Print this page
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/runtime/deoptimization.cpp
+++ new/src/share/vm/runtime/deoptimization.cpp
1 1 /*
2 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/systemDictionary.hpp"
27 27 #include "code/debugInfoRec.hpp"
28 28 #include "code/nmethod.hpp"
29 29 #include "code/pcDesc.hpp"
30 30 #include "code/scopeDesc.hpp"
31 31 #include "interpreter/bytecode.hpp"
32 32 #include "interpreter/interpreter.hpp"
33 33 #include "interpreter/oopMapCache.hpp"
34 34 #include "memory/allocation.inline.hpp"
35 35 #include "memory/oopFactory.hpp"
36 36 #include "memory/resourceArea.hpp"
37 37 #include "oops/methodOop.hpp"
38 38 #include "oops/oop.inline.hpp"
39 39 #include "prims/jvmtiThreadState.hpp"
40 40 #include "runtime/biasedLocking.hpp"
41 41 #include "runtime/compilationPolicy.hpp"
42 42 #include "runtime/deoptimization.hpp"
43 43 #include "runtime/interfaceSupport.hpp"
44 44 #include "runtime/sharedRuntime.hpp"
45 45 #include "runtime/signature.hpp"
46 46 #include "runtime/stubRoutines.hpp"
47 47 #include "runtime/thread.hpp"
48 48 #include "runtime/vframe.hpp"
49 49 #include "runtime/vframeArray.hpp"
50 50 #include "runtime/vframe_hp.hpp"
51 51 #include "utilities/events.hpp"
52 52 #include "utilities/xmlstream.hpp"
53 53 #ifdef TARGET_ARCH_x86
54 54 # include "vmreg_x86.inline.hpp"
55 55 #endif
56 56 #ifdef TARGET_ARCH_sparc
57 57 # include "vmreg_sparc.inline.hpp"
58 58 #endif
59 59 #ifdef TARGET_ARCH_zero
60 60 # include "vmreg_zero.inline.hpp"
61 61 #endif
62 62 #ifdef TARGET_ARCH_arm
63 63 # include "vmreg_arm.inline.hpp"
64 64 #endif
65 65 #ifdef TARGET_ARCH_ppc
66 66 # include "vmreg_ppc.inline.hpp"
67 67 #endif
68 68 #ifdef COMPILER2
69 69 #ifdef TARGET_ARCH_MODEL_x86_32
70 70 # include "adfiles/ad_x86_32.hpp"
71 71 #endif
72 72 #ifdef TARGET_ARCH_MODEL_x86_64
73 73 # include "adfiles/ad_x86_64.hpp"
74 74 #endif
75 75 #ifdef TARGET_ARCH_MODEL_sparc
76 76 # include "adfiles/ad_sparc.hpp"
77 77 #endif
78 78 #ifdef TARGET_ARCH_MODEL_zero
79 79 # include "adfiles/ad_zero.hpp"
80 80 #endif
81 81 #ifdef TARGET_ARCH_MODEL_arm
82 82 # include "adfiles/ad_arm.hpp"
83 83 #endif
84 84 #ifdef TARGET_ARCH_MODEL_ppc
85 85 # include "adfiles/ad_ppc.hpp"
86 86 #endif
87 87 #endif
88 88
89 89 bool DeoptimizationMarker::_is_active = false;
90 90
91 91 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
92 92 int caller_adjustment,
93 93 int caller_actual_parameters,
94 94 int number_of_frames,
95 95 intptr_t* frame_sizes,
96 96 address* frame_pcs,
97 97 BasicType return_type) {
98 98 _size_of_deoptimized_frame = size_of_deoptimized_frame;
99 99 _caller_adjustment = caller_adjustment;
100 100 _caller_actual_parameters = caller_actual_parameters;
101 101 _number_of_frames = number_of_frames;
102 102 _frame_sizes = frame_sizes;
103 103 _frame_pcs = frame_pcs;
104 104 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2);
105 105 _return_type = return_type;
106 106 _initial_fp = 0;
107 107 // PD (x86 only)
108 108 _counter_temp = 0;
109 109 _unpack_kind = 0;
110 110 _sender_sp_temp = 0;
111 111
112 112 _total_frame_sizes = size_of_frames();
113 113 }
114 114
115 115
116 116 Deoptimization::UnrollBlock::~UnrollBlock() {
117 117 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
118 118 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
119 119 FREE_C_HEAP_ARRAY(intptr_t, _register_block);
120 120 }
121 121
122 122
123 123 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
124 124 assert(register_number < RegisterMap::reg_count, "checking register number");
125 125 return &_register_block[register_number * 2];
126 126 }
127 127
128 128
129 129
130 130 int Deoptimization::UnrollBlock::size_of_frames() const {
131 131 // Acount first for the adjustment of the initial frame
132 132 int result = _caller_adjustment;
133 133 for (int index = 0; index < number_of_frames(); index++) {
134 134 result += frame_sizes()[index];
135 135 }
136 136 return result;
137 137 }
138 138
139 139
140 140 void Deoptimization::UnrollBlock::print() {
141 141 ttyLocker ttyl;
142 142 tty->print_cr("UnrollBlock");
143 143 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
144 144 tty->print( " frame_sizes: ");
145 145 for (int index = 0; index < number_of_frames(); index++) {
146 146 tty->print("%d ", frame_sizes()[index]);
147 147 }
148 148 tty->cr();
149 149 }
150 150
151 151
152 152 // In order to make fetch_unroll_info work properly with escape
153 153 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
154 154 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
155 155 // of previously eliminated objects occurs in realloc_objects, which is
156 156 // called from the method fetch_unroll_info_helper below.
157 157 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
158 158 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
159 159 // but makes the entry a little slower. There is however a little dance we have to
160 160 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
161 161
162 162 // fetch_unroll_info() is called at the beginning of the deoptimization
163 163 // handler. Note this fact before we start generating temporary frames
164 164 // that can confuse an asynchronous stack walker. This counter is
165 165 // decremented at the end of unpack_frames().
166 166 thread->inc_in_deopt_handler();
167 167
168 168 return fetch_unroll_info_helper(thread);
169 169 JRT_END
170 170
171 171
172 172 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
173 173 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
174 174
175 175 // Note: there is a safepoint safety issue here. No matter whether we enter
176 176 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
177 177 // the vframeArray is created.
178 178 //
179 179
180 180 // Allocate our special deoptimization ResourceMark
181 181 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
182 182 assert(thread->deopt_mark() == NULL, "Pending deopt!");
183 183 thread->set_deopt_mark(dmark);
184 184
185 185 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
186 186 RegisterMap map(thread, true);
187 187 RegisterMap dummy_map(thread, false);
188 188 // Now get the deoptee with a valid map
189 189 frame deoptee = stub_frame.sender(&map);
190 190 // Set the deoptee nmethod
191 191 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
192 192 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
193 193
194 194 if (VerifyStack) {
195 195 thread->validate_frame_layout();
196 196 }
197 197
198 198 // Create a growable array of VFrames where each VFrame represents an inlined
199 199 // Java frame. This storage is allocated with the usual system arena.
200 200 assert(deoptee.is_compiled_frame(), "Wrong frame type");
201 201 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
202 202 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
203 203 while (!vf->is_top()) {
204 204 assert(vf->is_compiled_frame(), "Wrong frame type");
205 205 chunk->push(compiledVFrame::cast(vf));
206 206 vf = vf->sender();
207 207 }
208 208 assert(vf->is_compiled_frame(), "Wrong frame type");
209 209 chunk->push(compiledVFrame::cast(vf));
210 210
211 211 #ifdef COMPILER2
212 212 // Reallocate the non-escaping objects and restore their fields. Then
213 213 // relock objects if synchronization on them was eliminated.
214 214 if (DoEscapeAnalysis) {
215 215 if (EliminateAllocations) {
216 216 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
217 217 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
218 218
219 219 // The flag return_oop() indicates call sites which return oop
220 220 // in compiled code. Such sites include java method calls,
221 221 // runtime calls (for example, used to allocate new objects/arrays
222 222 // on slow code path) and any other calls generated in compiled code.
223 223 // It is not guaranteed that we can get such information here only
224 224 // by analyzing bytecode in deoptimized frames. This is why this flag
225 225 // is set during method compilation (see Compile::Process_OopMap_Node()).
226 226 bool save_oop_result = chunk->at(0)->scope()->return_oop();
227 227 Handle return_value;
228 228 if (save_oop_result) {
229 229 // Reallocation may trigger GC. If deoptimization happened on return from
230 230 // call which returns oop we need to save it since it is not in oopmap.
231 231 oop result = deoptee.saved_oop_result(&map);
232 232 assert(result == NULL || result->is_oop(), "must be oop");
233 233 return_value = Handle(thread, result);
234 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
235 235 if (TraceDeoptimization) {
236 236 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
237 237 }
238 238 }
239 239 bool reallocated = false;
240 240 if (objects != NULL) {
241 241 JRT_BLOCK
242 242 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
243 243 JRT_END
244 244 }
245 245 if (reallocated) {
246 246 reassign_fields(&deoptee, &map, objects);
247 247 #ifndef PRODUCT
248 248 if (TraceDeoptimization) {
249 249 ttyLocker ttyl;
250 250 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
251 251 print_objects(objects);
252 252 }
253 253 #endif
254 254 }
255 255 if (save_oop_result) {
256 256 // Restore result.
