The other actions cause immediate removal of the present code.
1409 bool update_trap_state = true;
1410 bool make_not_entrant = false;
1411 bool make_not_compilable = false;
1412 bool reprofile = false;
1413 switch (action) {
1414 case Action_none:
1415 // Keep the old code.
1416 update_trap_state = false;
1417 break;
1418 case Action_maybe_recompile:
1419 // Do not need to invalidate the present code, but we can
1420 // initiate another
1421 // Start compiler without (necessarily) invalidating the nmethod.
1422 // The system will tolerate the old code, but new code should be
1423 // generated when possible.
1424 break;
1425 case Action_reinterpret:
1426 // Go back into the interpreter for a while, and then consider
1427 // recompiling form scratch.
1428 make_not_entrant = true;
1429 // Reset invocation counter for outer most method.
1430 // This will allow the interpreter to exercise the bytecodes
1431 // for a while before recompiling.
1432 // By contrast, Action_make_not_entrant is immediate.
1433 //
1434 // Note that the compiler will track null_check, null_assert,
1435 // range_check, and class_check events and log them as if they
1436 // had been traps taken from compiled code. This will update
1437 // the MDO trap history so that the next compilation will
1438 // properly detect hot trap sites.
1439 reprofile = true;
1440 break;
1441 case Action_make_not_entrant:
1442 // Request immediate recompilation, and get rid of the old code.
1443 // Make them not entrant, so next time they are called they get
1444 // recompiled. Unloaded classes are loaded now so recompile before next
1445 // time they are called. Same for uninitialized. The interpreter will
1446 // link the missing class, if any.
1447 make_not_entrant = true;
1448 break;
1449 case Action_make_not_compilable:
1450 // Give up on compiling this method at all.
1451 make_not_entrant = true;
1452 make_not_compilable = true;
1453 break;
1454 default:
1455 ShouldNotReachHere();
1456 }
1458 // Setting +ProfileTraps fixes the following, on all platforms:
1459 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1460 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1461 // recompile relies on a MethodData* to record heroic opt failures.
1463 // Whether the interpreter is producing MDO data or not, we also need
1464 // to use the MDO to detect hot deoptimization points and control
1465 // aggressive optimization.
1466 bool inc_recompile_count = false;
1467 ProfileData* pdata = NULL;
1468 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1469 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1470 uint this_trap_count = 0;
1471 bool maybe_prior_trap = false;
1472 bool maybe_prior_recompile = false;
1473 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1474 //outputs:
1475 this_trap_count,
1476 maybe_prior_trap,
1477 maybe_prior_recompile);
1478 // Because the interpreter also counts null, div0, range, and class
1479 // checks, these traps from compiled code are double-counted.
1480 // This is harmless; it just means that the PerXTrapLimit values
1481 // are in effect a little smaller than they look.
1483 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1484 if (per_bc_reason != Reason_none) {
1485 // Now take action based on the partially known per-BCI history.
1486 if (maybe_prior_trap
1487 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1488 // If there are too many traps at this BCI, force a recompile.
1489 // This will allow the compiler to see the limit overflow, and
1490 // take corrective action, if possible. The compiler generally
1491 // does not use the exact PerBytecodeTrapLimit value, but instead
1492 // changes its tactics if it sees any traps at all. This provides
1493 // a little hysteresis, delaying a recompile until a trap happens
1494 // several times.
1495 //
1496 // Actually, since there is only one bit of counter per BCI,
1497 // the possible per-BCI counts are {0,1,(per-method count)}.
1498 // This produces accurate results if in fact there is only
1499 // one hot trap site, but begins to get fuzzy if there are
1500 // many sites. For example, if there are ten sites each
1501 // trapping two or more times, they each get the blame for
1502 // all of their traps.
1503 make_not_entrant = true;
1504 }
1506 // Detect repeated recompilation at the same BCI, and enforce a limit.
1507 if (make_not_entrant && maybe_prior_recompile) {
1508 // More than one recompile at this point.
1509 inc_recompile_count = maybe_prior_trap;
1510 }
1511 } else {
1512 // For reasons which are not recorded per-bytecode, we simply
1513 // force recompiles unconditionally.
1514 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1515 make_not_entrant = true;
1516 }
1518 // Go back to the compiler if there are too many traps in this method.
1519 if (this_trap_count >= (uint)PerMethodTrapLimit) {
1520 // If there are too many traps in this method, force a recompile.
1521 // This will allow the compiler to see the limit overflow, and
1522 // take corrective action, if possible.
1523 // (This condition is an unlikely backstop only, because the
1524 // PerBytecodeTrapLimit is more likely to take effect first,
1525 // if it is applicable.)
