0
|
1 /*
|
|
2 * Copyright 1997-2007 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 */
|
|
24
|
|
25 #include "incls/_precompiled.incl"
|
|
26 #include "incls/_deoptimization.cpp.incl"
|
|
27
|
|
28 bool DeoptimizationMarker::_is_active = false;
|
|
29
|
|
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;
|
|
48
|
|
49 _total_frame_sizes = size_of_frames();
|
|
50 }
|
|
51
|
|
52
|
|
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 }
|
|
58
|
|
59
|
|
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 }
|
|
64
|
|
65
|
|
66
|
|
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 }
|
|
75
|
|
76
|
|
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 }
|
|
87
|
|
88
|
|
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
|
|
98
|
|
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();
|
|
104
|
|
105 return fetch_unroll_info_helper(thread);
|
|
106 JRT_END
|
|
107
|
|
108
|
|
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) {
|
|
111
|
|
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 //
|
|
116
|
|
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);
|
|
121
|
|
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);
|
|
127
|
|
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));
|
|
140
|
|
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 && EliminateAllocations) {
|
|
145 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
|
|
146 bool reallocated = false;
|
|
147 if (objects != NULL) {
|
|
148 JRT_BLOCK
|
|
149 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
|
|
150 JRT_END
|
|
151 }
|
|
152 if (reallocated) {
|
|
153 reassign_fields(&deoptee, &map, objects);
|
|
154 #ifndef PRODUCT
|
|
155 if (TraceDeoptimization) {
|
|
156 ttyLocker ttyl;
|
|
157 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
|
|
158 print_objects(objects);
|
|
159 }
|
|
160 #endif
|
|
161 }
|
|
162 for (int i = 0; i < chunk->length(); i++) {
|
|
163 GrowableArray<MonitorValue*>* monitors = chunk->at(i)->scope()->monitors();
|
|
164 if (monitors != NULL) {
|
|
165 relock_objects(&deoptee, &map, monitors);
|
|
166 #ifndef PRODUCT
|
|
167 if (TraceDeoptimization) {
|
|
168 ttyLocker ttyl;
|
|
169 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
|
|
170 for (int j = 0; i < monitors->length(); i++) {
|
|
171 MonitorValue* mv = monitors->at(i);
|
|
172 if (mv->eliminated()) {
|
|
173 StackValue* owner = StackValue::create_stack_value(&deoptee, &map, mv->owner());
|
|
174 tty->print_cr(" object <" INTPTR_FORMAT "> locked", owner->get_obj()());
|
|
175 }
|
|
176 }
|
|
177 }
|
|
178 #endif
|
|
179 }
|
|
180 }
|
|
181 }
|
|
182 #endif // COMPILER2
|
|
183 // Ensure that no safepoint is taken after pointers have been stored
|
|
184 // in fields of rematerialized objects. If a safepoint occurs from here on
|
|
185 // out the java state residing in the vframeArray will be missed.
|
|
186 No_Safepoint_Verifier no_safepoint;
|
|
187
|
|
188 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
|
|
189
|
|
190 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
|
|
191 thread->set_vframe_array_head(array);
|
|
192
|
|
193 // Now that the vframeArray has been created if we have any deferred local writes
|
|
194 // added by jvmti then we can free up that structure as the data is now in the
|
|
195 // vframeArray
|
|
196
|
|
197 if (thread->deferred_locals() != NULL) {
|
|
198 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
|
|
199 int i = 0;
|
|
200 do {
|
|
201 // Because of inlining we could have multiple vframes for a single frame
|
|
202 // and several of the vframes could have deferred writes. Find them all.
|
|
203 if (list->at(i)->id() == array->original().id()) {
|
|
204 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
|
|
205 list->remove_at(i);
|
|
206 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
|
|
207 delete dlv;
|
|
208 } else {
|
|
209 i++;
|
|
210 }
|
|
211 } while ( i < list->length() );
|
|
212 if (list->length() == 0) {
|
|
213 thread->set_deferred_locals(NULL);
|
|
214 // free the list and elements back to C heap.
|
|
215 delete list;
|
|
216 }
|
|
217
|
|
218 }
|
|
219
|
|
220 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
|
|
221 CodeBlob* cb = stub_frame.cb();
|
|
222 // Verify we have the right vframeArray
|
|
223 assert(cb->frame_size() >= 0, "Unexpected frame size");
|
|
224 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
|
|
225
|
|
226 #ifdef ASSERT
|
|
227 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
|
|
228 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp);
|
|
229 #endif
|
|
230 // This is a guarantee instead of an assert because if vframe doesn't match
|
|
231 // we will unpack the wrong deoptimized frame and wind up in strange places
|
|
232 // where it will be very difficult to figure out what went wrong. Better
|
|
233 // to die an early death here than some very obscure death later when the
|
|
234 // trail is cold.
|
|
235 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
|
|
236 // in that it will fail to detect a problem when there is one. This needs
|
|
237 // more work in tiger timeframe.
|
|
238 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
|
|
239
|
|
240 int number_of_frames = array->frames();
|
|
241
|
|
242 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
|
|
243 // virtual activation, which is the reverse of the elements in the vframes array.
|
|
244 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames);
|
|
245 // +1 because we always have an interpreter return address for the final slot.
|
|
246 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1);
|
|
247 int callee_parameters = 0;
|
|
248 int callee_locals = 0;
|
|
249 int popframe_extra_args = 0;
|
|
250 // Create an interpreter return address for the stub to use as its return
|
|
251 // address so the skeletal frames are perfectly walkable
|
|
252 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
|
|
253
|
|
254 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
|
|
255 // activation be put back on the expression stack of the caller for reexecution
|
|
256 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
|
|
257 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
|
|
258 }
|
|
259
|
|
260 //
|
|
261 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
|
|
262 // frame_sizes/frame_pcs[1] next oldest frame (int)
|
|
263 // frame_sizes/frame_pcs[n] youngest frame (int)
|
|
264 //
|
|
265 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
|
|
266 // owns the space for the return address to it's caller). Confusing ain't it.
|
|
267 //
|
|
268 // The vframe array can address vframes with indices running from
|
|
269 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
|
|
270 // When we create the skeletal frames we need the oldest frame to be in the zero slot
|
|
271 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
|
|
272 // so things look a little strange in this loop.
|
|
273 //
|
|
274 for (int index = 0; index < array->frames(); index++ ) {
|
|
275 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
|
|
276 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
|
|
277 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
|
|
278 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
|
|
279 callee_locals,
|
|
280 index == 0,
|
|
281 popframe_extra_args);
|
|
282 // This pc doesn't have to be perfect just good enough to identify the frame
|
|
283 // as interpreted so the skeleton frame will be walkable
|
|
284 // The correct pc will be set when the skeleton frame is completely filled out
|
|
285 // The final pc we store in the loop is wrong and will be overwritten below
|
|
286 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
|
|
287
|
|
288 callee_parameters = array->element(index)->method()->size_of_parameters();
|
|
289 callee_locals = array->element(index)->method()->max_locals();
|
|
290 popframe_extra_args = 0;
|
|
291 }
|
|
292
|
|
293 // Compute whether the root vframe returns a float or double value.
|
|
294 BasicType return_type;
|
|
295 {
|
|
296 HandleMark hm;
|
|
297 methodHandle method(thread, array->element(0)->method());
|
|
298 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci());
|
|
299 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL;
|
|
300 }
|
|
301
|
|
302 // Compute information for handling adapters and adjusting the frame size of the caller.
|
|
303 int caller_adjustment = 0;
|
|
304
|
|
305 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
|
|
306 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
|
|
307 // than simply use array->sender.pc(). This requires us to walk the current set of frames
|
|
308 //
|
|
309 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
|
|
310 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
|
|
311
|
|
312 // Compute the amount the oldest interpreter frame will have to adjust
|
|
313 // its caller's stack by. If the caller is a compiled frame then
|
|
314 // we pretend that the callee has no parameters so that the
|
|
315 // extension counts for the full amount of locals and not just
|
|
316 // locals-parms. This is because without a c2i adapter the parm
|
|
317 // area as created by the compiled frame will not be usable by
|
|
318 // the interpreter. (Depending on the calling convention there
|
|
319 // may not even be enough space).
|
|
320
|
|
321 // QQQ I'd rather see this pushed down into last_frame_adjust
|
|
322 // and have it take the sender (aka caller).
|
|
323
|
|
324 if (deopt_sender.is_compiled_frame()) {
|
|
325 caller_adjustment = last_frame_adjust(0, callee_locals);
|
|
326 } else if (callee_locals > callee_parameters) {
|
|
327 // The caller frame may need extending to accommodate
|
|
328 // non-parameter locals of the first unpacked interpreted frame.
|
|
329 // Compute that adjustment.
|
|
330 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
|
|
331 }
|
|
332
|
|
333
|
|
334 // If the sender is deoptimized the we must retrieve the address of the handler
|
|
335 // since the frame will "magically" show the original pc before the deopt
|
|
336 // and we'd undo the deopt.