257 257 deoptee.set_saved_oop_result(&map, return_value());
258 258 }
259 259 }
260 260 if (EliminateLocks) {
261 261 #ifndef PRODUCT
262 262 bool first = true;
263 263 #endif
264 264 for (int i = 0; i < chunk->length(); i++) {
265 265 compiledVFrame* cvf = chunk->at(i);
266 266 assert (cvf->scope() != NULL,"expect only compiled java frames");
267 267 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
268 268 if (monitors->is_nonempty()) {
269 269 relock_objects(monitors, thread);
270 270 #ifndef PRODUCT
271 271 if (TraceDeoptimization) {
272 272 ttyLocker ttyl;
273 273 for (int j = 0; j < monitors->length(); j++) {
274 274 MonitorInfo* mi = monitors->at(j);
275 275 if (mi->eliminated()) {
276 276 if (first) {
277 277 first = false;
278 278 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
279 279 }
280 280 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
281 281 }
282 282 }
283 283 }
284 284 #endif
285 285 }
286 286 }
287 287 }
288 288 }
289 289 #endif // COMPILER2
290 290 // Ensure that no safepoint is taken after pointers have been stored
291 291 // in fields of rematerialized objects. If a safepoint occurs from here on
292 292 // out the java state residing in the vframeArray will be missed.
293 293 No_Safepoint_Verifier no_safepoint;
294 294
295 295 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
296 296
297 297 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
298 298 thread->set_vframe_array_head(array);
299 299
300 300 // Now that the vframeArray has been created if we have any deferred local writes
301 301 // added by jvmti then we can free up that structure as the data is now in the
302 302 // vframeArray
303 303
304 304 if (thread->deferred_locals() != NULL) {
305 305 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
306 306 int i = 0;
307 307 do {
308 308 // Because of inlining we could have multiple vframes for a single frame
309 309 // and several of the vframes could have deferred writes. Find them all.
310 310 if (list->at(i)->id() == array->original().id()) {
311 311 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
312 312 list->remove_at(i);
313 313 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
314 314 delete dlv;
315 315 } else {
316 316 i++;
317 317 }
318 318 } while ( i < list->length() );
319 319 if (list->length() == 0) {
320 320 thread->set_deferred_locals(NULL);
321 321 // free the list and elements back to C heap.
322 322 delete list;
323 323 }
324 324
325 325 }
326 326
327 327 #ifndef SHARK
328 328 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
329 329 CodeBlob* cb = stub_frame.cb();
330 330 // Verify we have the right vframeArray
331 331 assert(cb->frame_size() >= 0, "Unexpected frame size");
332 332 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
333 333
334 334 // If the deopt call site is a MethodHandle invoke call site we have
335 335 // to adjust the unpack_sp.
336 336 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
337 337 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
338 338 unpack_sp = deoptee.unextended_sp();
339 339
340 340 #ifdef ASSERT
341 341 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
342 342 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp);
343 343 #endif
344 344 #else
345 345 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
346 346 #endif // !SHARK
347 347
348 348 // This is a guarantee instead of an assert because if vframe doesn't match
349 349 // we will unpack the wrong deoptimized frame and wind up in strange places
350 350 // where it will be very difficult to figure out what went wrong. Better
351 351 // to die an early death here than some very obscure death later when the
352 352 // trail is cold.
353 353 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
354 354 // in that it will fail to detect a problem when there is one. This needs
355 355 // more work in tiger timeframe.
356 356 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
357 357
358 358 int number_of_frames = array->frames();
359 359
360 360 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
361 361 // virtual activation, which is the reverse of the elements in the vframes array.
362 362 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames);
363 363 // +1 because we always have an interpreter return address for the final slot.
364 364 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1);
365 365 int callee_parameters = 0;
366 366 int callee_locals = 0;
367 367 int popframe_extra_args = 0;
368 368 // Create an interpreter return address for the stub to use as its return
369 369 // address so the skeletal frames are perfectly walkable
370 370 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
371 371
372 372 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
373 373 // activation be put back on the expression stack of the caller for reexecution
374 374 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
375 375 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
376 376 }
377 377
378 378 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
379 379 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
380 380 // than simply use array->sender.pc(). This requires us to walk the current set of frames
381 381 //
382 382 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
383 383 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
384 384
385 385 // It's possible that the number of paramters at the call site is
386 386 // different than number of arguments in the callee when method
387 387 // handles are used. If the caller is interpreted get the real
388 388 // value so that the proper amount of space can be added to it's
389 389 // frame.
390 390 int caller_actual_parameters = callee_parameters;
391 391 if (deopt_sender.is_interpreted_frame()) {
392 392 methodHandle method = deopt_sender.interpreter_frame_method();
393 393 Bytecode_invoke cur = Bytecode_invoke_check(method,
394 394 deopt_sender.interpreter_frame_bci());
395 395 Symbol* signature = method->constants()->signature_ref_at(cur.index());
396 396 ArgumentSizeComputer asc(signature);
397 397 caller_actual_parameters = asc.size() + (cur.has_receiver() ? 1 : 0);
398 398 }
399 399
400 400 //
401 401 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
402 402 // frame_sizes/frame_pcs[1] next oldest frame (int)
403 403 // frame_sizes/frame_pcs[n] youngest frame (int)
404 404 //
405 405 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
406 406 // owns the space for the return address to it's caller). Confusing ain't it.
407 407 //
408 408 // The vframe array can address vframes with indices running from
409 409 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
410 410 // When we create the skeletal frames we need the oldest frame to be in the zero slot
411 411 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
412 412 // so things look a little strange in this loop.
413 413 //
414 414 for (int index = 0; index < array->frames(); index++ ) {
415 415 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
416 416 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
417 417 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
418 418 int caller_parms = callee_parameters;
419 419 if (index == array->frames() - 1) {
420 420 // Use the value from the interpreted caller
421 421 caller_parms = caller_actual_parameters;
422 422 }
423 423 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms,
424 424 callee_parameters,
425 425 callee_locals,
426 426 index == 0,
427 427 popframe_extra_args);
428 428 // This pc doesn't have to be perfect just good enough to identify the frame
429 429 // as interpreted so the skeleton frame will be walkable
430 430 // The correct pc will be set when the skeleton frame is completely filled out
431 431 // The final pc we store in the loop is wrong and will be overwritten below
432 432 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
433 433
434 434 callee_parameters = array->element(index)->method()->size_of_parameters();
435 435 callee_locals = array->element(index)->method()->max_locals();
436 436 popframe_extra_args = 0;
437 437 }
438 438
439 439 // Compute whether the root vframe returns a float or double value.
440 440 BasicType return_type;
441 441 {
442 442 HandleMark hm;
443 443 methodHandle method(thread, array->element(0)->method());
444 444 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
445 445 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
446 446 }
447 447
448 448 // Compute information for handling adapters and adjusting the frame size of the caller.
449 449 int caller_adjustment = 0;
450 450
451 451 // Compute the amount the oldest interpreter frame will have to adjust
452 452 // its caller's stack by. If the caller is a compiled frame then
453 453 // we pretend that the callee has no parameters so that the
454 454 // extension counts for the full amount of locals and not just
455 455 // locals-parms. This is because without a c2i adapter the parm
456 456 // area as created by the compiled frame will not be usable by
457 457 // the interpreter. (Depending on the calling convention there
458 458 // may not even be enough space).
459 459
460 460 // QQQ I'd rather see this pushed down into last_frame_adjust
461 461 // and have it take the sender (aka caller).
462 462
463 463 if (deopt_sender.is_compiled_frame()) {
464 464 caller_adjustment = last_frame_adjust(0, callee_locals);
465 465 } else if (callee_locals > caller_actual_parameters) {
466 466 // The caller frame may need extending to accommodate
467 467 // non-parameter locals of the first unpacked interpreted frame.
468 468 // Compute that adjustment.
469 469 caller_adjustment = last_frame_adjust(caller_actual_parameters, callee_locals);
470 470 }
471 471
472 472 // If the sender is deoptimized the we must retrieve the address of the handler
473 473 // since the frame will "magically" show the original pc before the deopt
474 474 // and we'd undo the deopt.
475 475
476 476 frame_pcs[0] = deopt_sender.raw_pc();
477 477
478 478 #ifndef SHARK
479 479 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
480 480 #endif // SHARK
481 481
482 482 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
483 483 caller_adjustment * BytesPerWord,
484 484 caller_actual_parameters,
485 485 number_of_frames,
486 486 frame_sizes,
487 487 frame_pcs,
488 488 return_type);
489 489 // On some platforms, we need a way to pass fp to the unpacking code
490 490 // so the skeletal frames come out correct.
491 491 info->set_initial_fp((intptr_t) array->sender().fp());
492 492
493 493 if (array->frames() > 1) {
494 494 if (VerifyStack && TraceDeoptimization) {
495 495 tty->print_cr("Deoptimizing method containing inlining");
496 496 }
497 497 }
498 498
499 499 array->set_unroll_block(info);
500 500 return info;
501 501 }
502 502
503 503 // Called to cleanup deoptimization data structures in normal case
504 504 // after unpacking to stack and when stack overflow error occurs
505 505 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
506 506 vframeArray *array) {
507 507
508 508 // Get array if coming from exception
509 509 if (array == NULL) {
510 510 array = thread->vframe_array_head();
511 511 }
512 512 thread->set_vframe_array_head(NULL);
513 513
514 514 // Free the previous UnrollBlock
515 515 vframeArray* old_array = thread->vframe_array_last();
516 516 thread->set_vframe_array_last(array);
517 517
518 518 if (old_array != NULL) {
519 519 UnrollBlock* old_info = old_array->unroll_block();
520 520 old_array->set_unroll_block(NULL);
521 521 delete old_info;
522 522 delete old_array;
523 523 }
524 524
525 525 // Deallocate any resource creating in this routine and any ResourceObjs allocated
526 526 // inside the vframeArray (StackValueCollections)
527 527
528 528 delete thread->deopt_mark();
529 529 thread->set_deopt_mark(NULL);
530 530 thread->set_deopt_nmethod(NULL);
531 531
532 532
533 533 if (JvmtiExport::can_pop_frame()) {
534 534 #ifndef CC_INTERP
535 535 // Regardless of whether we entered this routine with the pending
536 536 // popframe condition bit set, we should always clear it now
537 537 thread->clear_popframe_condition();
538 538 #else
539 539 // C++ interpeter will clear has_pending_popframe when it enters
540 540 // with method_resume. For deopt_resume2 we clear it now.
