Mercurial > hg > truffle
annotate src/share/vm/runtime/deoptimization.cpp @ 2059:9508a52cbd32
Add deoptimization blob support.
author | Thomas Wuerthinger <wuerthinger@ssw.jku.at> |
---|---|
date | Wed, 19 Jan 2011 15:48:15 +0100 |
parents | 06f017f7daa7 |
children | d25d4ca69222 |
rev | line source |
---|---|
0 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1255
diff
changeset
|
2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1255
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1255
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1255
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 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/methodOop.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 COMPILER2 | |
63 #ifdef TARGET_ARCH_MODEL_x86_32 | |
64 # include "adfiles/ad_x86_32.hpp" | |
65 #endif | |
66 #ifdef TARGET_ARCH_MODEL_x86_64 | |
67 # include "adfiles/ad_x86_64.hpp" | |
68 #endif | |
69 #ifdef TARGET_ARCH_MODEL_sparc | |
70 # include "adfiles/ad_sparc.hpp" | |
71 #endif | |
72 #ifdef TARGET_ARCH_MODEL_zero | |
73 # include "adfiles/ad_zero.hpp" | |
74 #endif | |
75 #endif | |
0 | 76 |
77 bool DeoptimizationMarker::_is_active = false; | |
78 | |
79 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, | |
80 int caller_adjustment, | |
81 int number_of_frames, | |
82 intptr_t* frame_sizes, | |
83 address* frame_pcs, | |
84 BasicType return_type) { | |
85 _size_of_deoptimized_frame = size_of_deoptimized_frame; | |
86 _caller_adjustment = caller_adjustment; | |
87 _number_of_frames = number_of_frames; | |
88 _frame_sizes = frame_sizes; | |
89 _frame_pcs = frame_pcs; | |
90 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); | |
91 _return_type = return_type; | |
92 // PD (x86 only) | |
93 _counter_temp = 0; | |
94 _initial_fp = 0; | |
95 _unpack_kind = 0; | |
96 _sender_sp_temp = 0; | |
97 | |
98 _total_frame_sizes = size_of_frames(); | |
99 } | |
100 | |
101 | |
102 Deoptimization::UnrollBlock::~UnrollBlock() { | |
103 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); | |
104 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); | |
105 FREE_C_HEAP_ARRAY(intptr_t, _register_block); | |
106 } | |
107 | |
108 | |
109 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { | |
110 assert(register_number < RegisterMap::reg_count, "checking register number"); | |
111 return &_register_block[register_number * 2]; | |
112 } | |
113 | |
114 | |
115 | |
116 int Deoptimization::UnrollBlock::size_of_frames() const { | |
117 // Acount first for the adjustment of the initial frame | |
118 int result = _caller_adjustment; | |
119 for (int index = 0; index < number_of_frames(); index++) { | |
120 result += frame_sizes()[index]; | |
121 } | |
122 return result; | |
123 } | |
124 | |
125 | |
126 void Deoptimization::UnrollBlock::print() { | |
127 ttyLocker ttyl; | |
128 tty->print_cr("UnrollBlock"); | |
129 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); | |
130 tty->print( " frame_sizes: "); | |
131 for (int index = 0; index < number_of_frames(); index++) { | |
132 tty->print("%d ", frame_sizes()[index]); | |
133 } | |
134 tty->cr(); | |
135 } | |
136 | |
137 | |
138 // In order to make fetch_unroll_info work properly with escape | |
139 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and | |
140 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation | |
141 // of previously eliminated objects occurs in realloc_objects, which is | |
142 // called from the method fetch_unroll_info_helper below. | |
143 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) | |
144 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, | |
145 // but makes the entry a little slower. There is however a little dance we have to | |
146 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro | |
147 | |
148 // fetch_unroll_info() is called at the beginning of the deoptimization | |
149 // handler. Note this fact before we start generating temporary frames | |
150 // that can confuse an asynchronous stack walker. This counter is | |
151 // decremented at the end of unpack_frames(). | |
152 thread->inc_in_deopt_handler(); | |
153 | |
154 return fetch_unroll_info_helper(thread); | |
155 JRT_END | |
156 | |
157 | |
158 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) | |
159 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { | |
160 | |
161 // Note: there is a safepoint safety issue here. No matter whether we enter | |
162 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once | |
163 // the vframeArray is created. | |
164 // | |
165 | |
1937
4853c5cad3aa
More deoptmization tracing.
Thomas Wuerthinger <wuerthinger@ssw.jku.at>
parents:
1905
diff
changeset
|
166 if (TraceDeoptimization) { |
4853c5cad3aa
More deoptmization tracing.
Thomas Wuerthinger <wuerthinger@ssw.jku.at>
parents:
1905
diff
changeset
|
167 tty->print_cr("fetching unroll info"); |
4853c5cad3aa
More deoptmization tracing.
Thomas Wuerthinger <wuerthinger@ssw.jku.at>
parents:
1905
diff
changeset
|
168 } |
4853c5cad3aa
More deoptmization tracing.
Thomas Wuerthinger <wuerthinger@ssw.jku.at>
parents:
1905
diff
changeset
|
169 |
0 | 170 // Allocate our special deoptimization ResourceMark |
171 DeoptResourceMark* dmark = new DeoptResourceMark(thread); | |
172 assert(thread->deopt_mark() == NULL, "Pending deopt!"); | |
173 thread->set_deopt_mark(dmark); | |
174 | |
175 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect | |
176 RegisterMap map(thread, true); | |
177 RegisterMap dummy_map(thread, false); | |
178 // Now get the deoptee with a valid map | |
179 frame deoptee = stub_frame.sender(&map); | |
1814
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
180 // Set the deoptee nmethod |
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
181 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); |
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
182 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); |
0 | 183 |
184 // Create a growable array of VFrames where each VFrame represents an inlined | |
185 // Java frame. This storage is allocated with the usual system arena. | |
186 assert(deoptee.is_compiled_frame(), "Wrong frame type"); | |
187 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); | |
188 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); | |
189 while (!vf->is_top()) { | |
190 assert(vf->is_compiled_frame(), "Wrong frame type"); | |
191 chunk->push(compiledVFrame::cast(vf)); | |
192 vf = vf->sender(); | |
193 } | |
194 assert(vf->is_compiled_frame(), "Wrong frame type"); | |
195 chunk->push(compiledVFrame::cast(vf)); | |
196 | |
197 #ifdef COMPILER2 | |
198 // Reallocate the non-escaping objects and restore their fields. Then | |
199 // relock objects if synchronization on them was eliminated. | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
200 if (DoEscapeAnalysis) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
201 if (EliminateAllocations) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
202 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
203 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); |
1253
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
204 |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
205 // The flag return_oop() indicates call sites which return oop |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
206 // in compiled code. Such sites include java method calls, |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
207 // runtime calls (for example, used to allocate new objects/arrays |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
208 // on slow code path) and any other calls generated in compiled code. |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
209 // It is not guaranteed that we can get such information here only |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
210 // by analyzing bytecode in deoptimized frames. This is why this flag |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
211 // is set during method compilation (see Compile::Process_OopMap_Node()). |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
212 bool save_oop_result = chunk->at(0)->scope()->return_oop(); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
213 Handle return_value; |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
214 if (save_oop_result) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
215 // Reallocation may trigger GC. If deoptimization happened on return from |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
216 // call which returns oop we need to save it since it is not in oopmap. |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
217 oop result = deoptee.saved_oop_result(&map); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
218 assert(result == NULL || result->is_oop(), "must be oop"); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
219 return_value = Handle(thread, result); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
220 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
221 if (TraceDeoptimization) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
222 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
223 } |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
224 } |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
225 bool reallocated = false; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
226 if (objects != NULL) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
227 JRT_BLOCK |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
228 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
229 JRT_END |
0 | 230 } |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
231 if (reallocated) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
232 reassign_fields(&deoptee, &map, objects); |
0 | 233 #ifndef PRODUCT |
234 if (TraceDeoptimization) { | |
235 ttyLocker ttyl; | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
236 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
237 print_objects(objects); |
1253
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
238 } |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
239 #endif |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
240 } |
1253
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
241 if (save_oop_result) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
242 // Restore result. |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
243 deoptee.set_saved_oop_result(&map, return_value()); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
244 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
