Thu, 20 Jun 2013 16:30:44 -0700
8016586: PPC64 (part 3): basic changes for PPC64
Summary: added #includes needed for ppc64 port. Renamed _MODEL_ppc to _MODEL_ppc_32 and renamed corresponding old _ppc files to _ppc_32.
Reviewed-by: dholmes, kvn
1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "code/debugInfoRec.hpp"
28 #include "code/nmethod.hpp"
29 #include "code/pcDesc.hpp"
30 #include "code/scopeDesc.hpp"
31 #include "interpreter/bytecode.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/oopMapCache.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "memory/oopFactory.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/method.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "prims/jvmtiThreadState.hpp"
40 #include "runtime/biasedLocking.hpp"
41 #include "runtime/compilationPolicy.hpp"
42 #include "runtime/deoptimization.hpp"
43 #include "runtime/interfaceSupport.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.hpp"
47 #include "runtime/thread.hpp"
48 #include "runtime/vframe.hpp"
49 #include "runtime/vframeArray.hpp"
50 #include "runtime/vframe_hp.hpp"
51 #include "utilities/events.hpp"
52 #include "utilities/xmlstream.hpp"
53 #ifdef TARGET_ARCH_x86
54 # include "vmreg_x86.inline.hpp"
55 #endif
56 #ifdef TARGET_ARCH_sparc
57 # include "vmreg_sparc.inline.hpp"
58 #endif
59 #ifdef TARGET_ARCH_zero
60 # include "vmreg_zero.inline.hpp"
61 #endif
62 #ifdef TARGET_ARCH_arm
63 # include "vmreg_arm.inline.hpp"
64 #endif
65 #ifdef TARGET_ARCH_ppc
66 # include "vmreg_ppc.inline.hpp"
67 #endif
68 #ifdef COMPILER2
69 #ifdef TARGET_ARCH_MODEL_x86_32
70 # include "adfiles/ad_x86_32.hpp"
71 #endif
72 #ifdef TARGET_ARCH_MODEL_x86_64
73 # include "adfiles/ad_x86_64.hpp"
74 #endif
75 #ifdef TARGET_ARCH_MODEL_sparc
76 # include "adfiles/ad_sparc.hpp"
77 #endif
78 #ifdef TARGET_ARCH_MODEL_zero
79 # include "adfiles/ad_zero.hpp"
80 #endif
81 #ifdef TARGET_ARCH_MODEL_arm
82 # include "adfiles/ad_arm.hpp"
83 #endif
84 #ifdef TARGET_ARCH_MODEL_ppc_32
85 # include "adfiles/ad_ppc_32.hpp"
86 #endif
87 #ifdef TARGET_ARCH_MODEL_ppc_64
88 # include "adfiles/ad_ppc_64.hpp"
89 #endif
90 #endif // COMPILER2
92 bool DeoptimizationMarker::_is_active = false;
94 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame,
95 int caller_adjustment,
96 int caller_actual_parameters,
97 int number_of_frames,
98 intptr_t* frame_sizes,
99 address* frame_pcs,
100 BasicType return_type) {
101 _size_of_deoptimized_frame = size_of_deoptimized_frame;
102 _caller_adjustment = caller_adjustment;
103 _caller_actual_parameters = caller_actual_parameters;
104 _number_of_frames = number_of_frames;
105 _frame_sizes = frame_sizes;
106 _frame_pcs = frame_pcs;
107 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
108 _return_type = return_type;
109 _initial_info = 0;
110 // PD (x86 only)
111 _counter_temp = 0;
112 _unpack_kind = 0;
113 _sender_sp_temp = 0;
115 _total_frame_sizes = size_of_frames();
116 }
119 Deoptimization::UnrollBlock::~UnrollBlock() {
120 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
121 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
122 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
123 }
126 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
127 assert(register_number < RegisterMap::reg_count, "checking register number");
128 return &_register_block[register_number * 2];
129 }
133 int Deoptimization::UnrollBlock::size_of_frames() const {
134 // Acount first for the adjustment of the initial frame
135 int result = _caller_adjustment;
136 for (int index = 0; index < number_of_frames(); index++) {
137 result += frame_sizes()[index];
138 }
139 return result;
140 }
143 void Deoptimization::UnrollBlock::print() {
144 ttyLocker ttyl;
145 tty->print_cr("UnrollBlock");
146 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
147 tty->print( " frame_sizes: ");
148 for (int index = 0; index < number_of_frames(); index++) {
149 tty->print("%d ", frame_sizes()[index]);
150 }
151 tty->cr();
152 }
155 // In order to make fetch_unroll_info work properly with escape
156 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
157 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
158 // of previously eliminated objects occurs in realloc_objects, which is
159 // called from the method fetch_unroll_info_helper below.
160 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
161 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
162 // but makes the entry a little slower. There is however a little dance we have to
163 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
165 // fetch_unroll_info() is called at the beginning of the deoptimization
166 // handler. Note this fact before we start generating temporary frames
167 // that can confuse an asynchronous stack walker. This counter is
168 // decremented at the end of unpack_frames().
169 thread->inc_in_deopt_handler();
171 return fetch_unroll_info_helper(thread);
172 JRT_END
175 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
176 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
178 // Note: there is a safepoint safety issue here. No matter whether we enter
179 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
180 // the vframeArray is created.
181 //
183 // Allocate our special deoptimization ResourceMark
184 DeoptResourceMark* dmark = new DeoptResourceMark(thread);
185 assert(thread->deopt_mark() == NULL, "Pending deopt!");
186 thread->set_deopt_mark(dmark);
188 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
189 RegisterMap map(thread, true);
190 RegisterMap dummy_map(thread, false);
191 // Now get the deoptee with a valid map
192 frame deoptee = stub_frame.sender(&map);
193 // Set the deoptee nmethod
194 assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
195 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
197 if (VerifyStack) {
198 thread->validate_frame_layout();
199 }
201 // Create a growable array of VFrames where each VFrame represents an inlined
202 // Java frame. This storage is allocated with the usual system arena.
203 assert(deoptee.is_compiled_frame(), "Wrong frame type");
204 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
205 vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
206 while (!vf->is_top()) {
207 assert(vf->is_compiled_frame(), "Wrong frame type");
208 chunk->push(compiledVFrame::cast(vf));
209 vf = vf->sender();
210 }
211 assert(vf->is_compiled_frame(), "Wrong frame type");
212 chunk->push(compiledVFrame::cast(vf));
214 #ifdef COMPILER2
215 // Reallocate the non-escaping objects and restore their fields. Then
216 // relock objects if synchronization on them was eliminated.
217 if (DoEscapeAnalysis || EliminateNestedLocks) {
218 if (EliminateAllocations) {
219 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
220 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
222 // The flag return_oop() indicates call sites which return oop
223 // in compiled code. Such sites include java method calls,
224 // runtime calls (for example, used to allocate new objects/arrays
225 // on slow code path) and any other calls generated in compiled code.
226 // It is not guaranteed that we can get such information here only
227 // by analyzing bytecode in deoptimized frames. This is why this flag
228 // is set during method compilation (see Compile::Process_OopMap_Node()).
229 bool save_oop_result = chunk->at(0)->scope()->return_oop();
230 Handle return_value;
231 if (save_oop_result) {
232 // Reallocation may trigger GC. If deoptimization happened on return from
233 // call which returns oop we need to save it since it is not in oopmap.
234 oop result = deoptee.saved_oop_result(&map);
235 assert(result == NULL || result->is_oop(), "must be oop");
236 return_value = Handle(thread, result);
237 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
238 if (TraceDeoptimization) {
239 ttyLocker ttyl;
240 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread);
241 }
242 }
243 bool reallocated = false;
244 if (objects != NULL) {
245 JRT_BLOCK
246 reallocated = realloc_objects(thread, &deoptee, objects, THREAD);
247 JRT_END
248 }
249 if (reallocated) {
250 reassign_fields(&deoptee, &map, objects);
251 #ifndef PRODUCT
252 if (TraceDeoptimization) {
253 ttyLocker ttyl;
254 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
255 print_objects(objects);
256 }
257 #endif
258 }
259 if (save_oop_result) {
260 // Restore result.