541 541 if (thread->popframe_forcing_deopt_reexecution())
542 542 thread->clear_popframe_condition();
543 543 #endif /* CC_INTERP */
544 544 }
545 545
546 546 // unpack_frames() is called at the end of the deoptimization handler
547 547 // and (in C2) at the end of the uncommon trap handler. Note this fact
548 548 // so that an asynchronous stack walker can work again. This counter is
549 549 // incremented at the beginning of fetch_unroll_info() and (in C2) at
550 550 // the beginning of uncommon_trap().
551 551 thread->dec_in_deopt_handler();
552 552 }
553 553
554 554
555 555 // Return BasicType of value being returned
556 556 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
557 557
558 558 // We are already active int he special DeoptResourceMark any ResourceObj's we
559 559 // allocate will be freed at the end of the routine.
560 560
561 561 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
562 562 // but makes the entry a little slower. There is however a little dance we have to
563 563 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
564 564 ResetNoHandleMark rnhm; // No-op in release/product versions
565 565 HandleMark hm;
566 566
567 567 frame stub_frame = thread->last_frame();
568 568
569 569 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
570 570 // must point to the vframeArray for the unpack frame.
571 571 vframeArray* array = thread->vframe_array_head();
572 572
573 573 #ifndef PRODUCT
574 574 if (TraceDeoptimization) {
575 575 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
576 576 }
577 577 #endif
578 578
579 579 UnrollBlock* info = array->unroll_block();
580 580
581 581 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
582 582 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
583 583
584 584 BasicType bt = info->return_type();
585 585
586 586 // If we have an exception pending, claim that the return type is an oop
587 587 // so the deopt_blob does not overwrite the exception_oop.
588 588
589 589 if (exec_mode == Unpack_exception)
590 590 bt = T_OBJECT;
591 591
592 592 // Cleanup thread deopt data
593 593 cleanup_deopt_info(thread, array);
594 594
595 595 #ifndef PRODUCT
596 596 if (VerifyStack) {
597 597 ResourceMark res_mark;
598 598
599 599 thread->validate_frame_layout();
600 600
601 601 // Verify that the just-unpacked frames match the interpreter's
602 602 // notions of expression stack and locals
603 603 vframeArray* cur_array = thread->vframe_array_last();
604 604 RegisterMap rm(thread, false);
605 605 rm.set_include_argument_oops(false);
606 606 bool is_top_frame = true;
607 607 int callee_size_of_parameters = 0;
608 608 int callee_max_locals = 0;
609 609 for (int i = 0; i < cur_array->frames(); i++) {
610 610 vframeArrayElement* el = cur_array->element(i);
611 611 frame* iframe = el->iframe();
612 612 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
613 613
614 614 // Get the oop map for this bci
615 615 InterpreterOopMap mask;
616 616 int cur_invoke_parameter_size = 0;
617 617 bool try_next_mask = false;
618 618 int next_mask_expression_stack_size = -1;
619 619 int top_frame_expression_stack_adjustment = 0;
620 620 methodHandle mh(thread, iframe->interpreter_frame_method());
621 621 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
622 622 BytecodeStream str(mh);
623 623 str.set_start(iframe->interpreter_frame_bci());
624 624 int max_bci = mh->code_size();
625 625 // Get to the next bytecode if possible
626 626 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
627 627 // Check to see if we can grab the number of outgoing arguments
628 628 // at an uncommon trap for an invoke (where the compiler
629 629 // generates debug info before the invoke has executed)
630 630 Bytecodes::Code cur_code = str.next();
631 631 if (cur_code == Bytecodes::_invokevirtual ||
632 632 cur_code == Bytecodes::_invokespecial ||
633 633 cur_code == Bytecodes::_invokestatic ||
634 634 cur_code == Bytecodes::_invokeinterface) {
635 635 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
636 636 Symbol* signature = invoke.signature();
637 637 ArgumentSizeComputer asc(signature);
638 638 cur_invoke_parameter_size = asc.size();
639 639 if (cur_code != Bytecodes::_invokestatic) {
640 640 // Add in receiver
641 641 ++cur_invoke_parameter_size;
642 642 }
643 643 }
644 644 if (str.bci() < max_bci) {
645 645 Bytecodes::Code bc = str.next();
646 646 if (bc >= 0) {
647 647 // The interpreter oop map generator reports results before
648 648 // the current bytecode has executed except in the case of
649 649 // calls. It seems to be hard to tell whether the compiler
650 650 // has emitted debug information matching the "state before"
651 651 // a given bytecode or the state after, so we try both
652 652 switch (cur_code) {
653 653 case Bytecodes::_invokevirtual:
654 654 case Bytecodes::_invokespecial:
655 655 case Bytecodes::_invokestatic:
656 656 case Bytecodes::_invokeinterface:
657 657 case Bytecodes::_athrow:
658 658 break;
659 659 default: {
660 660 InterpreterOopMap next_mask;
661 661 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
662 662 next_mask_expression_stack_size = next_mask.expression_stack_size();
663 663 // Need to subtract off the size of the result type of
664 664 // the bytecode because this is not described in the
665 665 // debug info but returned to the interpreter in the TOS
666 666 // caching register
667 667 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
668 668 if (bytecode_result_type != T_ILLEGAL) {
669 669 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
670 670 }
671 671 assert(top_frame_expression_stack_adjustment >= 0, "");
672 672 try_next_mask = true;
673 673 break;
674 674 }
675 675 }
676 676 }
677 677 }
678 678
679 679 // Verify stack depth and oops in frame
680 680 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
681 681 if (!(
682 682 /* SPARC */
683 683 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
684 684 /* x86 */
685 685 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
686 686 (try_next_mask &&
687 687 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
688 688 top_frame_expression_stack_adjustment))) ||
689 689 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
690 690 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
691 691 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
692 692 )) {
693 693 ttyLocker ttyl;
694 694
695 695 // Print out some information that will help us debug the problem
696 696 tty->print_cr("Wrong number of expression stack elements during deoptimization");
697 697 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
698 698 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
699 699 iframe->interpreter_frame_expression_stack_size());
700 700 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
701 701 tty->print_cr(" try_next_mask = %d", try_next_mask);
702 702 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
703 703 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
704 704 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
705 705 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
706 706 tty->print_cr(" exec_mode = %d", exec_mode);
707 707 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
708 708 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
709 709 tty->print_cr(" Interpreted frames:");
710 710 for (int k = 0; k < cur_array->frames(); k++) {
711 711 vframeArrayElement* el = cur_array->element(k);
712 712 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
713 713 }
714 714 cur_array->print_on_2(tty);
715 715 guarantee(false, "wrong number of expression stack elements during deopt");
716 716 }
717 717 VerifyOopClosure verify;
718 718 iframe->oops_interpreted_do(&verify, &rm, false);
719 719 callee_size_of_parameters = mh->size_of_parameters();
720 720 callee_max_locals = mh->max_locals();
721 721 is_top_frame = false;
722 722 }
723 723 }
724 724 #endif /* !PRODUCT */
725 725
726 726
727 727 return bt;
728 728 JRT_END
729 729
730 730
731 731 int Deoptimization::deoptimize_dependents() {
732 732 Threads::deoptimized_wrt_marked_nmethods();
733 733 return 0;
734 734 }
735 735
736 736
737 737 #ifdef COMPILER2
738 738 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
739 739 Handle pending_exception(thread->pending_exception());
740 740 const char* exception_file = thread->exception_file();
741 741 int exception_line = thread->exception_line();
742 742 thread->clear_pending_exception();
743 743
744 744 for (int i = 0; i < objects->length(); i++) {
745 745 assert(objects->at(i)->is_object(), "invalid debug information");
746 746 ObjectValue* sv = (ObjectValue*) objects->at(i);
747 747
748 748 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
749 749 oop obj = NULL;
750 750
751 751 if (k->oop_is_instance()) {
752 752 instanceKlass* ik = instanceKlass::cast(k());
753 753 obj = ik->allocate_instance(CHECK_(false));
754 754 } else if (k->oop_is_typeArray()) {
755 755 typeArrayKlass* ak = typeArrayKlass::cast(k());
756 756 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
757 757 int len = sv->field_size() / type2size[ak->element_type()];
758 758 obj = ak->allocate(len, CHECK_(false));
759 759 } else if (k->oop_is_objArray()) {
760 760 objArrayKlass* ak = objArrayKlass::cast(k());
761 761 obj = ak->allocate(sv->field_size(), CHECK_(false));
762 762 }
763 763
764 764 assert(obj != NULL, "allocation failed");
765 765 assert(sv->value().is_null(), "redundant reallocation");
766 766 sv->set_value(obj);
767 767 }
768 768
769 769 if (pending_exception.not_null()) {
770 770 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
771 771 }
772 772
773 773 return true;
774 774 }
775 775
776 776 // This assumes that the fields are stored in ObjectValue in the same order
777 777 // they are yielded by do_nonstatic_fields.