|
|
337
|
|
338 frame_pcs[0] = deopt_sender.raw_pc();
|
|
339
|
|
340 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
|
|
341
|
|
342 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
|
|
343 caller_adjustment * BytesPerWord,
|
|
344 number_of_frames,
|
|
345 frame_sizes,
|
|
346 frame_pcs,
|
|
347 return_type);
|
|
348 #if defined(IA32) || defined(AMD64)
|
|
349 // We need a way to pass fp to the unpacking code so the skeletal frames
|
|
350 // come out correct. This is only needed for x86 because of c2 using ebp
|
|
351 // as an allocatable register. So this update is useless (and harmless)
|
|
352 // on the other platforms. It would be nice to do this in a different
|
|
353 // way but even the old style deoptimization had a problem with deriving
|
|
354 // this value. NEEDS_CLEANUP
|
|
355 // Note: now that c1 is using c2's deopt blob we must do this on all
|
|
356 // x86 based platforms
|
|
357 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes());
|
|
358 *fp_addr = array->sender().fp(); // was adapter_caller
|
|
359 #endif /* IA32 || AMD64 */
|
|
360
|
|
361 if (array->frames() > 1) {
|
|
362 if (VerifyStack && TraceDeoptimization) {
|
|
363 tty->print_cr("Deoptimizing method containing inlining");
|
|
364 }
|
|
365 }
|
|
366
|
|
367 array->set_unroll_block(info);
|
|
368 return info;
|
|
369 }
|
|
370
|
|
371 // Called to cleanup deoptimization data structures in normal case
|
|
372 // after unpacking to stack and when stack overflow error occurs
|
|
373 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
|
|
374 vframeArray *array) {
|
|
375
|
|
376 // Get array if coming from exception
|
|
377 if (array == NULL) {
|
|
378 array = thread->vframe_array_head();
|
|
379 }
|
|
380 thread->set_vframe_array_head(NULL);
|
|
381
|
|
382 // Free the previous UnrollBlock
|
|
383 vframeArray* old_array = thread->vframe_array_last();
|
|
384 thread->set_vframe_array_last(array);
|
|
385
|
|
386 if (old_array != NULL) {
|
|
387 UnrollBlock* old_info = old_array->unroll_block();
|
|
388 old_array->set_unroll_block(NULL);
|
|
389 delete old_info;
|
|
390 delete old_array;
|
|
391 }
|
|
392
|
|
393 // Deallocate any resource creating in this routine and any ResourceObjs allocated
|
|
394 // inside the vframeArray (StackValueCollections)
|
|
395
|
|
396 delete thread->deopt_mark();
|
|
397 thread->set_deopt_mark(NULL);
|
|
398
|
|
399
|
|
400 if (JvmtiExport::can_pop_frame()) {
|
|
401 #ifndef CC_INTERP
|
|
402 // Regardless of whether we entered this routine with the pending
|
|
403 // popframe condition bit set, we should always clear it now
|
|
404 thread->clear_popframe_condition();
|
|
405 #else
|
|
406 // C++ interpeter will clear has_pending_popframe when it enters
|
|
407 // with method_resume. For deopt_resume2 we clear it now.
|
|
408 if (thread->popframe_forcing_deopt_reexecution())
|
|
409 thread->clear_popframe_condition();
|
|
410 #endif /* CC_INTERP */
|
|
411 }
|
|
412
|
|
413 // unpack_frames() is called at the end of the deoptimization handler
|
|
414 // and (in C2) at the end of the uncommon trap handler. Note this fact
|
|
415 // so that an asynchronous stack walker can work again. This counter is
|
|
416 // incremented at the beginning of fetch_unroll_info() and (in C2) at
|
|
417 // the beginning of uncommon_trap().
|
|
418 thread->dec_in_deopt_handler();
|
|
419 }
|
|
420
|
|
421
|
|
422 // Return BasicType of value being returned
|
|
423 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
|
|
424
|
|
425 // We are already active int he special DeoptResourceMark any ResourceObj's we
|
|
426 // allocate will be freed at the end of the routine.
|
|
427
|
|
428 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
|
|
429 // but makes the entry a little slower. There is however a little dance we have to
|
|
430 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
|
|
431 ResetNoHandleMark rnhm; // No-op in release/product versions
|
|
432 HandleMark hm;
|
|
433
|
|
434 frame stub_frame = thread->last_frame();
|
|
435
|
|
436 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
|
|
437 // must point to the vframeArray for the unpack frame.
|
|
438 vframeArray* array = thread->vframe_array_head();
|
|
439
|
|
440 #ifndef PRODUCT
|
|
441 if (TraceDeoptimization) {
|
|
442 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
|
|
443 }
|
|
444 #endif
|
|
445
|
|
446 UnrollBlock* info = array->unroll_block();
|
|
447
|
|
448 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
|
|
449 array->unpack_to_stack(stub_frame, exec_mode);
|
|
450
|
|
451 BasicType bt = info->return_type();
|
|
452
|
|
453 // If we have an exception pending, claim that the return type is an oop
|
|
454 // so the deopt_blob does not overwrite the exception_oop.
|
|
455
|
|
456 if (exec_mode == Unpack_exception)
|
|
457 bt = T_OBJECT;
|
|
458
|
|
459 // Cleanup thread deopt data
|
|
460 cleanup_deopt_info(thread, array);
|
|
461
|
|
462 #ifndef PRODUCT
|
|
463 if (VerifyStack) {
|
|
464 ResourceMark res_mark;
|
|
465
|
|
466 // Verify that the just-unpacked frames match the interpreter's
|
|
467 // notions of expression stack and locals
|
|
468 vframeArray* cur_array = thread->vframe_array_last();
|
|
469 RegisterMap rm(thread, false);
|
|
470 rm.set_include_argument_oops(false);
|
|
471 bool is_top_frame = true;
|
|
472 int callee_size_of_parameters = 0;
|
|
473 int callee_max_locals = 0;
|
|
474 for (int i = 0; i < cur_array->frames(); i++) {
|
|
475 vframeArrayElement* el = cur_array->element(i);
|
|
476 frame* iframe = el->iframe();
|
|
477 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
|
|
478
|
|
479 // Get the oop map for this bci
|
|
480 InterpreterOopMap mask;
|
|
481 int cur_invoke_parameter_size = 0;
|
|
482 bool try_next_mask = false;
|
|
483 int next_mask_expression_stack_size = -1;
|
|
484 int top_frame_expression_stack_adjustment = 0;
|
|
485 methodHandle mh(thread, iframe->interpreter_frame_method());
|
|
486 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
|
|
487 BytecodeStream str(mh);
|
|
488 str.set_start(iframe->interpreter_frame_bci());
|
|
489 int max_bci = mh->code_size();
|
|
490 // Get to the next bytecode if possible
|
|
491 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
|
|
492 // Check to see if we can grab the number of outgoing arguments
|
|
493 // at an uncommon trap for an invoke (where the compiler
|
|
494 // generates debug info before the invoke has executed)
|
|
495 Bytecodes::Code cur_code = str.next();
|
|
496 if (cur_code == Bytecodes::_invokevirtual ||
|
|
497 cur_code == Bytecodes::_invokespecial ||
|
|
498 cur_code == Bytecodes::_invokestatic ||
|
|
499 cur_code == Bytecodes::_invokeinterface) {
|
|
500 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci());
|
|
501 symbolHandle signature(thread, invoke->signature());
|
|
502 ArgumentSizeComputer asc(signature);
|
|
503 cur_invoke_parameter_size = asc.size();
|
|
504 if (cur_code != Bytecodes::_invokestatic) {
|
|
505 // Add in receiver
|
|
506 ++cur_invoke_parameter_size;
|
|
507 }
|
|
508 }
|
|
509 if (str.bci() < max_bci) {
|
|
510 Bytecodes::Code bc = str.next();
|
|
511 if (bc >= 0) {
|
|
512 // The interpreter oop map generator reports results before
|
|
513 // the current bytecode has executed except in the case of
|
|
514 // calls. It seems to be hard to tell whether the compiler
|
|
515 // has emitted debug information matching the "state before"
|
|
516 // a given bytecode or the state after, so we try both
|
|
517 switch (cur_code) {
|
|
518 case Bytecodes::_invokevirtual:
|
|
519 case Bytecodes::_invokespecial:
|
|
520 case Bytecodes::_invokestatic:
|
|
521 case Bytecodes::_invokeinterface:
|
|
522 case Bytecodes::_athrow:
|
|
523 break;
|
|
524 default: {
|
|
525 InterpreterOopMap next_mask;
|
|
526 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
|
|
527 next_mask_expression_stack_size = next_mask.expression_stack_size();
|
|
528 // Need to subtract off the size of the result type of
|
|
529 // the bytecode because this is not described in the
|
|
530 // debug info but returned to the interpreter in the TOS
|
|
531 // caching register
|
|
532 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
|
|
533 if (bytecode_result_type != T_ILLEGAL) {
|
|
534 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
|
|
535 }
|
|
536 assert(top_frame_expression_stack_adjustment >= 0, "");
|
|
537 try_next_mask = true;
|
|
538 break;
|
|
539 }
|
|
540 }
|
|
541 }
|
|
542 }
|
|
543
|
|