245 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
246 if (EliminateLocks) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
247 #ifndef PRODUCT |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
248 bool first = true; |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
249 #endif |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
250 for (int i = 0; i < chunk->length(); i++) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
251 compiledVFrame* cvf = chunk->at(i); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
252 assert (cvf->scope() != NULL,"expect only compiled java frames"); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
253 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
254 if (monitors->is_nonempty()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
255 relock_objects(monitors, thread); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
256 #ifndef PRODUCT |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
257 if (TraceDeoptimization) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
258 ttyLocker ttyl; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
259 for (int j = 0; j < monitors->length(); j++) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
260 MonitorInfo* mi = monitors->at(j); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
261 if (mi->eliminated()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
262 if (first) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
263 first = false; |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
264 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
265 } |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
266 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
267 } |
0 | 268 } |
269 } | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
270 #endif |
0 | 271 } |
272 } | |
273 } | |
274 } | |
275 #endif // COMPILER2 | |
276 // Ensure that no safepoint is taken after pointers have been stored | |
277 // in fields of rematerialized objects. If a safepoint occurs from here on | |
278 // out the java state residing in the vframeArray will be missed. | |
279 No_Safepoint_Verifier no_safepoint; | |
280 | |
281 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); | |
282 | |
283 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; | |
284 thread->set_vframe_array_head(array); | |
285 | |
286 // Now that the vframeArray has been created if we have any deferred local writes | |
287 // added by jvmti then we can free up that structure as the data is now in the | |
288 // vframeArray | |
289 | |
290 if (thread->deferred_locals() != NULL) { | |
291 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); | |
292 int i = 0; | |
293 do { | |
294 // Because of inlining we could have multiple vframes for a single frame | |
295 // and several of the vframes could have deferred writes. Find them all. | |
296 if (list->at(i)->id() == array->original().id()) { | |
297 jvmtiDeferredLocalVariableSet* dlv = list->at(i); | |
298 list->remove_at(i); | |
299 // individual jvmtiDeferredLocalVariableSet are CHeapObj's | |
300 delete dlv; | |
301 } else { | |
302 i++; | |
303 } | |
304 } while ( i < list->length() ); | |
305 if (list->length() == 0) { | |
306 thread->set_deferred_locals(NULL); | |
307 // free the list and elements back to C heap. | |
308 delete list; | |
309 } | |
310 | |
311 } | |
312 | |
1692 | 313 #ifndef SHARK |
0 | 314 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. |
315 CodeBlob* cb = stub_frame.cb(); | |
316 // Verify we have the right vframeArray | |
317 assert(cb->frame_size() >= 0, "Unexpected frame size"); | |
318 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); | |
319 | |
1204 | 320 // If the deopt call site is a MethodHandle invoke call site we have |
321 // to adjust the unpack_sp. | |
322 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); | |
323 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) | |
324 unpack_sp = deoptee.unextended_sp(); | |
325 | |
0 | 326 #ifdef ASSERT |
327 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); | |
328 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); | |
329 #endif | |
1692 | 330 #else |
331 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); | |
332 #endif // !SHARK | |
333 | |
0 | 334 // This is a guarantee instead of an assert because if vframe doesn't match |
335 // we will unpack the wrong deoptimized frame and wind up in strange places | |
336 // where it will be very difficult to figure out what went wrong. Better | |
337 // to die an early death here than some very obscure death later when the | |
338 // trail is cold. | |
339 // Note: on ia64 this guarantee can be fooled by frames with no memory stack | |
340 // in that it will fail to detect a problem when there is one. This needs | |
341 // more work in tiger timeframe. | |
342 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); | |
343 | |
344 int number_of_frames = array->frames(); | |
345 | |
346 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost | |
347 // virtual activation, which is the reverse of the elements in the vframes array. | |
348 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); | |
349 // +1 because we always have an interpreter return address for the final slot. | |
350 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); | |
351 int callee_parameters = 0; | |
352 int callee_locals = 0; | |
353 int popframe_extra_args = 0; | |
354 // Create an interpreter return address for the stub to use as its return | |
355 // address so the skeletal frames are perfectly walkable | |
356 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); | |
357 | |
358 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost | |
359 // activation be put back on the expression stack of the caller for reexecution | |
360 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { | |
361 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); | |
362 } | |
363 | |
364 // | |
365 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) | |
366 // frame_sizes/frame_pcs[1] next oldest frame (int) | |
367 // frame_sizes/frame_pcs[n] youngest frame (int) | |
368 // | |
369 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame | |
370 // owns the space for the return address to it's caller). Confusing ain't it. | |
371 // | |
372 // The vframe array can address vframes with indices running from | |
373 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. | |
374 // When we create the skeletal frames we need the oldest frame to be in the zero slot | |
375 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. | |
376 // so things look a little strange in this loop. | |
377 // | |
378 for (int index = 0; index < array->frames(); index++ ) { | |
379 // frame[number_of_frames - 1 ] = on_stack_size(youngest) | |
380 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) | |
381 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) | |
382 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, | |
383 callee_locals, | |
384 index == 0, | |
385 popframe_extra_args); | |
386 // This pc doesn't have to be perfect just good enough to identify the frame | |
387 // as interpreted so the skeleton frame will be walkable | |
388 // The correct pc will be set when the skeleton frame is completely filled out | |
389 // The final pc we store in the loop is wrong and will be overwritten below | |
390 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; | |
391 | |
392 callee_parameters = array->element(index)->method()->size_of_parameters(); | |
393 callee_locals = array->element(index)->method()->max_locals(); | |
394 popframe_extra_args = 0; | |
395 } | |
396 | |
397 // Compute whether the root vframe returns a float or double value. | |
398 BasicType return_type; | |
399 { | |
400 HandleMark hm; | |
401 methodHandle method(thread, array->element(0)->method()); | |
402 Bytecode_invoke* invoke = Bytecode_invoke_at_check(method, array->element(0)->bci()); | |
403 return_type = (invoke != NULL) ? invoke->result_type(thread) : T_ILLEGAL; | |
404 } | |
405 | |
406 // Compute information for handling adapters and adjusting the frame size of the caller. | |
407 int caller_adjustment = 0; | |
408 | |
409 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized | |
410 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather | |
411 // than simply use array->sender.pc(). This requires us to walk the current set of frames | |
412 // | |
413 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame | |
414 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller | |
415 | |
416 // Compute the amount the oldest interpreter frame will have to adjust | |
417 // its caller's stack by. If the caller is a compiled frame then | |
418 // we pretend that the callee has no parameters so that the | |
419 // extension counts for the full amount of locals and not just | |
420 // locals-parms. This is because without a c2i adapter the parm | |
421 // area as created by the compiled frame will not be usable by | |
422 // the interpreter. (Depending on the calling convention there | |
423 // may not even be enough space). | |
424 | |
425 // QQQ I'd rather see this pushed down into last_frame_adjust | |
426 // and have it take the sender (aka caller). | |
427 | |
428 if (deopt_sender.is_compiled_frame()) { | |
429 caller_adjustment = last_frame_adjust(0, callee_locals); | |
430 } else if (callee_locals > callee_parameters) { | |
431 // The caller frame may need extending to accommodate | |
432 // non-parameter locals of the first unpacked interpreted frame. | |
433 // Compute that adjustment. | |
434 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); | |
435 } | |
436 | |
437 | |
438 // If the sender is deoptimized the we must retrieve the address of the handler | |
439 // since the frame will "magically" show the original pc before the deopt | |
440 // and we'd undo the deopt. | |
441 | |
442 frame_pcs[0] = deopt_sender.raw_pc(); | |
443 | |
1692 | 444 #ifndef SHARK |
0 | 445 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); |
1692 | 446 #endif // SHARK |
0 | 447 |
448 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, | |
449 caller_adjustment * BytesPerWord, | |
450 number_of_frames, | |
451 frame_sizes, | |
452 frame_pcs, | |
453 return_type); | |
454 #if defined(IA32) || defined(AMD64) | |
455 // We need a way to pass fp to the unpacking code so the skeletal frames | |
456 // come out correct. This is only needed for x86 because of c2 using ebp | |