778 778 class FieldReassigner: public FieldClosure {
779 779 frame* _fr;
780 780 RegisterMap* _reg_map;
781 781 ObjectValue* _sv;
782 782 instanceKlass* _ik;
783 783 oop _obj;
784 784
785 785 int _i;
786 786 public:
787 787 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
788 788 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
789 789
790 790 int i() const { return _i; }
791 791
792 792
793 793 void do_field(fieldDescriptor* fd) {
794 794 intptr_t val;
795 795 StackValue* value =
796 796 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
797 797 int offset = fd->offset();
798 798 switch (fd->field_type()) {
799 799 case T_OBJECT: case T_ARRAY:
800 800 assert(value->type() == T_OBJECT, "Agreement.");
801 801 _obj->obj_field_put(offset, value->get_obj()());
802 802 break;
803 803
804 804 case T_LONG: case T_DOUBLE: {
805 805 assert(value->type() == T_INT, "Agreement.");
806 806 StackValue* low =
807 807 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
808 808 #ifdef _LP64
809 809 jlong res = (jlong)low->get_int();
810 810 #else
811 811 #ifdef SPARC
812 812 // For SPARC we have to swap high and low words.
813 813 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
814 814 #else
815 815 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
816 816 #endif //SPARC
817 817 #endif
818 818 _obj->long_field_put(offset, res);
819 819 break;
820 820 }
821 821 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
822 822 case T_INT: case T_FLOAT: // 4 bytes.
823 823 assert(value->type() == T_INT, "Agreement.");
824 824 val = value->get_int();
825 825 _obj->int_field_put(offset, (jint)*((jint*)&val));
826 826 break;
827 827
828 828 case T_SHORT: case T_CHAR: // 2 bytes
829 829 assert(value->type() == T_INT, "Agreement.");
830 830 val = value->get_int();
831 831 _obj->short_field_put(offset, (jshort)*((jint*)&val));
832 832 break;
833 833
834 834 case T_BOOLEAN: case T_BYTE: // 1 byte
835 835 assert(value->type() == T_INT, "Agreement.");
836 836 val = value->get_int();
837 837 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
838 838 break;
839 839
840 840 default:
841 841 ShouldNotReachHere();
842 842 }
843 843 _i++;
844 844 }
845 845 };
846 846
847 847 // restore elements of an eliminated type array
848 848 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
849 849 int index = 0;
850 850 intptr_t val;
851 851
852 852 for (int i = 0; i < sv->field_size(); i++) {
853 853 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
854 854 switch(type) {
855 855 case T_LONG: case T_DOUBLE: {
856 856 assert(value->type() == T_INT, "Agreement.");
857 857 StackValue* low =
858 858 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
859 859 #ifdef _LP64
860 860 jlong res = (jlong)low->get_int();
861 861 #else
862 862 #ifdef SPARC
863 863 // For SPARC we have to swap high and low words.
864 864 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
865 865 #else
866 866 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
867 867 #endif //SPARC
868 868 #endif
869 869 obj->long_at_put(index, res);
870 870 break;
871 871 }
872 872
873 873 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
874 874 case T_INT: case T_FLOAT: // 4 bytes.
875 875 assert(value->type() == T_INT, "Agreement.");
876 876 val = value->get_int();
877 877 obj->int_at_put(index, (jint)*((jint*)&val));
878 878 break;
879 879
880 880 case T_SHORT: case T_CHAR: // 2 bytes
881 881 assert(value->type() == T_INT, "Agreement.");
882 882 val = value->get_int();
883 883 obj->short_at_put(index, (jshort)*((jint*)&val));
884 884 break;
885 885
886 886 case T_BOOLEAN: case T_BYTE: // 1 byte
887 887 assert(value->type() == T_INT, "Agreement.");
888 888 val = value->get_int();
889 889 obj->bool_at_put(index, (jboolean)*((jint*)&val));
890 890 break;
891 891
892 892 default:
893 893 ShouldNotReachHere();
894 894 }
895 895 index++;
896 896 }
897 897 }
898 898
899 899
900 900 // restore fields of an eliminated object array
901 901 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
902 902 for (int i = 0; i < sv->field_size(); i++) {
903 903 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
904 904 assert(value->type() == T_OBJECT, "object element expected");
905 905 obj->obj_at_put(i, value->get_obj()());
906 906 }
907 907 }
908 908
909 909
910 910 // restore fields of all eliminated objects and arrays
911 911 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
912 912 for (int i = 0; i < objects->length(); i++) {
913 913 ObjectValue* sv = (ObjectValue*) objects->at(i);
914 914 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
915 915 Handle obj = sv->value();
916 916 assert(obj.not_null(), "reallocation was missed");
917 917
918 918 if (k->oop_is_instance()) {
919 919 instanceKlass* ik = instanceKlass::cast(k());
920 920 FieldReassigner reassign(fr, reg_map, sv, obj());
921 921 ik->do_nonstatic_fields(&reassign);
922 922 } else if (k->oop_is_typeArray()) {
923 923 typeArrayKlass* ak = typeArrayKlass::cast(k());
924 924 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
925 925 } else if (k->oop_is_objArray()) {
926 926 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
927 927 }
928 928 }
929 929 }
930 930
931 931
932 932 // relock objects for which synchronization was eliminated
933 933 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
934 934 for (int i = 0; i < monitors->length(); i++) {
935 935 MonitorInfo* mon_info = monitors->at(i);
936 936 if (mon_info->eliminated()) {
937 937 assert(mon_info->owner() != NULL, "reallocation was missed");
938 938 Handle obj = Handle(mon_info->owner());
939 939 markOop mark = obj->mark();
940 940 if (UseBiasedLocking && mark->has_bias_pattern()) {
941 941 // New allocated objects may have the mark set to anonymously biased.
942 942 // Also the deoptimized method may called methods with synchronization
943 943 // where the thread-local object is bias locked to the current thread.
944 944 assert(mark->is_biased_anonymously() ||
945 945 mark->biased_locker() == thread, "should be locked to current thread");
946 946 // Reset mark word to unbiased prototype.
947 947 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
948 948 obj->set_mark(unbiased_prototype);
949 949 }
950 950 BasicLock* lock = mon_info->lock();
951 951 ObjectSynchronizer::slow_enter(obj, lock, thread);
952 952 }
953 953 assert(mon_info->owner()->is_locked(), "object must be locked now");
954 954 }
955 955 }
956 956
957 957
958 958 #ifndef PRODUCT
959 959 // print information about reallocated objects
960 960 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
961 961 fieldDescriptor fd;
962 962
963 963 for (int i = 0; i < objects->length(); i++) {
964 964 ObjectValue* sv = (ObjectValue*) objects->at(i);
965 965 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()());
966 966 Handle obj = sv->value();
967 967
968 968 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
969 969 k->as_klassOop()->print_value();
970 970 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
971 971 tty->cr();
972 972
973 973 if (Verbose) {
974 974 k->oop_print_on(obj(), tty);
975 975 }
976 976 }
977 977 }
978 978 #endif
979 979 #endif // COMPILER2
980 980
981 981 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
982 982
983 983 #ifndef PRODUCT
984 984 if (TraceDeoptimization) {
985 985 ttyLocker ttyl;
986 986 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
987 987 fr.print_on(tty);
988 988 tty->print_cr(" Virtual frames (innermost first):");
989 989 for (int index = 0; index < chunk->length(); index++) {
990 990 compiledVFrame* vf = chunk->at(index);
991 991 tty->print(" %2d - ", index);
992 992 vf->print_value();
993 993 int bci = chunk->at(index)->raw_bci();
994 994 const char* code_name;
995 995 if (bci == SynchronizationEntryBCI) {
996 996 code_name = "sync entry";
997 997 } else {
998 998 Bytecodes::Code code = vf->method()->code_at(bci);
999 999 code_name = Bytecodes::name(code);
1000 1000 }
1001 1001 tty->print(" - %s", code_name);
1002 1002 tty->print_cr(" @ bci %d ", bci);
1003 1003 if (Verbose) {
1004 1004 vf->print();
1005 1005 tty->cr();
1006 1006 }
1007 1007 }
1008 1008 }
1009 1009 #endif
1010 1010
1011 1011 // Register map for next frame (used for stack crawl). We capture
1012 1012 // the state of the deopt'ing frame's caller. Thus if we need to
1013 1013 // stuff a C2I adapter we can properly fill in the callee-save
1014 1014 // register locations.
1015 1015 frame caller = fr.sender(reg_map);
1016 1016 int frame_size = caller.sp() - fr.sp();
1017 1017
1018 1018 frame sender = caller;
1019 1019
1020 1020 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1021 1021 // the vframeArray containing the unpacking information is allocated in the C heap.