261 deoptee.set_saved_oop_result(&map, return_value());
262 }
263 }
264 if (EliminateLocks) {
265 #ifndef PRODUCT
266 bool first = true;
267 #endif
268 for (int i = 0; i < chunk->length(); i++) {
269 compiledVFrame* cvf = chunk->at(i);
270 assert (cvf->scope() != NULL,"expect only compiled java frames");
271 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
272 if (monitors->is_nonempty()) {
273 relock_objects(monitors, thread);
274 #ifndef PRODUCT
275 if (TraceDeoptimization) {
276 ttyLocker ttyl;
277 for (int j = 0; j < monitors->length(); j++) {
278 MonitorInfo* mi = monitors->at(j);
279 if (mi->eliminated()) {
280 if (first) {
281 first = false;
282 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
283 }
284 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner());
285 }
286 }
287 }
288 #endif
289 }
290 }
291 }
292 }
293 #endif // COMPILER2
294 // Ensure that no safepoint is taken after pointers have been stored
295 // in fields of rematerialized objects. If a safepoint occurs from here on
296 // out the java state residing in the vframeArray will be missed.
297 No_Safepoint_Verifier no_safepoint;
299 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk);
301 assert(thread->vframe_array_head() == NULL, "Pending deopt!");;
302 thread->set_vframe_array_head(array);
304 // Now that the vframeArray has been created if we have any deferred local writes
305 // added by jvmti then we can free up that structure as the data is now in the
306 // vframeArray
308 if (thread->deferred_locals() != NULL) {
309 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
310 int i = 0;
311 do {
312 // Because of inlining we could have multiple vframes for a single frame
313 // and several of the vframes could have deferred writes. Find them all.
314 if (list->at(i)->id() == array->original().id()) {
315 jvmtiDeferredLocalVariableSet* dlv = list->at(i);
316 list->remove_at(i);
317 // individual jvmtiDeferredLocalVariableSet are CHeapObj's
318 delete dlv;
319 } else {
320 i++;
321 }
322 } while ( i < list->length() );
323 if (list->length() == 0) {
324 thread->set_deferred_locals(NULL);
325 // free the list and elements back to C heap.
326 delete list;
327 }
329 }
331 #ifndef SHARK
332 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
333 CodeBlob* cb = stub_frame.cb();
334 // Verify we have the right vframeArray
335 assert(cb->frame_size() >= 0, "Unexpected frame size");
336 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
338 // If the deopt call site is a MethodHandle invoke call site we have
339 // to adjust the unpack_sp.
340 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
341 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
342 unpack_sp = deoptee.unextended_sp();
344 #ifdef ASSERT
345 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
346 #endif
347 #else
348 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
349 #endif // !SHARK
351 // This is a guarantee instead of an assert because if vframe doesn't match
352 // we will unpack the wrong deoptimized frame and wind up in strange places
353 // where it will be very difficult to figure out what went wrong. Better
354 // to die an early death here than some very obscure death later when the
355 // trail is cold.
356 // Note: on ia64 this guarantee can be fooled by frames with no memory stack
357 // in that it will fail to detect a problem when there is one. This needs
358 // more work in tiger timeframe.
359 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
361 int number_of_frames = array->frames();
363 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost
364 // virtual activation, which is the reverse of the elements in the vframes array.
365 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
366 // +1 because we always have an interpreter return address for the final slot.
367 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
368 int popframe_extra_args = 0;
369 // Create an interpreter return address for the stub to use as its return
370 // address so the skeletal frames are perfectly walkable
371 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
373 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
374 // activation be put back on the expression stack of the caller for reexecution
375 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
376 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
377 }
379 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
380 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
381 // than simply use array->sender.pc(). This requires us to walk the current set of frames
382 //
383 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
384 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller
386 // It's possible that the number of paramters at the call site is
387 // different than number of arguments in the callee when method
388 // handles are used. If the caller is interpreted get the real
389 // value so that the proper amount of space can be added to it's
390 // frame.
391 bool caller_was_method_handle = false;
392 if (deopt_sender.is_interpreted_frame()) {
393 methodHandle method = deopt_sender.interpreter_frame_method();
394 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
395 if (cur.is_invokedynamic() || cur.is_invokehandle()) {
396 // Method handle invokes may involve fairly arbitrary chains of
397 // calls so it's impossible to know how much actual space the
398 // caller has for locals.
399 caller_was_method_handle = true;
400 }
401 }
403 //
404 // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
405 // frame_sizes/frame_pcs[1] next oldest frame (int)
406 // frame_sizes/frame_pcs[n] youngest frame (int)
407 //
408 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
409 // owns the space for the return address to it's caller). Confusing ain't it.
410 //
411 // The vframe array can address vframes with indices running from
412 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame.
413 // When we create the skeletal frames we need the oldest frame to be in the zero slot
414 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
415 // so things look a little strange in this loop.
416 //
417 int callee_parameters = 0;
418 int callee_locals = 0;
419 for (int index = 0; index < array->frames(); index++ ) {
420 // frame[number_of_frames - 1 ] = on_stack_size(youngest)
421 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
422 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
423 int caller_parms = callee_parameters;
424 if ((index == array->frames() - 1) && caller_was_method_handle) {
425 caller_parms = 0;
426 }
427 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms,
428 callee_parameters,
429 callee_locals,
430 index == 0,
431 index == array->frames() - 1,
432 popframe_extra_args);
433 // This pc doesn't have to be perfect just good enough to identify the frame
434 // as interpreted so the skeleton frame will be walkable
435 // The correct pc will be set when the skeleton frame is completely filled out
436 // The final pc we store in the loop is wrong and will be overwritten below
437 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
439 callee_parameters = array->element(index)->method()->size_of_parameters();
440 callee_locals = array->element(index)->method()->max_locals();
441 popframe_extra_args = 0;
442 }
444 // Compute whether the root vframe returns a float or double value.
445 BasicType return_type;
446 {
447 HandleMark hm;
448 methodHandle method(thread, array->element(0)->method());
449 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
450 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
451 }
453 // Compute information for handling adapters and adjusting the frame size of the caller.
454 int caller_adjustment = 0;
456 // Compute the amount the oldest interpreter frame will have to adjust
457 // its caller's stack by. If the caller is a compiled frame then
458 // we pretend that the callee has no parameters so that the
459 // extension counts for the full amount of locals and not just
460 // locals-parms. This is because without a c2i adapter the parm
461 // area as created by the compiled frame will not be usable by
462 // the interpreter. (Depending on the calling convention there
463 // may not even be enough space).
465 // QQQ I'd rather see this pushed down into last_frame_adjust
466 // and have it take the sender (aka caller).
468 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
469 caller_adjustment = last_frame_adjust(0, callee_locals);
470 } else if (callee_locals > callee_parameters) {
471 // The caller frame may need extending to accommodate
472 // non-parameter locals of the first unpacked interpreted frame.
473 // Compute that adjustment.
474 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
475 }
477 // If the sender is deoptimized the we must retrieve the address of the handler
478 // since the frame will "magically" show the original pc before the deopt
479 // and we'd undo the deopt.
481 frame_pcs[0] = deopt_sender.raw_pc();
483 #ifndef SHARK
484 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
485 #endif // SHARK
487 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
488 caller_adjustment * BytesPerWord,
489 caller_was_method_handle ? 0 : callee_parameters,
490 number_of_frames,
491 frame_sizes,
492 frame_pcs,
493 return_type);
494 // On some platforms, we need a way to pass some platform dependent
495 // information to the unpacking code so the skeletal frames come out
496 // correct (initial fp value, unextended sp, ...)
497 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
499 if (array->frames() > 1) {
500 if (VerifyStack && TraceDeoptimization) {
501 ttyLocker ttyl;
502 tty->print_cr("Deoptimizing method containing inlining");
503 }
504 }
506 array->set_unroll_block(info);
507 return info;
508 }
510 // Called to cleanup deoptimization data structures in normal case
511 // after unpacking to stack and when stack overflow error occurs
512 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
513 vframeArray *array) {
515 // Get array if coming from exception
516 if (array == NULL) {
517 array = thread->vframe_array_head();
518 }
519 thread->set_vframe_array_head(NULL);
521 // Free the previous UnrollBlock
522 vframeArray* old_array = thread->vframe_array_last();
523 thread->set_vframe_array_last(array);
525 if (old_array != NULL) {
526 UnrollBlock* old_info = old_array->unroll_block();
527 old_array->set_unroll_block(NULL);
528 delete old_info;
529 delete old_array;
530 }
532 // Deallocate any resource creating in this routine and any ResourceObjs allocated
533 // inside the vframeArray (StackValueCollections)
535 delete thread->deopt_mark();
536 thread->set_deopt_mark(NULL);
537 thread->set_deopt_nmethod(NULL);
540 if (JvmtiExport::can_pop_frame()) {
541 #ifndef CC_INTERP
542 // Regardless of whether we entered this routine with the pending
543 // popframe condition bit set, we should always clear it now
544 thread->clear_popframe_condition();
545 #else
546 // C++ interpeter will clear has_pending_popframe when it enters
547 // with method_resume. For deopt_resume2 we clear it now.
548 if (thread->popframe_forcing_deopt_reexecution())
549 thread->clear_popframe_condition();
550 #endif /* CC_INTERP */
551 }
553 // unpack_frames() is called at the end of the deoptimization handler
554 // and (in C2) at the end of the uncommon trap handler. Note this fact
555 // so that an asynchronous stack walker can work again. This counter is
556 // incremented at the beginning of fetch_unroll_info() and (in C2) at
557 // the beginning of uncommon_trap().
558 thread->dec_in_deopt_handler();
559 }
562 // Return BasicType of value being returned
563 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
565 // We are already active int he special DeoptResourceMark any ResourceObj's we
566 // allocate will be freed at the end of the routine.
568 // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
569 // but makes the entry a little slower. There is however a little dance we have to
570 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
571 ResetNoHandleMark rnhm; // No-op in release/product versions
572 HandleMark hm;
574 frame stub_frame = thread->last_frame();
576 // Since the frame to unpack is the top frame of this thread, the vframe_array_head
577 // must point to the vframeArray for the unpack frame.
578 vframeArray* array = thread->vframe_array_head();
580 #ifndef PRODUCT
581 if (TraceDeoptimization) {
582 ttyLocker ttyl;
583 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
584 }
585 #endif
586 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
587 stub_frame.pc(), stub_frame.sp(), exec_mode);
589 UnrollBlock* info = array->unroll_block();
591 // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
592 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
594 BasicType bt = info->return_type();
596 // If we have an exception pending, claim that the return type is an oop
597 // so the deopt_blob does not overwrite the exception_oop.
599 if (exec_mode == Unpack_exception)
600 bt = T_OBJECT;
602 // Cleanup thread deopt data
603 cleanup_deopt_info(thread, array);
605 #ifndef PRODUCT
606 if (VerifyStack) {
607 ResourceMark res_mark;
609 thread->validate_frame_layout();
611 // Verify that the just-unpacked frames match the interpreter's
612 // notions of expression stack and locals
613 vframeArray* cur_array = thread->vframe_array_last();
614 RegisterMap rm(thread, false);
615 rm.set_include_argument_oops(false);
616 bool is_top_frame = true;
617 int callee_size_of_parameters = 0;
618 int callee_max_locals = 0;
619 for (int i = 0; i < cur_array->frames(); i++) {
620 vframeArrayElement* el = cur_array->element(i);
621 frame* iframe = el->iframe();
622 guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
624 // Get the oop map for this bci
625 InterpreterOopMap mask;
626 int cur_invoke_parameter_size = 0;
627 bool try_next_mask = false;
628 int next_mask_expression_stack_size = -1;
629 int top_frame_expression_stack_adjustment = 0;
630 methodHandle mh(thread, iframe->interpreter_frame_method());
631 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
632 BytecodeStream str(mh);
633 str.set_start(iframe->interpreter_frame_bci());
634 int max_bci = mh->code_size();
635 // Get to the next bytecode if possible
636 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
637 // Check to see if we can grab the number of outgoing arguments
638 // at an uncommon trap for an invoke (where the compiler
639 // generates debug info before the invoke has executed)
640 Bytecodes::Code cur_code = str.next();
641 if (cur_code == Bytecodes::_invokevirtual ||
642 cur_code == Bytecodes::_invokespecial ||
643 cur_code == Bytecodes::_invokestatic ||
644 cur_code == Bytecodes::_invokeinterface) {
645 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
646 Symbol* signature = invoke.signature();
647 ArgumentSizeComputer asc(signature);
648 cur_invoke_parameter_size = asc.size();
649 if (cur_code != Bytecodes::_invokestatic) {
650 // Add in receiver
651 ++cur_invoke_parameter_size;
652 }
653 }
654 if (str.bci() < max_bci) {
655 Bytecodes::Code bc = str.next();
656 if (bc >= 0) {
657 // The interpreter oop map generator reports results before
658 // the current bytecode has executed except in the case of
659 // calls. It seems to be hard to tell whether the compiler
660 // has emitted debug information matching the "state before"
661 // a given bytecode or the state after, so we try both
662 switch (cur_code) {
663 case Bytecodes::_invokevirtual:
664 case Bytecodes::_invokespecial:
665 case Bytecodes::_invokestatic:
666 case Bytecodes::_invokeinterface:
667 case Bytecodes::_athrow:
668 break;
669 default: {
670 InterpreterOopMap next_mask;
671 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
672 next_mask_expression_stack_size = next_mask.expression_stack_size();
673 // Need to subtract off the size of the result type of
674 // the bytecode because this is not described in the
675 // debug info but returned to the interpreter in the TOS
676 // caching register
677 BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
678 if (bytecode_result_type != T_ILLEGAL) {
679 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
680 }
681 assert(top_frame_expression_stack_adjustment >= 0, "");
682 try_next_mask = true;
683 break;
684 }
685 }
686 }
687 }
689 // Verify stack depth and oops in frame
690 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
691 if (!(
692 /* SPARC */
693 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
694 /* x86 */
695 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
696 (try_next_mask &&
697 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
698 top_frame_expression_stack_adjustment))) ||
699 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
700 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) &&
701 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
702 )) {
703 ttyLocker ttyl;
705 // Print out some information that will help us debug the problem
706 tty->print_cr("Wrong number of expression stack elements during deoptimization");
707 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
708 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements",
709 iframe->interpreter_frame_expression_stack_size());
710 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
711 tty->print_cr(" try_next_mask = %d", try_next_mask);
712 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
713 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters);
714 tty->print_cr(" callee_max_locals = %d", callee_max_locals);
715 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
716 tty->print_cr(" exec_mode = %d", exec_mode);
717 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
718 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
719 tty->print_cr(" Interpreted frames:");
720 for (int k = 0; k < cur_array->frames(); k++) {
721 vframeArrayElement* el = cur_array->element(k);
722 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
723 }
724 cur_array->print_on_2(tty);
725 guarantee(false, "wrong number of expression stack elements during deopt");
726 }
727 VerifyOopClosure verify;
728 iframe->oops_interpreted_do(&verify, NULL, &rm, false);
729 callee_size_of_parameters = mh->size_of_parameters();
730 callee_max_locals = mh->max_locals();
731 is_top_frame = false;
732 }
733 }
734 #endif /* !PRODUCT */
737 return bt;
738 JRT_END
741 int Deoptimization::deoptimize_dependents() {
742 Threads::deoptimized_wrt_marked_nmethods();
743 return 0;
744 }
747 #ifdef COMPILER2
748 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
749 Handle pending_exception(thread->pending_exception());
750 const char* exception_file = thread->exception_file();
751 int exception_line = thread->exception_line();
752 thread->clear_pending_exception();
754 for (int i = 0; i < objects->length(); i++) {
755 assert(objects->at(i)->is_object(), "invalid debug information");
756 ObjectValue* sv = (ObjectValue*) objects->at(i);
758 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
759 oop obj = NULL;
761 if (k->oop_is_instance()) {
762 InstanceKlass* ik = InstanceKlass::cast(k());
763 obj = ik->allocate_instance(CHECK_(false));
764 } else if (k->oop_is_typeArray()) {
765 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
766 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
767 int len = sv->field_size() / type2size[ak->element_type()];
768 obj = ak->allocate(len, CHECK_(false));
769 } else if (k->oop_is_objArray()) {
770 ObjArrayKlass* ak = ObjArrayKlass::cast(k());
771 obj = ak->allocate(sv->field_size(), CHECK_(false));
772 }
774 assert(obj != NULL, "allocation failed");
775 assert(sv->value().is_null(), "redundant reallocation");
776 sv->set_value(obj);
777 }
779 if (pending_exception.not_null()) {
780 thread->set_pending_exception(pending_exception(), exception_file, exception_line);
781 }
783 return true;
784 }
786 // This assumes that the fields are stored in ObjectValue in the same order
787 // they are yielded by do_nonstatic_fields.
788 class FieldReassigner: public FieldClosure {
789 frame* _fr;
790 RegisterMap* _reg_map;
791 ObjectValue* _sv;
792 InstanceKlass* _ik;
793 oop _obj;
795 int _i;
796 public:
797 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
798 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
800 int i() const { return _i; }
803 void do_field(fieldDescriptor* fd) {
804 intptr_t val;
805 StackValue* value =
806 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
807 int offset = fd->offset();
808 switch (fd->field_type()) {
809 case T_OBJECT: case T_ARRAY:
810 assert(value->type() == T_OBJECT, "Agreement.");
811 _obj->obj_field_put(offset, value->get_obj()());
812 break;
814 case T_LONG: case T_DOUBLE: {
815 assert(value->type() == T_INT, "Agreement.");
816 StackValue* low =
817 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
818 #ifdef _LP64
819 jlong res = (jlong)low->get_int();
820 #else
821 #ifdef SPARC
822 // For SPARC we have to swap high and low words.
823 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
824 #else
825 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
826 #endif //SPARC
827 #endif
828 _obj->long_field_put(offset, res);
829 break;
830 }
831 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
832 case T_INT: case T_FLOAT: // 4 bytes.
833 assert(value->type() == T_INT, "Agreement.");
834 val = value->get_int();
835 _obj->int_field_put(offset, (jint)*((jint*)&val));
836 break;
838 case T_SHORT: case T_CHAR: // 2 bytes
839 assert(value->type() == T_INT, "Agreement.");
840 val = value->get_int();
841 _obj->short_field_put(offset, (jshort)*((jint*)&val));
842 break;
844 case T_BOOLEAN: case T_BYTE: // 1 byte
845 assert(value->type() == T_INT, "Agreement.");
846 val = value->get_int();
847 _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
848 break;
850 default:
851 ShouldNotReachHere();
852 }
853 _i++;
854 }
855 };
857 // restore elements of an eliminated type array
858 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
859 int index = 0;
860 intptr_t val;
862 for (int i = 0; i < sv->field_size(); i++) {
863 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
864 switch(type) {
865 case T_LONG: case T_DOUBLE: {
866 assert(value->type() == T_INT, "Agreement.");
867 StackValue* low =
868 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
869 #ifdef _LP64
870 jlong res = (jlong)low->get_int();
871 #else
872 #ifdef SPARC
873 // For SPARC we have to swap high and low words.
874 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
875 #else
876 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
877 #endif //SPARC
878 #endif
879 obj->long_at_put(index, res);
880 break;
881 }
883 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
884 case T_INT: case T_FLOAT: // 4 bytes.
885 assert(value->type() == T_INT, "Agreement.");
886 val = value->get_int();
887 obj->int_at_put(index, (jint)*((jint*)&val));
888 break;
890 case T_SHORT: case T_CHAR: // 2 bytes
891 assert(value->type() == T_INT, "Agreement.");
892 val = value->get_int();
893 obj->short_at_put(index, (jshort)*((jint*)&val));
894 break;
896 case T_BOOLEAN: case T_BYTE: // 1 byte
897 assert(value->type() == T_INT, "Agreement.");
898 val = value->get_int();
899 obj->bool_at_put(index, (jboolean)*((jint*)&val));
900 break;
902 default:
903 ShouldNotReachHere();
904 }
905 index++;
906 }
907 }
910 // restore fields of an eliminated object array
911 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
912 for (int i = 0; i < sv->field_size(); i++) {
913 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
914 assert(value->type() == T_OBJECT, "object element expected");
915 obj->obj_at_put(i, value->get_obj()());
916 }
917 }
920 // restore fields of all eliminated objects and arrays
921 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) {
922 for (int i = 0; i < objects->length(); i++) {
923 ObjectValue* sv = (ObjectValue*) objects->at(i);
924 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
925 Handle obj = sv->value();
926 assert(obj.not_null(), "reallocation was missed");
928 if (k->oop_is_instance()) {
929 InstanceKlass* ik = InstanceKlass::cast(k());
930 FieldReassigner reassign(fr, reg_map, sv, obj());
931 ik->do_nonstatic_fields(&reassign);
932 } else if (k->oop_is_typeArray()) {
933 TypeArrayKlass* ak = TypeArrayKlass::cast(k());
934 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
935 } else if (k->oop_is_objArray()) {
936 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
937 }
938 }
939 }
942 // relock objects for which synchronization was eliminated
943 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) {
944 for (int i = 0; i < monitors->length(); i++) {
945 MonitorInfo* mon_info = monitors->at(i);
946 if (mon_info->eliminated()) {
947 assert(mon_info->owner() != NULL, "reallocation was missed");
948 Handle obj = Handle(mon_info->owner());
949 markOop mark = obj->mark();
950 if (UseBiasedLocking && mark->has_bias_pattern()) {
951 // New allocated objects may have the mark set to anonymously biased.
952 // Also the deoptimized method may called methods with synchronization
953 // where the thread-local object is bias locked to the current thread.
954 assert(mark->is_biased_anonymously() ||
955 mark->biased_locker() == thread, "should be locked to current thread");
956 // Reset mark word to unbiased prototype.
957 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
958 obj->set_mark(unbiased_prototype);
959 }
960 BasicLock* lock = mon_info->lock();
961 ObjectSynchronizer::slow_enter(obj, lock, thread);
962 }
963 assert(mon_info->owner()->is_locked(), "object must be locked now");
964 }
965 }
968 #ifndef PRODUCT
969 // print information about reallocated objects
970 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) {
971 fieldDescriptor fd;
973 for (int i = 0; i < objects->length(); i++) {
974 ObjectValue* sv = (ObjectValue*) objects->at(i);
975 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
976 Handle obj = sv->value();
978 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()());
979 k->print_value();
980 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
981 tty->cr();
983 if (Verbose) {
984 k->oop_print_on(obj(), tty);
985 }
986 }
987 }
988 #endif
989 #endif // COMPILER2
991 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) {
992 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
994 #ifndef PRODUCT
995 if (TraceDeoptimization) {
996 ttyLocker ttyl;
997 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
998 fr.print_on(tty);
999 tty->print_cr(" Virtual frames (innermost first):");
1000 for (int index = 0; index < chunk->length(); index++) {
1001 compiledVFrame* vf = chunk->at(index);
1002 tty->print(" %2d - ", index);
1003 vf->print_value();
1004 int bci = chunk->at(index)->raw_bci();
1005 const char* code_name;
1006 if (bci == SynchronizationEntryBCI) {
1007 code_name = "sync entry";
1008 } else {
1009 Bytecodes::Code code = vf->method()->code_at(bci);
1010 code_name = Bytecodes::name(code);
1011 }
1012 tty->print(" - %s", code_name);
1013 tty->print_cr(" @ bci %d ", bci);
1014 if (Verbose) {
1015 vf->print();
1016 tty->cr();
1017 }
1018 }
1019 }
1020 #endif
1022 // Register map for next frame (used for stack crawl). We capture
1023 // the state of the deopt'ing frame's caller. Thus if we need to
1024 // stuff a C2I adapter we can properly fill in the callee-save
1025 // register locations.
1026 frame caller = fr.sender(reg_map);
1027 int frame_size = caller.sp() - fr.sp();
1029 frame sender = caller;
1031 // Since the Java thread being deoptimized will eventually adjust it's own stack,
1032 // the vframeArray containing the unpacking information is allocated in the C heap.
1033 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1034 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr);
1036 // Compare the vframeArray to the collected vframes
1037 assert(array->structural_compare(thread, chunk), "just checking");
1039 #ifndef PRODUCT
1040 if (TraceDeoptimization) {
1041 ttyLocker ttyl;
1042 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array);
1043 }
1044 #endif // PRODUCT
1046 return array;
1047 }
1050 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1051 GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1052 for (int i = 0; i < monitors->length(); i++) {
1053 MonitorInfo* mon_info = monitors->at(i);
1054 if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1055 objects_to_revoke->append(Handle(mon_info->owner()));
1056 }
1057 }
1058 }
1061 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1062 if (!UseBiasedLocking) {
1063 return;
1064 }
1066 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1068 // Unfortunately we don't have a RegisterMap available in most of
1069 // the places we want to call this routine so we need to walk the
1070 // stack again to update the register map.
1071 if (map == NULL || !map->update_map()) {
1072 StackFrameStream sfs(thread, true);
1073 bool found = false;
1074 while (!found && !sfs.is_done()) {
1075 frame* cur = sfs.current();
1076 sfs.next();
1077 found = cur->id() == fr.id();
1078 }
1079 assert(found, "frame to be deoptimized not found on target thread's stack");
1080 map = sfs.register_map();
1081 }
1083 vframe* vf = vframe::new_vframe(&fr, map, thread);
1084 compiledVFrame* cvf = compiledVFrame::cast(vf);
1085 // Revoke monitors' biases in all scopes
1086 while (!cvf->is_top()) {
1087 collect_monitors(cvf, objects_to_revoke);
1088 cvf = compiledVFrame::cast(cvf->sender());
1089 }
1090 collect_monitors(cvf, objects_to_revoke);
1092 if (SafepointSynchronize::is_at_safepoint()) {
1093 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1094 } else {
1095 BiasedLocking::revoke(objects_to_revoke);
1096 }
1097 }
1100 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1101 if (!UseBiasedLocking) {
1102 return;
1103 }
1105 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1106 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1107 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1108 if (jt->has_last_Java_frame()) {
1109 StackFrameStream sfs(jt, true);
1110 while (!sfs.is_done()) {
1111 frame* cur = sfs.current();
1112 if (cb->contains(cur->pc())) {
1113 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1114 compiledVFrame* cvf = compiledVFrame::cast(vf);
1115 // Revoke monitors' biases in all scopes
1116 while (!cvf->is_top()) {
1117 collect_monitors(cvf, objects_to_revoke);
1118 cvf = compiledVFrame::cast(cvf->sender());
1119 }
1120 collect_monitors(cvf, objects_to_revoke);
1121 }
1122 sfs.next();
1123 }
1124 }
1125 }
1126 BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1127 }
1130 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1131 assert(fr.can_be_deoptimized(), "checking frame type");
1133 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1135 // Patch the nmethod so that when execution returns to it we will
1136 // deopt the execution state and return to the interpreter.
1137 fr.deoptimize(thread);
1138 }
1140 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1141 // Deoptimize only if the frame comes from compile code.
1142 // Do not deoptimize the frame which is already patched
1143 // during the execution of the loops below.
1144 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1145 return;
1146 }
1147 ResourceMark rm;
1148 DeoptimizationMarker dm;
1149 if (UseBiasedLocking) {
1150 revoke_biases_of_monitors(thread, fr, map);
1151 }
1152 deoptimize_single_frame(thread, fr);
1154 }
1157 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1158 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1159 "can only deoptimize other thread at a safepoint");
1160 // Compute frame and register map based on thread and sp.
1161 RegisterMap reg_map(thread, UseBiasedLocking);
1162 frame fr = thread->last_frame();
1163 while (fr.id() != id) {
1164 fr = fr.sender(®_map);
1165 }
1166 deoptimize(thread, fr, ®_map);
1167 }
1170 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1171 if (thread == Thread::current()) {
1172 Deoptimization::deoptimize_frame_internal(thread, id);
1173 } else {
1174 VM_DeoptimizeFrame deopt(thread, id);
1175 VMThread::execute(&deopt);
1176 }
1177 }
1180 // JVMTI PopFrame support
1181 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1182 {
1183 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1184 }
1185 JRT_END
1188 #if defined(COMPILER2) || defined(SHARK)
1189 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1190 // in case of an unresolved klass entry, load the class.
1191 if (constant_pool->tag_at(index).is_unresolved_klass()) {
1192 Klass* tk = constant_pool->klass_at(index, CHECK);
1193 return;
1194 }
1196 if (!constant_pool->tag_at(index).is_symbol()) return;
1198 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1199 Symbol* symbol = constant_pool->symbol_at(index);
1201 // class name?
1202 if (symbol->byte_at(0) != '(') {
1203 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1204 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1205 return;
1206 }
1208 // then it must be a signature!
1209 ResourceMark rm(THREAD);
1210 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1211 if (ss.is_object()) {
1212 Symbol* class_name = ss.as_symbol(CHECK);
1213 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1214 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1215 }
1216 }
1217 }
1220 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1221 EXCEPTION_MARK;
1222 load_class_by_index(constant_pool, index, THREAD);
1223 if (HAS_PENDING_EXCEPTION) {
1224 // Exception happened during classloading. We ignore the exception here, since it
1225 // is going to be rethrown since the current activation is going to be deoptimzied and
1226 // the interpreter will re-execute the bytecode.
1227 CLEAR_PENDING_EXCEPTION;
1228 }
1229 }
1231 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1232 HandleMark hm;
1234 // uncommon_trap() is called at the beginning of the uncommon trap
1235 // handler. Note this fact before we start generating temporary frames
1236 // that can confuse an asynchronous stack walker. This counter is
1237 // decremented at the end of unpack_frames().
1238 thread->inc_in_deopt_handler();
1240 // We need to update the map if we have biased locking.
1241 RegisterMap reg_map(thread, UseBiasedLocking);
1242 frame stub_frame = thread->last_frame();
1243 frame fr = stub_frame.sender(®_map);
1244 // Make sure the calling nmethod is not getting deoptimized and removed
1245 // before we are done with it.
1246 nmethodLocker nl(fr.pc());
1248 // Log a message
1249 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1250 trap_request, fr.pc());
1252 {
1253 ResourceMark rm;
1255 // Revoke biases of any monitors in the frame to ensure we can migrate them
1256 revoke_biases_of_monitors(thread, fr, ®_map);
1258 DeoptReason reason = trap_request_reason(trap_request);
1259 DeoptAction action = trap_request_action(trap_request);
1260 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1262 vframe* vf = vframe::new_vframe(&fr, ®_map, thread);
1263 compiledVFrame* cvf = compiledVFrame::cast(vf);
1265 nmethod* nm = cvf->code();
1267 ScopeDesc* trap_scope = cvf->scope();
1268 methodHandle trap_method = trap_scope->method();
1269 int trap_bci = trap_scope->bci();
1270 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci);
1272 // Record this event in the histogram.
1273 gather_statistics(reason, action, trap_bc);
1275 // Ensure that we can record deopt. history:
1276 bool create_if_missing = ProfileTraps;
1278 MethodData* trap_mdo =
1279 get_method_data(thread, trap_method, create_if_missing);
1281 // Log a message
1282 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1283 trap_reason_name(reason), trap_action_name(action), fr.pc(),
1284 trap_method->name_and_sig_as_C_string(), trap_bci);
1286 // Print a bunch of diagnostics, if requested.
1287 if (TraceDeoptimization || LogCompilation) {
1288 ResourceMark rm;
1289 ttyLocker ttyl;
1290 char buf[100];
1291 if (xtty != NULL) {
1292 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s",
1293 os::current_thread_id(),
1294 format_trap_request(buf, sizeof(buf), trap_request));
1295 nm->log_identity(xtty);
1296 }
1297 Symbol* class_name = NULL;
1298 bool unresolved = false;
1299 if (unloaded_class_index >= 0) {
1300 constantPoolHandle constants (THREAD, trap_method->constants());
1301 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1302 class_name = constants->klass_name_at(unloaded_class_index);
1303 unresolved = true;
1304 if (xtty != NULL)
1305 xtty->print(" unresolved='1'");
1306 } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1307 class_name = constants->symbol_at(unloaded_class_index);
1308 }
1309 if (xtty != NULL)
1310 xtty->name(class_name);
1311 }
1312 if (xtty != NULL && trap_mdo != NULL) {
1313 // Dump the relevant MDO state.
1314 // This is the deopt count for the current reason, any previous
1315 // reasons or recompiles seen at this point.
1316 int dcnt = trap_mdo->trap_count(reason);
1317 if (dcnt != 0)
1318 xtty->print(" count='%d'", dcnt);
1319 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1320 int dos = (pdata == NULL)? 0: pdata->trap_state();
1321 if (dos != 0) {
1322 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1323 if (trap_state_is_recompiled(dos)) {
1324 int recnt2 = trap_mdo->overflow_recompile_count();
1325 if (recnt2 != 0)
1326 xtty->print(" recompiles2='%d'", recnt2);
1327 }
1328 }
1329 }
1330 if (xtty != NULL) {
1331 xtty->stamp();
1332 xtty->end_head();
1333 }
1334 if (TraceDeoptimization) { // make noise on the tty
1335 tty->print("Uncommon trap occurred in");
1336 nm->method()->print_short_name(tty);
1337 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1338 fr.pc(),
1339 os::current_thread_id(),
1340 trap_reason_name(reason),
1341 trap_action_name(action),
1342 unloaded_class_index);
1343 if (class_name != NULL) {
1344 tty->print(unresolved ? " unresolved class: " : " symbol: ");
1345 class_name->print_symbol_on(tty);
1346 }
1347 tty->cr();
1348 }
1349 if (xtty != NULL) {
1350 // Log the precise location of the trap.
1351 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1352 xtty->begin_elem("jvms bci='%d'", sd->bci());
1353 xtty->method(sd->method());
1354 xtty->end_elem();
1355 if (sd->is_top()) break;
1356 }
1357 xtty->tail("uncommon_trap");
1358 }
1359 }
1360 // (End diagnostic printout.)
1362 // Load class if necessary
1363 if (unloaded_class_index >= 0) {
1364 constantPoolHandle constants(THREAD, trap_method->constants());
1365 load_class_by_index(constants, unloaded_class_index);
1366 }
1368 // Flush the nmethod if necessary and desirable.
1369 //
1370 // We need to avoid situations where we are re-flushing the nmethod
1371 // because of a hot deoptimization site. Repeated flushes at the same
1372 // point need to be detected by the compiler and avoided. If the compiler
1373 // cannot avoid them (or has a bug and "refuses" to avoid them), this
1374 // module must take measures to avoid an infinite cycle of recompilation
1375 // and deoptimization. There are several such measures:
1376 //
1377 // 1. If a recompilation is ordered a second time at some site X
1378 // and for the same reason R, the action is adjusted to 'reinterpret',
1379 // to give the interpreter time to exercise the method more thoroughly.
1380 // If this happens, the method's overflow_recompile_count is incremented.
1381 //
1382 // 2. If the compiler fails to reduce the deoptimization rate, then
1383 // the method's overflow_recompile_count will begin to exceed the set
1384 // limit PerBytecodeRecompilationCutoff. If this happens, the action
1385 // is adjusted to 'make_not_compilable', and the method is abandoned
1386 // to the interpreter. This is a performance hit for hot methods,
1387 // but is better than a disastrous infinite cycle of recompilations.
1388 // (Actually, only the method containing the site X is abandoned.)
1389 //
1390 // 3. In parallel with the previous measures, if the total number of
1391 // recompilations of a method exceeds the much larger set limit
1392 // PerMethodRecompilationCutoff, the method is abandoned.
1393 // This should only happen if the method is very large and has
1394 // many "lukewarm" deoptimizations. The code which enforces this
1395 // limit is elsewhere (class nmethod, class Method).
1396 //
1397 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1398 // to recompile at each bytecode independently of the per-BCI cutoff.
1399 //
1400 // The decision to update code is up to the compiler, and is encoded
1401 // in the Action_xxx code. If the compiler requests Action_none
1402 // no trap state is changed, no compiled code is changed, and the
1403 // computation suffers along in the interpreter.
1404 //
1405 // The other action codes specify various tactics for decompilation
1406 // and recompilation. Action_maybe_recompile is the loosest, and
1407 // allows the compiled code to stay around until enough traps are seen,
1408 // and until the compiler gets around to recompiling the trapping method.
1409 //
1410 // The other actions cause immediate removal of the present code.
1412 bool update_trap_state = true;
1413 bool make_not_entrant = false;
1414 bool make_not_compilable = false;
1415 bool reprofile = false;
1416 switch (action) {
1417 case Action_none:
1418 // Keep the old code.
1419 update_trap_state = false;
1420 break;
1421 case Action_maybe_recompile:
1422 // Do not need to invalidate the present code, but we can
1423 // initiate another
1424 // Start compiler without (necessarily) invalidating the nmethod.
1425 // The system will tolerate the old code, but new code should be
1426 // generated when possible.
1427 break;
1428 case Action_reinterpret:
1429 // Go back into the interpreter for a while, and then consider
1430 // recompiling form scratch.
1431 make_not_entrant = true;
1432 // Reset invocation counter for outer most method.
1433 // This will allow the interpreter to exercise the bytecodes
1434 // for a while before recompiling.
1435 // By contrast, Action_make_not_entrant is immediate.
1436 //
1437 // Note that the compiler will track null_check, null_assert,
1438 // range_check, and class_check events and log them as if they
1439 // had been traps taken from compiled code. This will update
1440 // the MDO trap history so that the next compilation will
1441 // properly detect hot trap sites.
1442 reprofile = true;
1443 break;
1444 case Action_make_not_entrant:
1445 // Request immediate recompilation, and get rid of the old code.
1446 // Make them not entrant, so next time they are called they get
1447 // recompiled. Unloaded classes are loaded now so recompile before next
1448 // time they are called. Same for uninitialized. The interpreter will
1449 // link the missing class, if any.
1450 make_not_entrant = true;
1451 break;
1452 case Action_make_not_compilable:
1453 // Give up on compiling this method at all.
1454 make_not_entrant = true;
1455 make_not_compilable = true;
1456 break;
1457 default:
1458 ShouldNotReachHere();
1459 }
1461 // Setting +ProfileTraps fixes the following, on all platforms:
1462 // 4852688: ProfileInterpreter is off by default for ia64. The result is
1463 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1464 // recompile relies on a MethodData* to record heroic opt failures.
1466 // Whether the interpreter is producing MDO data or not, we also need
1467 // to use the MDO to detect hot deoptimization points and control
1468 // aggressive optimization.
1469 bool inc_recompile_count = false;
1470 ProfileData* pdata = NULL;
1471 if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1472 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1473 uint this_trap_count = 0;
1474 bool maybe_prior_trap = false;
1475 bool maybe_prior_recompile = false;
1476 pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1477 //outputs:
1478 this_trap_count,
1479 maybe_prior_trap,
1480 maybe_prior_recompile);
1481 // Because the interpreter also counts null, div0, range, and class
1482 // checks, these traps from compiled code are double-counted.
1483 // This is harmless; it just means that the PerXTrapLimit values
1484 // are in effect a little smaller than they look.
1486 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1487 if (per_bc_reason != Reason_none) {
1488 // Now take action based on the partially known per-BCI history.
1489 if (maybe_prior_trap
1490 && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1491 // If there are too many traps at this BCI, force a recompile.
1492 // This will allow the compiler to see the limit overflow, and
1493 // take corrective action, if possible. The compiler generally
1494 // does not use the exact PerBytecodeTrapLimit value, but instead
1495 // changes its tactics if it sees any traps at all. This provides
1496 // a little hysteresis, delaying a recompile until a trap happens
1497 // several times.
1498 //
1499 // Actually, since there is only one bit of counter per BCI,
1500 // the possible per-BCI counts are {0,1,(per-method count)}.
1501 // This produces accurate results if in fact there is only
1502 // one hot trap site, but begins to get fuzzy if there are
1503 // many sites. For example, if there are ten sites each
1504 // trapping two or more times, they each get the blame for
1505 // all of their traps.
1506 make_not_entrant = true;
1507 }
1509 // Detect repeated recompilation at the same BCI, and enforce a limit.
1510 if (make_not_entrant && maybe_prior_recompile) {
1511 // More than one recompile at this point.
1512 inc_recompile_count = maybe_prior_trap;
1513 }
1514 } else {
1515 // For reasons which are not recorded per-bytecode, we simply
1516 // force recompiles unconditionally.
1517 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1518 make_not_entrant = true;
1519 }
1521 // Go back to the compiler if there are too many traps in this method.
1522 if (this_trap_count >= (uint)PerMethodTrapLimit) {
1523 // If there are too many traps in this method, force a recompile.
1524 // This will allow the compiler to see the limit overflow, and
1525 // take corrective action, if possible.
1526 // (This condition is an unlikely backstop only, because the
1527 // PerBytecodeTrapLimit is more likely to take effect first,
1528 // if it is applicable.)
1529 make_not_entrant = true;
1530 }
1532 // Here's more hysteresis: If there has been a recompile at
1533 // this trap point already, run the method in the interpreter
1534 // for a while to exercise it more thoroughly.
1535 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1536 reprofile = true;
1537 }
1539 }
1541 // Take requested actions on the method:
1543 // Recompile
1544 if (make_not_entrant) {
1545 if (!nm->make_not_entrant()) {
1546 return; // the call did not change nmethod's state
1547 }
1549 if (pdata != NULL) {
1550 // Record the recompilation event, if any.
1551 int tstate0 = pdata->trap_state();
1552 int tstate1 = trap_state_set_recompiled(tstate0, true);
1553 if (tstate1 != tstate0)
1554 pdata->set_trap_state(tstate1);
1555 }
1556 }
1558 if (inc_recompile_count) {
1559 trap_mdo->inc_overflow_recompile_count();
1560 if ((uint)trap_mdo->overflow_recompile_count() >
1561 (uint)PerBytecodeRecompilationCutoff) {
1562 // Give up on the method containing the bad BCI.
1563 if (trap_method() == nm->method()) {
1564 make_not_compilable = true;
1565 } else {
1566 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1567 // But give grace to the enclosing nm->method().
1568 }
1569 }
1570 }
1572 // Reprofile
1573 if (reprofile) {
1574 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1575 }
1577 // Give up compiling
1578 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1579 assert(make_not_entrant, "consistent");
1580 nm->method()->set_not_compilable(CompLevel_full_optimization);
1581 }
1583 } // Free marked resources
1585 }
1586 JRT_END
1588 MethodData*
1589 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1590 bool create_if_missing) {
1591 Thread* THREAD = thread;
1592 MethodData* mdo = m()->method_data();
1593 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1594 // Build an MDO. Ignore errors like OutOfMemory;
1595 // that simply means we won't have an MDO to update.
1596 Method::build_interpreter_method_data(m, THREAD);
1597 if (HAS_PENDING_EXCEPTION) {
1598 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1599 CLEAR_PENDING_EXCEPTION;
1600 }
1601 mdo = m()->method_data();
1602 }
1603 return mdo;
1604 }
1606 ProfileData*
1607 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1608 int trap_bci,
1609 Deoptimization::DeoptReason reason,
1610 //outputs:
1611 uint& ret_this_trap_count,
1612 bool& ret_maybe_prior_trap,
1613 bool& ret_maybe_prior_recompile) {
1614 uint prior_trap_count = trap_mdo->trap_count(reason);
1615 uint this_trap_count = trap_mdo->inc_trap_count(reason);
1617 // If the runtime cannot find a place to store trap history,
1618 // it is estimated based on the general condition of the method.
1619 // If the method has ever been recompiled, or has ever incurred
1620 // a trap with the present reason , then this BCI is assumed
1621 // (pessimistically) to be the culprit.
1622 bool maybe_prior_trap = (prior_trap_count != 0);
1623 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1624 ProfileData* pdata = NULL;
1627 // For reasons which are recorded per bytecode, we check per-BCI data.
1628 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1629 if (per_bc_reason != Reason_none) {
1630 // Find the profile data for this BCI. If there isn't one,
1631 // try to allocate one from the MDO's set of spares.
1632 // This will let us detect a repeated trap at this point.
1633 pdata = trap_mdo->allocate_bci_to_data(trap_bci);
1635 if (pdata != NULL) {
1636 // Query the trap state of this profile datum.
1637 int tstate0 = pdata->trap_state();
1638 if (!trap_state_has_reason(tstate0, per_bc_reason))
1639 maybe_prior_trap = false;
1640 if (!trap_state_is_recompiled(tstate0))
1641 maybe_prior_recompile = false;
1643 // Update the trap state of this profile datum.
1644 int tstate1 = tstate0;
1645 // Record the reason.
1646 tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1647 // Store the updated state on the MDO, for next time.
1648 if (tstate1 != tstate0)
1649 pdata->set_trap_state(tstate1);
1650 } else {
1651 if (LogCompilation && xtty != NULL) {
1652 ttyLocker ttyl;
1653 // Missing MDP? Leave a small complaint in the log.
1654 xtty->elem("missing_mdp bci='%d'", trap_bci);
1655 }
1656 }
1657 }
1659 // Return results:
1660 ret_this_trap_count = this_trap_count;
1661 ret_maybe_prior_trap = maybe_prior_trap;
1662 ret_maybe_prior_recompile = maybe_prior_recompile;
1663 return pdata;
1664 }
1666 void
1667 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1668 ResourceMark rm;
1669 // Ignored outputs:
1670 uint ignore_this_trap_count;
1671 bool ignore_maybe_prior_trap;
1672 bool ignore_maybe_prior_recompile;
1673 query_update_method_data(trap_mdo, trap_bci,
1674 (DeoptReason)reason,
1675 ignore_this_trap_count,
1676 ignore_maybe_prior_trap,
1677 ignore_maybe_prior_recompile);
1678 }
1680 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1682 // Still in Java no safepoints
1683 {
1684 // This enters VM and may safepoint
1685 uncommon_trap_inner(thread, trap_request);
1686 }
1687 return fetch_unroll_info_helper(thread);
1688 }
1690 // Local derived constants.
1691 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1692 const int DS_REASON_MASK = DataLayout::trap_mask >> 1;
1693 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1695 //---------------------------trap_state_reason---------------------------------
1696 Deoptimization::DeoptReason
1697 Deoptimization::trap_state_reason(int trap_state) {
1698 // This assert provides the link between the width of DataLayout::trap_bits
1699 // and the encoding of "recorded" reasons. It ensures there are enough
1700 // bits to store all needed reasons in the per-BCI MDO profile.
1701 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1702 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1703 trap_state -= recompile_bit;
1704 if (trap_state == DS_REASON_MASK) {
1705 return Reason_many;
1706 } else {
1707 assert((int)Reason_none == 0, "state=0 => Reason_none");
1708 return (DeoptReason)trap_state;
1709 }
1710 }
1711 //-------------------------trap_state_has_reason-------------------------------
1712 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1713 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1714 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1715 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1716 trap_state -= recompile_bit;
1717 if (trap_state == DS_REASON_MASK) {
1718 return -1; // true, unspecifically (bottom of state lattice)
1719 } else if (trap_state == reason) {
1720 return 1; // true, definitely
1721 } else if (trap_state == 0) {
1722 return 0; // false, definitely (top of state lattice)
1723 } else {
1724 return 0; // false, definitely
1725 }
1726 }
1727 //-------------------------trap_state_add_reason-------------------------------
1728 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1729 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1730 int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1731 trap_state -= recompile_bit;
1732 if (trap_state == DS_REASON_MASK) {
1733 return trap_state + recompile_bit; // already at state lattice bottom
1734 } else if (trap_state == reason) {
1735 return trap_state + recompile_bit; // the condition is already true
1736 } else if (trap_state == 0) {
1737 return reason + recompile_bit; // no condition has yet been true
1738 } else {
1739 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom
1740 }
1741 }
1742 //-----------------------trap_state_is_recompiled------------------------------
1743 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1744 return (trap_state & DS_RECOMPILE_BIT) != 0;
1745 }
1746 //-----------------------trap_state_set_recompiled-----------------------------
1747 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1748 if (z) return trap_state | DS_RECOMPILE_BIT;
1749 else return trap_state & ~DS_RECOMPILE_BIT;
1750 }
1751 //---------------------------format_trap_state---------------------------------
1752 // This is used for debugging and diagnostics, including hotspot.log output.
1753 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1754 int trap_state) {
1755 DeoptReason reason = trap_state_reason(trap_state);
1756 bool recomp_flag = trap_state_is_recompiled(trap_state);
1757 // Re-encode the state from its decoded components.
1758 int decoded_state = 0;
1759 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1760 decoded_state = trap_state_add_reason(decoded_state, reason);
1761 if (recomp_flag)
1762 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1763 // If the state re-encodes properly, format it symbolically.
1764 // Because this routine is used for debugging and diagnostics,
1765 // be robust even if the state is a strange value.
1766 size_t len;
1767 if (decoded_state != trap_state) {
1768 // Random buggy state that doesn't decode??
1769 len = jio_snprintf(buf, buflen, "#%d", trap_state);
1770 } else {
1771 len = jio_snprintf(buf, buflen, "%s%s",
1772 trap_reason_name(reason),
1773 recomp_flag ? " recompiled" : "");
1774 }
1775 if (len >= buflen)
1776 buf[buflen-1] = '\0';
1777 return buf;
1778 }
1781 //--------------------------------statics--------------------------------------
1782 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1783 = Deoptimization::Action_reinterpret;
1784 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1785 // Note: Keep this in sync. with enum DeoptReason.
1786 "none",
1787 "null_check",
1788 "null_assert",
1789 "range_check",
1790 "class_check",
1791 "array_check",
1792 "intrinsic",
1793 "bimorphic",
1794 "unloaded",
1795 "uninitialized",
1796 "unreached",
1797 "unhandled",
1798 "constraint",
1799 "div0_check",
1800 "age",
1801 "predicate",
1802 "loop_limit_check"
1803 };
1804 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1805 // Note: Keep this in sync. with enum DeoptAction.
1806 "none",
1807 "maybe_recompile",
1808 "reinterpret",
1809 "make_not_entrant",
1810 "make_not_compilable"
1811 };
1813 const char* Deoptimization::trap_reason_name(int reason) {
1814 if (reason == Reason_many) return "many";
1815 if ((uint)reason < Reason_LIMIT)
1816 return _trap_reason_name[reason];
1817 static char buf[20];
1818 sprintf(buf, "reason%d", reason);
1819 return buf;
1820 }
1821 const char* Deoptimization::trap_action_name(int action) {
1822 if ((uint)action < Action_LIMIT)
1823 return _trap_action_name[action];
1824 static char buf[20];
1825 sprintf(buf, "action%d", action);
1826 return buf;
1827 }
1829 // This is used for debugging and diagnostics, including hotspot.log output.
1830 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1831 int trap_request) {
1832 jint unloaded_class_index = trap_request_index(trap_request);
1833 const char* reason = trap_reason_name(trap_request_reason(trap_request));
1834 const char* action = trap_action_name(trap_request_action(trap_request));
1835 size_t len;
1836 if (unloaded_class_index < 0) {
1837 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1838 reason, action);
1839 } else {
1840 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1841 reason, action, unloaded_class_index);
1842 }
1843 if (len >= buflen)
1844 buf[buflen-1] = '\0';
1845 return buf;
1846 }
1848 juint Deoptimization::_deoptimization_hist
1849 [Deoptimization::Reason_LIMIT]
1850 [1 + Deoptimization::Action_LIMIT]
1851 [Deoptimization::BC_CASE_LIMIT]
1852 = {0};
1854 enum {
1855 LSB_BITS = 8,
1856 LSB_MASK = right_n_bits(LSB_BITS)
1857 };
1859 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1860 Bytecodes::Code bc) {
1861 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1862 assert(action >= 0 && action < Action_LIMIT, "oob");
1863 _deoptimization_hist[Reason_none][0][0] += 1; // total
1864 _deoptimization_hist[reason][0][0] += 1; // per-reason total
1865 juint* cases = _deoptimization_hist[reason][1+action];
1866 juint* bc_counter_addr = NULL;
1867 juint bc_counter = 0;
1868 // Look for an unused counter, or an exact match to this BC.
1869 if (bc != Bytecodes::_illegal) {
1870 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1871 juint* counter_addr = &cases[bc_case];
1872 juint counter = *counter_addr;
1873 if ((counter == 0 && bc_counter_addr == NULL)
1874 || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1875 // this counter is either free or is already devoted to this BC
1876 bc_counter_addr = counter_addr;
1877 bc_counter = counter | bc;
1878 }
1879 }
1880 }
1881 if (bc_counter_addr == NULL) {
1882 // Overflow, or no given bytecode.
1883 bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1884 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB
1885 }
1886 *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1887 }
1889 jint Deoptimization::total_deoptimization_count() {
1890 return _deoptimization_hist[Reason_none][0][0];
1891 }
1893 jint Deoptimization::deoptimization_count(DeoptReason reason) {
1894 assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1895 return _deoptimization_hist[reason][0][0];
1896 }
1898 void Deoptimization::print_statistics() {
1899 juint total = total_deoptimization_count();
1900 juint account = total;
1901 if (total != 0) {
1902 ttyLocker ttyl;
1903 if (xtty != NULL) xtty->head("statistics type='deoptimization'");
1904 tty->print_cr("Deoptimization traps recorded:");
1905 #define PRINT_STAT_LINE(name, r) \
1906 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
1907 PRINT_STAT_LINE("total", total);
1908 // For each non-zero entry in the histogram, print the reason,
1909 // the action, and (if specifically known) the type of bytecode.
1910 for (int reason = 0; reason < Reason_LIMIT; reason++) {
1911 for (int action = 0; action < Action_LIMIT; action++) {
1912 juint* cases = _deoptimization_hist[reason][1+action];
1913 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1914 juint counter = cases[bc_case];
1915 if (counter != 0) {
1916 char name[1*K];
1917 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
1918 if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
1919 bc = Bytecodes::_illegal;
1920 sprintf(name, "%s/%s/%s",
1921 trap_reason_name(reason),
1922 trap_action_name(action),
1923 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
1924 juint r = counter >> LSB_BITS;
1925 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
1926 account -= r;
1927 }
1928 }
1929 }
1930 }
1931 if (account != 0) {
1932 PRINT_STAT_LINE("unaccounted", account);
1933 }
1934 #undef PRINT_STAT_LINE
1935 if (xtty != NULL) xtty->tail("statistics");
1936 }
1937 }
1938 #else // COMPILER2 || SHARK
1941 // Stubs for C1 only system.
1942 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1943 return false;
1944 }
1946 const char* Deoptimization::trap_reason_name(int reason) {
1947 return "unknown";
1948 }
1950 void Deoptimization::print_statistics() {
1951 // no output
1952 }
1954 void
1955 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1956 // no udpate
1957 }
1959 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1960 return 0;
1961 }
1963 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1964 Bytecodes::Code bc) {
1965 // no update
1966 }
1968 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1969 int trap_state) {
1970 jio_snprintf(buf, buflen, "#%d", trap_state);
1971 return buf;
1972 }
1974 #endif // COMPILER2 || SHARK