457 // as an allocatable register. So this update is useless (and harmless) | |
458 // on the other platforms. It would be nice to do this in a different | |
459 // way but even the old style deoptimization had a problem with deriving | |
460 // this value. NEEDS_CLEANUP | |
461 // Note: now that c1 is using c2's deopt blob we must do this on all | |
462 // x86 based platforms | |
463 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); | |
464 *fp_addr = array->sender().fp(); // was adapter_caller | |
465 #endif /* IA32 || AMD64 */ | |
466 | |
467 if (array->frames() > 1) { | |
468 if (VerifyStack && TraceDeoptimization) { | |
469 tty->print_cr("Deoptimizing method containing inlining"); | |
470 } | |
471 } | |
472 | |
473 array->set_unroll_block(info); | |
474 return info; | |
475 } | |
476 | |
477 // Called to cleanup deoptimization data structures in normal case | |
478 // after unpacking to stack and when stack overflow error occurs | |
479 void Deoptimization::cleanup_deopt_info(JavaThread *thread, | |
480 vframeArray *array) { | |
481 | |
482 // Get array if coming from exception | |
483 if (array == NULL) { | |
484 array = thread->vframe_array_head(); | |
485 } | |
486 thread->set_vframe_array_head(NULL); | |
487 | |
488 // Free the previous UnrollBlock | |
489 vframeArray* old_array = thread->vframe_array_last(); | |
490 thread->set_vframe_array_last(array); | |
491 | |
492 if (old_array != NULL) { | |
493 UnrollBlock* old_info = old_array->unroll_block(); | |
494 old_array->set_unroll_block(NULL); | |
495 delete old_info; | |
496 delete old_array; | |
497 } | |
498 | |
499 // Deallocate any resource creating in this routine and any ResourceObjs allocated | |
500 // inside the vframeArray (StackValueCollections) | |
501 | |
502 delete thread->deopt_mark(); | |
503 thread->set_deopt_mark(NULL); | |
1814
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
504 thread->set_deopt_nmethod(NULL); |
0 | 505 |
506 | |
507 if (JvmtiExport::can_pop_frame()) { | |
508 #ifndef CC_INTERP | |
509 // Regardless of whether we entered this routine with the pending | |
510 // popframe condition bit set, we should always clear it now | |
511 thread->clear_popframe_condition(); | |
512 #else | |
513 // C++ interpeter will clear has_pending_popframe when it enters | |
514 // with method_resume. For deopt_resume2 we clear it now. | |
515 if (thread->popframe_forcing_deopt_reexecution()) | |
516 thread->clear_popframe_condition(); | |
517 #endif /* CC_INTERP */ | |
518 } | |
519 | |
520 // unpack_frames() is called at the end of the deoptimization handler | |
521 // and (in C2) at the end of the uncommon trap handler. Note this fact | |
522 // so that an asynchronous stack walker can work again. This counter is | |
523 // incremented at the beginning of fetch_unroll_info() and (in C2) at | |
524 // the beginning of uncommon_trap(). | |
525 thread->dec_in_deopt_handler(); | |
526 } | |
527 | |
528 | |
529 // Return BasicType of value being returned | |
530 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) | |
531 | |
532 // We are already active int he special DeoptResourceMark any ResourceObj's we | |
533 // allocate will be freed at the end of the routine. | |
534 | |
535 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, | |
536 // but makes the entry a little slower. There is however a little dance we have to | |
537 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro | |
538 ResetNoHandleMark rnhm; // No-op in release/product versions | |
539 HandleMark hm; | |
540 | |
541 frame stub_frame = thread->last_frame(); | |
542 | |
543 // Since the frame to unpack is the top frame of this thread, the vframe_array_head | |
544 // must point to the vframeArray for the unpack frame. | |
545 vframeArray* array = thread->vframe_array_head(); | |
546 | |
547 #ifndef PRODUCT | |
548 if (TraceDeoptimization) { | |
549 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); | |
550 } | |
551 #endif | |
552 | |
553 UnrollBlock* info = array->unroll_block(); | |
554 | |
555 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. | |
556 array->unpack_to_stack(stub_frame, exec_mode); | |
557 | |
558 BasicType bt = info->return_type(); | |
559 | |
560 // If we have an exception pending, claim that the return type is an oop | |
561 // so the deopt_blob does not overwrite the exception_oop. | |
562 | |
563 if (exec_mode == Unpack_exception) | |
564 bt = T_OBJECT; | |
565 | |
566 // Cleanup thread deopt data | |
567 cleanup_deopt_info(thread, array); | |
568 | |
569 #ifndef PRODUCT | |
570 if (VerifyStack) { | |
571 ResourceMark res_mark; | |
572 | |
573 // Verify that the just-unpacked frames match the interpreter's | |
574 // notions of expression stack and locals | |
575 vframeArray* cur_array = thread->vframe_array_last(); | |
576 RegisterMap rm(thread, false); | |
577 rm.set_include_argument_oops(false); | |
578 bool is_top_frame = true; | |
579 int callee_size_of_parameters = 0; | |
580 int callee_max_locals = 0; | |
581 for (int i = 0; i < cur_array->frames(); i++) { | |
582 vframeArrayElement* el = cur_array->element(i); | |
583 frame* iframe = el->iframe(); | |
584 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); | |
585 | |
586 // Get the oop map for this bci | |
587 InterpreterOopMap mask; | |
588 int cur_invoke_parameter_size = 0; | |
589 bool try_next_mask = false; | |
590 int next_mask_expression_stack_size = -1; | |
591 int top_frame_expression_stack_adjustment = 0; | |
592 methodHandle mh(thread, iframe->interpreter_frame_method()); | |
593 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); | |
594 BytecodeStream str(mh); | |
595 str.set_start(iframe->interpreter_frame_bci()); | |
596 int max_bci = mh->code_size(); | |
597 // Get to the next bytecode if possible | |
598 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); | |
599 // Check to see if we can grab the number of outgoing arguments | |
600 // at an uncommon trap for an invoke (where the compiler | |
601 // generates debug info before the invoke has executed) | |
602 Bytecodes::Code cur_code = str.next(); | |
603 if (cur_code == Bytecodes::_invokevirtual || | |
604 cur_code == Bytecodes::_invokespecial || | |
605 cur_code == Bytecodes::_invokestatic || | |
606 cur_code == Bytecodes::_invokeinterface) { | |
607 Bytecode_invoke* invoke = Bytecode_invoke_at(mh, iframe->interpreter_frame_bci()); | |
608 symbolHandle signature(thread, invoke->signature()); | |
609 ArgumentSizeComputer asc(signature); | |
610 cur_invoke_parameter_size = asc.size(); | |
611 if (cur_code != Bytecodes::_invokestatic) { | |
612 // Add in receiver | |
613 ++cur_invoke_parameter_size; | |
614 } | |
615 } | |
616 if (str.bci() < max_bci) { | |
617 Bytecodes::Code bc = str.next(); | |
618 if (bc >= 0) { | |
619 // The interpreter oop map generator reports results before | |
620 // the current bytecode has executed except in the case of | |
621 // calls. It seems to be hard to tell whether the compiler | |
622 // has emitted debug information matching the "state before" | |
623 // a given bytecode or the state after, so we try both | |
624 switch (cur_code) { | |
625 case Bytecodes::_invokevirtual: | |
626 case Bytecodes::_invokespecial: | |
627 case Bytecodes::_invokestatic: | |
628 case Bytecodes::_invokeinterface: | |
629 case Bytecodes::_athrow: | |
630 break; | |
631 default: { | |
632 InterpreterOopMap next_mask; | |
633 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); | |
634 next_mask_expression_stack_size = next_mask.expression_stack_size(); | |
635 // Need to subtract off the size of the result type of | |
636 // the bytecode because this is not described in the | |
637 // debug info but returned to the interpreter in the TOS | |
638 // caching register | |
639 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); | |
640 if (bytecode_result_type != T_ILLEGAL) { | |
641 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; | |
642 } | |
643 assert(top_frame_expression_stack_adjustment >= 0, ""); | |
644 try_next_mask = true; | |
645 break; | |
646 } | |
647 } | |
648 } | |
649 } | |
650 | |
651 // Verify stack depth and oops in frame | |
652 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) | |
653 if (!( | |
654 /* SPARC */ | |
655 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || | |
656 /* x86 */ | |
657 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || | |
658 (try_next_mask && | |
659 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - | |
660 top_frame_expression_stack_adjustment))) || | |
661 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || | |
662 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && | |
663 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) | |
664 )) { | |
665 ttyLocker ttyl; | |
666 | |
667 // Print out some information that will help us debug the problem | |
668 tty->print_cr("Wrong number of expression stack elements during deoptimization"); | |
669 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); | |
670 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", | |
671 iframe->interpreter_frame_expression_stack_size()); | |
672 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); | |
673 tty->print_cr(" try_next_mask = %d", try_next_mask); | |
674 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); | |
675 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); | |
676 tty->print_cr(" callee_max_locals = %d", callee_max_locals); | |
677 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); | |
678 tty->print_cr(" exec_mode = %d", exec_mode); | |
679 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); | |
680 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); | |
681 tty->print_cr(" Interpreted frames:"); | |
682 for (int k = 0; k < cur_array->frames(); k++) { | |
683 vframeArrayElement* el = cur_array->element(k); | |
684 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); | |
685 } | |
686 cur_array->print_on_2(tty); | |
687 guarantee(false, "wrong number of expression stack elements during deopt"); | |
688 } | |
689 VerifyOopClosure verify; | |
690 iframe->oops_interpreted_do(&verify, &rm, false); | |
691 callee_size_of_parameters = mh->size_of_parameters(); | |
692 callee_max_locals = mh->max_locals(); | |
693 is_top_frame = false; | |
694 } | |
695 } | |
696 #endif /* !PRODUCT */ | |
697 | |
698 | |
699 return bt; | |
700 JRT_END | |
701 | |
702 | |
703 int Deoptimization::deoptimize_dependents() { | |
704 Threads::deoptimized_wrt_marked_nmethods(); | |
705 return 0; | |
706 } | |