1022 1022 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1023 1023 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
1024 1024
1025 1025 // Compare the vframeArray to the collected vframes
1026 1026 assert(array->structural_compare(thread, chunk), "just checking");
1027 1027 Events::log("# vframes = %d", (intptr_t)chunk->length());
1028 1028
1029 1029 #ifndef PRODUCT
1030 1030 if (TraceDeoptimization) {
1031 1031 ttyLocker ttyl;
1032 1032 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1033 1033 }
1034 1034 #endif // PRODUCT
1035 1035
1036 1036 return array;
1037 1037 }
1038 1038
1039 1039
1040 1040 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1041 1041 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1042 1042 for (int i = 0; i < monitors->length(); i++) {
1043 1043 MonitorInfo* mon_info = monitors->at(i);
1044 1044 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1045 1045 objects_to_revoke->append(Handle(mon_info->owner()));
1046 1046 }
1047 1047 }
1048 1048 }
1049 1049
1050 1050
1051 1051 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1052 1052 if (!UseBiasedLocking) {
1053 1053 return;
1054 1054 }
1055 1055
1056 1056 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1057 1057
1058 1058 // Unfortunately we don't have a RegisterMap available in most of
1059 1059 // the places we want to call this routine so we need to walk the
1060 1060 // stack again to update the register map.
1061 1061 if (map == NULL || !map->update_map()) {
1062 1062 StackFrameStream sfs(thread, true);
1063 1063 bool found = false;
1064 1064 while (!found && !sfs.is_done()) {
1065 1065 frame* cur = sfs.current();
1066 1066 sfs.next();
1067 1067 found = cur->id() == fr.id();
1068 1068 }
1069 1069 assert(found, "frame to be deoptimized not found on target thread's stack");
1070 1070 map = sfs.register_map();
1071 1071 }
1072 1072
1073 1073 vframe* vf = vframe::new_vframe(&fr, map, thread);
1074 1074 compiledVFrame* cvf = compiledVFrame::cast(vf);
1075 1075 // Revoke monitors' biases in all scopes
1076 1076 while (!cvf->is_top()) {
1077 1077 collect_monitors(cvf, objects_to_revoke);
1078 1078 cvf = compiledVFrame::cast(cvf->sender());
1079 1079 }
1080 1080 collect_monitors(cvf, objects_to_revoke);
1081 1081
1082 1082 if (SafepointSynchronize::is_at_safepoint()) {
1083 1083 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1084 1084 } else {
1085 1085 BiasedLocking::revoke(objects_to_revoke);
1086 1086 }
1087 1087 }
1088 1088
1089 1089
1090 1090 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1091 1091 if (!UseBiasedLocking) {
1092 1092 return;
1093 1093 }
1094 1094
1095 1095 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1096 1096 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1097 1097 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1098 1098 if (jt->has_last_Java_frame()) {
1099 1099 StackFrameStream sfs(jt, true);
1100 1100 while (!sfs.is_done()) {
1101 1101 frame* cur = sfs.current();
1102 1102 if (cb->contains(cur->pc())) {
1103 1103 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1104 1104 compiledVFrame* cvf = compiledVFrame::cast(vf);
1105 1105 // Revoke monitors' biases in all scopes
1106 1106 while (!cvf->is_top()) {
1107 1107 collect_monitors(cvf, objects_to_revoke);
1108 1108 cvf = compiledVFrame::cast(cvf->sender());
1109 1109 }
1110 1110 collect_monitors(cvf, objects_to_revoke);
1111 1111 }
1112 1112 sfs.next();
1113 1113 }
1114 1114 }
1115 1115 }
1116 1116 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1117 1117 }
1118 1118
1119 1119
1120 1120 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1121 1121 assert(fr.can_be_deoptimized(), "checking frame type");
1122 1122
1123 1123 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1124 1124
1125 1125 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id());
1126 1126
1127 1127 // Patch the nmethod so that when execution returns to it we will
1128 1128 // deopt the execution state and return to the interpreter.
1129 1129 fr.deoptimize(thread);
1130 1130 }
1131 1131
1132 1132 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1133 1133 // Deoptimize only if the frame comes from compile code.
1134 1134 // Do not deoptimize the frame which is already patched
1135 1135 // during the execution of the loops below.
1136 1136 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1137 1137 return;
1138 1138 }
1139 1139 ResourceMark rm;
1140 1140 DeoptimizationMarker dm;
1141 1141 if (UseBiasedLocking) {
1142 1142 revoke_biases_of_monitors(thread, fr, map);
1143 1143 }
1144 1144 deoptimize_single_frame(thread, fr);
1145 1145
1146 1146 }
1147 1147
1148 1148
1149 1149 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1150 1150 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1151 1151 "can only deoptimize other thread at a safepoint");
1152 1152 // Compute frame and register map based on thread and sp.
1153 1153 RegisterMap reg_map(thread, UseBiasedLocking);
1154 1154 frame fr = thread->last_frame();
1155 1155 while (fr.id() != id) {
1156 1156 fr = fr.sender(®_map);
1157 1157 }
1158 1158 deoptimize(thread, fr, ®_map);
1159 1159 }
1160 1160
1161 1161
1162 1162 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1163 1163 if (thread == Thread::current()) {
1164 1164 Deoptimization::deoptimize_frame_internal(thread, id);
1165 1165 } else {
1166 1166 VM_DeoptimizeFrame deopt(thread, id);
1167 1167 VMThread::execute(&deopt);
1168 1168 }
1169 1169 }
1170 1170
1171 1171
1172 1172 // JVMTI PopFrame support
1173 1173 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1174 1174 {
1175 1175 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1176 1176 }
1177 1177 JRT_END
1178 1178
1179 1179
1180 1180 #if defined(COMPILER2) || defined(SHARK)
1181 1181 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1182 1182 // in case of an unresolved klass entry, load the class.
1183 1183 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1184 1184 klassOop tk = constant_pool->klass_at(index, CHECK);
1185 1185 return;
1186 1186 }
1187 1187
1188 1188 if (!constant_pool->tag_at(index).is_symbol()) return;
1189 1189
1190 1190 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader());
1191 1191 Symbol* symbol = constant_pool->symbol_at(index);
1192 1192
1193 1193 // class name?
1194 1194 if (symbol->byte_at(0) != '(') {
1195 1195 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1196 1196 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1197 1197 return;
1198 1198 }
1199 1199
1200 1200 // then it must be a signature!
1201 1201 ResourceMark rm(THREAD);
1202 1202 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1203 1203 if (ss.is_object()) {
1204 1204 Symbol* class_name = ss.as_symbol(CHECK);
1205 1205 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain());
1206 1206 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1207 1207 }
1208 1208 }
1209 1209 }
1210 1210
1211 1211
1212 1212 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1213 1213 EXCEPTION_MARK;
1214 1214 load_class_by_index(constant_pool, index, THREAD);
1215 1215 if (HAS_PENDING_EXCEPTION) {
1216 1216 // Exception happened during classloading. We ignore the exception here, since it
1217 1217 // is going to be rethrown since the current activation is going to be deoptimzied and
1218 1218 // the interpreter will re-execute the bytecode.
1219 1219 CLEAR_PENDING_EXCEPTION;
1220 1220 }
1221 1221 }
1222 1222
1223 1223 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1224 1224 HandleMark hm;
1225 1225
1226 1226 // uncommon_trap() is called at the beginning of the uncommon trap
1227 1227 // handler. Note this fact before we start generating temporary frames
1228 1228 // that can confuse an asynchronous stack walker. This counter is
1229 1229 // decremented at the end of unpack_frames().
1230 1230 thread->inc_in_deopt_handler();
1231 1231
1232 1232 // We need to update the map if we have biased locking.
1233 1233 RegisterMap reg_map(thread, UseBiasedLocking);
1234 1234 frame stub_frame = thread->last_frame();
1235 1235 frame fr = stub_frame.sender(®_map);
1236 1236 // Make sure the calling nmethod is not getting deoptimized and removed
1237 1237 // before we are done with it.
1238 1238 nmethodLocker nl(fr.pc());
1239 1239
1240 1240 {
1241 1241 ResourceMark rm;
1242 1242
1243 1243 // Revoke biases of any monitors in the frame to ensure we can migrate them
1244 1244 revoke_biases_of_monitors(thread, fr, ®_map);
1245 1245
1246 1246 DeoptReason reason = trap_request_reason(trap_request);
1247 1247 DeoptAction action = trap_request_action(trap_request);
1248 1248 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1249 1249
1250 1250 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request);
1251 1251 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1252 1252 compiledVFrame* cvf = compiledVFrame::cast(vf);
1253 1253
1254 1254 nmethod* nm = cvf->code();
1255 1255
1256 1256 ScopeDesc* trap_scope = cvf->scope();
1257 1257 methodHandle trap_method = trap_scope->method();
1258 1258 int trap_bci = trap_scope->bci();
1259 1259 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1260 1260
1261 1261 // Record this event in the histogram.
1262 1262 gather_statistics(reason, action, trap_bc);
1263 1263
1264 1264 // Ensure that we can record deopt. history:
1265 1265 bool create_if_missing = ProfileTraps;
1266 1266
1267 1267 methodDataHandle trap_mdo
1268 1268 (THREAD, get_method_data(thread, trap_method, create_if_missing));
1269 1269
1270 1270 // Print a bunch of diagnostics, if requested.
1271 1271 if (TraceDeoptimization || LogCompilation) {
1272 1272 ResourceMark rm;
1273 1273 ttyLocker ttyl;
1274 1274 char buf[100];
1275 1275 if (xtty != NULL) {
1276 1276 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1277 1277 os::current_thread_id(),
1278 1278 format_trap_request(buf, sizeof(buf), trap_request));
1279 1279 nm->log_identity(xtty);
1280 1280 }
1281 1281 Symbol* class_name = NULL;
1282 1282 bool unresolved = false;
1283 1283 if (unloaded_class_index >= 0) {
1284 1284 constantPoolHandle constants (THREAD, trap_method->constants());
1285 1285 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1286 1286 class_name = constants->klass_name_at(unloaded_class_index);
1287 1287 unresolved = true;
1288 1288 if (xtty != NULL)
1289 1289 xtty->print(" unresolved='1'");
1290 1290 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1291 1291 class_name = constants->symbol_at(unloaded_class_index);
1292 1292 }
1293 1293 if (xtty != NULL)
1294 1294 xtty->name(class_name);
1295 1295 }
1296 1296 if (xtty != NULL && trap_mdo.not_null()) {
1297 1297 // Dump the relevant MDO state.