1526 make_not_entrant = true;
1527 }
1529 // Here's more hysteresis: If there has been a recompile at
1530 // this trap point already, run the method in the interpreter
1531 // for a while to exercise it more thoroughly.
1532 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1533 reprofile = true;
1534 }
1536 }
1538 // Take requested actions on the method:
1540 // Recompile
1541 if (make_not_entrant) {
1542 if (!nm->make_not_entrant()) {
1543 return; // the call did not change nmethod's state
1544 }
1546 if (pdata != NULL) {
1547 // Record the recompilation event, if any.
1548 int tstate0 = pdata->trap_state();
1549 int tstate1 = trap_state_set_recompiled(tstate0, true);
1550 if (tstate1 != tstate0)
1551 pdata->set_trap_state(tstate1);
1552 }
1553 }
1555 if (inc_recompile_count) {
1556 trap_mdo->inc_overflow_recompile_count();
1557 if ((uint)trap_mdo->overflow_recompile_count() >
1558 (uint)PerBytecodeRecompilationCutoff) {
1559 // Give up on the method containing the bad BCI.
1560 if (trap_method() == nm->method()) {
1561 make_not_compilable = true;
1562 } else {
1563 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1564 // But give grace to the enclosing nm->method().
1565 }
1566 }
1567 }
1569 // Reprofile
1570 if (reprofile) {
1571 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1572 }
1574 // Give up compiling
1575 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1576 assert(make_not_entrant, "consistent");
1577 nm->method()->set_not_compilable(CompLevel_full_optimization);
1578 }
1580 } // Free marked resources
1582 }
1583 JRT_END
1585 MethodData*
1586 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1587 bool create_if_missing) {
1588 Thread* THREAD = thread;
1589 MethodData* mdo = m()->method_data();
1590 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1591 // Build an MDO. Ignore errors like OutOfMemory;
1592 // that simply means we won't have an MDO to update.
1593 Method::build_interpreter_method_data(m, THREAD);
1594 if (HAS_PENDING_EXCEPTION) {
1595 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1596 CLEAR_PENDING_EXCEPTION;
1597 }
1598 mdo = m()->method_data();
1599 }
1600 return mdo;
1601 }
1603 ProfileData*
1604 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1605 int trap_bci,
1606 Deoptimization::DeoptReason reason,
1607 //outputs:
1608 uint& ret_this_trap_count,
1609 bool& ret_maybe_prior_trap,
1610 bool& ret_maybe_prior_recompile) {
1611 uint prior_trap_count = trap_mdo->trap_count(reason);
1612 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1614 // If the runtime cannot find a place to store trap history,
1615 // it is estimated based on the general condition of the method.
1616 // If the method has ever been recompiled, or has ever incurred
1617 // a trap with the present reason , then this BCI is assumed
1618 // (pessimistically) to be the culprit.
1619 bool maybe_prior_trap = (prior_trap_count != 0);
1620 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1621 ProfileData* pdata = NULL;
1624 // For reasons which are recorded per bytecode, we check per-BCI data.
1625 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1626 if (per_bc_reason != Reason_none) {
1627 // Find the profile data for this BCI. If there isn't one,
1628 // try to allocate one from the MDO's set of spares.
1629 // This will let us detect a repeated trap at this point.
1630 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
1632 if (pdata != NULL) {
1633 // Query the trap state of this profile datum.
1634 int tstate0 = pdata->trap_state();
1635 if (!trap_state_has_reason(tstate0, per_bc_reason))
1636 maybe_prior_trap = false;
1637 if (!trap_state_is_recompiled(tstate0))
1638 maybe_prior_recompile = false;
1640 // Update the trap state of this profile datum.
1641 int tstate1 = tstate0;
1642 // Record the reason.
1643 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1644 // Store the updated state on the MDO, for next time.
1645 if (tstate1 != tstate0)
1646 pdata->set_trap_state(tstate1);
1647 } else {
1648 if (LogCompilation && xtty != NULL) {
1649 ttyLocker ttyl;
1650 // Missing MDP? Leave a small complaint in the log.
1651 xtty->elem("missing_mdp bci='%d'", trap_bci);
1652 }
1653 }
1654 }
1656 // Return results:
1657 ret_this_trap_count = this_trap_count;
1658 ret_maybe_prior_trap = maybe_prior_trap;
1659 ret_maybe_prior_recompile = maybe_prior_recompile;
1660 return pdata;
1661 }
1663 void
1664 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1665 ResourceMark rm;
1666 // Ignored outputs:
1667 uint ignore_this_trap_count;
1668 bool ignore_maybe_prior_trap;
1669 bool ignore_maybe_prior_recompile;
1670 query_update_method_data(trap_mdo, trap_bci,
1671 (DeoptReason)reason,
1672 ignore_this_trap_count,
1673 ignore_maybe_prior_trap,
1674 ignore_maybe_prior_recompile);
1675 }
1677 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1679 // Still in Java no safepoints
1680 {
1681 // This enters VM and may safepoint
1682 uncommon_trap_inner(thread, trap_request);
1683 }
1684 return fetch_unroll_info_helper(thread);
1685 }
1687 // Local derived constants.
1688 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1689 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1690 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1692 //---------------------------trap_state_reason---------------------------------
1693 Deoptimization::DeoptReason
1694 Deoptimization::trap_state_reason(int trap_state) {
1695 // This assert provides the link between the width of DataLayout::trap_bits
1696 // and the encoding of "recorded" reasons. It ensures there are enough
1697 // bits to store all needed reasons in the per-BCI MDO profile.
1698 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1699 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1700 trap_state -= recompile_bit;
1701 if (trap_state == DS_REASON_MASK) {
1702 return Reason_many;
1703 } else {
1704 assert((int)Reason_none == 0, "state=0 => Reason_none");
1705 return (DeoptReason)trap_state;
1706 }
1707 }
1708 //-------------------------trap_state_has_reason-------------------------------
1709 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1710 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1711 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1712 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1713 trap_state -= recompile_bit;
1714 if (trap_state == DS_REASON_MASK) {
1715 return -1; // true, unspecifically (bottom of state lattice)
1716 } else if (trap_state == reason) {
1717 return 1; // true, definitely
1718 } else if (trap_state == 0) {
1719 return 0; // false, definitely (top of state lattice)
1720 } else {
1721 return 0; // false, definitely
1722 }
1723 }
1724 //-------------------------trap_state_add_reason-------------------------------
1725 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1726 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1727 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1728 trap_state -= recompile_bit;
1729 if (trap_state == DS_REASON_MASK) {
1730 return trap_state + recompile_bit; // already at state lattice bottom
1731 } else if (trap_state == reason) {
1732 return trap_state + recompile_bit; // the condition is already true
1733 } else if (trap_state == 0) {
1734 return reason + recompile_bit; // no condition has yet been true
1735 } else {
1736 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1737 }
1738 }
1739 //-----------------------trap_state_is_recompiled------------------------------
1740 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1741 return (trap_state & DS_RECOMPILE_BIT) != 0;
1742 }
1743 //-----------------------trap_state_set_recompiled-----------------------------
1744 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1745 if (z) return trap_state | DS_RECOMPILE_BIT;
1746 else return trap_state & ~DS_RECOMPILE_BIT;
1747 }
1748 //---------------------------format_trap_state---------------------------------
1749 // This is used for debugging and diagnostics, including hotspot.log output.
1750 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1751 int trap_state) {
1752 DeoptReason reason = trap_state_reason(trap_state);
1753 bool recomp_flag = trap_state_is_recompiled(trap_state);
1754 // Re-encode the state from its decoded components.
1755 int decoded_state = 0;
1756 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1757 decoded_state = trap_state_add_reason(decoded_state, reason);
1758 if (recomp_flag)
1759 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1760 // If the state re-encodes properly, format it symbolically.
1761 // Because this routine is used for debugging and diagnostics,
1762 // be robust even if the state is a strange value.
1763 size_t len;
1764 if (decoded_state != trap_state) {
1765 // Random buggy state that doesn't decode??
1766 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1767 } else {
1768 len = jio_snprintf(buf, buflen, "%s%s",
1769 trap_reason_name(reason),
1770 recomp_flag ? " recompiled" : "");
1771 }
1772 if (len >= buflen)
1773 buf[buflen-1] = '\0';
1774 return buf;
1775 }
1778 //--------------------------------statics--------------------------------------
1779 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1780 = Deoptimization::Action_reinterpret;
1781 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1782 // Note: Keep this in sync. with enum DeoptReason.
1783 "none",
1784 "null_check",
1785 "null_assert",
1786 "range_check",
1787 "class_check",
1788 "array_check",
1789 "intrinsic",
1790 "bimorphic",
1791 "unloaded",
1792 "uninitialized",
1793 "unreached",
1794 "unhandled",
1795 "constraint",
1796 "div0_check",
1797 "age",
1798 "predicate",
1799 "loop_limit_check"
1800 };
1801 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1802 // Note: Keep this in sync. with enum DeoptAction.
1803 "none",
1804 "maybe_recompile",
1805 "reinterpret",
1806 "make_not_entrant",
1807 "make_not_compilable"
1808 };
1810 const char* Deoptimization::trap_reason_name(int reason) {
1811 if (reason == Reason_many) return "many";
1812 if ((uint)reason < Reason_LIMIT)
1813 return _trap_reason_name[reason];
1814 static char buf[20];
1815 sprintf(buf, "reason%d", reason);
1816 return buf;
1817 }
1818 const char* Deoptimization::trap_action_name(int action) {
1819 if ((uint)action < Action_LIMIT)
1820 return _trap_action_name[action];
1821 static char buf[20];
1822 sprintf(buf, "action%d", action);
1823 return buf;
1824 }
1826 // This is used for debugging and diagnostics, including hotspot.log output.
1827 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1828 int trap_request) {
1829 jint unloaded_class_index = trap_request_index(trap_request);
1830 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1831 const char* action = trap_action_name(trap_request_action(trap_request));
1832 size_t len;
1833 if (unloaded_class_index < 0) {
1834 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1835 reason, action);
1836 } else {
1837 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1838 reason, action, unloaded_class_index);
1839 }
1840 if (len >= buflen)
1841 buf[buflen-1] = '\0';
1842 return buf;
1843 }
1845 juint Deoptimization::_deoptimization_hist
1846 [Deoptimization::Reason_LIMIT]
1847 [1 + Deoptimization::Action_LIMIT]
1848 [Deoptimization::BC_CASE_LIMIT]
1849 = {0};
1851 enum {
1852 LSB_BITS = 8,
1853 LSB_MASK = right_n_bits(LSB_BITS)
1854 };
1856 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1857 Bytecodes::Code bc) {
1858 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1859 assert(action >= 0 && action < Action_LIMIT, "oob");
1860 _deoptimization_hist[Reason_none][0][0] += 1; // total
1861 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1862 juint* cases = _deoptimization_hist[reason][1+action];
1863 juint* bc_counter_addr = NULL;
1864 juint bc_counter = 0;
1865 // Look for an unused counter, or an exact match to this BC.
1866 if (bc != Bytecodes::_illegal) {
1867 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1868 juint* counter_addr = &cases[bc_case];
1869 juint counter = *counter_addr;
1870 if ((counter == 0 && bc_counter_addr == NULL)
1871 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1872 // this counter is either free or is already devoted to this BC
1873 bc_counter_addr = counter_addr;
1874 bc_counter = counter | bc;
1875 }
1876 }
1877 }
1878 if (bc_counter_addr == NULL) {
1879 // Overflow, or no given bytecode.
1880 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1881 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1882 }
1883 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1884 }
1886 jint Deoptimization::total_deoptimization_count() {
1887 return _deoptimization_hist[Reason_none][0][0];
1888 }
1890 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1891 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1892 return _deoptimization_hist[reason][0][0];
1893 }
1895 void Deoptimization::print_statistics() {
1896 juint total = total_deoptimization_count();
1897 juint account = total;
1898 if (total != 0) {
1899 ttyLocker ttyl;
1900 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1901 tty->print_cr("Deoptimization traps recorded:");
1902 #define PRINT_STAT_LINE(name, r) \
1903 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1904 PRINT_STAT_LINE("total", total);
1905 // For each non-zero entry in the histogram, print the reason,
1906 // the action, and (if specifically known) the type of bytecode.
1907 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1908 for (int action = 0; action < Action_LIMIT; action++) {
1909 juint* cases = _deoptimization_hist[reason][1+action];
1910 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1911 juint counter = cases[bc_case];
1912 if (counter != 0) {
1913 char name[1*K];
1914 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1915 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1916 bc = Bytecodes::_illegal;
1917 sprintf(name, "%s/%s/%s",
1918 trap_reason_name(reason),
1919 trap_action_name(action),
1920 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1921 juint r = counter >> LSB_BITS;
1922 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1923 account -= r;
1924 }
1925 }
1926 }
1927 }
1928 if (account != 0) {
1929 PRINT_STAT_LINE("unaccounted", account);
1930 }
1931 #undef PRINT_STAT_LINE
1932 if (xtty != NULL) xtty->tail("statistics");
1933 }
1934 }
1935 #else // COMPILER2 || SHARK
1938 // Stubs for C1 only system.
1939 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1940 return false;
1941 }
1943 const char* Deoptimization::trap_reason_name(int reason) {
1944 return "unknown";
1945 }
1947 void Deoptimization::print_statistics() {
1948 // no output
1949 }
1951 void
1952 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1953 // no udpate
1954 }
1956 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1957 return 0;
1958 }
1960 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1961 Bytecodes::Code bc) {
1962 // no update
1963 }
1965 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1966 int trap_state) {
1967 jio_snprintf(buf, buflen, "#%d", trap_state);
1968 return buf;
1969 }
1971 #endif // COMPILER2 || SHARK