707 | |
708 | |
709 #ifdef COMPILER2 | |
710 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { | |
711 Handle pending_exception(thread->pending_exception()); | |
712 const char* exception_file = thread->exception_file(); | |
713 int exception_line = thread->exception_line(); | |
714 thread->clear_pending_exception(); | |
715 | |
716 for (int i = 0; i < objects->length(); i++) { | |
717 assert(objects->at(i)->is_object(), "invalid debug information"); | |
718 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
719 | |
720 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
721 oop obj = NULL; | |
722 | |
723 if (k->oop_is_instance()) { | |
724 instanceKlass* ik = instanceKlass::cast(k()); | |
725 obj = ik->allocate_instance(CHECK_(false)); | |
726 } else if (k->oop_is_typeArray()) { | |
727 typeArrayKlass* ak = typeArrayKlass::cast(k()); | |
728 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); | |
729 int len = sv->field_size() / type2size[ak->element_type()]; | |
730 obj = ak->allocate(len, CHECK_(false)); | |
731 } else if (k->oop_is_objArray()) { | |
732 objArrayKlass* ak = objArrayKlass::cast(k()); | |
733 obj = ak->allocate(sv->field_size(), CHECK_(false)); | |
734 } | |
735 | |
736 assert(obj != NULL, "allocation failed"); | |
737 assert(sv->value().is_null(), "redundant reallocation"); | |
738 sv->set_value(obj); | |
739 } | |
740 | |
741 if (pending_exception.not_null()) { | |
742 thread->set_pending_exception(pending_exception(), exception_file, exception_line); | |
743 } | |
744 | |
745 return true; | |
746 } | |
747 | |
748 // This assumes that the fields are stored in ObjectValue in the same order | |
749 // they are yielded by do_nonstatic_fields. | |
750 class FieldReassigner: public FieldClosure { | |
751 frame* _fr; | |
752 RegisterMap* _reg_map; | |
753 ObjectValue* _sv; | |
754 instanceKlass* _ik; | |
755 oop _obj; | |
756 | |
757 int _i; | |
758 public: | |
759 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : | |
760 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} | |
761 | |
762 int i() const { return _i; } | |
763 | |
764 | |
765 void do_field(fieldDescriptor* fd) { | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
766 intptr_t val; |
0 | 767 StackValue* value = |
768 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); | |
769 int offset = fd->offset(); | |
770 switch (fd->field_type()) { | |
771 case T_OBJECT: case T_ARRAY: | |
772 assert(value->type() == T_OBJECT, "Agreement."); | |
773 _obj->obj_field_put(offset, value->get_obj()()); | |
774 break; | |
775 | |
776 case T_LONG: case T_DOUBLE: { | |
777 assert(value->type() == T_INT, "Agreement."); | |
778 StackValue* low = | |
779 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
780 #ifdef _LP64 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
781 jlong res = (jlong)low->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
782 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
783 #ifdef SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
784 // For SPARC we have to swap high and low words. |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
785 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
786 #else |
0 | 787 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
788 #endif //SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
789 #endif |
0 | 790 _obj->long_field_put(offset, res); |
791 break; | |
792 } | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
793 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
0 | 794 case T_INT: case T_FLOAT: // 4 bytes. |
795 assert(value->type() == T_INT, "Agreement."); | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
796 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
797 _obj->int_field_put(offset, (jint)*((jint*)&val)); |
0 | 798 break; |
799 | |
800 case T_SHORT: case T_CHAR: // 2 bytes | |
801 assert(value->type() == T_INT, "Agreement."); | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
802 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
803 _obj->short_field_put(offset, (jshort)*((jint*)&val)); |
0 | 804 break; |
805 | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
806 case T_BOOLEAN: case T_BYTE: // 1 byte |
0 | 807 assert(value->type() == T_INT, "Agreement."); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
808 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
809 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); |
0 | 810 break; |
811 | |
812 default: | |
813 ShouldNotReachHere(); | |
814 } | |
815 _i++; | |
816 } | |
817 }; | |
818 | |
819 // restore elements of an eliminated type array | |
820 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { | |
821 int index = 0; | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
822 intptr_t val; |
0 | 823 |
824 for (int i = 0; i < sv->field_size(); i++) { | |
825 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); | |
826 switch(type) { | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
827 case T_LONG: case T_DOUBLE: { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
828 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
829 StackValue* low = |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
830 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
831 #ifdef _LP64 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
832 jlong res = (jlong)low->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
833 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
834 #ifdef SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
835 // For SPARC we have to swap high and low words. |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
836 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
837 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
838 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
839 #endif //SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
840 #endif |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
841 obj->long_at_put(index, res); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
842 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
843 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
844 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
845 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
846 case T_INT: case T_FLOAT: // 4 bytes. |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
847 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
848 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
849 obj->int_at_put(index, (jint)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
850 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
851 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
852 case T_SHORT: case T_CHAR: // 2 bytes |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
853 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
854 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
855 obj->short_at_put(index, (jshort)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
856 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
857 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
858 case T_BOOLEAN: case T_BYTE: // 1 byte |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
859 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
860 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
861 obj->bool_at_put(index, (jboolean)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
862 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
863 |
0 | 864 default: |
865 ShouldNotReachHere(); | |
866 } | |
867 index++; | |
868 } | |
869 } | |
870 | |
871 | |
872 // restore fields of an eliminated object array | |
873 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { | |
874 for (int i = 0; i < sv->field_size(); i++) { | |
875 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); | |
876 assert(value->type() == T_OBJECT, "object element expected"); | |
877 obj->obj_at_put(i, value->get_obj()()); | |
878 } | |
879 } | |
880 | |
881 | |
882 // restore fields of all eliminated objects and arrays | |
883 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { | |
884 for (int i = 0; i < objects->length(); i++) { | |
885 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
886 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
887 Handle obj = sv->value(); | |
888 assert(obj.not_null(), "reallocation was missed"); | |
889 | |
890 if (k->oop_is_instance()) { | |
891 instanceKlass* ik = instanceKlass::cast(k()); | |
892 FieldReassigner reassign(fr, reg_map, sv, obj()); | |
893 ik->do_nonstatic_fields(&reassign); | |
894 } else if (k->oop_is_typeArray()) { | |
895 typeArrayKlass* ak = typeArrayKlass::cast(k()); | |
896 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); | |
897 } else if (k->oop_is_objArray()) { | |
898 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); | |
899 } | |
900 } | |
901 } | |
902 | |
903 | |
904 // relock objects for which synchronization was eliminated | |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
905 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { |
0 | 906 for (int i = 0; i < monitors->length(); i++) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
907 MonitorInfo* mon_info = monitors->at(i); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
908 if (mon_info->eliminated()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
909 assert(mon_info->owner() != NULL, "reallocation was missed"); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
910 Handle obj = Handle(mon_info->owner()); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
911 markOop mark = obj->mark(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
912 if (UseBiasedLocking && mark->has_bias_pattern()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
913 // New allocated objects may have the mark set to anonymously biased. |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
914 // Also the deoptimized method may called methods with synchronization |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
915 // where the thread-local object is bias locked to the current thread. |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
916 assert(mark->is_biased_anonymously() || |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
917 mark->biased_locker() == thread, "should be locked to current thread"); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
918 // Reset mark word to unbiased prototype. |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
919 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