1298 1298 // This is the deopt count for the current reason, any previous
1299 1299 // reasons or recompiles seen at this point.
1300 1300 int dcnt = trap_mdo->trap_count(reason);
1301 1301 if (dcnt != 0)
1302 1302 xtty->print(" count='%d'", dcnt);
1303 1303 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1304 1304 int dos = (pdata == NULL)? 0: pdata->trap_state();
1305 1305 if (dos != 0) {
1306 1306 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1307 1307 if (trap_state_is_recompiled(dos)) {
1308 1308 int recnt2 = trap_mdo->overflow_recompile_count();
1309 1309 if (recnt2 != 0)
1310 1310 xtty->print(" recompiles2='%d'", recnt2);
1311 1311 }
1312 1312 }
1313 1313 }
1314 1314 if (xtty != NULL) {
1315 1315 xtty->stamp();
1316 1316 xtty->end_head();
1317 1317 }
1318 1318 if (TraceDeoptimization) { // make noise on the tty
1319 1319 tty->print("Uncommon trap occurred in");
1320 1320 nm->method()->print_short_name(tty);
1321 1321 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d",
1322 1322 fr.pc(),
1323 1323 (int) os::current_thread_id(),
1324 1324 trap_reason_name(reason),
1325 1325 trap_action_name(action),
1326 1326 unloaded_class_index);
1327 1327 if (class_name != NULL) {
1328 1328 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1329 1329 class_name->print_symbol_on(tty);
1330 1330 }
1331 1331 tty->cr();
1332 1332 }
1333 1333 if (xtty != NULL) {
1334 1334 // Log the precise location of the trap.
1335 1335 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1336 1336 xtty->begin_elem("jvms bci='%d'", sd->bci());
1337 1337 xtty->method(sd->method());
1338 1338 xtty->end_elem();
1339 1339 if (sd->is_top()) break;
1340 1340 }
1341 1341 xtty->tail("uncommon_trap");
1342 1342 }
1343 1343 }
1344 1344 // (End diagnostic printout.)
1345 1345
1346 1346 // Load class if necessary
1347 1347 if (unloaded_class_index >= 0) {
1348 1348 constantPoolHandle constants(THREAD, trap_method->constants());
1349 1349 load_class_by_index(constants, unloaded_class_index);
1350 1350 }
1351 1351
1352 1352 // Flush the nmethod if necessary and desirable.
1353 1353 //
1354 1354 // We need to avoid situations where we are re-flushing the nmethod
1355 1355 // because of a hot deoptimization site. Repeated flushes at the same
1356 1356 // point need to be detected by the compiler and avoided. If the compiler
1357 1357 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1358 1358 // module must take measures to avoid an infinite cycle of recompilation
1359 1359 // and deoptimization. There are several such measures:
1360 1360 //
1361 1361 // 1. If a recompilation is ordered a second time at some site X
1362 1362 // and for the same reason R, the action is adjusted to 'reinterpret',
1363 1363 // to give the interpreter time to exercise the method more thoroughly.
1364 1364 // If this happens, the method's overflow_recompile_count is incremented.
1365 1365 //
1366 1366 // 2. If the compiler fails to reduce the deoptimization rate, then
1367 1367 // the method's overflow_recompile_count will begin to exceed the set
1368 1368 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1369 1369 // is adjusted to 'make_not_compilable', and the method is abandoned
1370 1370 // to the interpreter. This is a performance hit for hot methods,
1371 1371 // but is better than a disastrous infinite cycle of recompilations.
1372 1372 // (Actually, only the method containing the site X is abandoned.)
1373 1373 //
1374 1374 // 3. In parallel with the previous measures, if the total number of
1375 1375 // recompilations of a method exceeds the much larger set limit
1376 1376 // PerMethodRecompilationCutoff, the method is abandoned.
1377 1377 // This should only happen if the method is very large and has
1378 1378 // many "lukewarm" deoptimizations. The code which enforces this
1379 1379 // limit is elsewhere (class nmethod, class methodOopDesc).
1380 1380 //
1381 1381 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1382 1382 // to recompile at each bytecode independently of the per-BCI cutoff.
1383 1383 //
1384 1384 // The decision to update code is up to the compiler, and is encoded
1385 1385 // in the Action_xxx code. If the compiler requests Action_none
1386 1386 // no trap state is changed, no compiled code is changed, and the
1387 1387 // computation suffers along in the interpreter.
1388 1388 //
1389 1389 // The other action codes specify various tactics for decompilation
1390 1390 // and recompilation. Action_maybe_recompile is the loosest, and
1391 1391 // allows the compiled code to stay around until enough traps are seen,
1392 1392 // and until the compiler gets around to recompiling the trapping method.
1393 1393 //
1394 1394 // The other actions cause immediate removal of the present code.
1395 1395
1396 1396 bool update_trap_state = true;
1397 1397 bool make_not_entrant = false;
1398 1398 bool make_not_compilable = false;
1399 1399 bool reprofile = false;
1400 1400 switch (action) {
1401 1401 case Action_none:
1402 1402 // Keep the old code.
1403 1403 update_trap_state = false;
1404 1404 break;
1405 1405 case Action_maybe_recompile:
1406 1406 // Do not need to invalidate the present code, but we can
1407 1407 // initiate another
1408 1408 // Start compiler without (necessarily) invalidating the nmethod.
1409 1409 // The system will tolerate the old code, but new code should be
1410 1410 // generated when possible.
1411 1411 break;
1412 1412 case Action_reinterpret:
1413 1413 // Go back into the interpreter for a while, and then consider
1414 1414 // recompiling form scratch.
1415 1415 make_not_entrant = true;
1416 1416 // Reset invocation counter for outer most method.
1417 1417 // This will allow the interpreter to exercise the bytecodes
1418 1418 // for a while before recompiling.
1419 1419 // By contrast, Action_make_not_entrant is immediate.
1420 1420 //
1421 1421 // Note that the compiler will track null_check, null_assert,
1422 1422 // range_check, and class_check events and log them as if they
1423 1423 // had been traps taken from compiled code. This will update
1424 1424 // the MDO trap history so that the next compilation will
1425 1425 // properly detect hot trap sites.
1426 1426 reprofile = true;
1427 1427 break;
1428 1428 case Action_make_not_entrant:
1429 1429 // Request immediate recompilation, and get rid of the old code.
1430 1430 // Make them not entrant, so next time they are called they get
1431 1431 // recompiled. Unloaded classes are loaded now so recompile before next
1432 1432 // time they are called. Same for uninitialized. The interpreter will
1433 1433 // link the missing class, if any.
1434 1434 make_not_entrant = true;
1435 1435 break;
1436 1436 case Action_make_not_compilable:
1437 1437 // Give up on compiling this method at all.
1438 1438 make_not_entrant = true;
1439 1439 make_not_compilable = true;
1440 1440 break;
1441 1441 default:
1442 1442 ShouldNotReachHere();
1443 1443 }
1444 1444
1445 1445 // Setting +ProfileTraps fixes the following, on all platforms:
1446 1446 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1447 1447 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1448 1448 // recompile relies on a methodDataOop to record heroic opt failures.
1449 1449
1450 1450 // Whether the interpreter is producing MDO data or not, we also need
1451 1451 // to use the MDO to detect hot deoptimization points and control
1452 1452 // aggressive optimization.
1453 1453 bool inc_recompile_count = false;
1454 1454 ProfileData* pdata = NULL;
1455 1455 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) {
1456 1456 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity");
1457 1457 uint this_trap_count = 0;
1458 1458 bool maybe_prior_trap = false;
1459 1459 bool maybe_prior_recompile = false;
1460 1460 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1461 1461 //outputs:
1462 1462 this_trap_count,
1463 1463 maybe_prior_trap,
1464 1464 maybe_prior_recompile);
1465 1465 // Because the interpreter also counts null, div0, range, and class
1466 1466 // checks, these traps from compiled code are double-counted.
1467 1467 // This is harmless; it just means that the PerXTrapLimit values
1468 1468 // are in effect a little smaller than they look.
1469 1469
1470 1470 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1471 1471 if (per_bc_reason != Reason_none) {
1472 1472 // Now take action based on the partially known per-BCI history.
1473 1473 if (maybe_prior_trap
1474 1474 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1475 1475 // If there are too many traps at this BCI, force a recompile.
1476 1476 // This will allow the compiler to see the limit overflow, and
1477 1477 // take corrective action, if possible. The compiler generally
1478 1478 // does not use the exact PerBytecodeTrapLimit value, but instead
1479 1479 // changes its tactics if it sees any traps at all. This provides
1480 1480 // a little hysteresis, delaying a recompile until a trap happens
↓ open down ↓ |
1480 lines elided |
↑ open up ↑ |
1481 1481 // several times.
1482 1482 //
1483 1483 // Actually, since there is only one bit of counter per BCI,
1484 1484 // the possible per-BCI counts are {0,1,(per-method count)}.
1485 1485 // This produces accurate results if in fact there is only
1486 1486 // one hot trap site, but begins to get fuzzy if there are
1487 1487 // many sites. For example, if there are ten sites each
1488 1488 // trapping two or more times, they each get the blame for
1489 1489 // all of their traps.