544 // Verify stack depth and oops in frame
|
|
545 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
|
|
546 if (!(
|
|
547 /* SPARC */
|
|
548 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
|
|
549 /* x86 */
|
|
550 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
|
|
551 (try_next_mask &&
|
|
552 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
|
|
553 top_frame_expression_stack_adjustment))) ||
|
|
554 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
|
|
555 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
|
|
556 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
|
|
557 )) {
|
|
558 ttyLocker ttyl;
|
|
559
|
|
560 // Print out some information that will help us debug the problem
|
|
561 tty->print_cr("Wrong number of expression stack elements during deoptimization");
|
|
562 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
|
|
563 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
|
|
564 iframe->interpreter_frame_expression_stack_size());
|
|
565 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
|
|
566 tty->print_cr(" try_next_mask = %d", try_next_mask);
|
|
567 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
|
|
568 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
|
|
569 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
|
|
570 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
|
|
571 tty->print_cr(" exec_mode = %d", exec_mode);
|
|
572 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
|
|
573 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
|
|
574 tty->print_cr(" Interpreted frames:");
|
|
575 for (int k = 0; k < cur_array->frames(); k++) {
|
|
576 vframeArrayElement* el = cur_array->element(k);
|
|
577 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
|
|
578 }
|
|
579 cur_array->print_on_2(tty);
|
|
580 guarantee(false, "wrong number of expression stack elements during deopt");
|
|
581 }
|
|
582 VerifyOopClosure verify;
|
|
583 iframe->oops_interpreted_do(&verify, &rm, false);
|
|
584 callee_size_of_parameters = mh->size_of_parameters();
|
|
585 callee_max_locals = mh->max_locals();
|
|
586 is_top_frame = false;
|
|
587 }
|
|
588 }
|
|
589 #endif /* !PRODUCT */
|
|
590
|
|
591
|
|
592 return bt;
|
|
593 JRT_END
|
|
594
|
|
595
|
|
596 int Deoptimization::deoptimize_dependents() {
|
|
597 Threads::deoptimized_wrt_marked_nmethods();
|
|
598 return 0;
|
|
599 }
|
|
600
|
|
601
|
|
602 #ifdef COMPILER2
|
|
603 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
|
|
604 Handle pending_exception(thread->pending_exception());
|
|
605 const char* exception_file = thread->exception_file();
|
|
606 int exception_line = thread->exception_line();
|
|
607 thread->clear_pending_exception();
|
|
608
|
|
609 for (int i = 0; i < objects->length(); i++) {
|
|
610 assert(objects->at(i)->is_object(), "invalid debug information");
|
|
611 ObjectValue* sv = (ObjectValue*) objects->at(i);
|
|
612
|
|
613 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
|
|
614 oop obj = NULL;
|
|
615
|
|
616 if (k->oop_is_instance()) {
|
|
617 instanceKlass* ik = instanceKlass::cast(k());
|
|
618 obj = ik->allocate_instance(CHECK_(false));
|
|
619 } else if (k->oop_is_typeArray()) {
|
|
620 typeArrayKlass* ak = typeArrayKlass::cast(k());
|
|
621 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
|
|
622 int len = sv->field_size() / type2size[ak->element_type()];
|
|
623 obj = ak->allocate(len, CHECK_(false));
|
|
624 } else if (k->oop_is_objArray()) {
|
|
625 objArrayKlass* ak = objArrayKlass::cast(k());
|
|
626 obj = ak->allocate(sv->field_size(), CHECK_(false));
|
|
627 }
|
|
628
|
|
629 assert(obj != NULL, "allocation failed");
|
|
630 assert(sv->value().is_null(), "redundant reallocation");
|
|
631 sv->set_value(obj);
|
|
632 }
|
|
633
|
|
634 if (pending_exception.not_null()) {
|
|
635 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
|
|
636 }
|
|
637
|
|
638 return true;
|
|
639 }
|
|
640
|
|
641 // This assumes that the fields are stored in ObjectValue in the same order
|
|
642 // they are yielded by do_nonstatic_fields.
|
|
643 class FieldReassigner: public FieldClosure {
|
|
644 frame* _fr;
|
|
645 RegisterMap* _reg_map;
|
|
646 ObjectValue* _sv;
|
|
647 instanceKlass* _ik;
|
|
648 oop _obj;
|
|
649
|
|
650 int _i;
|
|
651 public:
|
|
652 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
|
|
653 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
|
|
654
|
|
655 int i() const { return _i; }
|
|
656
|
|
657
|
|
658 void do_field(fieldDescriptor* fd) {
|
|
659 StackValue* value =
|
|
660 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
|
|
661 int offset = fd->offset();
|
|
662 switch (fd->field_type()) {
|
|
663 case T_OBJECT: case T_ARRAY:
|
|
664 assert(value->type() == T_OBJECT, "Agreement.");
|
|
665 _obj->obj_field_put(offset, value->get_obj()());
|
|
666 break;
|
|
667
|
|
668 case T_LONG: case T_DOUBLE: {
|
|
669 assert(value->type() == T_INT, "Agreement.");
|
|
670 StackValue* low =
|
|
671 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
|
|
672 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
|
|
673 _obj->long_field_put(offset, res);
|
|
674 break;
|
|
675 }
|
|
676
|
|
677 case T_INT: case T_FLOAT: // 4 bytes.
|
|
678 assert(value->type() == T_INT, "Agreement.");
|
|
679 _obj->int_field_put(offset, (jint)value->get_int());
|
|
680 break;
|
|
681
|
|
682 case T_SHORT: case T_CHAR: // 2 bytes
|
|
683 assert(value->type() == T_INT, "Agreement.");
|
|
684 _obj->short_field_put(offset, (jshort)value->get_int());
|
|
685 break;
|
|
686
|
|
687 case T_BOOLEAN: // 1 byte
|
|
688 assert(value->type() == T_INT, "Agreement.");
|
|
689 _obj->bool_field_put(offset, (jboolean)value->get_int());
|
|
690 break;
|
|
691
|
|
692 default:
|
|
693 ShouldNotReachHere();
|
|
694 }
|
|
695 _i++;
|
|
696 }
|
|
697 };
|
|
698
|
|
699 // restore elements of an eliminated type array
|
|
700 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
|
|
701 StackValue* low;
|
|
702 jlong lval;
|
|
703 int index = 0;
|
|
704
|
|
705 for (int i = 0; i < sv->field_size(); i++) {
|
|
706 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
|
|
707 switch(type) {
|
|
708 case T_BOOLEAN: obj->bool_at_put (index, (jboolean) value->get_int()); break;
|
|
709 case T_BYTE: obj->byte_at_put (index, (jbyte) value->get_int()); break;
|
|
710 case T_CHAR: obj->char_at_put (index, (jchar) value->get_int()); break;
|
|
711 case T_SHORT: obj->short_at_put(index, (jshort) value->get_int()); break;
|
|
712 case T_INT: obj->int_at_put (index, (jint) value->get_int()); break;
|
|
713 case T_FLOAT: obj->float_at_put(index, (jfloat) value->get_int()); break;
|
|
714 case T_LONG:
|
|
715 case T_DOUBLE:
|
|
716 low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
|
|
717 lval = jlong_from((jint)value->get_int(), (jint)low->get_int());
|
|
718 sv->value()->long_field_put(index, lval);
|
|
719 break;
|
|
720 default:
|
|
721 ShouldNotReachHere();
|
|
722 }
|
|
723 index++;
|
|
724 }
|
|
725 }
|
|
726
|
|
727
|
|
728 // restore fields of an eliminated object array
|
|
729 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
|
|
730 for (int i = 0; i < sv->field_size(); i++) {
|
|
731 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
|
|
732 assert(value->type() == T_OBJECT, "object element expected");
|
|
733 obj->obj_at_put(i, value->get_obj()());
|
|
734 }
|
|
735 }
|
|
736
|
|
737
|
|
738 // restore fields of all eliminated objects and arrays
|
|
739 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
|
|
740 for (int i = 0; i < objects->length(); i++) {
|
|
741 ObjectValue* sv = (ObjectValue*) objects->at(i);
|
|
742 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
|
|
743 Handle obj = sv->value();
|
|
744 assert(obj.not_null(), "reallocation was missed");
|
|
745
|
|
746 if (k->oop_is_instance()) {
|
|
747 instanceKlass* ik = instanceKlass::cast(k());
|
|
748 FieldReassigner reassign(fr, reg_map, sv, obj());
|
|
749 ik->do_nonstatic_fields(&reassign);
|
|
750 } else if (k->oop_is_typeArray()) {
|
|
751 typeArrayKlass* ak = typeArrayKlass::cast(k());
|
|
752 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
|
|
753 } else if (k->oop_is_objArray()) {
|
|
754 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
|
|
755 }
|
|
756 }
|
|
757 }
|
|
758
|
|
759
|
|
760 // relock objects for which synchronization was eliminated
|
|
761 void Deoptimization::relock_objects(frame* fr, RegisterMap* reg_map, GrowableArray<MonitorValue*>* monitors) {
|
|
762 for (int i = 0; i < monitors->length(); i++) {
|
|
763 MonitorValue* mv = monitors->at(i);
|
|
764 StackValue* owner = StackValue::create_stack_value(fr, reg_map, mv->owner());
|
|
765 if (mv->eliminated()) {
|
|
766 Handle obj = owner->get_obj();
|
|
767 assert(obj.not_null(), "reallocation was missed");
|
|
768 BasicLock* lock = StackValue::resolve_monitor_lock(fr, mv->basic_lock());
|
|
769 lock->set_displaced_header(obj->mark());
|
|
770 obj->set_mark((markOop) lock);
|
|
771 }
|
|
772 assert(owner->get_obj()->is_locked(), "object must be locked now");
|
|
773 }
|
|
774 }
|
|
775
|
|
776
|
|
777 #ifndef PRODUCT
|
|
778 // print information about reallocated objects
|
|
779 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
|
|
780 fieldDescriptor fd;
|
|
781
|
|
782 for (int i = 0; i < objects->length(); i++) {
|
|
783 ObjectValue* sv = (ObjectValue*) objects->at(i);
|
|
784 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
|
|
785 Handle obj = sv->value();
|
|
786
|
|
787 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
|
|
788 k->as_klassOop()->print_value();
|
|
789 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
|
|
790 tty->cr();
|
|
791
|
|
792 if (Verbose) {
|
|
793 k->oop_print_on(obj(), tty);
|
|
794 }
|
|
795 }
|
|
796 }
|
|
797 #endif
|
|
798 #endif // COMPILER2
|
|
799
|
|
800 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
|
|
801
|
|
802 #ifndef PRODUCT
|
|
803 if (TraceDeoptimization) {
|
|
804 ttyLocker ttyl;
|
|
805 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
|
|
806 fr.print_on(tty);
|
|
807 tty->print_cr(" Virtual frames (innermost first):");
|
|
808 for (int index = 0; index < chunk->length(); index++) {
|
|
809 compiledVFrame* vf = chunk->at(index);
|
|
810 tty->print(" %2d - ", index);
|
|
811 vf->print_value();
|
|
812 int bci = chunk->at(index)->raw_bci();
|
|
813 const char* code_name;
|
|
814 if (bci == SynchronizationEntryBCI) {
|
|
815 code_name = "sync entry";
|
|
816 } else {
|
|
817 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci);
|
|
818 code_name = Bytecodes::name(code);
|
|
819 }
|
|
820 tty->print(" - %s", code_name);
|
|
821 tty->print_cr(" @ bci %d ", bci);
|
|
822 if (Verbose) {
|
|
823 vf->print();
|
|
824 tty->cr();
|
|
825 }
|
|
826 }
|
|
827 }
|
|
828 #endif
|
|
829
|
|
830 // Register map for next frame (used for stack crawl). We capture
|
|
831 // the state of the deopt'ing frame's caller. Thus if we need to
|
|
832 // stuff a C2I adapter we can properly fill in the callee-save
|
|
833 // register locations.
|
|
834 frame caller = fr.sender(reg_map);
|
|
835 int frame_size = caller.sp() - fr.sp();
|
|
836
|
|
837 frame sender = caller;
|
|
838
|
|
839 // Since the Java thread being deoptimized will eventually adjust it's own stack,
|
|
840 // the vframeArray containing the unpacking information is allocated in the C heap.
|
|
841 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
|
|
842 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
|
|
843
|
|
844 // Compare the vframeArray to the collected vframes
|
|
845 assert(array->structural_compare(thread, chunk), "just checking");
|
|
846 Events::log("# vframes = %d", (intptr_t)chunk->length());
|
|
847
|
|
848 #ifndef PRODUCT
|
|
849 if (TraceDeoptimization) {
|
|
850 ttyLocker ttyl;
|
|
851 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
|
|
852 if (Verbose) {
|
|
853 int count = 0;
|
|
854 // this used to leak deoptimizedVFrame like it was going out of style!!!
|
|
855 for (int index = 0; index < array->frames(); index++ ) {
|
|
856 vframeArrayElement* e = array->element(index);
|
|
857 e->print(tty);
|
|
858
|
|
859 /*
|
|
860 No printing yet.
|
|
861 array->vframe_at(index)->print_activation(count++);
|
|
862 // better as...
|
|
863 array->print_activation_for(index, count++);
|
|
864 */
|
|
865 }
|
|
866 }
|
|
867 }
|
|
868 #endif // PRODUCT
|
|
869
|
|
870 return array;
|
|
871 }
|
|
872
|
|
873
|
|
874 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
|
|
875 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
|
|
876 for (int i = 0; i < monitors->length(); i++) {
|
|
877 MonitorInfo* mon_info = monitors->at(i);
|
|
878 if (mon_info->owner() != NULL) {
|
|
879 objects_to_revoke->append(Handle(mon_info->owner()));
|
|
880 }
|
|
881 }
|
|
882 }
|
|
883
|
|
884
|
|
885 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
|
|
886 if (!UseBiasedLocking) {
|
|
887 return;
|
|
888 }
|
|
889
|
|
890 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
|
|
891
|
|
892 // Unfortunately we don't have a RegisterMap available in most of
|
|
893 // the places we want to call this routine so we need to walk the
|
|
894 // stack again to update the register map.
|
|
895 if (map == NULL || !map->update_map()) {
|
|
896 StackFrameStream sfs(thread, true);
|
|
897 bool found = false;
|
|
898 while (!found && !sfs.is_done()) {
|
|
899 frame* cur = sfs.current();
|
|
900 sfs.next();
|
|
901 found = cur->id() == fr.id();
|
|
902 }
|
|
903 assert(found, "frame to be deoptimized not found on target thread's stack");
|
|
904 map = sfs.register_map();
|
|
905 }
|
|
906
|
|
907 vframe* vf = vframe::new_vframe(&fr, map, thread);
|
|
908 compiledVFrame* cvf = compiledVFrame::cast(vf);
|
|
909 // Revoke monitors' biases in all scopes
|
|
910 while (!cvf->is_top()) {
|
|
911 collect_monitors(cvf, objects_to_revoke);
|
|
912 cvf = compiledVFrame::cast(cvf->sender());
|
|
913 }
|
|
914 collect_monitors(cvf, objects_to_revoke);
|
|
915
|
|
916 if (SafepointSynchronize::is_at_safepoint()) {
|
|
917 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
|
|
918 } else {
|
|
919 BiasedLocking::revoke(objects_to_revoke);
|
|
920 }
|
|
921 }
|
|
922
|
|
923
|
|
924 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
|
|
925 if (!UseBiasedLocking) {
|
|
926 return;
|
|
927 }
|
|
928
|
|
929 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
|
|
930 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
|
|
931 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
|
|
932 if (jt->has_last_Java_frame()) {
|
|
933 StackFrameStream sfs(jt, true);
|
|
934 while (!sfs.is_done()) {
|
|
935 frame* cur = sfs.current();
|
|
936 if (cb->contains(cur->pc())) {
|
|
937 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
|
|
938 compiledVFrame* cvf = compiledVFrame::cast(vf);
|
|
939 // Revoke monitors' biases in all scopes
|
|
940 while (!cvf->is_top()) {
|
|
941 collect_monitors(cvf, objects_to_revoke);
|
|
942 cvf = compiledVFrame::cast(cvf->sender());
|
|
943 }
|
|
944 collect_monitors(cvf, objects_to_revoke);
|
|
945 }
|
|
946 sfs.next();
|
|
947 }
|
|
948 }
|
|
949 }
|
|
950 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
|
|
951 }
|
|
952
|
|
953
|
|
954 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
|
|
955 assert(fr.can_be_deoptimized(), "checking frame type");
|
|
956
|
|
957 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
|
|
958
|
|
959 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id());
|
|
960
|
|
961 // Patch the nmethod so that when execution returns to it we will
|
|
962 // deopt the execution state and return to the interpreter.
|
|
963 fr.deoptimize(thread);
|
|
964 }
|
|
965
|
|
966 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
|
|
967 // Deoptimize only if the frame comes from compile code.
|
|
968 // Do not deoptimize the frame which is already patched
|
|
969 // during the execution of the loops below.
|
|
970 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
|
|
971 return;
|
|
972 }
|
|
973 ResourceMark rm;
|
|
974 DeoptimizationMarker dm;
|
|
975 if (UseBiasedLocking) {
|
|
976 revoke_biases_of_monitors(thread, fr, map);
|
|
977 }
|
|
978 deoptimize_single_frame(thread, fr);
|
|
979
|
|
980 }
|
|
981
|
|
982
|
|
983 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
|
|
984 // Compute frame and register map based on thread and sp.
|
|
985 RegisterMap reg_map(thread, UseBiasedLocking);
|
|
986 frame fr = thread->last_frame();
|
|
987 while (fr.id() != id) {
|
|
988 fr = fr.sender(®_map);
|
|
989 }
|
|
990 deoptimize(thread, fr, ®_map);
|
|
991 }
|
|
992
|
|
993
|
|
994 // JVMTI PopFrame support
|
|
995 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
|
|
996 {
|
|
997 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
|
|
998 }
|
|
999 JRT_END
|
|
1000
|
|
1001
|
|
1002 #ifdef COMPILER2
|
|
1003 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
|
|
1004 // in case of an unresolved klass entry, load the class.
|
|
1005 if (constant_pool->tag_at(index).is_unresolved_klass()) {
|
|
1006 klassOop tk = constant_pool->klass_at(index, CHECK);
|
|
1007 return;
|
|
1008 }
|
|
1009
|
|
1010 if (!constant_pool->tag_at(index).is_symbol()) return;
|
|
1011
|
|
1012 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
|
|
1013 symbolHandle symbol (THREAD, constant_pool->symbol_at(index));
|
|
1014
|
|
1015 // class name?
|
|
1016 if (symbol->byte_at(0) != '(') {
|
|
1017 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
|
|
1018 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
|
|
1019 return;
|
|
1020 }
|
|
1021
|
|
1022 // then it must be a signature!
|
|
1023 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
|
|
1024 if (ss.is_object()) {
|
|
1025 symbolOop s = ss.as_symbol(CHECK);
|
|
1026 symbolHandle class_name (THREAD, s);
|
|
1027 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
|
|
1028 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
|
|
1029 }
|
|
1030 }
|
|
1031 }
|
|
1032
|
|
1033
|
|
1034 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
|
|
1035 EXCEPTION_MARK;
|
|
1036 load_class_by_index(constant_pool, index, THREAD);
|
|
1037 if (HAS_PENDING_EXCEPTION) {
|
|
1038 // Exception happened during classloading. We ignore the exception here, since it
|
|
1039 // is going to be rethrown since the current activation is going to be deoptimzied and
|
|
1040 // the interpreter will re-execute the bytecode.
|
|
1041 CLEAR_PENDING_EXCEPTION;
|
|
1042 }
|
|
1043 }
|
|
1044
|
|
1045 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
|
|
1046 HandleMark hm;
|
|
1047
|
|
1048 // uncommon_trap() is called at the beginning of the uncommon trap
|
|
1049 // handler. Note this fact before we start generating temporary frames
|
|
1050 // that can confuse an asynchronous stack walker. This counter is
|
|
1051 // decremented at the end of unpack_frames().
|
|
1052 thread->inc_in_deopt_handler();
|
|
1053
|
|
1054 // We need to update the map if we have biased locking.
|
|
1055 RegisterMap reg_map(thread, UseBiasedLocking);
|
|
1056 frame stub_frame = thread->last_frame();
|
|
1057 frame fr = stub_frame.sender(®_map);
|
|
1058 // Make sure the calling nmethod is not getting deoptimized and removed
|
|
1059 // before we are done with it.
|
|
1060 nmethodLocker nl(fr.pc());
|
|
1061
|
|
1062 {
|
|
1063 ResourceMark rm;
|
|
1064
|
|
1065 // Revoke biases of any monitors in the frame to ensure we can migrate them
|
|
1066 revoke_biases_of_monitors(thread, fr, ®_map);
|
|
1067
|
|
1068 DeoptReason reason = trap_request_reason(trap_request);
|
|
1069 DeoptAction action = trap_request_action(trap_request);
|
|
1070 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
|
|
1071
|
|
1072 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request);
|
|
1073 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
|
|
1074 compiledVFrame* cvf = compiledVFrame::cast(vf);
|
|
1075
|
|
1076 nmethod* nm = cvf->code();
|
|
1077
|
|
1078 ScopeDesc* trap_scope = cvf->scope();
|
|
1079 methodHandle trap_method = trap_scope->method();
|
|
1080 int trap_bci = trap_scope->bci();
|
|
1081 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code();
|
|
1082
|
|
1083 // Record this event in the histogram.
|
|
1084 gather_statistics(reason, action, trap_bc);
|
|
1085
|
|
1086 // Ensure that we can record deopt. history:
|
|
1087 bool create_if_missing = ProfileTraps;
|
|
1088
|
|
1089 methodDataHandle trap_mdo
|
|
1090 (THREAD, get_method_data(thread, trap_method, create_if_missing));
|
|
1091
|
|
1092 // Print a bunch of diagnostics, if requested.
|
|
1093 if (TraceDeoptimization || LogCompilation) {
|
|
1094 ResourceMark rm;
|
|
1095 ttyLocker ttyl;
|
|
1096 char buf[100];
|
|
1097 if (xtty != NULL) {
|
|
1098 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
|
|
1099 os::current_thread_id(),
|
|
1100 format_trap_request(buf, sizeof(buf), trap_request));
|
|
1101 nm->log_identity(xtty);
|
|
1102 }
|
|
1103 symbolHandle class_name;
|
|
1104 bool unresolved = false;
|
|
1105 if (unloaded_class_index >= 0) {
|
|
1106 constantPoolHandle constants (THREAD, trap_method->constants());
|
|
1107 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
|
|
1108 class_name = symbolHandle(THREAD,
|
|
1109 constants->klass_name_at(unloaded_class_index));
|
|
1110 unresolved = true;
|
|
1111 if (xtty != NULL)
|
|
1112 xtty->print(" unresolved='1'");
|
|
1113 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
|
|
1114 class_name = symbolHandle(THREAD,
|
|
1115 constants->symbol_at(unloaded_class_index));
|
|
1116 }
|
|
1117 if (xtty != NULL)
|
|
1118 xtty->name(class_name);
|
|
1119 }
|
|
1120 if (xtty != NULL && trap_mdo.not_null()) {
|
|
1121 // Dump the relevant MDO state.
|
|
1122 // This is the deopt count for the current reason, any previous
|
|
1123 // reasons or recompiles seen at this point.
|
|
1124 int dcnt = trap_mdo->trap_count(reason);
|
|
1125 if (dcnt != 0)
|
|
1126 xtty->print(" count='%d'", dcnt);
|
|
1127 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
|
|
1128 int dos = (pdata == NULL)? 0: pdata->trap_state();
|
|
1129 if (dos != 0) {
|
|
1130 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
|
|
1131 if (trap_state_is_recompiled(dos)) {
|
|
1132 int recnt2 = trap_mdo->overflow_recompile_count();
|
|
1133 if (recnt2 != 0)
|
|
1134 xtty->print(" recompiles2='%d'", recnt2);
|
|
1135 }
|
|
1136 }
|
|
1137 }
|
|
1138 if (xtty != NULL) {
|
|
1139 xtty->stamp();
|
|
1140 xtty->end_head();
|
|
1141 }
|
|
1142 if (TraceDeoptimization) { // make noise on the tty
|
|
1143 tty->print("Uncommon trap occurred in");
|
|
1144 nm->method()->print_short_name(tty);
|
|
1145 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
|
|
1146 fr.pc(),
|
|
1147 (int) os::current_thread_id(),
|
|
1148 trap_reason_name(reason),
|
|
1149 trap_action_name(action),
|
|
1150 unloaded_class_index);
|
|
1151 if (class_name.not_null()) {
|
|
1152 tty->print(unresolved ? " unresolved class: " : " symbol: ");
|
|
1153 class_name->print_symbol_on(tty);
|
|
1154 }
|
|
1155 tty->cr();
|
|
1156 }
|
|
1157 if (xtty != NULL) {
|
|
1158 // Log the precise location of the trap.
|
|
1159 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
|
|
1160 xtty->begin_elem("jvms bci='%d'", sd->bci());
|
|
1161 xtty->method(sd->method());
|
|
1162 xtty->end_elem();
|
|
1163 if (sd->is_top()) break;
|
|
1164 }
|
|
1165 xtty->tail("uncommon_trap");
|
|
1166 }
|
|
1167 }
|
|
1168 // (End diagnostic printout.)
|
|
1169
|
|
1170 // Load class if necessary
|
|
1171 if (unloaded_class_index >= 0) {
|
|
1172 constantPoolHandle constants(THREAD, trap_method->constants());
|
|
1173 load_class_by_index(constants, unloaded_class_index);
|
|
1174 }
|
|
1175
|
|
1176 // Flush the nmethod if necessary and desirable.
|
|
1177 //
|
|
1178 // We need to avoid situations where we are re-flushing the nmethod
|
|
1179 // because of a hot deoptimization site. Repeated flushes at the same
|
|
1180 // point need to be detected by the compiler and avoided. If the compiler
|
|
1181 // cannot avoid them (or has a bug and "refuses" to avoid them), this
|
|
1182 // module must take measures to avoid an infinite cycle of recompilation
|
|
1183 // and deoptimization. There are several such measures:
|
|
1184 //
|
|
1185 // 1. If a recompilation is ordered a second time at some site X
|
|
1186 // and for the same reason R, the action is adjusted to 'reinterpret',
|
|
1187 // to give the interpreter time to exercise the method more thoroughly.
|
|
1188 // If this happens, the method's overflow_recompile_count is incremented.
|
|
1189 //
|
|
1190 // 2. If the compiler fails to reduce the deoptimization rate, then
|
|
1191 // the method's overflow_recompile_count will begin to exceed the set
|
|
1192 // limit PerBytecodeRecompilationCutoff. If this happens, the action
|
|
1193 // is adjusted to 'make_not_compilable', and the method is abandoned
|
|
1194 // to the interpreter. This is a performance hit for hot methods,
|
|
1195 // but is better than a disastrous infinite cycle of recompilations.
|
|
1196 // (Actually, only the method containing the site X is abandoned.)
|
|
1197 //
|
|
1198 // 3. In parallel with the previous measures, if the total number of
|
|
1199 // recompilations of a method exceeds the much larger set limit
|
|
1200 // PerMethodRecompilationCutoff, the method is abandoned.
|
|
1201 // This should only happen if the method is very large and has
|
|
1202 // many "lukewarm" deoptimizations. The code which enforces this
|
|
1203 // limit is elsewhere (class nmethod, class methodOopDesc).
|
|
1204 //
|
|
1205 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
|
|
1206 // to recompile at each bytecode independently of the per-BCI cutoff.
|
|
1207 //
|
|
1208 // The decision to update code is up to the compiler, and is encoded
|
|
1209 // in the Action_xxx code. If the compiler requests Action_none
|
|
1210 // no trap state is changed, no compiled code is changed, and the
|
|
1211 // computation suffers along in the interpreter.
|
|
1212 //
|
|
1213 // The other action codes specify various tactics for decompilation
|
|
1214 // and recompilation. Action_maybe_recompile is the loosest, and
|
|
1215 // allows the compiled code to stay around until enough traps are seen,
|
|
1216 // and until the compiler gets around to recompiling the trapping method.
|
|
1217 //
|
|
1218 // The other actions cause immediate removal of the present code.
|
|
1219
|
|
1220 bool update_trap_state = true;
|
|
1221 bool make_not_entrant = false;
|
|
1222 bool make_not_compilable = false;
|
|
1223 bool reset_counters = false;
|
|
1224 switch (action) {
|
|
1225 case Action_none:
|
|
1226 // Keep the old code.
|
|
1227 update_trap_state = false;
|
|
1228 break;
|
|
1229 case Action_maybe_recompile:
|
|
1230 // Do not need to invalidate the present code, but we can
|
|
1231 // initiate another
|
|
1232 // Start compiler without (necessarily) invalidating the nmethod.
|
|
1233 // The system will tolerate the old code, but new code should be
|
|
1234 // generated when possible.
|
|
1235 break;
|
|
1236 case Action_reinterpret:
|
|
1237 // Go back into the interpreter for a while, and then consider
|
|
1238 // recompiling form scratch.
|
|
1239 make_not_entrant = true;
|
|
1240 // Reset invocation counter for outer most method.
|
|
1241 // This will allow the interpreter to exercise the bytecodes
|
|
1242 // for a while before recompiling.
|
|
1243 // By contrast, Action_make_not_entrant is immediate.
|
|
1244 //
|
|
1245 // Note that the compiler will track null_check, null_assert,
|
|
1246 // range_check, and class_check events and log them as if they
|
|
1247 // had been traps taken from compiled code. This will update
|
|
1248 // the MDO trap history so that the next compilation will
|
|
1249 // properly detect hot trap sites.
|
|
1250 reset_counters = true;
|
|
1251 break;
|
|
1252 case Action_make_not_entrant:
|
|
1253 // Request immediate recompilation, and get rid of the old code.
|
|
1254 // Make them not entrant, so next time they are called they get
|
|
1255 // recompiled. Unloaded classes are loaded now so recompile before next
|
|
1256 // time they are called. Same for uninitialized. The interpreter will
|
|
1257 // link the missing class, if any.
|
|
1258 make_not_entrant = true;
|
|
1259 break;
|
|
1260 case Action_make_not_compilable:
|
|
1261 // Give up on compiling this method at all.
|
|
1262 make_not_entrant = true;
|
|
1263 make_not_compilable = true;
|
|
1264 break;
|
|
1265 default:
|
|
1266 ShouldNotReachHere();
|
|
1267 }
|
|
1268
|
|
1269 // Setting +ProfileTraps fixes the following, on all platforms:
|
|
1270 // 4852688: ProfileInterpreter is off by default for ia64. The result is
|
|
1271 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
|
|
1272 // recompile relies on a methodDataOop to record heroic opt failures.
|
|
1273
|
|
1274 // Whether the interpreter is producing MDO data or not, we also need
|
|
1275 // to use the MDO to detect hot deoptimization points and control
|
|
1276 // aggressive optimization.
|
|
1277 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
|
|
1278 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
|
|
1279 uint this_trap_count = 0;
|
|
1280 bool maybe_prior_trap = false;
|
|
1281 bool maybe_prior_recompile = false;
|
|
1282 ProfileData* pdata
|
|
1283 = query_update_method_data(trap_mdo, trap_bci, reason,
|
|
1284 //outputs:
|
|
1285 this_trap_count,
|
|
1286 maybe_prior_trap,
|
|
1287 maybe_prior_recompile);
|
|
1288 // Because the interpreter also counts null, div0, range, and class
|
|
1289 // checks, these traps from compiled code are double-counted.
|
|
1290 // This is harmless; it just means that the PerXTrapLimit values
|
|
1291 // are in effect a little smaller than they look.
|
|
1292
|
|
1293 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
|
|
1294 if (per_bc_reason != Reason_none) {
|
|
1295 // Now take action based on the partially known per-BCI history.
|
|
1296 if (maybe_prior_trap
|
|
1297 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
|
|
1298 // If there are too many traps at this BCI, force a recompile.
|
|
1299 // This will allow the compiler to see the limit overflow, and
|
|
1300 // take corrective action, if possible. The compiler generally
|
|
1301 // does not use the exact PerBytecodeTrapLimit value, but instead
|
|
1302 // changes its tactics if it sees any traps at all. This provides
|
|
1303 // a little hysteresis, delaying a recompile until a trap happens
|
|
1304 // several times.
|
|
1305 //
|
|
1306 // Actually, since there is only one bit of counter per BCI,
|
|
1307 // the possible per-BCI counts are {0,1,(per-method count)}.
|
|
1308 // This produces accurate results if in fact there is only
|
|
1309 // one hot trap site, but begins to get fuzzy if there are
|
|
1310 // many sites. For example, if there are ten sites each
|
|
1311 // trapping two or more times, they each get the blame for
|
|
1312 // all of their traps.
|
|
1313 make_not_entrant = true;
|
|
1314 }
|
|
1315
|
|
1316 // Detect repeated recompilation at the same BCI, and enforce a limit.
|
|
1317 if (make_not_entrant && maybe_prior_recompile) {
|
|
1318 // More than one recompile at this point.
|
|
1319 trap_mdo->inc_overflow_recompile_count();
|
|
1320 if (maybe_prior_trap
|
|
1321 && ((uint)trap_mdo->overflow_recompile_count()
|
|
1322 > (uint)PerBytecodeRecompilationCutoff)) {
|
|
1323 // Give up on the method containing the bad BCI.
|
|
1324 if (trap_method() == nm->method()) {
|
|
1325 make_not_compilable = true;
|
|
1326 } else {
|
|
1327 trap_method->set_not_compilable();
|
|
1328 // But give grace to the enclosing nm->method().
|
|
1329 }
|
|
1330 }
|
|
1331 }
|
|
1332 } else {
|
|
1333 // For reasons which are not recorded per-bytecode, we simply
|
|
1334 // force recompiles unconditionally.
|
|
1335 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
|
|
1336 make_not_entrant = true;
|
|
1337 }
|
|
1338
|
|
1339 // Go back to the compiler if there are too many traps in this method.
|
|
1340 if (this_trap_count >= (uint)PerMethodTrapLimit) {
|
|
1341 // If there are too many traps in this method, force a recompile.
|
|
1342 // This will allow the compiler to see the limit overflow, and
|
|
1343 // take corrective action, if possible.
|
|
1344 // (This condition is an unlikely backstop only, because the
|
|
1345 // PerBytecodeTrapLimit is more likely to take effect first,
|
|
1346 // if it is applicable.)
|
|
1347 make_not_entrant = true;
|
|
1348 }
|
|
1349
|
|
1350 // Here's more hysteresis: If there has been a recompile at
|
|
1351 // this trap point already, run the method in the interpreter
|
|
1352 // for a while to exercise it more thoroughly.
|
|
1353 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
|
|
1354 reset_counters = true;
|
|
1355 }
|
|
1356
|
|
1357 if (make_not_entrant && pdata != NULL) {
|
|
1358 // Record the recompilation event, if any.
|
|
1359 int tstate0 = pdata->trap_state();
|
|
1360 int tstate1 = trap_state_set_recompiled(tstate0, true);
|
|
1361 if (tstate1 != tstate0)
|
|
1362 pdata->set_trap_state(tstate1);
|
|
1363 }
|
|
1364 }
|
|
1365
|
|
1366 // Take requested actions on the method:
|
|
1367
|
|
1368 // Reset invocation counters
|
|
1369 if (reset_counters) {
|
|
1370 if (nm->is_osr_method())
|
|
1371 reset_invocation_counter(trap_scope, CompileThreshold);
|
|
1372 else
|
|
1373 reset_invocation_counter(trap_scope);
|
|
1374 }
|
|
1375
|
|
1376 // Recompile
|
|
1377 if (make_not_entrant) {
|
|
1378 nm->make_not_entrant();
|
|
1379 }
|
|
1380
|
|
1381 // Give up compiling
|
|
1382 if (make_not_compilable) {
|
|
1383 assert(make_not_entrant, "consistent");
|
|
1384 nm->method()->set_not_compilable();
|
|
1385 }
|
|
1386
|
|
1387 } // Free marked resources
|
|
1388
|
|
1389 }
|
|
1390 JRT_END
|
|
1391
|
|
1392 methodDataOop
|
|
1393 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
|
|
1394 bool create_if_missing) {
|
|
1395 Thread* THREAD = thread;
|
|
1396 methodDataOop mdo = m()->method_data();
|
|
1397 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
|
|
1398 // Build an MDO. Ignore errors like OutOfMemory;
|
|
1399 // that simply means we won't have an MDO to update.
|
|
1400 methodOopDesc::build_interpreter_method_data(m, THREAD);
|
|
1401 if (HAS_PENDING_EXCEPTION) {
|
|
1402 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
|
|
1403 CLEAR_PENDING_EXCEPTION;
|
|
1404 }
|
|
1405 mdo = m()->method_data();
|
|
1406 }
|
|
1407 return mdo;
|
|
1408 }
|
|
1409
|
|
1410 ProfileData*
|
|
1411 Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
|
|
1412 int trap_bci,
|
|
1413 Deoptimization::DeoptReason reason,
|
|
1414 //outputs:
|
|
1415 uint& ret_this_trap_count,
|
|
1416 bool& ret_maybe_prior_trap,
|
|
1417 bool& ret_maybe_prior_recompile) {
|
|
1418 uint prior_trap_count = trap_mdo->trap_count(reason);
|
|
1419 uint this_trap_count = trap_mdo->inc_trap_count(reason);
|
|
1420
|
|
1421 // If the runtime cannot find a place to store trap history,
|
|
1422 // it is estimated based on the general condition of the method.
|
|
1423 // If the method has ever been recompiled, or has ever incurred
|
|
1424 // a trap with the present reason , then this BCI is assumed
|
|
1425 // (pessimistically) to be the culprit.
|
|
1426 bool maybe_prior_trap = (prior_trap_count != 0);
|
|
1427 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
|
|
1428 ProfileData* pdata = NULL;
|
|
1429
|
|
1430
|
|
1431 // For reasons which are recorded per bytecode, we check per-BCI data.
|
|
1432 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
|
|
1433 if (per_bc_reason != Reason_none) {
|
|
1434 // Find the profile data for this BCI. If there isn't one,
|
|
1435 // try to allocate one from the MDO's set of spares.
|
|
1436 // This will let us detect a repeated trap at this point.
|
|
1437 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
|
|
1438
|
|
1439 if (pdata != NULL) {
|
|
1440 // Query the trap state of this profile datum.
|
|
1441 int tstate0 = pdata->trap_state();
|
|
1442 if (!trap_state_has_reason(tstate0, per_bc_reason))
|
|
1443 maybe_prior_trap = false;
|
|
1444 if (!trap_state_is_recompiled(tstate0))
|
|
1445 maybe_prior_recompile = false;
|
|
1446
|
|
1447 // Update the trap state of this profile datum.
|
|
1448 int tstate1 = tstate0;
|
|
1449 // Record the reason.
|
|
1450 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
|
|
1451 // Store the updated state on the MDO, for next time.
|
|
1452 if (tstate1 != tstate0)
|
|
1453 pdata->set_trap_state(tstate1);
|
|
1454 } else {
|
|
1455 if (LogCompilation && xtty != NULL)
|
|
1456 // Missing MDP? Leave a small complaint in the log.
|
|
1457 xtty->elem("missing_mdp bci='%d'", trap_bci);
|
|
1458 }
|
|
1459 }
|
|
1460
|
|
1461 // Return results:
|
|
1462 ret_this_trap_count = this_trap_count;
|
|
1463 ret_maybe_prior_trap = maybe_prior_trap;
|
|
1464 ret_maybe_prior_recompile = maybe_prior_recompile;
|
|
1465 return pdata;
|
|
1466 }
|
|
1467
|
|
1468 void
|
|
1469 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
|
|
1470 ResourceMark rm;
|
|
1471 // Ignored outputs:
|
|
1472 uint ignore_this_trap_count;
|
|
1473 bool ignore_maybe_prior_trap;
|
|
1474 bool ignore_maybe_prior_recompile;
|
|
1475 query_update_method_data(trap_mdo, trap_bci,
|
|
1476 (DeoptReason)reason,
|
|
1477 ignore_this_trap_count,
|
|
1478 ignore_maybe_prior_trap,
|
|
1479 ignore_maybe_prior_recompile);
|
|
1480 }
|
|
1481
|
|
1482 void Deoptimization::reset_invocation_counter(ScopeDesc* trap_scope, jint top_count) {
|
|
1483 ScopeDesc* sd = trap_scope;
|
|
1484 for (; !sd->is_top(); sd = sd->sender()) {
|
|
1485 // Reset ICs of inlined methods, since they can trigger compilations also.
|
|
1486 sd->method()->invocation_counter()->reset();
|
|
1487 }
|
|
1488 InvocationCounter* c = sd->method()->invocation_counter();
|
|
1489 if (top_count != _no_count) {
|
|
1490 // It was an OSR method, so bump the count higher.
|
|
1491 c->set(c->state(), top_count);
|
|
1492 } else {
|
|
1493 c->reset();
|
|
1494 }
|
|
1495 sd->method()->backedge_counter()->reset();
|
|
1496 }
|
|
1497
|
|
1498 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
|
|
1499
|
|
1500 // Still in Java no safepoints
|
|
1501 {
|
|
1502 // This enters VM and may safepoint
|
|
1503 uncommon_trap_inner(thread, trap_request);
|
|
1504 }
|
|
1505 return fetch_unroll_info_helper(thread);
|
|
1506 }
|
|
1507
|
|
1508 // Local derived constants.
|
|
1509 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
|
|
1510 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
|
|
1511 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
|
|
1512
|
|
1513 //---------------------------trap_state_reason---------------------------------
|
|
1514 Deoptimization::DeoptReason
|
|
1515 Deoptimization::trap_state_reason(int trap_state) {
|
|
1516 // This assert provides the link between the width of DataLayout::trap_bits
|
|
1517 // and the encoding of "recorded" reasons. It ensures there are enough
|
|
1518 // bits to store all needed reasons in the per-BCI MDO profile.
|
|
1519 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
|
|
1520 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
|
|
1521 trap_state -= recompile_bit;
|
|
1522 if (trap_state == DS_REASON_MASK) {
|
|
1523 return Reason_many;
|
|
1524 } else {
|
|
1525 assert((int)Reason_none == 0, "state=0 => Reason_none");
|
|
1526 return (DeoptReason)trap_state;
|
|
1527 }
|
|
1528 }
|
|
1529 //-------------------------trap_state_has_reason-------------------------------
|
|
1530 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
|
|
1531 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
|
|
1532 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
|
|
1533 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
|
|
1534 trap_state -= recompile_bit;
|
|
1535 if (trap_state == DS_REASON_MASK) {
|
|
1536 return -1; // true, unspecifically (bottom of state lattice)
|
|
1537 } else if (trap_state == reason) {
|
|
1538 return 1; // true, definitely
|
|
1539 } else if (trap_state == 0) {
|
|
1540 return 0; // false, definitely (top of state lattice)
|
|
1541 } else {
|
|
1542 return 0; // false, definitely
|
|
1543 }
|
|
1544 }
|
|
1545 //-------------------------trap_state_add_reason-------------------------------
|
|
1546 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
|
|
1547 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
|
|
1548 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
|
|
1549 trap_state -= recompile_bit;
|
|
1550 if (trap_state == DS_REASON_MASK) {
|
|
1551 return trap_state + recompile_bit; // already at state lattice bottom
|
|
1552 } else if (trap_state == reason) {
|
|
1553 return trap_state + recompile_bit; // the condition is already true
|
|
1554 } else if (trap_state == 0) {
|
|
1555 return reason + recompile_bit; // no condition has yet been true
|
|
1556 } else {
|
|
1557 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
|
|
1558 }
|
|
1559 }
|
|
1560 //-----------------------trap_state_is_recompiled------------------------------
|
|
1561 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
|
|
1562 return (trap_state & DS_RECOMPILE_BIT) != 0;
|
|
1563 }
|
|
1564 //-----------------------trap_state_set_recompiled-----------------------------
|
|
1565 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
|
|
1566 if (z) return trap_state | DS_RECOMPILE_BIT;
|
|
1567 else return trap_state & ~DS_RECOMPILE_BIT;
|
|
1568 }
|
|
1569 //---------------------------format_trap_state---------------------------------
|
|
1570 // This is used for debugging and diagnostics, including hotspot.log output.
|
|
1571 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
|
|
1572 int trap_state) {
|
|
1573 DeoptReason reason = trap_state_reason(trap_state);
|
|
1574 bool recomp_flag = trap_state_is_recompiled(trap_state);
|
|
1575 // Re-encode the state from its decoded components.
|
|
1576 int decoded_state = 0;
|
|
1577 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
|
|
1578 decoded_state = trap_state_add_reason(decoded_state, reason);
|
|
1579 if (recomp_flag)
|
|
1580 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
|
|
1581 // If the state re-encodes properly, format it symbolically.
|
|
1582 // Because this routine is used for debugging and diagnostics,
|
|
1583 // be robust even if the state is a strange value.
|
|
1584 size_t len;
|
|
1585 if (decoded_state != trap_state) {
|
|
1586 // Random buggy state that doesn't decode??
|
|
1587 len = jio_snprintf(buf, buflen, "#%d", trap_state);
|
|
1588 } else {
|
|
1589 len = jio_snprintf(buf, buflen, "%s%s",
|
|
1590 trap_reason_name(reason),
|
|
1591 recomp_flag ? " recompiled" : "");
|
|
1592 }
|
|
1593 if (len >= buflen)
|
|
1594 buf[buflen-1] = '\0';
|
|
1595 return buf;
|
|
1596 }
|
|
1597
|
|
1598
|
|
1599 //--------------------------------statics--------------------------------------
|
|
1600 Deoptimization::DeoptAction Deoptimization::_unloaded_action
|
|
1601 = Deoptimization::Action_reinterpret;
|
|
1602 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
|
|
1603 // Note: Keep this in sync. with enum DeoptReason.
|
|
1604 "none",
|
|
1605 "null_check",
|
|
1606 "null_assert",
|
|
1607 "range_check",
|
|
1608 "class_check",
|
|
1609 "array_check",
|
|
1610 "intrinsic",
|
|
1611 "unloaded",
|
|
1612 "uninitialized",
|
|
1613 "unreached",
|
|
1614 "unhandled",
|
|
1615 "constraint",
|
|
1616 "div0_check",
|
|
1617 "age"
|
|
1618 };
|
|
1619 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
|
|
1620 // Note: Keep this in sync. with enum DeoptAction.
|
|
1621 "none",
|
|
1622 "maybe_recompile",
|
|
1623 "reinterpret",
|
|
1624 "make_not_entrant",
|
|
1625 "make_not_compilable"
|
|
1626 };
|
|
1627
|
|
1628 const char* Deoptimization::trap_reason_name(int reason) {
|
|
1629 if (reason == Reason_many) return "many";
|
|
1630 if ((uint)reason < Reason_LIMIT)
|
|
1631 return _trap_reason_name[reason];
|
|
1632 static char buf[20];
|
|
1633 sprintf(buf, "reason%d", reason);
|
|
1634 return buf;
|
|
1635 }
|
|
1636 const char* Deoptimization::trap_action_name(int action) {
|
|
1637 if ((uint)action < Action_LIMIT)
|
|
1638 return _trap_action_name[action];
|
|
1639 static char buf[20];
|
|
1640 sprintf(buf, "action%d", action);
|
|
1641 return buf;
|
|
1642 }
|
|
1643
|
|
1644 // This is used for debugging and diagnostics, including hotspot.log output.
|
|
1645 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
|
|
1646 int trap_request) {
|
|
1647 jint unloaded_class_index = trap_request_index(trap_request);
|
|
1648 const char* reason = trap_reason_name(trap_request_reason(trap_request));
|
|
1649 const char* action = trap_action_name(trap_request_action(trap_request));
|
|
1650 size_t len;
|
|
1651 if (unloaded_class_index < 0) {
|
|
1652 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
|
|
1653 reason, action);
|
|
1654 } else {
|
|
1655 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
|
|
1656 reason, action, unloaded_class_index);
|
|
1657 }
|
|
1658 if (len >= buflen)
|
|
1659 buf[buflen-1] = '\0';
|
|
1660 return buf;
|
|
1661 }
|
|
1662
|
|
1663 juint Deoptimization::_deoptimization_hist
|
|
1664 [Deoptimization::Reason_LIMIT]
|
|
1665 [1 + Deoptimization::Action_LIMIT]
|
|
1666 [Deoptimization::BC_CASE_LIMIT]
|
|
1667 = {0};
|
|
1668
|
|
1669 enum {
|
|
1670 LSB_BITS = 8,
|
|
1671 LSB_MASK = right_n_bits(LSB_BITS)
|
|
1672 };
|
|
1673
|
|
1674 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
|
|
1675 Bytecodes::Code bc) {
|
|
1676 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
|
|
1677 assert(action >= 0 && action < Action_LIMIT, "oob");
|
|
1678 _deoptimization_hist[Reason_none][0][0] += 1; // total
|
|
1679 _deoptimization_hist[reason][0][0] += 1; // per-reason total
|
|
1680 juint* cases = _deoptimization_hist[reason][1+action];
|
|
1681 juint* bc_counter_addr = NULL;
|
|
1682 juint bc_counter = 0;
|
|
1683 // Look for an unused counter, or an exact match to this BC.
|
|
1684 if (bc != Bytecodes::_illegal) {
|
|
1685 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
|
|
1686 juint* counter_addr = &cases[bc_case];
|
|
1687 juint counter = *counter_addr;
|
|
1688 if ((counter == 0 && bc_counter_addr == NULL)
|
|
1689 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
|
|
1690 // this counter is either free or is already devoted to this BC
|
|
1691 bc_counter_addr = counter_addr;
|
|
1692 bc_counter = counter | bc;
|
|
1693 }
|
|
1694 }
|
|
1695 }
|
|
1696 if (bc_counter_addr == NULL) {
|
|
1697 // Overflow, or no given bytecode.
|
|
1698 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
|
|
1699 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
|
|
1700 }
|
|
1701 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
|
|
1702 }
|
|
1703
|
|
1704 jint Deoptimization::total_deoptimization_count() {
|
|
1705 return _deoptimization_hist[Reason_none][0][0];
|
|
1706 }
|
|
1707
|
|
1708 jint Deoptimization::deoptimization_count(DeoptReason reason) {
|
|
1709 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
|
|
1710 return _deoptimization_hist[reason][0][0];
|
|
1711 }
|
|
1712
|
|
1713 void Deoptimization::print_statistics() {
|
|
1714 juint total = total_deoptimization_count();
|
|
1715 juint account = total;
|
|
1716 if (total != 0) {
|
|
1717 ttyLocker ttyl;
|
|
1718 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
|
|
1719 tty->print_cr("Deoptimization traps recorded:");
|
|
1720 #define PRINT_STAT_LINE(name, r) \
|
|
1721 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
|
|
1722 PRINT_STAT_LINE("total", total);
|
|
1723 // For each non-zero entry in the histogram, print the reason,
|
|
1724 // the action, and (if specifically known) the type of bytecode.
|
|
1725 for (int reason = 0; reason < Reason_LIMIT; reason++) {
|
|
1726 for (int action = 0; action < Action_LIMIT; action++) {
|
|
1727 juint* cases = _deoptimization_hist[reason][1+action];
|
|
1728 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
|
|
1729 juint counter = cases[bc_case];
|
|
1730 if (counter != 0) {
|
|
1731 char name[1*K];
|
|
1732 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
|
|
1733 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
|
|
1734 bc = Bytecodes::_illegal;
|
|
1735 sprintf(name, "%s/%s/%s",
|
|
1736 trap_reason_name(reason),
|
|
1737 trap_action_name(action),
|
|
1738 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
|
|
1739 juint r = counter >> LSB_BITS;
|
|
1740 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
|
|
1741 account -= r;
|
|
1742 }
|
|
1743 }
|
|
1744 }
|
|
1745 }
|
|
1746 if (account != 0) {
|
|
1747 PRINT_STAT_LINE("unaccounted", account);
|
|
1748 }
|
|
1749 #undef PRINT_STAT_LINE
|
|
1750 if (xtty != NULL) xtty->tail("statistics");
|
|
1751 }
|
|
1752 }
|
|
1753 #else // COMPILER2
|
|
1754
|
|
1755
|
|
1756 // Stubs for C1 only system.
|
|
1757 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
|
|
1758 return false;
|
|
1759 }
|
|
1760
|
|
1761 const char* Deoptimization::trap_reason_name(int reason) {
|
|
1762 return "unknown";
|
|
1763 }
|
|
1764
|
|
1765 void Deoptimization::print_statistics() {
|
|
1766 // no output
|
|
1767 }
|
|
1768
|
|
1769 void
|
|
1770 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
|
|
1771 // no udpate
|
|
1772 }
|
|
1773
|
|
1774 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
|
|
1775 return 0;
|
|
1776 }
|
|
1777
|
|
1778 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
|
|
1779 Bytecodes::Code bc) {
|
|
1780 // no update
|
|
1781 }
|
|
1782
|
|
1783 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
|
|
1784 int trap_state) {
|
|
1785 jio_snprintf(buf, buflen, "#%d", trap_state);
|
|
1786 return buf;
|
|
1787 }
|
|
1788
|
|
1789 #endif // COMPILER2
|