920 obj->set_mark(unbiased_prototype); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
921 } |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
922 BasicLock* lock = mon_info->lock(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
923 ObjectSynchronizer::slow_enter(obj, lock, thread); |
0 | 924 } |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
925 assert(mon_info->owner()->is_locked(), "object must be locked now"); |
0 | 926 } |
927 } | |
928 | |
929 | |
930 #ifndef PRODUCT | |
931 // print information about reallocated objects | |
932 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { | |
933 fieldDescriptor fd; | |
934 | |
935 for (int i = 0; i < objects->length(); i++) { | |
936 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
937 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
938 Handle obj = sv->value(); | |
939 | |
940 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); | |
941 k->as_klassOop()->print_value(); | |
942 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); | |
943 tty->cr(); | |
944 | |
945 if (Verbose) { | |
946 k->oop_print_on(obj(), tty); | |
947 } | |
948 } | |
949 } | |
950 #endif | |
951 #endif // COMPILER2 | |
952 | |
953 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { | |
954 | |
955 #ifndef PRODUCT | |
956 if (TraceDeoptimization) { | |
957 ttyLocker ttyl; | |
958 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); | |
959 fr.print_on(tty); | |
960 tty->print_cr(" Virtual frames (innermost first):"); | |
961 for (int index = 0; index < chunk->length(); index++) { | |
962 compiledVFrame* vf = chunk->at(index); | |
963 tty->print(" %2d - ", index); | |
964 vf->print_value(); | |
965 int bci = chunk->at(index)->raw_bci(); | |
966 const char* code_name; | |
967 if (bci == SynchronizationEntryBCI) { | |
968 code_name = "sync entry"; | |
969 } else { | |
970 Bytecodes::Code code = Bytecodes::code_at(vf->method(), bci); | |
971 code_name = Bytecodes::name(code); | |
972 } | |
973 tty->print(" - %s", code_name); | |
974 tty->print_cr(" @ bci %d ", bci); | |
975 if (Verbose) { | |
976 vf->print(); | |
977 tty->cr(); | |
978 } | |
979 } | |
980 } | |
981 #endif | |
982 | |
983 // Register map for next frame (used for stack crawl). We capture | |
984 // the state of the deopt'ing frame's caller. Thus if we need to | |
985 // stuff a C2I adapter we can properly fill in the callee-save | |
986 // register locations. | |
987 frame caller = fr.sender(reg_map); | |
988 int frame_size = caller.sp() - fr.sp(); | |
989 | |
990 frame sender = caller; | |
991 | |
992 // Since the Java thread being deoptimized will eventually adjust it's own stack, | |
993 // the vframeArray containing the unpacking information is allocated in the C heap. | |
994 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). | |
995 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); | |
996 | |
997 // Compare the vframeArray to the collected vframes | |
998 assert(array->structural_compare(thread, chunk), "just checking"); | |
999 Events::log("# vframes = %d", (intptr_t)chunk->length()); | |
1000 | |
1001 #ifndef PRODUCT | |
1002 if (TraceDeoptimization) { | |
1003 ttyLocker ttyl; | |
1004 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); | |
1005 } | |
1006 #endif // PRODUCT | |
1007 | |
1008 return array; | |
1009 } | |
1010 | |
1011 | |
1012 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { | |
1013 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); | |
1014 for (int i = 0; i < monitors->length(); i++) { | |
1015 MonitorInfo* mon_info = monitors->at(i); | |
818
b109e761e927
6837472: com/sun/jdi/MonitorFrameInfo.java fails with AggressiveOpts in 6u14
kvn
parents:
196
diff
changeset
|
1016 if (!mon_info->eliminated() && mon_info->owner() != NULL) { |
0 | 1017 objects_to_revoke->append(Handle(mon_info->owner())); |
1018 } | |
1019 } | |
1020 } | |
1021 | |
1022 | |
1023 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { | |
1024 if (!UseBiasedLocking) { | |
1025 return; | |
1026 } | |
1027 | |
1028 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); | |
1029 | |
1030 // Unfortunately we don't have a RegisterMap available in most of | |
1031 // the places we want to call this routine so we need to walk the | |
1032 // stack again to update the register map. | |
1033 if (map == NULL || !map->update_map()) { | |
1034 StackFrameStream sfs(thread, true); | |
1035 bool found = false; | |
1036 while (!found && !sfs.is_done()) { | |
1037 frame* cur = sfs.current(); | |
1038 sfs.next(); | |
1039 found = cur->id() == fr.id(); | |
1040 } | |
1041 assert(found, "frame to be deoptimized not found on target thread's stack"); | |
1042 map = sfs.register_map(); | |
1043 } | |
1044 | |
1045 vframe* vf = vframe::new_vframe(&fr, map, thread); | |
1046 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1047 // Revoke monitors' biases in all scopes | |
1048 while (!cvf->is_top()) { | |
1049 collect_monitors(cvf, objects_to_revoke); | |
1050 cvf = compiledVFrame::cast(cvf->sender()); | |
1051 } | |
1052 collect_monitors(cvf, objects_to_revoke); | |
1053 | |
1054 if (SafepointSynchronize::is_at_safepoint()) { | |
1055 BiasedLocking::revoke_at_safepoint(objects_to_revoke); | |
1056 } else { | |
1057 BiasedLocking::revoke(objects_to_revoke); | |
1058 } | |
1059 } | |
1060 | |
1061 | |
1062 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { | |
1063 if (!UseBiasedLocking) { | |
1064 return; | |
1065 } | |
1066 | |
1067 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); | |
1068 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); | |
1069 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { | |
1070 if (jt->has_last_Java_frame()) { | |
1071 StackFrameStream sfs(jt, true); | |
1072 while (!sfs.is_done()) { | |
1073 frame* cur = sfs.current(); | |
1074 if (cb->contains(cur->pc())) { | |
1075 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); | |
1076 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1077 // Revoke monitors' biases in all scopes | |
1078 while (!cvf->is_top()) { | |
1079 collect_monitors(cvf, objects_to_revoke); | |
1080 cvf = compiledVFrame::cast(cvf->sender()); | |
1081 } | |
1082 collect_monitors(cvf, objects_to_revoke); | |
1083 } | |
1084 sfs.next(); | |
1085 } | |
1086 } | |
1087 } | |
1088 BiasedLocking::revoke_at_safepoint(objects_to_revoke); | |
1089 } | |
1090 | |
1091 | |
1092 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { | |
1093 assert(fr.can_be_deoptimized(), "checking frame type"); | |
1094 | |
1095 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); | |
1096 | |
1097 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); | |
1098 | |
1099 // Patch the nmethod so that when execution returns to it we will | |
1100 // deopt the execution state and return to the interpreter. | |
1101 fr.deoptimize(thread); | |
1102 } | |
1103 | |
1104 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { | |
1105 // Deoptimize only if the frame comes from compile code. | |
1106 // Do not deoptimize the frame which is already patched | |
1107 // during the execution of the loops below. | |
1108 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { | |
1109 return; | |
1110 } | |
1111 ResourceMark rm; | |
1112 DeoptimizationMarker dm; | |
1113 if (UseBiasedLocking) { | |
1114 revoke_biases_of_monitors(thread, fr, map); | |
1115 } | |
1116 deoptimize_single_frame(thread, fr); | |
1117 | |
1118 } | |
1119 | |
1120 | |
1905
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1121 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1122 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1123 "can only deoptimize other thread at a safepoint"); |
0 | 1124 // Compute frame and register map based on thread and sp. |
1125 RegisterMap reg_map(thread, UseBiasedLocking); | |
1126 frame fr = thread->last_frame(); | |
1127 while (fr.id() != id) { | |
1128 fr = fr.sender(®_map); | |
1129 } | |
1130 deoptimize(thread, fr, ®_map); | |
1131 } | |
1132 | |
1133 | |
1905
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1134 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1135 if (thread == Thread::current()) { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1136 Deoptimization::deoptimize_frame_internal(thread, id); |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1137 } else { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1138 VM_DeoptimizeFrame deopt(thread, id); |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1139 VMThread::execute(&deopt); |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1140 } |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1141 } |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1142 |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1143 |
0 | 1144 // JVMTI PopFrame support |
1145 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) | |
1146 { | |
1147 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); | |
1148 } | |
1149 JRT_END | |
1150 | |
1151 | |
1152 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { | |
1153 // in case of an unresolved klass entry, load the class. | |
1154 if (constant_pool->tag_at(index).is_unresolved_klass()) { | |
1155 klassOop tk = constant_pool->klass_at(index, CHECK); | |
1156 return; | |
1157 } | |
1158 | |
1159 if (!constant_pool->tag_at(index).is_symbol()) return; | |
1160 | |
1161 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); | |
1162 symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); | |
1163 | |
1164 // class name? | |
1165 if (symbol->byte_at(0) != '(') { | |
1166 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); | |
1167 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); | |
1168 return; | |
1169 } | |
1170 | |
1171 // then it must be a signature! | |
1172 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { | |
1173 if (ss.is_object()) { | |
1174 symbolOop s = ss.as_symbol(CHECK); | |
1175 symbolHandle class_name (THREAD, s); | |
1176 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); | |
1177 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); | |
1178 } | |
1179 } | |
1180 } | |
1181 | |
1182 | |
1183 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { | |
1184 EXCEPTION_MARK; | |
1185 load_class_by_index(constant_pool, index, THREAD); | |
1186 if (HAS_PENDING_EXCEPTION) { | |
1187 // Exception happened during classloading. We ignore the exception here, since it | |
1188 // is going to be rethrown since the current activation is going to be deoptimzied and | |
1189 // the interpreter will re-execute the bytecode. | |
1190 CLEAR_PENDING_EXCEPTION; | |
1191 } | |
1192 } | |
1193 | |
1194 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { | |
1195 HandleMark hm; | |
1196 | |
1197 // uncommon_trap() is called at the beginning of the uncommon trap | |
1198 // handler. Note this fact before we start generating temporary frames | |
1199 // that can confuse an asynchronous stack walker. This counter is | |
1200 // decremented at the end of unpack_frames(). | |
1201 thread->inc_in_deopt_handler(); | |
1202 | |
1203 // We need to update the map if we have biased locking. | |
1204 RegisterMap reg_map(thread, UseBiasedLocking); | |
1205 frame stub_frame = thread->last_frame(); | |
1206 frame fr = stub_frame.sender(®_map); | |
1207 // Make sure the calling nmethod is not getting deoptimized and removed | |
1208 // before we are done with it. | |
1209 nmethodLocker nl(fr.pc()); | |
1210 | |
1211 { | |
1212 ResourceMark rm; | |
1213 | |
1214 // Revoke biases of any monitors in the frame to ensure we can migrate them | |
1215 revoke_biases_of_monitors(thread, fr, ®_map); | |
1216 | |
1217 DeoptReason reason = trap_request_reason(trap_request); | |
1218 DeoptAction action = trap_request_action(trap_request); | |
1219 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 | |
1220 | |
1221 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); | |
1222 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); | |
1223 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1224 | |
1225 nmethod* nm = cvf->code(); | |
1226 | |
1227 ScopeDesc* trap_scope = cvf->scope(); | |
1228 methodHandle trap_method = trap_scope->method(); | |
1229 int trap_bci = trap_scope->bci(); | |
1230 Bytecodes::Code trap_bc = Bytecode_at(trap_method->bcp_from(trap_bci))->java_code(); | |
1231 | |
1232 // Record this event in the histogram. | |
1233 gather_statistics(reason, action, trap_bc); | |
1234 | |
1235 // Ensure that we can record deopt. history: | |
1236 bool create_if_missing = ProfileTraps; | |
1237 | |
1238 methodDataHandle trap_mdo | |
1239 (THREAD, get_method_data(thread, trap_method, create_if_missing)); | |
1240 | |
1241 // Print a bunch of diagnostics, if requested. | |
1242 if (TraceDeoptimization || LogCompilation) { | |
1243 ResourceMark rm; | |
1244 ttyLocker ttyl; | |
1245 char buf[100]; | |
1246 if (xtty != NULL) { | |
1247 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", | |
1248 os::current_thread_id(), | |
1249 format_trap_request(buf, sizeof(buf), trap_request)); | |
1250 nm->log_identity(xtty); | |
1251 } | |
1252 symbolHandle class_name; | |
1253 bool unresolved = false; | |
1254 if (unloaded_class_index >= 0) { | |
1255 constantPoolHandle constants (THREAD, trap_method->constants()); | |
1256 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { | |
1257 class_name = symbolHandle(THREAD, | |
1258 constants->klass_name_at(unloaded_class_index)); | |
1259 unresolved = true; | |
1260 if (xtty != NULL) | |
1261 xtty->print(" unresolved='1'"); | |
1262 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { | |
1263 class_name = symbolHandle(THREAD, | |
1264 constants->symbol_at(unloaded_class_index)); | |
1265 } | |
1266 if (xtty != NULL) | |
1267 xtty->name(class_name); | |
1268 } | |
1269 if (xtty != NULL && trap_mdo.not_null()) { | |
1270 // Dump the relevant MDO state. | |
1271 // This is the deopt count for the current reason, any previous | |
1272 // reasons or recompiles seen at this point. | |
1273 int dcnt = trap_mdo->trap_count(reason); | |
1274 if (dcnt != 0) | |
1275 xtty->print(" count='%d'", dcnt); | |
1276 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); | |
1277 int dos = (pdata == NULL)? 0: pdata->trap_state(); | |
1278 if (dos != 0) { | |
1279 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); | |
1280 if (trap_state_is_recompiled(dos)) { | |
1281 int recnt2 = trap_mdo->overflow_recompile_count(); | |
1282 if (recnt2 != 0) | |
1283 xtty->print(" recompiles2='%d'", recnt2); | |
1284 } | |
1285 } | |
1286 } | |
1287 if (xtty != NULL) { | |
1288 xtty->stamp(); | |
1289 xtty->end_head(); | |
1290 } | |
1291 if (TraceDeoptimization) { // make noise on the tty | |
1292 tty->print("Uncommon trap occurred in"); | |
1293 nm->method()->print_short_name(tty); | |
1294 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", | |
1295 fr.pc(), | |
1296 (int) os::current_thread_id(), | |
1297 trap_reason_name(reason), | |
1298 trap_action_name(action), | |
1299 unloaded_class_index); | |
1300 if (class_name.not_null()) { | |
1301 tty->print(unresolved ? " unresolved class: " : " symbol: "); | |
1302 class_name->print_symbol_on(tty); | |
1303 } | |
1304 tty->cr(); | |
1305 } | |
1306 if (xtty != NULL) { | |
1307 // Log the precise location of the trap. | |
1308 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { | |
1309 xtty->begin_elem("jvms bci='%d'", sd->bci()); | |
1310 xtty->method(sd->method()); | |
1311 xtty->end_elem(); | |
1312 if (sd->is_top()) break; | |
1313 } | |
1314 xtty->tail("uncommon_trap"); | |
1315 } | |
1316 } | |
1317 // (End diagnostic printout.) | |
1318 | |
1319 // Load class if necessary | |
1320 if (unloaded_class_index >= 0) { | |
1321 constantPoolHandle constants(THREAD, trap_method->constants()); | |
1322 load_class_by_index(constants, unloaded_class_index); | |
1323 } | |
1324 | |
1325 // Flush the nmethod if necessary and desirable. | |
1326 // | |
1327 // We need to avoid situations where we are re-flushing the nmethod | |
1328 // because of a hot deoptimization site. Repeated flushes at the same | |
1329 // point need to be detected by the compiler and avoided. If the compiler | |
1330 // cannot avoid them (or has a bug and "refuses" to avoid them), this | |
1331 // module must take measures to avoid an infinite cycle of recompilation | |
1332 // and deoptimization. There are several such measures: | |
1333 // | |
1334 // 1. If a recompilation is ordered a second time at some site X | |
1335 // and for the same reason R, the action is adjusted to 'reinterpret', | |
1336 // to give the interpreter time to exercise the method more thoroughly. | |
1337 // If this happens, the method's overflow_recompile_count is incremented. | |
1338 // | |
1339 // 2. If the compiler fails to reduce the deoptimization rate, then | |
1340 // the method's overflow_recompile_count will begin to exceed the set | |
1341 // limit PerBytecodeRecompilationCutoff. If this happens, the action | |
1342 // is adjusted to 'make_not_compilable', and the method is abandoned | |
1343 // to the interpreter. This is a performance hit for hot methods, | |
1344 // but is better than a disastrous infinite cycle of recompilations. | |
1345 // (Actually, only the method containing the site X is abandoned.) | |
1346 // | |
1347 // 3. In parallel with the previous measures, if the total number of | |
1348 // recompilations of a method exceeds the much larger set limit | |
1349 // PerMethodRecompilationCutoff, the method is abandoned. | |
1350 // This should only happen if the method is very large and has | |
1351 // many "lukewarm" deoptimizations. The code which enforces this | |
1352 // limit is elsewhere (class nmethod, class methodOopDesc). | |
1353 // | |
1354 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance | |
1355 // to recompile at each bytecode independently of the per-BCI cutoff. | |
1356 // | |
1357 // The decision to update code is up to the compiler, and is encoded | |
1358 // in the Action_xxx code. If the compiler requests Action_none | |
1359 // no trap state is changed, no compiled code is changed, and the | |
1360 // computation suffers along in the interpreter. | |
1361 // | |
1362 // The other action codes specify various tactics for decompilation | |
1363 // and recompilation. Action_maybe_recompile is the loosest, and | |
1364 // allows the compiled code to stay around until enough traps are seen, | |
1365 // and until the compiler gets around to recompiling the trapping method. | |
1366 // | |
1367 // The other actions cause immediate removal of the present code. | |
1368 | |
1369 bool update_trap_state = true; | |
1370 bool make_not_entrant = false; | |
1371 bool make_not_compilable = false; | |
1783 | 1372 bool reprofile = false; |
0 | 1373 switch (action) { |
1374 case Action_none: | |
1375 // Keep the old code. | |
1376 update_trap_state = false; | |
1377 break; | |
1378 case Action_maybe_recompile: | |
1379 // Do not need to invalidate the present code, but we can | |
1380 // initiate another | |
1381 // Start compiler without (necessarily) invalidating the nmethod. | |
1382 // The system will tolerate the old code, but new code should be | |
1383 // generated when possible. | |
1384 break; | |
1385 case Action_reinterpret: | |
1386 // Go back into the interpreter for a while, and then consider | |
1387 // recompiling form scratch. | |
1388 make_not_entrant = true; | |
1389 // Reset invocation counter for outer most method. | |
1390 // This will allow the interpreter to exercise the bytecodes | |
1391 // for a while before recompiling. | |
1392 // By contrast, Action_make_not_entrant is immediate. | |
1393 // | |
1394 // Note that the compiler will track null_check, null_assert, | |
1395 // range_check, and class_check events and log them as if they | |
1396 // had been traps taken from compiled code. This will update | |
1397 // the MDO trap history so that the next compilation will | |
1398 // properly detect hot trap sites. | |
1783 | 1399 reprofile = true; |
0 | 1400 break; |
1401 case Action_make_not_entrant: | |
1402 // Request immediate recompilation, and get rid of the old code. | |
1403 // Make them not entrant, so next time they are called they get | |
1404 // recompiled. Unloaded classes are loaded now so recompile before next | |
1405 // time they are called. Same for uninitialized. The interpreter will | |
1406 // link the missing class, if any. | |
1407 make_not_entrant = true; | |
1408 break; | |
1409 case Action_make_not_compilable: | |
1410 // Give up on compiling this method at all. | |
1411 make_not_entrant = true; | |
1412 make_not_compilable = true; | |
1413 break; | |
1414 default: | |
1415 ShouldNotReachHere(); | |
1416 } | |
1417 | |
1418 // Setting +ProfileTraps fixes the following, on all platforms: | |
1419 // 4852688: ProfileInterpreter is off by default for ia64. The result is | |
1420 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the | |
1421 // recompile relies on a methodDataOop to record heroic opt failures. | |
1422 | |
1423 // Whether the interpreter is producing MDO data or not, we also need | |
1424 // to use the MDO to detect hot deoptimization points and control | |
1425 // aggressive optimization. | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1426 bool inc_recompile_count = false; |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1427 ProfileData* pdata = NULL; |
0 | 1428 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { |
1429 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); | |
1430 uint this_trap_count = 0; | |
1431 bool maybe_prior_trap = false; | |
1432 bool maybe_prior_recompile = false; | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1433 pdata = query_update_method_data(trap_mdo, trap_bci, reason, |
0 | 1434 //outputs: |
1435 this_trap_count, | |
1436 maybe_prior_trap, | |
1437 maybe_prior_recompile); | |
1438 // Because the interpreter also counts null, div0, range, and class | |
1439 // checks, these traps from compiled code are double-counted. | |
1440 // This is harmless; it just means that the PerXTrapLimit values | |
1441 // are in effect a little smaller than they look. | |
1442 | |
1443 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); | |
1444 if (per_bc_reason != Reason_none) { | |
1445 // Now take action based on the partially known per-BCI history. | |
1446 if (maybe_prior_trap | |
1447 && this_trap_count >= (uint)PerBytecodeTrapLimit) { | |
1448 // If there are too many traps at this BCI, force a recompile. | |
1449 // This will allow the compiler to see the limit overflow, and | |
1450 // take corrective action, if possible. The compiler generally | |
1451 // does not use the exact PerBytecodeTrapLimit value, but instead | |
1452 // changes its tactics if it sees any traps at all. This provides | |
1453 // a little hysteresis, delaying a recompile until a trap happens | |
1454 // several times. | |
1455 // | |
1456 // Actually, since there is only one bit of counter per BCI, | |
1457 // the possible per-BCI counts are {0,1,(per-method count)}. | |
1458 // This produces accurate results if in fact there is only | |
1459 // one hot trap site, but begins to get fuzzy if there are | |
1460 // many sites. For example, if there are ten sites each | |
1461 // trapping two or more times, they each get the blame for | |
1462 // all of their traps. | |
1463 make_not_entrant = true; | |
1464 } | |
1465 | |
1466 // Detect repeated recompilation at the same BCI, and enforce a limit. | |
1467 if (make_not_entrant && maybe_prior_recompile) { | |
1468 // More than one recompile at this point. | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1469 inc_recompile_count = maybe_prior_trap; |
0 | 1470 } |
1471 } else { | |
1472 // For reasons which are not recorded per-bytecode, we simply | |
1473 // force recompiles unconditionally. | |
1474 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) | |
1475 make_not_entrant = true; | |
1476 } | |
1477 | |
1478 // Go back to the compiler if there are too many traps in this method. | |
1479 if (this_trap_count >= (uint)PerMethodTrapLimit) { | |
1480 // If there are too many traps in this method, force a recompile. | |
1481 // This will allow the compiler to see the limit overflow, and | |
1482 // take corrective action, if possible. | |
1483 // (This condition is an unlikely backstop only, because the | |
1484 // PerBytecodeTrapLimit is more likely to take effect first, | |
1485 // if it is applicable.) | |
1486 make_not_entrant = true; | |
1487 } | |
1488 | |
1489 // Here's more hysteresis: If there has been a recompile at | |
1490 // this trap point already, run the method in the interpreter | |
1491 // for a while to exercise it more thoroughly. | |
1492 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { | |
1783 | 1493 reprofile = true; |
0 | 1494 } |
1495 | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1496 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1497 |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1498 // Take requested actions on the method: |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1499 |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1500 // Recompile |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1501 if (make_not_entrant) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1502 if (!nm->make_not_entrant()) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1503 return; // the call did not change nmethod's state |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1504 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1505 |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1506 if (pdata != NULL) { |
0 | 1507 // Record the recompilation event, if any. |
1508 int tstate0 = pdata->trap_state(); | |
1509 int tstate1 = trap_state_set_recompiled(tstate0, true); | |
1510 if (tstate1 != tstate0) | |
1511 pdata->set_trap_state(tstate1); | |
1512 } | |
1513 } | |
1514 | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1515 if (inc_recompile_count) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1516 trap_mdo->inc_overflow_recompile_count(); |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1517 if ((uint)trap_mdo->overflow_recompile_count() > |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1518 (uint)PerBytecodeRecompilationCutoff) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1519 // Give up on the method containing the bad BCI. |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1520 if (trap_method() == nm->method()) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1521 make_not_compilable = true; |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1522 } else { |
1783 | 1523 trap_method->set_not_compilable(CompLevel_full_optimization); |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1524 // But give grace to the enclosing nm->method(). |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1525 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1526 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1527 } |
0 | 1528 |
1783 | 1529 // Reprofile |
1530 if (reprofile) { | |
1531 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); | |
0 | 1532 } |
1533 | |
1534 // Give up compiling | |
1783 | 1535 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { |
0 | 1536 assert(make_not_entrant, "consistent"); |
1783 | 1537 nm->method()->set_not_compilable(CompLevel_full_optimization); |
0 | 1538 } |
1539 | |
1540 } // Free marked resources | |
1541 | |
1542 } | |
1543 JRT_END | |
1544 | |
1545 methodDataOop | |
1546 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, | |
1547 bool create_if_missing) { | |
1548 Thread* THREAD = thread; | |
1549 methodDataOop mdo = m()->method_data(); | |
1550 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { | |
1551 // Build an MDO. Ignore errors like OutOfMemory; | |
1552 // that simply means we won't have an MDO to update. | |
1553 methodOopDesc::build_interpreter_method_data(m, THREAD); | |
1554 if (HAS_PENDING_EXCEPTION) { | |
1555 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); | |
1556 CLEAR_PENDING_EXCEPTION; | |
1557 } | |
1558 mdo = m()->method_data(); | |
1559 } | |
1560 return mdo; | |
1561 } | |
1562 | |
1563 ProfileData* | |
1564 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, | |
1565 int trap_bci, | |
1566 Deoptimization::DeoptReason reason, | |
1567 //outputs: | |
1568 uint& ret_this_trap_count, | |
1569 bool& ret_maybe_prior_trap, | |
1570 bool& ret_maybe_prior_recompile) { | |
1571 uint prior_trap_count = trap_mdo->trap_count(reason); | |
1572 uint this_trap_count = trap_mdo->inc_trap_count(reason); | |
1573 | |
1574 // If the runtime cannot find a place to store trap history, | |
1575 // it is estimated based on the general condition of the method. | |
1576 // If the method has ever been recompiled, or has ever incurred | |
1577 // a trap with the present reason , then this BCI is assumed | |
1578 // (pessimistically) to be the culprit. | |
1579 bool maybe_prior_trap = (prior_trap_count != 0); | |
1580 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); | |
1581 ProfileData* pdata = NULL; | |
1582 | |
1583 | |
1584 // For reasons which are recorded per bytecode, we check per-BCI data. | |
1585 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); | |
1586 if (per_bc_reason != Reason_none) { | |
1587 // Find the profile data for this BCI. If there isn't one, | |
1588 // try to allocate one from the MDO's set of spares. | |
1589 // This will let us detect a repeated trap at this point. | |
1590 pdata = trap_mdo->allocate_bci_to_data(trap_bci); | |
1591 | |
1592 if (pdata != NULL) { | |
1593 // Query the trap state of this profile datum. | |
1594 int tstate0 = pdata->trap_state(); | |
1595 if (!trap_state_has_reason(tstate0, per_bc_reason)) | |
1596 maybe_prior_trap = false; | |
1597 if (!trap_state_is_recompiled(tstate0)) | |
1598 maybe_prior_recompile = false; | |
1599 | |
1600 // Update the trap state of this profile datum. | |
1601 int tstate1 = tstate0; | |
1602 // Record the reason. | |
1603 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); | |
1604 // Store the updated state on the MDO, for next time. | |
1605 if (tstate1 != tstate0) | |
1606 pdata->set_trap_state(tstate1); | |
1607 } else { | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1608 if (LogCompilation && xtty != NULL) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1609 ttyLocker ttyl; |
0 | 1610 // Missing MDP? Leave a small complaint in the log. |
1611 xtty->elem("missing_mdp bci='%d'", trap_bci); | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1612 } |
0 | 1613 } |
1614 } | |
1615 | |
1616 // Return results: | |
1617 ret_this_trap_count = this_trap_count; | |
1618 ret_maybe_prior_trap = maybe_prior_trap; | |
1619 ret_maybe_prior_recompile = maybe_prior_recompile; | |
1620 return pdata; | |
1621 } | |
1622 | |
1623 void | |
1624 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { | |
1625 ResourceMark rm; | |
1626 // Ignored outputs: | |
1627 uint ignore_this_trap_count; | |
1628 bool ignore_maybe_prior_trap; | |
1629 bool ignore_maybe_prior_recompile; | |
1630 query_update_method_data(trap_mdo, trap_bci, | |
1631 (DeoptReason)reason, | |
1632 ignore_this_trap_count, | |
1633 ignore_maybe_prior_trap, | |
1634 ignore_maybe_prior_recompile); | |
1635 } | |
1636 | |
1637 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { | |
1638 | |
1639 // Still in Java no safepoints | |
1640 { | |
1641 // This enters VM and may safepoint | |
1642 uncommon_trap_inner(thread, trap_request); | |
1643 } | |
1644 return fetch_unroll_info_helper(thread); | |
1645 } | |
1646 | |
1647 // Local derived constants. | |
1648 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. | |
1649 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; | |
1650 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; | |
1651 | |
1652 //---------------------------trap_state_reason--------------------------------- | |
1653 Deoptimization::DeoptReason | |
1654 Deoptimization::trap_state_reason(int trap_state) { | |
1655 // This assert provides the link between the width of DataLayout::trap_bits | |
1656 // and the encoding of "recorded" reasons. It ensures there are enough | |
1657 // bits to store all needed reasons in the per-BCI MDO profile. | |
1658 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); | |
1659 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1660 trap_state -= recompile_bit; | |
1661 if (trap_state == DS_REASON_MASK) { | |
1662 return Reason_many; | |
1663 } else { | |
1664 assert((int)Reason_none == 0, "state=0 => Reason_none"); | |
1665 return (DeoptReason)trap_state; | |
1666 } | |
1667 } | |
1668 //-------------------------trap_state_has_reason------------------------------- | |
1669 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { | |
1670 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); | |
1671 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); | |
1672 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1673 trap_state -= recompile_bit; | |
1674 if (trap_state == DS_REASON_MASK) { | |
1675 return -1; // true, unspecifically (bottom of state lattice) | |
1676 } else if (trap_state == reason) { | |
1677 return 1; // true, definitely | |
1678 } else if (trap_state == 0) { | |
1679 return 0; // false, definitely (top of state lattice) | |
1680 } else { | |
1681 return 0; // false, definitely | |
1682 } | |
1683 } | |
1684 //-------------------------trap_state_add_reason------------------------------- | |
1685 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { | |
1686 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); | |
1687 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1688 trap_state -= recompile_bit; | |
1689 if (trap_state == DS_REASON_MASK) { | |
1690 return trap_state + recompile_bit; // already at state lattice bottom | |
1691 } else if (trap_state == reason) { | |
1692 return trap_state + recompile_bit; // the condition is already true | |
1693 } else if (trap_state == 0) { | |
1694 return reason + recompile_bit; // no condition has yet been true | |
1695 } else { | |
1696 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom | |
1697 } | |
1698 } | |
1699 //-----------------------trap_state_is_recompiled------------------------------ | |
1700 bool Deoptimization::trap_state_is_recompiled(int trap_state) { | |
1701 return (trap_state & DS_RECOMPILE_BIT) != 0; | |
1702 } | |
1703 //-----------------------trap_state_set_recompiled----------------------------- | |
1704 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { | |
1705 if (z) return trap_state | DS_RECOMPILE_BIT; | |
1706 else return trap_state & ~DS_RECOMPILE_BIT; | |
1707 } | |
1708 //---------------------------format_trap_state--------------------------------- | |
1709 // This is used for debugging and diagnostics, including hotspot.log output. | |
1710 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, | |
1711 int trap_state) { | |
1712 DeoptReason reason = trap_state_reason(trap_state); | |
1713 bool recomp_flag = trap_state_is_recompiled(trap_state); | |
1714 // Re-encode the state from its decoded components. | |
1715 int decoded_state = 0; | |
1716 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) | |
1717 decoded_state = trap_state_add_reason(decoded_state, reason); | |
1718 if (recomp_flag) | |
1719 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); | |
1720 // If the state re-encodes properly, format it symbolically. | |
1721 // Because this routine is used for debugging and diagnostics, | |
1722 // be robust even if the state is a strange value. | |
1723 size_t len; | |
1724 if (decoded_state != trap_state) { | |
1725 // Random buggy state that doesn't decode?? | |
1726 len = jio_snprintf(buf, buflen, "#%d", trap_state); | |
1727 } else { | |
1728 len = jio_snprintf(buf, buflen, "%s%s", | |
1729 trap_reason_name(reason), | |
1730 recomp_flag ? " recompiled" : ""); | |
1731 } | |
1732 if (len >= buflen) | |
1733 buf[buflen-1] = '\0'; | |
1734 return buf; | |
1735 } | |
1736 | |
1737 | |
1738 //--------------------------------statics-------------------------------------- | |
1739 Deoptimization::DeoptAction Deoptimization::_unloaded_action | |
1740 = Deoptimization::Action_reinterpret; | |
1741 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { | |
1742 // Note: Keep this in sync. with enum DeoptReason. | |
1743 "none", | |
1744 "null_check", | |
1745 "null_assert", | |
1746 "range_check", | |
1747 "class_check", | |
1748 "array_check", | |
1749 "intrinsic", | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1750 "bimorphic", |
0 | 1751 "unloaded", |
1752 "uninitialized", | |
1753 "unreached", | |
1754 "unhandled", | |
1755 "constraint", | |
1756 "div0_check", | |
1172 | 1757 "age", |
1758 "predicate" | |
0 | 1759 }; |
1760 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { | |
1761 // Note: Keep this in sync. with enum DeoptAction. | |
1762 "none", | |
1763 "maybe_recompile", | |
1764 "reinterpret", | |
1765 "make_not_entrant", | |
1766 "make_not_compilable" | |
1767 }; | |
1768 | |
1769 const char* Deoptimization::trap_reason_name(int reason) { | |
1770 if (reason == Reason_many) return "many"; | |
1771 if ((uint)reason < Reason_LIMIT) | |
1772 return _trap_reason_name[reason]; | |
1773 static char buf[20]; | |
1774 sprintf(buf, "reason%d", reason); | |
1775 return buf; | |
1776 } | |
1777 const char* Deoptimization::trap_action_name(int action) { | |
1778 if ((uint)action < Action_LIMIT) | |
1779 return _trap_action_name[action]; | |
1780 static char buf[20]; | |
1781 sprintf(buf, "action%d", action); | |
1782 return buf; | |
1783 } | |
1784 | |
1785 // This is used for debugging and diagnostics, including hotspot.log output. | |
1786 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, | |
1787 int trap_request) { | |
1788 jint unloaded_class_index = trap_request_index(trap_request); | |
1789 const char* reason = trap_reason_name(trap_request_reason(trap_request)); | |
1790 const char* action = trap_action_name(trap_request_action(trap_request)); | |
1791 size_t len; | |
1792 if (unloaded_class_index < 0) { | |
1793 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", | |
1794 reason, action); | |
1795 } else { | |
1796 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", | |
1797 reason, action, unloaded_class_index); | |
1798 } | |
1799 if (len >= buflen) | |
1800 buf[buflen-1] = '\0'; | |
1801 return buf; | |
1802 } | |
1803 | |
1804 juint Deoptimization::_deoptimization_hist | |
1805 [Deoptimization::Reason_LIMIT] | |
1806 [1 + Deoptimization::Action_LIMIT] | |
1807 [Deoptimization::BC_CASE_LIMIT] | |
1808 = {0}; | |
1809 | |
1810 enum { | |
1811 LSB_BITS = 8, | |
1812 LSB_MASK = right_n_bits(LSB_BITS) | |
1813 }; | |
1814 | |
1815 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, | |
1816 Bytecodes::Code bc) { | |
1817 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); | |
1818 assert(action >= 0 && action < Action_LIMIT, "oob"); | |
1819 _deoptimization_hist[Reason_none][0][0] += 1; // total | |
1820 _deoptimization_hist[reason][0][0] += 1; // per-reason total | |
1821 juint* cases = _deoptimization_hist[reason][1+action]; | |
1822 juint* bc_counter_addr = NULL; | |
1823 juint bc_counter = 0; | |
1824 // Look for an unused counter, or an exact match to this BC. | |
1825 if (bc != Bytecodes::_illegal) { | |
1826 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { | |
1827 juint* counter_addr = &cases[bc_case]; | |
1828 juint counter = *counter_addr; | |
1829 if ((counter == 0 && bc_counter_addr == NULL) | |
1830 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { | |
1831 // this counter is either free or is already devoted to this BC | |
1832 bc_counter_addr = counter_addr; | |
1833 bc_counter = counter | bc; | |
1834 } | |
1835 } | |
1836 } | |
1837 if (bc_counter_addr == NULL) { | |
1838 // Overflow, or no given bytecode. | |
1839 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; | |
1840 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB | |
1841 } | |
1842 *bc_counter_addr = bc_counter + (1 << LSB_BITS); | |
1843 } | |
1844 | |
1845 jint Deoptimization::total_deoptimization_count() { | |
1846 return _deoptimization_hist[Reason_none][0][0]; | |
1847 } | |
1848 | |
1849 jint Deoptimization::deoptimization_count(DeoptReason reason) { | |
1850 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); | |
1851 return _deoptimization_hist[reason][0][0]; | |
1852 } | |
1853 | |
1854 void Deoptimization::print_statistics() { | |
1855 juint total = total_deoptimization_count(); | |
1856 juint account = total; | |
1857 if (total != 0) { | |
1858 ttyLocker ttyl; | |
1859 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); | |
1860 tty->print_cr("Deoptimization traps recorded:"); | |
1861 #define PRINT_STAT_LINE(name, r) \ | |
1862 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); | |
1863 PRINT_STAT_LINE("total", total); | |
1864 // For each non-zero entry in the histogram, print the reason, | |
1865 // the action, and (if specifically known) the type of bytecode. | |
1866 for (int reason = 0; reason < Reason_LIMIT; reason++) { | |
1867 for (int action = 0; action < Action_LIMIT; action++) { | |
1868 juint* cases = _deoptimization_hist[reason][1+action]; | |
1869 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { | |
1870 juint counter = cases[bc_case]; | |
1871 if (counter != 0) { | |
1872 char name[1*K]; | |
1873 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); | |
1874 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) | |
1875 bc = Bytecodes::_illegal; | |
1876 sprintf(name, "%s/%s/%s", | |
1877 trap_reason_name(reason), | |
1878 trap_action_name(action), | |
1879 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); | |
1880 juint r = counter >> LSB_BITS; | |
1881 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); | |
1882 account -= r; | |
1883 } | |
1884 } | |
1885 } | |
1886 } | |
1887 if (account != 0) { | |
1888 PRINT_STAT_LINE("unaccounted", account); | |
1889 } | |
1890 #undef PRINT_STAT_LINE | |
1891 if (xtty != NULL) xtty->tail("statistics"); | |
1892 } | |
1893 } |