1490 1490 make_not_entrant = true;
1491 + if (per_bc_reason == Reason_unhandled) {
1492 + make_not_compilable = true;
1493 + }
1491 1494 }
1492 1495
1493 1496 // Detect repeated recompilation at the same BCI, and enforce a limit.
1494 1497 if (make_not_entrant && maybe_prior_recompile) {
1495 1498 // More than one recompile at this point.
1496 1499 inc_recompile_count = maybe_prior_trap;
1497 1500 }
1498 1501 } else {
1499 1502 // For reasons which are not recorded per-bytecode, we simply
1500 1503 // force recompiles unconditionally.
1501 1504 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1502 1505 make_not_entrant = true;
1503 1506 }
1504 1507
1505 1508 // Go back to the compiler if there are too many traps in this method.
1506 1509 if (this_trap_count >= (uint)PerMethodTrapLimit) {
1507 1510 // If there are too many traps in this method, force a recompile.
1508 1511 // This will allow the compiler to see the limit overflow, and
1509 1512 // take corrective action, if possible.
1510 1513 // (This condition is an unlikely backstop only, because the
1511 1514 // PerBytecodeTrapLimit is more likely to take effect first,
1512 1515 // if it is applicable.)
1513 1516 make_not_entrant = true;
1514 1517 }
1515 1518
1516 1519 // Here's more hysteresis: If there has been a recompile at
1517 1520 // this trap point already, run the method in the interpreter
1518 1521 // for a while to exercise it more thoroughly.
1519 1522 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1520 1523 reprofile = true;
1521 1524 }
1522 1525
1523 1526 }
1524 1527
1525 1528 // Take requested actions on the method:
1526 1529
1527 1530 // Recompile
1528 1531 if (make_not_entrant) {
1529 1532 if (!nm->make_not_entrant()) {
1530 1533 return; // the call did not change nmethod's state
1531 1534 }
1532 1535
1533 1536 if (pdata != NULL) {
1534 1537 // Record the recompilation event, if any.
1535 1538 int tstate0 = pdata->trap_state();
1536 1539 int tstate1 = trap_state_set_recompiled(tstate0, true);
1537 1540 if (tstate1 != tstate0)
1538 1541 pdata->set_trap_state(tstate1);
1539 1542 }
1540 1543 }
1541 1544
1542 1545 if (inc_recompile_count) {
1543 1546 trap_mdo->inc_overflow_recompile_count();
1544 1547 if ((uint)trap_mdo->overflow_recompile_count() >
1545 1548 (uint)PerBytecodeRecompilationCutoff) {
1546 1549 // Give up on the method containing the bad BCI.
1547 1550 if (trap_method() == nm->method()) {
1548 1551 make_not_compilable = true;
1549 1552 } else {
1550 1553 trap_method->set_not_compilable(CompLevel_full_optimization);
1551 1554 // But give grace to the enclosing nm->method().
1552 1555 }
1553 1556 }
1554 1557 }
1555 1558
1556 1559 // Reprofile
1557 1560 if (reprofile) {
1558 1561 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1559 1562 }
1560 1563
1561 1564 // Give up compiling
1562 1565 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1563 1566 assert(make_not_entrant, "consistent");
1564 1567 nm->method()->set_not_compilable(CompLevel_full_optimization);
1565 1568 }
1566 1569
1567 1570 } // Free marked resources
1568 1571
1569 1572 }
1570 1573 JRT_END
1571 1574
1572 1575 methodDataOop
1573 1576 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1574 1577 bool create_if_missing) {
1575 1578 Thread* THREAD = thread;
1576 1579 methodDataOop mdo = m()->method_data();
1577 1580 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1578 1581 // Build an MDO. Ignore errors like OutOfMemory;
1579 1582 // that simply means we won't have an MDO to update.
1580 1583 methodOopDesc::build_interpreter_method_data(m, THREAD);
1581 1584 if (HAS_PENDING_EXCEPTION) {
1582 1585 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1583 1586 CLEAR_PENDING_EXCEPTION;
1584 1587 }
1585 1588 mdo = m()->method_data();
1586 1589 }
1587 1590 return mdo;
1588 1591 }
1589 1592
1590 1593 ProfileData*
1591 1594 Deoptimization::query_update_method_data(methodDataHandle trap_mdo,
1592 1595 int trap_bci,
1593 1596 Deoptimization::DeoptReason reason,
1594 1597 //outputs:
1595 1598 uint& ret_this_trap_count,
1596 1599 bool& ret_maybe_prior_trap,
1597 1600 bool& ret_maybe_prior_recompile) {
1598 1601 uint prior_trap_count = trap_mdo->trap_count(reason);
1599 1602 uint this_trap_count = trap_mdo->inc_trap_count(reason);
↓ open down ↓ |
99 lines elided |
↑ open up ↑ |
1600 1603
1601 1604 // If the runtime cannot find a place to store trap history,
1602 1605 // it is estimated based on the general condition of the method.
1603 1606 // If the method has ever been recompiled, or has ever incurred
1604 1607 // a trap with the present reason , then this BCI is assumed
1605 1608 // (pessimistically) to be the culprit.
1606 1609 bool maybe_prior_trap = (prior_trap_count != 0);
1607 1610 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1608 1611 ProfileData* pdata = NULL;
1609 1612
1610 -
1611 1613 // For reasons which are recorded per bytecode, we check per-BCI data.
1612 1614 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1613 1615 if (per_bc_reason != Reason_none) {
1614 1616 // Find the profile data for this BCI. If there isn't one,
1615 1617 // try to allocate one from the MDO's set of spares.
1616 1618 // This will let us detect a repeated trap at this point.
1617 1619 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
1618 1620
1619 1621 if (pdata != NULL) {
1620 1622 // Query the trap state of this profile datum.
1621 1623 int tstate0 = pdata->trap_state();
1622 1624 if (!trap_state_has_reason(tstate0, per_bc_reason))
1623 1625 maybe_prior_trap = false;
1624 1626 if (!trap_state_is_recompiled(tstate0))
1625 1627 maybe_prior_recompile = false;
1626 1628
1627 1629 // Update the trap state of this profile datum.
1628 1630 int tstate1 = tstate0;
1629 1631 // Record the reason.
1630 1632 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1631 1633 // Store the updated state on the MDO, for next time.
1632 1634 if (tstate1 != tstate0)
1633 1635 pdata->set_trap_state(tstate1);
1634 1636 } else {
1635 1637 if (LogCompilation && xtty != NULL) {
1636 1638 ttyLocker ttyl;
1637 1639 // Missing MDP? Leave a small complaint in the log.
1638 1640 xtty->elem("missing_mdp bci='%d'", trap_bci);
1639 1641 }
1640 1642 }
1641 1643 }
1642 1644
1643 1645 // Return results:
1644 1646 ret_this_trap_count = this_trap_count;
1645 1647 ret_maybe_prior_trap = maybe_prior_trap;
1646 1648 ret_maybe_prior_recompile = maybe_prior_recompile;
1647 1649 return pdata;
1648 1650 }
1649 1651
1650 1652 void
1651 1653 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1652 1654 ResourceMark rm;
1653 1655 // Ignored outputs:
1654 1656 uint ignore_this_trap_count;
1655 1657 bool ignore_maybe_prior_trap;
1656 1658 bool ignore_maybe_prior_recompile;
1657 1659 query_update_method_data(trap_mdo, trap_bci,
1658 1660 (DeoptReason)reason,
1659 1661 ignore_this_trap_count,
1660 1662 ignore_maybe_prior_trap,
1661 1663 ignore_maybe_prior_recompile);
1662 1664 }
1663 1665
1664 1666 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1665 1667
1666 1668 // Still in Java no safepoints
1667 1669 {
1668 1670 // This enters VM and may safepoint
1669 1671 uncommon_trap_inner(thread, trap_request);
1670 1672 }
1671 1673 return fetch_unroll_info_helper(thread);
1672 1674 }
1673 1675
1674 1676 // Local derived constants.
1675 1677 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1676 1678 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1677 1679 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1678 1680
1679 1681 //---------------------------trap_state_reason---------------------------------
1680 1682 Deoptimization::DeoptReason
1681 1683 Deoptimization::trap_state_reason(int trap_state) {
1682 1684 // This assert provides the link between the width of DataLayout::trap_bits
1683 1685 // and the encoding of "recorded" reasons. It ensures there are enough
1684 1686 // bits to store all needed reasons in the per-BCI MDO profile.
1685 1687 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1686 1688 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1687 1689 trap_state -= recompile_bit;
1688 1690 if (trap_state == DS_REASON_MASK) {
1689 1691 return Reason_many;
1690 1692 } else {
1691 1693 assert((int)Reason_none == 0, "state=0 => Reason_none");
1692 1694 return (DeoptReason)trap_state;
1693 1695 }
1694 1696 }
1695 1697 //-------------------------trap_state_has_reason-------------------------------
1696 1698 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1697 1699 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1698 1700 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1699 1701 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1700 1702 trap_state -= recompile_bit;
1701 1703 if (trap_state == DS_REASON_MASK) {
1702 1704 return -1; // true, unspecifically (bottom of state lattice)
1703 1705 } else if (trap_state == reason) {
1704 1706 return 1; // true, definitely
1705 1707 } else if (trap_state == 0) {
1706 1708 return 0; // false, definitely (top of state lattice)
1707 1709 } else {
1708 1710 return 0; // false, definitely
1709 1711 }
1710 1712 }
1711 1713 //-------------------------trap_state_add_reason-------------------------------
1712 1714 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1713 1715 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1714 1716 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1715 1717 trap_state -= recompile_bit;
1716 1718 if (trap_state == DS_REASON_MASK) {
1717 1719 return trap_state + recompile_bit; // already at state lattice bottom
1718 1720 } else if (trap_state == reason) {
1719 1721 return trap_state + recompile_bit; // the condition is already true
1720 1722 } else if (trap_state == 0) {
1721 1723 return reason + recompile_bit; // no condition has yet been true
1722 1724 } else {
1723 1725 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1724 1726 }
1725 1727 }
1726 1728 //-----------------------trap_state_is_recompiled------------------------------
1727 1729 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1728 1730 return (trap_state & DS_RECOMPILE_BIT) != 0;
1729 1731 }
1730 1732 //-----------------------trap_state_set_recompiled-----------------------------
1731 1733 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1732 1734 if (z) return trap_state | DS_RECOMPILE_BIT;
1733 1735 else return trap_state & ~DS_RECOMPILE_BIT;
1734 1736 }
1735 1737 //---------------------------format_trap_state---------------------------------
1736 1738 // This is used for debugging and diagnostics, including hotspot.log output.
1737 1739 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1738 1740 int trap_state) {
1739 1741 DeoptReason reason = trap_state_reason(trap_state);
1740 1742 bool recomp_flag = trap_state_is_recompiled(trap_state);
1741 1743 // Re-encode the state from its decoded components.
1742 1744 int decoded_state = 0;
1743 1745 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1744 1746 decoded_state = trap_state_add_reason(decoded_state, reason);
1745 1747 if (recomp_flag)
1746 1748 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1747 1749 // If the state re-encodes properly, format it symbolically.
1748 1750 // Because this routine is used for debugging and diagnostics,
1749 1751 // be robust even if the state is a strange value.
1750 1752 size_t len;
1751 1753 if (decoded_state != trap_state) {
1752 1754 // Random buggy state that doesn't decode??
1753 1755 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1754 1756 } else {
1755 1757 len = jio_snprintf(buf, buflen, "%s%s",
1756 1758 trap_reason_name(reason),
1757 1759 recomp_flag ? " recompiled" : "");
1758 1760 }
1759 1761 if (len >= buflen)
1760 1762 buf[buflen-1] = '\0';
1761 1763 return buf;
1762 1764 }
1763 1765
1764 1766
1765 1767 //--------------------------------statics--------------------------------------
1766 1768 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1767 1769 = Deoptimization::Action_reinterpret;
1768 1770 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1769 1771 // Note: Keep this in sync. with enum DeoptReason.
1770 1772 "none",
1771 1773 "null_check",
1772 1774 "null_assert",
1773 1775 "range_check",
1774 1776 "class_check",
1775 1777 "array_check",
1776 1778 "intrinsic",
1777 1779 "bimorphic",
1778 1780 "unloaded",
1779 1781 "uninitialized",
1780 1782 "unreached",
1781 1783 "unhandled",
1782 1784 "constraint",
1783 1785 "div0_check",
1784 1786 "age",
1785 1787 "predicate",
1786 1788 "loop_limit_check"
1787 1789 };
1788 1790 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1789 1791 // Note: Keep this in sync. with enum DeoptAction.
1790 1792 "none",
1791 1793 "maybe_recompile",
1792 1794 "reinterpret",
1793 1795 "make_not_entrant",
1794 1796 "make_not_compilable"
1795 1797 };
1796 1798
1797 1799 const char* Deoptimization::trap_reason_name(int reason) {
1798 1800 if (reason == Reason_many) return "many";
1799 1801 if ((uint)reason < Reason_LIMIT)
1800 1802 return _trap_reason_name[reason];
1801 1803 static char buf[20];
1802 1804 sprintf(buf, "reason%d", reason);
1803 1805 return buf;
1804 1806 }
1805 1807 const char* Deoptimization::trap_action_name(int action) {
1806 1808 if ((uint)action < Action_LIMIT)
1807 1809 return _trap_action_name[action];
1808 1810 static char buf[20];
1809 1811 sprintf(buf, "action%d", action);
1810 1812 return buf;
1811 1813 }
1812 1814
1813 1815 // This is used for debugging and diagnostics, including hotspot.log output.
1814 1816 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1815 1817 int trap_request) {
1816 1818 jint unloaded_class_index = trap_request_index(trap_request);
1817 1819 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1818 1820 const char* action = trap_action_name(trap_request_action(trap_request));
1819 1821 size_t len;
1820 1822 if (unloaded_class_index < 0) {
1821 1823 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1822 1824 reason, action);
1823 1825 } else {
1824 1826 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1825 1827 reason, action, unloaded_class_index);
1826 1828 }
1827 1829 if (len >= buflen)
1828 1830 buf[buflen-1] = '\0';
1829 1831 return buf;
1830 1832 }
1831 1833
1832 1834 juint Deoptimization::_deoptimization_hist
1833 1835 [Deoptimization::Reason_LIMIT]
1834 1836 [1 + Deoptimization::Action_LIMIT]
1835 1837 [Deoptimization::BC_CASE_LIMIT]
1836 1838 = {0};
1837 1839
1838 1840 enum {
1839 1841 LSB_BITS = 8,
1840 1842 LSB_MASK = right_n_bits(LSB_BITS)
1841 1843 };
1842 1844
1843 1845 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1844 1846 Bytecodes::Code bc) {
1845 1847 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1846 1848 assert(action >= 0 && action < Action_LIMIT, "oob");
1847 1849 _deoptimization_hist[Reason_none][0][0] += 1; // total
1848 1850 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1849 1851 juint* cases = _deoptimization_hist[reason][1+action];
1850 1852 juint* bc_counter_addr = NULL;
1851 1853 juint bc_counter = 0;
1852 1854 // Look for an unused counter, or an exact match to this BC.
1853 1855 if (bc != Bytecodes::_illegal) {
1854 1856 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1855 1857 juint* counter_addr = &cases[bc_case];
1856 1858 juint counter = *counter_addr;
1857 1859 if ((counter == 0 && bc_counter_addr == NULL)
1858 1860 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1859 1861 // this counter is either free or is already devoted to this BC
1860 1862 bc_counter_addr = counter_addr;
1861 1863 bc_counter = counter | bc;
1862 1864 }
1863 1865 }
1864 1866 }
1865 1867 if (bc_counter_addr == NULL) {
1866 1868 // Overflow, or no given bytecode.
1867 1869 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1868 1870 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1869 1871 }
1870 1872 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1871 1873 }
1872 1874
1873 1875 jint Deoptimization::total_deoptimization_count() {
1874 1876 return _deoptimization_hist[Reason_none][0][0];
1875 1877 }
1876 1878
1877 1879 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1878 1880 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1879 1881 return _deoptimization_hist[reason][0][0];
1880 1882 }
1881 1883
1882 1884 void Deoptimization::print_statistics() {
1883 1885 juint total = total_deoptimization_count();
1884 1886 juint account = total;
1885 1887 if (total != 0) {
1886 1888 ttyLocker ttyl;
1887 1889 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1888 1890 tty->print_cr("Deoptimization traps recorded:");
1889 1891 #define PRINT_STAT_LINE(name, r) \
1890 1892 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1891 1893 PRINT_STAT_LINE("total", total);
1892 1894 // For each non-zero entry in the histogram, print the reason,
1893 1895 // the action, and (if specifically known) the type of bytecode.
1894 1896 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1895 1897 for (int action = 0; action < Action_LIMIT; action++) {
1896 1898 juint* cases = _deoptimization_hist[reason][1+action];
1897 1899 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1898 1900 juint counter = cases[bc_case];
1899 1901 if (counter != 0) {
1900 1902 char name[1*K];
1901 1903 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1902 1904 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1903 1905 bc = Bytecodes::_illegal;
1904 1906 sprintf(name, "%s/%s/%s",
1905 1907 trap_reason_name(reason),
1906 1908 trap_action_name(action),
1907 1909 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1908 1910 juint r = counter >> LSB_BITS;
1909 1911 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1910 1912 account -= r;
1911 1913 }
1912 1914 }
1913 1915 }
1914 1916 }
1915 1917 if (account != 0) {
1916 1918 PRINT_STAT_LINE("unaccounted", account);
1917 1919 }
1918 1920 #undef PRINT_STAT_LINE
1919 1921 if (xtty != NULL) xtty->tail("statistics");
1920 1922 }
1921 1923 }
1922 1924 #else // COMPILER2 || SHARK
1923 1925
1924 1926
1925 1927 // Stubs for C1 only system.
1926 1928 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1927 1929 return false;
1928 1930 }
1929 1931
1930 1932 const char* Deoptimization::trap_reason_name(int reason) {
1931 1933 return "unknown";
1932 1934 }
1933 1935
1934 1936 void Deoptimization::print_statistics() {
1935 1937 // no output
1936 1938 }
1937 1939
1938 1940 void
1939 1941 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) {
1940 1942 // no udpate
1941 1943 }
1942 1944
1943 1945 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1944 1946 return 0;
1945 1947 }
1946 1948
1947 1949 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1948 1950 Bytecodes::Code bc) {
1949 1951 // no update
1950 1952 }
1951 1953
1952 1954 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1953 1955 int trap_state) {
1954 1956 jio_snprintf(buf, buflen, "#%d", trap_state);
1955 1957 return buf;
1956 1958 }
1957 1959
1958 1960 #endif // COMPILER2 || SHARK
↓ open down ↓ |
338 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX