Mercurial > hg > graal-jvmci-8
annotate src/share/vm/runtime/deoptimization.cpp @ 2174:234761c55641
6608385: G1: need to support parallel reference processing
Summary: Implement support for ParallelRefProcEnabled in the reference processing that takes place at the end of G1 concurrent marking.
Reviewed-by: tonyp, ysr
author | johnc |
---|---|
date | Tue, 25 Jan 2011 10:56:22 -0800 |
parents | 8012aa3ccede |
children | 3582bf76420e |
rev | line source |
---|---|
0 | 1 /* |
2142 | 2 * Copyright (c) 1997, 2011, 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 | |
166 // Allocate our special deoptimization ResourceMark | |
167 DeoptResourceMark* dmark = new DeoptResourceMark(thread); | |
168 assert(thread->deopt_mark() == NULL, "Pending deopt!"); | |
169 thread->set_deopt_mark(dmark); | |
170 | |
171 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect | |
172 RegisterMap map(thread, true); | |
173 RegisterMap dummy_map(thread, false); | |
174 // Now get the deoptee with a valid map | |
175 frame deoptee = stub_frame.sender(&map); | |
1814
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
176 // Set the deoptee nmethod |
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
177 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); |
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
178 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); |
0 | 179 |
180 // Create a growable array of VFrames where each VFrame represents an inlined | |
181 // Java frame. This storage is allocated with the usual system arena. | |
182 assert(deoptee.is_compiled_frame(), "Wrong frame type"); | |
183 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); | |
184 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); | |
185 while (!vf->is_top()) { | |
186 assert(vf->is_compiled_frame(), "Wrong frame type"); | |
187 chunk->push(compiledVFrame::cast(vf)); | |
188 vf = vf->sender(); | |
189 } | |
190 assert(vf->is_compiled_frame(), "Wrong frame type"); | |
191 chunk->push(compiledVFrame::cast(vf)); | |
192 | |
193 #ifdef COMPILER2 | |
194 // Reallocate the non-escaping objects and restore their fields. Then | |
195 // relock objects if synchronization on them was eliminated. | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
196 if (DoEscapeAnalysis) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
197 if (EliminateAllocations) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
198 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
|
199 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
|
200 |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
201 // 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
|
202 // 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
|
203 // 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
|
204 // 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
|
205 // 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
|
206 // 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
|
207 // 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
|
208 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
|
209 Handle return_value; |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
210 if (save_oop_result) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
211 // 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
|
212 // 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
|
213 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
|
214 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
|
215 return_value = Handle(thread, result); |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
216 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
|
217 if (TraceDeoptimization) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
218 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
|
219 } |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
220 } |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
221 bool reallocated = false; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
222 if (objects != NULL) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
223 JRT_BLOCK |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
224 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
225 JRT_END |
0 | 226 } |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
227 if (reallocated) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
228 reassign_fields(&deoptee, &map, objects); |
0 | 229 #ifndef PRODUCT |
230 if (TraceDeoptimization) { | |
231 ttyLocker ttyl; | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
232 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
|
233 print_objects(objects); |
1253
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
234 } |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
235 #endif |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
236 } |
1253
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
237 if (save_oop_result) { |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
238 // Restore result. |
f70b0d9ab095
6910618: C2: Error: assert(d->is_oop(),"JVM_ArrayCopy: dst not an oop")
kvn
parents:
1206
diff
changeset
|
239 deoptee.set_saved_oop_result(&map, return_value()); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
240 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
241 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
242 if (EliminateLocks) { |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
243 #ifndef PRODUCT |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
244 bool first = true; |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
245 #endif |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
246 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
|
247 compiledVFrame* cvf = chunk->at(i); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
248 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
|
249 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
250 if (monitors->is_nonempty()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
251 relock_objects(monitors, thread); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
252 #ifndef PRODUCT |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
253 if (TraceDeoptimization) { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
254 ttyLocker ttyl; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
255 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
|
256 MonitorInfo* mi = monitors->at(j); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
257 if (mi->eliminated()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
258 if (first) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
259 first = false; |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
260 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
|
261 } |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
262 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
|
263 } |
0 | 264 } |
265 } | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
266 #endif |
0 | 267 } |
268 } | |
269 } | |
270 } | |
271 #endif // COMPILER2 | |
272 // Ensure that no safepoint is taken after pointers have been stored | |
273 // in fields of rematerialized objects. If a safepoint occurs from here on | |
274 // out the java state residing in the vframeArray will be missed. | |
275 No_Safepoint_Verifier no_safepoint; | |
276 | |
277 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); | |
278 | |
279 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; | |
280 thread->set_vframe_array_head(array); | |
281 | |
282 // Now that the vframeArray has been created if we have any deferred local writes | |
283 // added by jvmti then we can free up that structure as the data is now in the | |
284 // vframeArray | |
285 | |
286 if (thread->deferred_locals() != NULL) { | |
287 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); | |
288 int i = 0; | |
289 do { | |
290 // Because of inlining we could have multiple vframes for a single frame | |
291 // and several of the vframes could have deferred writes. Find them all. | |
292 if (list->at(i)->id() == array->original().id()) { | |
293 jvmtiDeferredLocalVariableSet* dlv = list->at(i); | |
294 list->remove_at(i); | |
295 // individual jvmtiDeferredLocalVariableSet are CHeapObj's | |
296 delete dlv; | |
297 } else { | |
298 i++; | |
299 } | |
300 } while ( i < list->length() ); | |
301 if (list->length() == 0) { | |
302 thread->set_deferred_locals(NULL); | |
303 // free the list and elements back to C heap. | |
304 delete list; | |
305 } | |
306 | |
307 } | |
308 | |
1692 | 309 #ifndef SHARK |
0 | 310 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. |
311 CodeBlob* cb = stub_frame.cb(); | |
312 // Verify we have the right vframeArray | |
313 assert(cb->frame_size() >= 0, "Unexpected frame size"); | |
314 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); | |
315 | |
1204 | 316 // If the deopt call site is a MethodHandle invoke call site we have |
317 // to adjust the unpack_sp. | |
318 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); | |
319 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) | |
320 unpack_sp = deoptee.unextended_sp(); | |
321 | |
0 | 322 #ifdef ASSERT |
323 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); | |
324 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); | |
325 #endif | |
1692 | 326 #else |
327 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); | |
328 #endif // !SHARK | |
329 | |
0 | 330 // This is a guarantee instead of an assert because if vframe doesn't match |
331 // we will unpack the wrong deoptimized frame and wind up in strange places | |
332 // where it will be very difficult to figure out what went wrong. Better | |
333 // to die an early death here than some very obscure death later when the | |
334 // trail is cold. | |
335 // Note: on ia64 this guarantee can be fooled by frames with no memory stack | |
336 // in that it will fail to detect a problem when there is one. This needs | |
337 // more work in tiger timeframe. | |
338 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); | |
339 | |
340 int number_of_frames = array->frames(); | |
341 | |
342 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost | |
343 // virtual activation, which is the reverse of the elements in the vframes array. | |
344 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); | |
345 // +1 because we always have an interpreter return address for the final slot. | |
346 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); | |
347 int callee_parameters = 0; | |
348 int callee_locals = 0; | |
349 int popframe_extra_args = 0; | |
350 // Create an interpreter return address for the stub to use as its return | |
351 // address so the skeletal frames are perfectly walkable | |
352 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); | |
353 | |
354 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost | |
355 // activation be put back on the expression stack of the caller for reexecution | |
356 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { | |
357 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); | |
358 } | |
359 | |
360 // | |
361 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) | |
362 // frame_sizes/frame_pcs[1] next oldest frame (int) | |
363 // frame_sizes/frame_pcs[n] youngest frame (int) | |
364 // | |
365 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame | |
366 // owns the space for the return address to it's caller). Confusing ain't it. | |
367 // | |
368 // The vframe array can address vframes with indices running from | |
369 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. | |
370 // When we create the skeletal frames we need the oldest frame to be in the zero slot | |
371 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. | |
372 // so things look a little strange in this loop. | |
373 // | |
374 for (int index = 0; index < array->frames(); index++ ) { | |
375 // frame[number_of_frames - 1 ] = on_stack_size(youngest) | |
376 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) | |
377 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) | |
378 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, | |
379 callee_locals, | |
380 index == 0, | |
381 popframe_extra_args); | |
382 // This pc doesn't have to be perfect just good enough to identify the frame | |
383 // as interpreted so the skeleton frame will be walkable | |
384 // The correct pc will be set when the skeleton frame is completely filled out | |
385 // The final pc we store in the loop is wrong and will be overwritten below | |
386 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; | |
387 | |
388 callee_parameters = array->element(index)->method()->size_of_parameters(); | |
389 callee_locals = array->element(index)->method()->max_locals(); | |
390 popframe_extra_args = 0; | |
391 } | |
392 | |
393 // Compute whether the root vframe returns a float or double value. | |
394 BasicType return_type; | |
395 { | |
396 HandleMark hm; | |
397 methodHandle method(thread, array->element(0)->method()); | |
2142 | 398 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); |
399 return_type = invoke.is_valid() ? invoke.result_type(thread) : T_ILLEGAL; | |
0 | 400 } |
401 | |
402 // Compute information for handling adapters and adjusting the frame size of the caller. | |
403 int caller_adjustment = 0; | |
404 | |
405 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized | |
406 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather | |
407 // than simply use array->sender.pc(). This requires us to walk the current set of frames | |
408 // | |
409 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame | |
410 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller | |
411 | |
412 // Compute the amount the oldest interpreter frame will have to adjust | |
413 // its caller's stack by. If the caller is a compiled frame then | |
414 // we pretend that the callee has no parameters so that the | |
415 // extension counts for the full amount of locals and not just | |
416 // locals-parms. This is because without a c2i adapter the parm | |
417 // area as created by the compiled frame will not be usable by | |
418 // the interpreter. (Depending on the calling convention there | |
419 // may not even be enough space). | |
420 | |
421 // QQQ I'd rather see this pushed down into last_frame_adjust | |
422 // and have it take the sender (aka caller). | |
423 | |
424 if (deopt_sender.is_compiled_frame()) { | |
425 caller_adjustment = last_frame_adjust(0, callee_locals); | |
426 } else if (callee_locals > callee_parameters) { | |
427 // The caller frame may need extending to accommodate | |
428 // non-parameter locals of the first unpacked interpreted frame. | |
429 // Compute that adjustment. | |
430 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); | |
431 } | |
432 | |
433 | |
434 // If the sender is deoptimized the we must retrieve the address of the handler | |
435 // since the frame will "magically" show the original pc before the deopt | |
436 // and we'd undo the deopt. | |
437 | |
438 frame_pcs[0] = deopt_sender.raw_pc(); | |
439 | |
1692 | 440 #ifndef SHARK |
0 | 441 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); |
1692 | 442 #endif // SHARK |
0 | 443 |
444 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, | |
445 caller_adjustment * BytesPerWord, | |
446 number_of_frames, | |
447 frame_sizes, | |
448 frame_pcs, | |
449 return_type); | |
450 #if defined(IA32) || defined(AMD64) | |
451 // We need a way to pass fp to the unpacking code so the skeletal frames | |
452 // come out correct. This is only needed for x86 because of c2 using ebp | |
453 // as an allocatable register. So this update is useless (and harmless) | |
454 // on the other platforms. It would be nice to do this in a different | |
455 // way but even the old style deoptimization had a problem with deriving | |
456 // this value. NEEDS_CLEANUP | |
457 // Note: now that c1 is using c2's deopt blob we must do this on all | |
458 // x86 based platforms | |
459 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); | |
460 *fp_addr = array->sender().fp(); // was adapter_caller | |
461 #endif /* IA32 || AMD64 */ | |
462 | |
463 if (array->frames() > 1) { | |
464 if (VerifyStack && TraceDeoptimization) { | |
465 tty->print_cr("Deoptimizing method containing inlining"); | |
466 } | |
467 } | |
468 | |
469 array->set_unroll_block(info); | |
470 return info; | |
471 } | |
472 | |
473 // Called to cleanup deoptimization data structures in normal case | |
474 // after unpacking to stack and when stack overflow error occurs | |
475 void Deoptimization::cleanup_deopt_info(JavaThread *thread, | |
476 vframeArray *array) { | |
477 | |
478 // Get array if coming from exception | |
479 if (array == NULL) { | |
480 array = thread->vframe_array_head(); | |
481 } | |
482 thread->set_vframe_array_head(NULL); | |
483 | |
484 // Free the previous UnrollBlock | |
485 vframeArray* old_array = thread->vframe_array_last(); | |
486 thread->set_vframe_array_last(array); | |
487 | |
488 if (old_array != NULL) { | |
489 UnrollBlock* old_info = old_array->unroll_block(); | |
490 old_array->set_unroll_block(NULL); | |
491 delete old_info; | |
492 delete old_array; | |
493 } | |
494 | |
495 // Deallocate any resource creating in this routine and any ResourceObjs allocated | |
496 // inside the vframeArray (StackValueCollections) | |
497 | |
498 delete thread->deopt_mark(); | |
499 thread->set_deopt_mark(NULL); | |
1814
fd5d4527cdf5
6986270: guarantee(*bcp != Bytecodes::_monitorenter || exec_mode != Deoptimization::Unpack_exception) fails
iveresov
parents:
1783
diff
changeset
|
500 thread->set_deopt_nmethod(NULL); |
0 | 501 |
502 | |
503 if (JvmtiExport::can_pop_frame()) { | |
504 #ifndef CC_INTERP | |
505 // Regardless of whether we entered this routine with the pending | |
506 // popframe condition bit set, we should always clear it now | |
507 thread->clear_popframe_condition(); | |
508 #else | |
509 // C++ interpeter will clear has_pending_popframe when it enters | |
510 // with method_resume. For deopt_resume2 we clear it now. | |
511 if (thread->popframe_forcing_deopt_reexecution()) | |
512 thread->clear_popframe_condition(); | |
513 #endif /* CC_INTERP */ | |
514 } | |
515 | |
516 // unpack_frames() is called at the end of the deoptimization handler | |
517 // and (in C2) at the end of the uncommon trap handler. Note this fact | |
518 // so that an asynchronous stack walker can work again. This counter is | |
519 // incremented at the beginning of fetch_unroll_info() and (in C2) at | |
520 // the beginning of uncommon_trap(). | |
521 thread->dec_in_deopt_handler(); | |
522 } | |
523 | |
524 | |
525 // Return BasicType of value being returned | |
526 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) | |
527 | |
528 // We are already active int he special DeoptResourceMark any ResourceObj's we | |
529 // allocate will be freed at the end of the routine. | |
530 | |
531 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, | |
532 // but makes the entry a little slower. There is however a little dance we have to | |
533 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro | |
534 ResetNoHandleMark rnhm; // No-op in release/product versions | |
535 HandleMark hm; | |
536 | |
537 frame stub_frame = thread->last_frame(); | |
538 | |
539 // Since the frame to unpack is the top frame of this thread, the vframe_array_head | |
540 // must point to the vframeArray for the unpack frame. | |
541 vframeArray* array = thread->vframe_array_head(); | |
542 | |
543 #ifndef PRODUCT | |
544 if (TraceDeoptimization) { | |
545 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); | |
546 } | |
547 #endif | |
548 | |
549 UnrollBlock* info = array->unroll_block(); | |
550 | |
551 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. | |
552 array->unpack_to_stack(stub_frame, exec_mode); | |
553 | |
554 BasicType bt = info->return_type(); | |
555 | |
556 // If we have an exception pending, claim that the return type is an oop | |
557 // so the deopt_blob does not overwrite the exception_oop. | |
558 | |
559 if (exec_mode == Unpack_exception) | |
560 bt = T_OBJECT; | |
561 | |
562 // Cleanup thread deopt data | |
563 cleanup_deopt_info(thread, array); | |
564 | |
565 #ifndef PRODUCT | |
566 if (VerifyStack) { | |
567 ResourceMark res_mark; | |
568 | |
569 // Verify that the just-unpacked frames match the interpreter's | |
570 // notions of expression stack and locals | |
571 vframeArray* cur_array = thread->vframe_array_last(); | |
572 RegisterMap rm(thread, false); | |
573 rm.set_include_argument_oops(false); | |
574 bool is_top_frame = true; | |
575 int callee_size_of_parameters = 0; | |
576 int callee_max_locals = 0; | |
577 for (int i = 0; i < cur_array->frames(); i++) { | |
578 vframeArrayElement* el = cur_array->element(i); | |
579 frame* iframe = el->iframe(); | |
580 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); | |
581 | |
582 // Get the oop map for this bci | |
583 InterpreterOopMap mask; | |
584 int cur_invoke_parameter_size = 0; | |
585 bool try_next_mask = false; | |
586 int next_mask_expression_stack_size = -1; | |
587 int top_frame_expression_stack_adjustment = 0; | |
588 methodHandle mh(thread, iframe->interpreter_frame_method()); | |
589 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); | |
590 BytecodeStream str(mh); | |
591 str.set_start(iframe->interpreter_frame_bci()); | |
592 int max_bci = mh->code_size(); | |
593 // Get to the next bytecode if possible | |
594 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); | |
595 // Check to see if we can grab the number of outgoing arguments | |
596 // at an uncommon trap for an invoke (where the compiler | |
597 // generates debug info before the invoke has executed) | |
598 Bytecodes::Code cur_code = str.next(); | |
599 if (cur_code == Bytecodes::_invokevirtual || | |
600 cur_code == Bytecodes::_invokespecial || | |
601 cur_code == Bytecodes::_invokestatic || | |
602 cur_code == Bytecodes::_invokeinterface) { | |
2142 | 603 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); |
604 symbolHandle signature(thread, invoke.signature()); | |
0 | 605 ArgumentSizeComputer asc(signature); |
606 cur_invoke_parameter_size = asc.size(); | |
607 if (cur_code != Bytecodes::_invokestatic) { | |
608 // Add in receiver | |
609 ++cur_invoke_parameter_size; | |
610 } | |
611 } | |
612 if (str.bci() < max_bci) { | |
613 Bytecodes::Code bc = str.next(); | |
614 if (bc >= 0) { | |
615 // The interpreter oop map generator reports results before | |
616 // the current bytecode has executed except in the case of | |
617 // calls. It seems to be hard to tell whether the compiler | |
618 // has emitted debug information matching the "state before" | |
619 // a given bytecode or the state after, so we try both | |
620 switch (cur_code) { | |
621 case Bytecodes::_invokevirtual: | |
622 case Bytecodes::_invokespecial: | |
623 case Bytecodes::_invokestatic: | |
624 case Bytecodes::_invokeinterface: | |
625 case Bytecodes::_athrow: | |
626 break; | |
627 default: { | |
628 InterpreterOopMap next_mask; | |
629 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); | |
630 next_mask_expression_stack_size = next_mask.expression_stack_size(); | |
631 // Need to subtract off the size of the result type of | |
632 // the bytecode because this is not described in the | |
633 // debug info but returned to the interpreter in the TOS | |
634 // caching register | |
635 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); | |
636 if (bytecode_result_type != T_ILLEGAL) { | |
637 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; | |
638 } | |
639 assert(top_frame_expression_stack_adjustment >= 0, ""); | |
640 try_next_mask = true; | |
641 break; | |
642 } | |
643 } | |
644 } | |
645 } | |
646 | |
647 // Verify stack depth and oops in frame | |
648 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) | |
649 if (!( | |
650 /* SPARC */ | |
651 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || | |
652 /* x86 */ | |
653 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || | |
654 (try_next_mask && | |
655 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - | |
656 top_frame_expression_stack_adjustment))) || | |
657 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || | |
658 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && | |
659 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) | |
660 )) { | |
661 ttyLocker ttyl; | |
662 | |
663 // Print out some information that will help us debug the problem | |
664 tty->print_cr("Wrong number of expression stack elements during deoptimization"); | |
665 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); | |
666 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", | |
667 iframe->interpreter_frame_expression_stack_size()); | |
668 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); | |
669 tty->print_cr(" try_next_mask = %d", try_next_mask); | |
670 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); | |
671 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); | |
672 tty->print_cr(" callee_max_locals = %d", callee_max_locals); | |
673 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); | |
674 tty->print_cr(" exec_mode = %d", exec_mode); | |
675 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); | |
676 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); | |
677 tty->print_cr(" Interpreted frames:"); | |
678 for (int k = 0; k < cur_array->frames(); k++) { | |
679 vframeArrayElement* el = cur_array->element(k); | |
680 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); | |
681 } | |
682 cur_array->print_on_2(tty); | |
683 guarantee(false, "wrong number of expression stack elements during deopt"); | |
684 } | |
685 VerifyOopClosure verify; | |
686 iframe->oops_interpreted_do(&verify, &rm, false); | |
687 callee_size_of_parameters = mh->size_of_parameters(); | |
688 callee_max_locals = mh->max_locals(); | |
689 is_top_frame = false; | |
690 } | |
691 } | |
692 #endif /* !PRODUCT */ | |
693 | |
694 | |
695 return bt; | |
696 JRT_END | |
697 | |
698 | |
699 int Deoptimization::deoptimize_dependents() { | |
700 Threads::deoptimized_wrt_marked_nmethods(); | |
701 return 0; | |
702 } | |
703 | |
704 | |
705 #ifdef COMPILER2 | |
706 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { | |
707 Handle pending_exception(thread->pending_exception()); | |
708 const char* exception_file = thread->exception_file(); | |
709 int exception_line = thread->exception_line(); | |
710 thread->clear_pending_exception(); | |
711 | |
712 for (int i = 0; i < objects->length(); i++) { | |
713 assert(objects->at(i)->is_object(), "invalid debug information"); | |
714 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
715 | |
716 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
717 oop obj = NULL; | |
718 | |
719 if (k->oop_is_instance()) { | |
720 instanceKlass* ik = instanceKlass::cast(k()); | |
721 obj = ik->allocate_instance(CHECK_(false)); | |
722 } else if (k->oop_is_typeArray()) { | |
723 typeArrayKlass* ak = typeArrayKlass::cast(k()); | |
724 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); | |
725 int len = sv->field_size() / type2size[ak->element_type()]; | |
726 obj = ak->allocate(len, CHECK_(false)); | |
727 } else if (k->oop_is_objArray()) { | |
728 objArrayKlass* ak = objArrayKlass::cast(k()); | |
729 obj = ak->allocate(sv->field_size(), CHECK_(false)); | |
730 } | |
731 | |
732 assert(obj != NULL, "allocation failed"); | |
733 assert(sv->value().is_null(), "redundant reallocation"); | |
734 sv->set_value(obj); | |
735 } | |
736 | |
737 if (pending_exception.not_null()) { | |
738 thread->set_pending_exception(pending_exception(), exception_file, exception_line); | |
739 } | |
740 | |
741 return true; | |
742 } | |
743 | |
744 // This assumes that the fields are stored in ObjectValue in the same order | |
745 // they are yielded by do_nonstatic_fields. | |
746 class FieldReassigner: public FieldClosure { | |
747 frame* _fr; | |
748 RegisterMap* _reg_map; | |
749 ObjectValue* _sv; | |
750 instanceKlass* _ik; | |
751 oop _obj; | |
752 | |
753 int _i; | |
754 public: | |
755 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : | |
756 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} | |
757 | |
758 int i() const { return _i; } | |
759 | |
760 | |
761 void do_field(fieldDescriptor* fd) { | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
762 intptr_t val; |
0 | 763 StackValue* value = |
764 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); | |
765 int offset = fd->offset(); | |
766 switch (fd->field_type()) { | |
767 case T_OBJECT: case T_ARRAY: | |
768 assert(value->type() == T_OBJECT, "Agreement."); | |
769 _obj->obj_field_put(offset, value->get_obj()()); | |
770 break; | |
771 | |
772 case T_LONG: case T_DOUBLE: { | |
773 assert(value->type() == T_INT, "Agreement."); | |
774 StackValue* low = | |
775 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
|
776 #ifdef _LP64 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
777 jlong res = (jlong)low->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
778 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
779 #ifdef SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
780 // 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
|
781 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
|
782 #else |
0 | 783 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
|
784 #endif //SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
785 #endif |
0 | 786 _obj->long_field_put(offset, res); |
787 break; | |
788 } | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
789 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. |
0 | 790 case T_INT: case T_FLOAT: // 4 bytes. |
791 assert(value->type() == T_INT, "Agreement."); | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
792 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
793 _obj->int_field_put(offset, (jint)*((jint*)&val)); |
0 | 794 break; |
795 | |
796 case T_SHORT: case T_CHAR: // 2 bytes | |
797 assert(value->type() == T_INT, "Agreement."); | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
798 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
799 _obj->short_field_put(offset, (jshort)*((jint*)&val)); |
0 | 800 break; |
801 | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
802 case T_BOOLEAN: case T_BYTE: // 1 byte |
0 | 803 assert(value->type() == T_INT, "Agreement."); |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
804 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
805 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); |
0 | 806 break; |
807 | |
808 default: | |
809 ShouldNotReachHere(); | |
810 } | |
811 _i++; | |
812 } | |
813 }; | |
814 | |
815 // restore elements of an eliminated type array | |
816 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { | |
817 int index = 0; | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
818 intptr_t val; |
0 | 819 |
820 for (int i = 0; i < sv->field_size(); i++) { | |
821 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); | |
822 switch(type) { | |
44
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
823 case T_LONG: case T_DOUBLE: { |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
824 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
825 StackValue* low = |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
826 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
|
827 #ifdef _LP64 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
828 jlong res = (jlong)low->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
829 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
830 #ifdef SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
831 // 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
|
832 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
|
833 #else |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
834 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
|
835 #endif //SPARC |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
836 #endif |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
837 obj->long_at_put(index, res); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
838 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
839 } |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
840 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
841 // 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
|
842 case T_INT: case T_FLOAT: // 4 bytes. |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
843 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
844 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
845 obj->int_at_put(index, (jint)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
846 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
847 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
848 case T_SHORT: case T_CHAR: // 2 bytes |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
849 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
850 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
851 obj->short_at_put(index, (jshort)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
852 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
853 |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
854 case T_BOOLEAN: case T_BYTE: // 1 byte |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
855 assert(value->type() == T_INT, "Agreement."); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
856 val = value->get_int(); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
857 obj->bool_at_put(index, (jboolean)*((jint*)&val)); |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
858 break; |
52fed2ec0afb
6667620: (Escape Analysis) fix deoptimization for scalar replaced objects
kvn
parents:
0
diff
changeset
|
859 |
0 | 860 default: |
861 ShouldNotReachHere(); | |
862 } | |
863 index++; | |
864 } | |
865 } | |
866 | |
867 | |
868 // restore fields of an eliminated object array | |
869 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { | |
870 for (int i = 0; i < sv->field_size(); i++) { | |
871 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); | |
872 assert(value->type() == T_OBJECT, "object element expected"); | |
873 obj->obj_at_put(i, value->get_obj()()); | |
874 } | |
875 } | |
876 | |
877 | |
878 // restore fields of all eliminated objects and arrays | |
879 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { | |
880 for (int i = 0; i < objects->length(); i++) { | |
881 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
882 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
883 Handle obj = sv->value(); | |
884 assert(obj.not_null(), "reallocation was missed"); | |
885 | |
886 if (k->oop_is_instance()) { | |
887 instanceKlass* ik = instanceKlass::cast(k()); | |
888 FieldReassigner reassign(fr, reg_map, sv, obj()); | |
889 ik->do_nonstatic_fields(&reassign); | |
890 } else if (k->oop_is_typeArray()) { | |
891 typeArrayKlass* ak = typeArrayKlass::cast(k()); | |
892 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); | |
893 } else if (k->oop_is_objArray()) { | |
894 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); | |
895 } | |
896 } | |
897 } | |
898 | |
899 | |
900 // 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
|
901 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { |
0 | 902 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
|
903 MonitorInfo* mon_info = monitors->at(i); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
904 if (mon_info->eliminated()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
905 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
|
906 Handle obj = Handle(mon_info->owner()); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
907 markOop mark = obj->mark(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
908 if (UseBiasedLocking && mark->has_bias_pattern()) { |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
909 // 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
|
910 // 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
|
911 // 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
|
912 assert(mark->is_biased_anonymously() || |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
913 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
|
914 // Reset mark word to unbiased prototype. |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
915 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
|
916 obj->set_mark(unbiased_prototype); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
917 } |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
918 BasicLock* lock = mon_info->lock(); |
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
919 ObjectSynchronizer::slow_enter(obj, lock, thread); |
0 | 920 } |
83
d3cd40645d0d
6681646: Relocking of a scalar replaced object during deoptimization is broken
kvn
parents:
44
diff
changeset
|
921 assert(mon_info->owner()->is_locked(), "object must be locked now"); |
0 | 922 } |
923 } | |
924 | |
925 | |
926 #ifndef PRODUCT | |
927 // print information about reallocated objects | |
928 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { | |
929 fieldDescriptor fd; | |
930 | |
931 for (int i = 0; i < objects->length(); i++) { | |
932 ObjectValue* sv = (ObjectValue*) objects->at(i); | |
933 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); | |
934 Handle obj = sv->value(); | |
935 | |
936 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); | |
937 k->as_klassOop()->print_value(); | |
938 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); | |
939 tty->cr(); | |
940 | |
941 if (Verbose) { | |
942 k->oop_print_on(obj(), tty); | |
943 } | |
944 } | |
945 } | |
946 #endif | |
947 #endif // COMPILER2 | |
948 | |
949 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { | |
950 | |
951 #ifndef PRODUCT | |
952 if (TraceDeoptimization) { | |
953 ttyLocker ttyl; | |
954 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); | |
955 fr.print_on(tty); | |
956 tty->print_cr(" Virtual frames (innermost first):"); | |
957 for (int index = 0; index < chunk->length(); index++) { | |
958 compiledVFrame* vf = chunk->at(index); | |
959 tty->print(" %2d - ", index); | |
960 vf->print_value(); | |
961 int bci = chunk->at(index)->raw_bci(); | |
962 const char* code_name; | |
963 if (bci == SynchronizationEntryBCI) { | |
964 code_name = "sync entry"; | |
965 } else { | |
2142 | 966 Bytecodes::Code code = vf->method()->code_at(bci); |
0 | 967 code_name = Bytecodes::name(code); |
968 } | |
969 tty->print(" - %s", code_name); | |
970 tty->print_cr(" @ bci %d ", bci); | |
971 if (Verbose) { | |
972 vf->print(); | |
973 tty->cr(); | |
974 } | |
975 } | |
976 } | |
977 #endif | |
978 | |
979 // Register map for next frame (used for stack crawl). We capture | |
980 // the state of the deopt'ing frame's caller. Thus if we need to | |
981 // stuff a C2I adapter we can properly fill in the callee-save | |
982 // register locations. | |
983 frame caller = fr.sender(reg_map); | |
984 int frame_size = caller.sp() - fr.sp(); | |
985 | |
986 frame sender = caller; | |
987 | |
988 // Since the Java thread being deoptimized will eventually adjust it's own stack, | |
989 // the vframeArray containing the unpacking information is allocated in the C heap. | |
990 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). | |
991 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); | |
992 | |
993 // Compare the vframeArray to the collected vframes | |
994 assert(array->structural_compare(thread, chunk), "just checking"); | |
995 Events::log("# vframes = %d", (intptr_t)chunk->length()); | |
996 | |
997 #ifndef PRODUCT | |
998 if (TraceDeoptimization) { | |
999 ttyLocker ttyl; | |
1000 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); | |
1001 } | |
1002 #endif // PRODUCT | |
1003 | |
1004 return array; | |
1005 } | |
1006 | |
1007 | |
1008 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { | |
1009 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); | |
1010 for (int i = 0; i < monitors->length(); i++) { | |
1011 MonitorInfo* mon_info = monitors->at(i); | |
818
b109e761e927
6837472: com/sun/jdi/MonitorFrameInfo.java fails with AggressiveOpts in 6u14
kvn
parents:
196
diff
changeset
|
1012 if (!mon_info->eliminated() && mon_info->owner() != NULL) { |
0 | 1013 objects_to_revoke->append(Handle(mon_info->owner())); |
1014 } | |
1015 } | |
1016 } | |
1017 | |
1018 | |
1019 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { | |
1020 if (!UseBiasedLocking) { | |
1021 return; | |
1022 } | |
1023 | |
1024 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); | |
1025 | |
1026 // Unfortunately we don't have a RegisterMap available in most of | |
1027 // the places we want to call this routine so we need to walk the | |
1028 // stack again to update the register map. | |
1029 if (map == NULL || !map->update_map()) { | |
1030 StackFrameStream sfs(thread, true); | |
1031 bool found = false; | |
1032 while (!found && !sfs.is_done()) { | |
1033 frame* cur = sfs.current(); | |
1034 sfs.next(); | |
1035 found = cur->id() == fr.id(); | |
1036 } | |
1037 assert(found, "frame to be deoptimized not found on target thread's stack"); | |
1038 map = sfs.register_map(); | |
1039 } | |
1040 | |
1041 vframe* vf = vframe::new_vframe(&fr, map, thread); | |
1042 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1043 // Revoke monitors' biases in all scopes | |
1044 while (!cvf->is_top()) { | |
1045 collect_monitors(cvf, objects_to_revoke); | |
1046 cvf = compiledVFrame::cast(cvf->sender()); | |
1047 } | |
1048 collect_monitors(cvf, objects_to_revoke); | |
1049 | |
1050 if (SafepointSynchronize::is_at_safepoint()) { | |
1051 BiasedLocking::revoke_at_safepoint(objects_to_revoke); | |
1052 } else { | |
1053 BiasedLocking::revoke(objects_to_revoke); | |
1054 } | |
1055 } | |
1056 | |
1057 | |
1058 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { | |
1059 if (!UseBiasedLocking) { | |
1060 return; | |
1061 } | |
1062 | |
1063 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); | |
1064 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); | |
1065 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { | |
1066 if (jt->has_last_Java_frame()) { | |
1067 StackFrameStream sfs(jt, true); | |
1068 while (!sfs.is_done()) { | |
1069 frame* cur = sfs.current(); | |
1070 if (cb->contains(cur->pc())) { | |
1071 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); | |
1072 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1073 // Revoke monitors' biases in all scopes | |
1074 while (!cvf->is_top()) { | |
1075 collect_monitors(cvf, objects_to_revoke); | |
1076 cvf = compiledVFrame::cast(cvf->sender()); | |
1077 } | |
1078 collect_monitors(cvf, objects_to_revoke); | |
1079 } | |
1080 sfs.next(); | |
1081 } | |
1082 } | |
1083 } | |
1084 BiasedLocking::revoke_at_safepoint(objects_to_revoke); | |
1085 } | |
1086 | |
1087 | |
1088 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { | |
1089 assert(fr.can_be_deoptimized(), "checking frame type"); | |
1090 | |
1091 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); | |
1092 | |
1093 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); | |
1094 | |
1095 // Patch the nmethod so that when execution returns to it we will | |
1096 // deopt the execution state and return to the interpreter. | |
1097 fr.deoptimize(thread); | |
1098 } | |
1099 | |
1100 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { | |
1101 // Deoptimize only if the frame comes from compile code. | |
1102 // Do not deoptimize the frame which is already patched | |
1103 // during the execution of the loops below. | |
1104 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { | |
1105 return; | |
1106 } | |
1107 ResourceMark rm; | |
1108 DeoptimizationMarker dm; | |
1109 if (UseBiasedLocking) { | |
1110 revoke_biases_of_monitors(thread, fr, map); | |
1111 } | |
1112 deoptimize_single_frame(thread, fr); | |
1113 | |
1114 } | |
1115 | |
1116 | |
1905
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1117 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
|
1118 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
|
1119 "can only deoptimize other thread at a safepoint"); |
0 | 1120 // Compute frame and register map based on thread and sp. |
1121 RegisterMap reg_map(thread, UseBiasedLocking); | |
1122 frame fr = thread->last_frame(); | |
1123 while (fr.id() != id) { | |
1124 fr = fr.sender(®_map); | |
1125 } | |
1126 deoptimize(thread, fr, ®_map); | |
1127 } | |
1128 | |
1129 | |
1905
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1130 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
|
1131 if (thread == Thread::current()) { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1132 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
|
1133 } else { |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1134 VM_DeoptimizeFrame deopt(thread, id); |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1135 VMThread::execute(&deopt); |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1136 } |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1137 } |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1138 |
ce6848d0666d
6968367: can_post_on_exceptions is still using VM_DeoptimizeFrame in some places
never
parents:
1814
diff
changeset
|
1139 |
0 | 1140 // JVMTI PopFrame support |
1141 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) | |
1142 { | |
1143 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); | |
1144 } | |
1145 JRT_END | |
1146 | |
1147 | |
1692 | 1148 #if defined(COMPILER2) || defined(SHARK) |
0 | 1149 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { |
1150 // in case of an unresolved klass entry, load the class. | |
1151 if (constant_pool->tag_at(index).is_unresolved_klass()) { | |
1152 klassOop tk = constant_pool->klass_at(index, CHECK); | |
1153 return; | |
1154 } | |
1155 | |
1156 if (!constant_pool->tag_at(index).is_symbol()) return; | |
1157 | |
1158 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); | |
1159 symbolHandle symbol (THREAD, constant_pool->symbol_at(index)); | |
1160 | |
1161 // class name? | |
1162 if (symbol->byte_at(0) != '(') { | |
1163 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); | |
1164 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); | |
1165 return; | |
1166 } | |
1167 | |
1168 // then it must be a signature! | |
1169 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { | |
1170 if (ss.is_object()) { | |
1171 symbolOop s = ss.as_symbol(CHECK); | |
1172 symbolHandle class_name (THREAD, s); | |
1173 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); | |
1174 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); | |
1175 } | |
1176 } | |
1177 } | |
1178 | |
1179 | |
1180 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { | |
1181 EXCEPTION_MARK; | |
1182 load_class_by_index(constant_pool, index, THREAD); | |
1183 if (HAS_PENDING_EXCEPTION) { | |
1184 // Exception happened during classloading. We ignore the exception here, since it | |
1185 // is going to be rethrown since the current activation is going to be deoptimzied and | |
1186 // the interpreter will re-execute the bytecode. | |
1187 CLEAR_PENDING_EXCEPTION; | |
1188 } | |
1189 } | |
1190 | |
1191 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { | |
1192 HandleMark hm; | |
1193 | |
1194 // uncommon_trap() is called at the beginning of the uncommon trap | |
1195 // handler. Note this fact before we start generating temporary frames | |
1196 // that can confuse an asynchronous stack walker. This counter is | |
1197 // decremented at the end of unpack_frames(). | |
1198 thread->inc_in_deopt_handler(); | |
1199 | |
1200 // We need to update the map if we have biased locking. | |
1201 RegisterMap reg_map(thread, UseBiasedLocking); | |
1202 frame stub_frame = thread->last_frame(); | |
1203 frame fr = stub_frame.sender(®_map); | |
1204 // Make sure the calling nmethod is not getting deoptimized and removed | |
1205 // before we are done with it. | |
1206 nmethodLocker nl(fr.pc()); | |
1207 | |
1208 { | |
1209 ResourceMark rm; | |
1210 | |
1211 // Revoke biases of any monitors in the frame to ensure we can migrate them | |
1212 revoke_biases_of_monitors(thread, fr, ®_map); | |
1213 | |
1214 DeoptReason reason = trap_request_reason(trap_request); | |
1215 DeoptAction action = trap_request_action(trap_request); | |
1216 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 | |
1217 | |
1218 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); | |
1219 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); | |
1220 compiledVFrame* cvf = compiledVFrame::cast(vf); | |
1221 | |
1222 nmethod* nm = cvf->code(); | |
1223 | |
1224 ScopeDesc* trap_scope = cvf->scope(); | |
1225 methodHandle trap_method = trap_scope->method(); | |
1226 int trap_bci = trap_scope->bci(); | |
2142 | 1227 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); |
0 | 1228 |
1229 // Record this event in the histogram. | |
1230 gather_statistics(reason, action, trap_bc); | |
1231 | |
1232 // Ensure that we can record deopt. history: | |
1233 bool create_if_missing = ProfileTraps; | |
1234 | |
1235 methodDataHandle trap_mdo | |
1236 (THREAD, get_method_data(thread, trap_method, create_if_missing)); | |
1237 | |
1238 // Print a bunch of diagnostics, if requested. | |
1239 if (TraceDeoptimization || LogCompilation) { | |
1240 ResourceMark rm; | |
1241 ttyLocker ttyl; | |
1242 char buf[100]; | |
1243 if (xtty != NULL) { | |
1244 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", | |
1245 os::current_thread_id(), | |
1246 format_trap_request(buf, sizeof(buf), trap_request)); | |
1247 nm->log_identity(xtty); | |
1248 } | |
1249 symbolHandle class_name; | |
1250 bool unresolved = false; | |
1251 if (unloaded_class_index >= 0) { | |
1252 constantPoolHandle constants (THREAD, trap_method->constants()); | |
1253 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { | |
1254 class_name = symbolHandle(THREAD, | |
1255 constants->klass_name_at(unloaded_class_index)); | |
1256 unresolved = true; | |
1257 if (xtty != NULL) | |
1258 xtty->print(" unresolved='1'"); | |
1259 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { | |
1260 class_name = symbolHandle(THREAD, | |
1261 constants->symbol_at(unloaded_class_index)); | |
1262 } | |
1263 if (xtty != NULL) | |
1264 xtty->name(class_name); | |
1265 } | |
1266 if (xtty != NULL && trap_mdo.not_null()) { | |
1267 // Dump the relevant MDO state. | |
1268 // This is the deopt count for the current reason, any previous | |
1269 // reasons or recompiles seen at this point. | |
1270 int dcnt = trap_mdo->trap_count(reason); | |
1271 if (dcnt != 0) | |
1272 xtty->print(" count='%d'", dcnt); | |
1273 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); | |
1274 int dos = (pdata == NULL)? 0: pdata->trap_state(); | |
1275 if (dos != 0) { | |
1276 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); | |
1277 if (trap_state_is_recompiled(dos)) { | |
1278 int recnt2 = trap_mdo->overflow_recompile_count(); | |
1279 if (recnt2 != 0) | |
1280 xtty->print(" recompiles2='%d'", recnt2); | |
1281 } | |
1282 } | |
1283 } | |
1284 if (xtty != NULL) { | |
1285 xtty->stamp(); | |
1286 xtty->end_head(); | |
1287 } | |
1288 if (TraceDeoptimization) { // make noise on the tty | |
1289 tty->print("Uncommon trap occurred in"); | |
1290 nm->method()->print_short_name(tty); | |
1291 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", | |
1292 fr.pc(), | |
1293 (int) os::current_thread_id(), | |
1294 trap_reason_name(reason), | |
1295 trap_action_name(action), | |
1296 unloaded_class_index); | |
1297 if (class_name.not_null()) { | |
1298 tty->print(unresolved ? " unresolved class: " : " symbol: "); | |
1299 class_name->print_symbol_on(tty); | |
1300 } | |
1301 tty->cr(); | |
1302 } | |
1303 if (xtty != NULL) { | |
1304 // Log the precise location of the trap. | |
1305 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { | |
1306 xtty->begin_elem("jvms bci='%d'", sd->bci()); | |
1307 xtty->method(sd->method()); | |
1308 xtty->end_elem(); | |
1309 if (sd->is_top()) break; | |
1310 } | |
1311 xtty->tail("uncommon_trap"); | |
1312 } | |
1313 } | |
1314 // (End diagnostic printout.) | |
1315 | |
1316 // Load class if necessary | |
1317 if (unloaded_class_index >= 0) { | |
1318 constantPoolHandle constants(THREAD, trap_method->constants()); | |
1319 load_class_by_index(constants, unloaded_class_index); | |
1320 } | |
1321 | |
1322 // Flush the nmethod if necessary and desirable. | |
1323 // | |
1324 // We need to avoid situations where we are re-flushing the nmethod | |
1325 // because of a hot deoptimization site. Repeated flushes at the same | |
1326 // point need to be detected by the compiler and avoided. If the compiler | |
1327 // cannot avoid them (or has a bug and "refuses" to avoid them), this | |
1328 // module must take measures to avoid an infinite cycle of recompilation | |
1329 // and deoptimization. There are several such measures: | |
1330 // | |
1331 // 1. If a recompilation is ordered a second time at some site X | |
1332 // and for the same reason R, the action is adjusted to 'reinterpret', | |
1333 // to give the interpreter time to exercise the method more thoroughly. | |
1334 // If this happens, the method's overflow_recompile_count is incremented. | |
1335 // | |
1336 // 2. If the compiler fails to reduce the deoptimization rate, then | |
1337 // the method's overflow_recompile_count will begin to exceed the set | |
1338 // limit PerBytecodeRecompilationCutoff. If this happens, the action | |
1339 // is adjusted to 'make_not_compilable', and the method is abandoned | |
1340 // to the interpreter. This is a performance hit for hot methods, | |
1341 // but is better than a disastrous infinite cycle of recompilations. | |
1342 // (Actually, only the method containing the site X is abandoned.) | |
1343 // | |
1344 // 3. In parallel with the previous measures, if the total number of | |
1345 // recompilations of a method exceeds the much larger set limit | |
1346 // PerMethodRecompilationCutoff, the method is abandoned. | |
1347 // This should only happen if the method is very large and has | |
1348 // many "lukewarm" deoptimizations. The code which enforces this | |
1349 // limit is elsewhere (class nmethod, class methodOopDesc). | |
1350 // | |
1351 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance | |
1352 // to recompile at each bytecode independently of the per-BCI cutoff. | |
1353 // | |
1354 // The decision to update code is up to the compiler, and is encoded | |
1355 // in the Action_xxx code. If the compiler requests Action_none | |
1356 // no trap state is changed, no compiled code is changed, and the | |
1357 // computation suffers along in the interpreter. | |
1358 // | |
1359 // The other action codes specify various tactics for decompilation | |
1360 // and recompilation. Action_maybe_recompile is the loosest, and | |
1361 // allows the compiled code to stay around until enough traps are seen, | |
1362 // and until the compiler gets around to recompiling the trapping method. | |
1363 // | |
1364 // The other actions cause immediate removal of the present code. | |
1365 | |
1366 bool update_trap_state = true; | |
1367 bool make_not_entrant = false; | |
1368 bool make_not_compilable = false; | |
1783 | 1369 bool reprofile = false; |
0 | 1370 switch (action) { |
1371 case Action_none: | |
1372 // Keep the old code. | |
1373 update_trap_state = false; | |
1374 break; | |
1375 case Action_maybe_recompile: | |
1376 // Do not need to invalidate the present code, but we can | |
1377 // initiate another | |
1378 // Start compiler without (necessarily) invalidating the nmethod. | |
1379 // The system will tolerate the old code, but new code should be | |
1380 // generated when possible. | |
1381 break; | |
1382 case Action_reinterpret: | |
1383 // Go back into the interpreter for a while, and then consider | |
1384 // recompiling form scratch. | |
1385 make_not_entrant = true; | |
1386 // Reset invocation counter for outer most method. | |
1387 // This will allow the interpreter to exercise the bytecodes | |
1388 // for a while before recompiling. | |
1389 // By contrast, Action_make_not_entrant is immediate. | |
1390 // | |
1391 // Note that the compiler will track null_check, null_assert, | |
1392 // range_check, and class_check events and log them as if they | |
1393 // had been traps taken from compiled code. This will update | |
1394 // the MDO trap history so that the next compilation will | |
1395 // properly detect hot trap sites. | |
1783 | 1396 reprofile = true; |
0 | 1397 break; |
1398 case Action_make_not_entrant: | |
1399 // Request immediate recompilation, and get rid of the old code. | |
1400 // Make them not entrant, so next time they are called they get | |
1401 // recompiled. Unloaded classes are loaded now so recompile before next | |
1402 // time they are called. Same for uninitialized. The interpreter will | |
1403 // link the missing class, if any. | |
1404 make_not_entrant = true; | |
1405 break; | |
1406 case Action_make_not_compilable: | |
1407 // Give up on compiling this method at all. | |
1408 make_not_entrant = true; | |
1409 make_not_compilable = true; | |
1410 break; | |
1411 default: | |
1412 ShouldNotReachHere(); | |
1413 } | |
1414 | |
1415 // Setting +ProfileTraps fixes the following, on all platforms: | |
1416 // 4852688: ProfileInterpreter is off by default for ia64. The result is | |
1417 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the | |
1418 // recompile relies on a methodDataOop to record heroic opt failures. | |
1419 | |
1420 // Whether the interpreter is producing MDO data or not, we also need | |
1421 // to use the MDO to detect hot deoptimization points and control | |
1422 // aggressive optimization. | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1423 bool inc_recompile_count = false; |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1424 ProfileData* pdata = NULL; |
0 | 1425 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { |
1426 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); | |
1427 uint this_trap_count = 0; | |
1428 bool maybe_prior_trap = false; | |
1429 bool maybe_prior_recompile = false; | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1430 pdata = query_update_method_data(trap_mdo, trap_bci, reason, |
0 | 1431 //outputs: |
1432 this_trap_count, | |
1433 maybe_prior_trap, | |
1434 maybe_prior_recompile); | |
1435 // Because the interpreter also counts null, div0, range, and class | |
1436 // checks, these traps from compiled code are double-counted. | |
1437 // This is harmless; it just means that the PerXTrapLimit values | |
1438 // are in effect a little smaller than they look. | |
1439 | |
1440 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); | |
1441 if (per_bc_reason != Reason_none) { | |
1442 // Now take action based on the partially known per-BCI history. | |
1443 if (maybe_prior_trap | |
1444 && this_trap_count >= (uint)PerBytecodeTrapLimit) { | |
1445 // If there are too many traps at this BCI, force a recompile. | |
1446 // This will allow the compiler to see the limit overflow, and | |
1447 // take corrective action, if possible. The compiler generally | |
1448 // does not use the exact PerBytecodeTrapLimit value, but instead | |
1449 // changes its tactics if it sees any traps at all. This provides | |
1450 // a little hysteresis, delaying a recompile until a trap happens | |
1451 // several times. | |
1452 // | |
1453 // Actually, since there is only one bit of counter per BCI, | |
1454 // the possible per-BCI counts are {0,1,(per-method count)}. | |
1455 // This produces accurate results if in fact there is only | |
1456 // one hot trap site, but begins to get fuzzy if there are | |
1457 // many sites. For example, if there are ten sites each | |
1458 // trapping two or more times, they each get the blame for | |
1459 // all of their traps. | |
1460 make_not_entrant = true; | |
1461 } | |
1462 | |
1463 // Detect repeated recompilation at the same BCI, and enforce a limit. | |
1464 if (make_not_entrant && maybe_prior_recompile) { | |
1465 // More than one recompile at this point. | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1466 inc_recompile_count = maybe_prior_trap; |
0 | 1467 } |
1468 } else { | |
1469 // For reasons which are not recorded per-bytecode, we simply | |
1470 // force recompiles unconditionally. | |
1471 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) | |
1472 make_not_entrant = true; | |
1473 } | |
1474 | |
1475 // Go back to the compiler if there are too many traps in this method. | |
1476 if (this_trap_count >= (uint)PerMethodTrapLimit) { | |
1477 // If there are too many traps in this method, force a recompile. | |
1478 // This will allow the compiler to see the limit overflow, and | |
1479 // take corrective action, if possible. | |
1480 // (This condition is an unlikely backstop only, because the | |
1481 // PerBytecodeTrapLimit is more likely to take effect first, | |
1482 // if it is applicable.) | |
1483 make_not_entrant = true; | |
1484 } | |
1485 | |
1486 // Here's more hysteresis: If there has been a recompile at | |
1487 // this trap point already, run the method in the interpreter | |
1488 // for a while to exercise it more thoroughly. | |
1489 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { | |
1783 | 1490 reprofile = true; |
0 | 1491 } |
1492 | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1493 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1494 |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1495 // Take requested actions on the method: |
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 // Recompile |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1498 if (make_not_entrant) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1499 if (!nm->make_not_entrant()) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1500 return; // the call did not change nmethod's state |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1501 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1502 |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1503 if (pdata != NULL) { |
0 | 1504 // Record the recompilation event, if any. |
1505 int tstate0 = pdata->trap_state(); | |
1506 int tstate1 = trap_state_set_recompiled(tstate0, true); | |
1507 if (tstate1 != tstate0) | |
1508 pdata->set_trap_state(tstate1); | |
1509 } | |
1510 } | |
1511 | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1512 if (inc_recompile_count) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1513 trap_mdo->inc_overflow_recompile_count(); |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1514 if ((uint)trap_mdo->overflow_recompile_count() > |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1515 (uint)PerBytecodeRecompilationCutoff) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1516 // Give up on the method containing the bad BCI. |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1517 if (trap_method() == nm->method()) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1518 make_not_compilable = true; |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1519 } else { |
1783 | 1520 trap_method->set_not_compilable(CompLevel_full_optimization); |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1521 // But give grace to the enclosing nm->method(). |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1522 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1523 } |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1524 } |
0 | 1525 |
1783 | 1526 // Reprofile |
1527 if (reprofile) { | |
1528 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); | |
0 | 1529 } |
1530 | |
1531 // Give up compiling | |
1783 | 1532 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { |
0 | 1533 assert(make_not_entrant, "consistent"); |
1783 | 1534 nm->method()->set_not_compilable(CompLevel_full_optimization); |
0 | 1535 } |
1536 | |
1537 } // Free marked resources | |
1538 | |
1539 } | |
1540 JRT_END | |
1541 | |
1542 methodDataOop | |
1543 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, | |
1544 bool create_if_missing) { | |
1545 Thread* THREAD = thread; | |
1546 methodDataOop mdo = m()->method_data(); | |
1547 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { | |
1548 // Build an MDO. Ignore errors like OutOfMemory; | |
1549 // that simply means we won't have an MDO to update. | |
1550 methodOopDesc::build_interpreter_method_data(m, THREAD); | |
1551 if (HAS_PENDING_EXCEPTION) { | |
1552 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); | |
1553 CLEAR_PENDING_EXCEPTION; | |
1554 } | |
1555 mdo = m()->method_data(); | |
1556 } | |
1557 return mdo; | |
1558 } | |
1559 | |
1560 ProfileData* | |
1561 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, | |
1562 int trap_bci, | |
1563 Deoptimization::DeoptReason reason, | |
1564 //outputs: | |
1565 uint& ret_this_trap_count, | |
1566 bool& ret_maybe_prior_trap, | |
1567 bool& ret_maybe_prior_recompile) { | |
1568 uint prior_trap_count = trap_mdo->trap_count(reason); | |
1569 uint this_trap_count = trap_mdo->inc_trap_count(reason); | |
1570 | |
1571 // If the runtime cannot find a place to store trap history, | |
1572 // it is estimated based on the general condition of the method. | |
1573 // If the method has ever been recompiled, or has ever incurred | |
1574 // a trap with the present reason , then this BCI is assumed | |
1575 // (pessimistically) to be the culprit. | |
1576 bool maybe_prior_trap = (prior_trap_count != 0); | |
1577 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); | |
1578 ProfileData* pdata = NULL; | |
1579 | |
1580 | |
1581 // For reasons which are recorded per bytecode, we check per-BCI data. | |
1582 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); | |
1583 if (per_bc_reason != Reason_none) { | |
1584 // Find the profile data for this BCI. If there isn't one, | |
1585 // try to allocate one from the MDO's set of spares. | |
1586 // This will let us detect a repeated trap at this point. | |
1587 pdata = trap_mdo->allocate_bci_to_data(trap_bci); | |
1588 | |
1589 if (pdata != NULL) { | |
1590 // Query the trap state of this profile datum. | |
1591 int tstate0 = pdata->trap_state(); | |
1592 if (!trap_state_has_reason(tstate0, per_bc_reason)) | |
1593 maybe_prior_trap = false; | |
1594 if (!trap_state_is_recompiled(tstate0)) | |
1595 maybe_prior_recompile = false; | |
1596 | |
1597 // Update the trap state of this profile datum. | |
1598 int tstate1 = tstate0; | |
1599 // Record the reason. | |
1600 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); | |
1601 // Store the updated state on the MDO, for next time. | |
1602 if (tstate1 != tstate0) | |
1603 pdata->set_trap_state(tstate1); | |
1604 } else { | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1605 if (LogCompilation && xtty != NULL) { |
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1606 ttyLocker ttyl; |
0 | 1607 // Missing MDP? Leave a small complaint in the log. |
1608 xtty->elem("missing_mdp bci='%d'", trap_bci); | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1609 } |
0 | 1610 } |
1611 } | |
1612 | |
1613 // Return results: | |
1614 ret_this_trap_count = this_trap_count; | |
1615 ret_maybe_prior_trap = maybe_prior_trap; | |
1616 ret_maybe_prior_recompile = maybe_prior_recompile; | |
1617 return pdata; | |
1618 } | |
1619 | |
1620 void | |
1621 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { | |
1622 ResourceMark rm; | |
1623 // Ignored outputs: | |
1624 uint ignore_this_trap_count; | |
1625 bool ignore_maybe_prior_trap; | |
1626 bool ignore_maybe_prior_recompile; | |
1627 query_update_method_data(trap_mdo, trap_bci, | |
1628 (DeoptReason)reason, | |
1629 ignore_this_trap_count, | |
1630 ignore_maybe_prior_trap, | |
1631 ignore_maybe_prior_recompile); | |
1632 } | |
1633 | |
1634 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { | |
1635 | |
1636 // Still in Java no safepoints | |
1637 { | |
1638 // This enters VM and may safepoint | |
1639 uncommon_trap_inner(thread, trap_request); | |
1640 } | |
1641 return fetch_unroll_info_helper(thread); | |
1642 } | |
1643 | |
1644 // Local derived constants. | |
1645 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. | |
1646 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; | |
1647 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; | |
1648 | |
1649 //---------------------------trap_state_reason--------------------------------- | |
1650 Deoptimization::DeoptReason | |
1651 Deoptimization::trap_state_reason(int trap_state) { | |
1652 // This assert provides the link between the width of DataLayout::trap_bits | |
1653 // and the encoding of "recorded" reasons. It ensures there are enough | |
1654 // bits to store all needed reasons in the per-BCI MDO profile. | |
1655 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); | |
1656 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1657 trap_state -= recompile_bit; | |
1658 if (trap_state == DS_REASON_MASK) { | |
1659 return Reason_many; | |
1660 } else { | |
1661 assert((int)Reason_none == 0, "state=0 => Reason_none"); | |
1662 return (DeoptReason)trap_state; | |
1663 } | |
1664 } | |
1665 //-------------------------trap_state_has_reason------------------------------- | |
1666 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { | |
1667 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); | |
1668 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); | |
1669 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1670 trap_state -= recompile_bit; | |
1671 if (trap_state == DS_REASON_MASK) { | |
1672 return -1; // true, unspecifically (bottom of state lattice) | |
1673 } else if (trap_state == reason) { | |
1674 return 1; // true, definitely | |
1675 } else if (trap_state == 0) { | |
1676 return 0; // false, definitely (top of state lattice) | |
1677 } else { | |
1678 return 0; // false, definitely | |
1679 } | |
1680 } | |
1681 //-------------------------trap_state_add_reason------------------------------- | |
1682 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { | |
1683 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); | |
1684 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); | |
1685 trap_state -= recompile_bit; | |
1686 if (trap_state == DS_REASON_MASK) { | |
1687 return trap_state + recompile_bit; // already at state lattice bottom | |
1688 } else if (trap_state == reason) { | |
1689 return trap_state + recompile_bit; // the condition is already true | |
1690 } else if (trap_state == 0) { | |
1691 return reason + recompile_bit; // no condition has yet been true | |
1692 } else { | |
1693 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom | |
1694 } | |
1695 } | |
1696 //-----------------------trap_state_is_recompiled------------------------------ | |
1697 bool Deoptimization::trap_state_is_recompiled(int trap_state) { | |
1698 return (trap_state & DS_RECOMPILE_BIT) != 0; | |
1699 } | |
1700 //-----------------------trap_state_set_recompiled----------------------------- | |
1701 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { | |
1702 if (z) return trap_state | DS_RECOMPILE_BIT; | |
1703 else return trap_state & ~DS_RECOMPILE_BIT; | |
1704 } | |
1705 //---------------------------format_trap_state--------------------------------- | |
1706 // This is used for debugging and diagnostics, including hotspot.log output. | |
1707 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, | |
1708 int trap_state) { | |
1709 DeoptReason reason = trap_state_reason(trap_state); | |
1710 bool recomp_flag = trap_state_is_recompiled(trap_state); | |
1711 // Re-encode the state from its decoded components. | |
1712 int decoded_state = 0; | |
1713 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) | |
1714 decoded_state = trap_state_add_reason(decoded_state, reason); | |
1715 if (recomp_flag) | |
1716 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); | |
1717 // If the state re-encodes properly, format it symbolically. | |
1718 // Because this routine is used for debugging and diagnostics, | |
1719 // be robust even if the state is a strange value. | |
1720 size_t len; | |
1721 if (decoded_state != trap_state) { | |
1722 // Random buggy state that doesn't decode?? | |
1723 len = jio_snprintf(buf, buflen, "#%d", trap_state); | |
1724 } else { | |
1725 len = jio_snprintf(buf, buflen, "%s%s", | |
1726 trap_reason_name(reason), | |
1727 recomp_flag ? " recompiled" : ""); | |
1728 } | |
1729 if (len >= buflen) | |
1730 buf[buflen-1] = '\0'; | |
1731 return buf; | |
1732 } | |
1733 | |
1734 | |
1735 //--------------------------------statics-------------------------------------- | |
1736 Deoptimization::DeoptAction Deoptimization::_unloaded_action | |
1737 = Deoptimization::Action_reinterpret; | |
1738 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { | |
1739 // Note: Keep this in sync. with enum DeoptReason. | |
1740 "none", | |
1741 "null_check", | |
1742 "null_assert", | |
1743 "range_check", | |
1744 "class_check", | |
1745 "array_check", | |
1746 "intrinsic", | |
1206
87684f1a88b5
6614597: Performance variability in jvm2008 xml.validation
kvn
parents:
1204
diff
changeset
|
1747 "bimorphic", |
0 | 1748 "unloaded", |
1749 "uninitialized", | |
1750 "unreached", | |
1751 "unhandled", | |
1752 "constraint", | |
1753 "div0_check", | |
1172 | 1754 "age", |
1755 "predicate" | |
0 | 1756 }; |
1757 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { | |
1758 // Note: Keep this in sync. with enum DeoptAction. | |
1759 "none", | |
1760 "maybe_recompile", | |
1761 "reinterpret", | |
1762 "make_not_entrant", | |
1763 "make_not_compilable" | |
1764 }; | |
1765 | |
1766 const char* Deoptimization::trap_reason_name(int reason) { | |
1767 if (reason == Reason_many) return "many"; | |
1768 if ((uint)reason < Reason_LIMIT) | |
1769 return _trap_reason_name[reason]; | |
1770 static char buf[20]; | |
1771 sprintf(buf, "reason%d", reason); | |
1772 return buf; | |
1773 } | |
1774 const char* Deoptimization::trap_action_name(int action) { | |
1775 if ((uint)action < Action_LIMIT) | |
1776 return _trap_action_name[action]; | |
1777 static char buf[20]; | |
1778 sprintf(buf, "action%d", action); | |
1779 return buf; | |
1780 } | |
1781 | |
1782 // This is used for debugging and diagnostics, including hotspot.log output. | |
1783 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, | |
1784 int trap_request) { | |
1785 jint unloaded_class_index = trap_request_index(trap_request); | |
1786 const char* reason = trap_reason_name(trap_request_reason(trap_request)); | |
1787 const char* action = trap_action_name(trap_request_action(trap_request)); | |
1788 size_t len; | |
1789 if (unloaded_class_index < 0) { | |
1790 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", | |
1791 reason, action); | |
1792 } else { | |
1793 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", | |
1794 reason, action, unloaded_class_index); | |
1795 } | |
1796 if (len >= buflen) | |
1797 buf[buflen-1] = '\0'; | |
1798 return buf; | |
1799 } | |
1800 | |
1801 juint Deoptimization::_deoptimization_hist | |
1802 [Deoptimization::Reason_LIMIT] | |
1803 [1 + Deoptimization::Action_LIMIT] | |
1804 [Deoptimization::BC_CASE_LIMIT] | |
1805 = {0}; | |
1806 | |
1807 enum { | |
1808 LSB_BITS = 8, | |
1809 LSB_MASK = right_n_bits(LSB_BITS) | |
1810 }; | |
1811 | |
1812 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, | |
1813 Bytecodes::Code bc) { | |
1814 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); | |
1815 assert(action >= 0 && action < Action_LIMIT, "oob"); | |
1816 _deoptimization_hist[Reason_none][0][0] += 1; // total | |
1817 _deoptimization_hist[reason][0][0] += 1; // per-reason total | |
1818 juint* cases = _deoptimization_hist[reason][1+action]; | |
1819 juint* bc_counter_addr = NULL; | |
1820 juint bc_counter = 0; | |
1821 // Look for an unused counter, or an exact match to this BC. | |
1822 if (bc != Bytecodes::_illegal) { | |
1823 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { | |
1824 juint* counter_addr = &cases[bc_case]; | |
1825 juint counter = *counter_addr; | |
1826 if ((counter == 0 && bc_counter_addr == NULL) | |
1827 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { | |
1828 // this counter is either free or is already devoted to this BC | |
1829 bc_counter_addr = counter_addr; | |
1830 bc_counter = counter | bc; | |
1831 } | |
1832 } | |
1833 } | |
1834 if (bc_counter_addr == NULL) { | |
1835 // Overflow, or no given bytecode. | |
1836 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; | |
1837 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB | |
1838 } | |
1839 *bc_counter_addr = bc_counter + (1 << LSB_BITS); | |
1840 } | |
1841 | |
1842 jint Deoptimization::total_deoptimization_count() { | |
1843 return _deoptimization_hist[Reason_none][0][0]; | |
1844 } | |
1845 | |
1846 jint Deoptimization::deoptimization_count(DeoptReason reason) { | |
1847 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); | |
1848 return _deoptimization_hist[reason][0][0]; | |
1849 } | |
1850 | |
1851 void Deoptimization::print_statistics() { | |
1852 juint total = total_deoptimization_count(); | |
1853 juint account = total; | |
1854 if (total != 0) { | |
1855 ttyLocker ttyl; | |
1856 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); | |
1857 tty->print_cr("Deoptimization traps recorded:"); | |
1858 #define PRINT_STAT_LINE(name, r) \ | |
1859 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); | |
1860 PRINT_STAT_LINE("total", total); | |
1861 // For each non-zero entry in the histogram, print the reason, | |
1862 // the action, and (if specifically known) the type of bytecode. | |
1863 for (int reason = 0; reason < Reason_LIMIT; reason++) { | |
1864 for (int action = 0; action < Action_LIMIT; action++) { | |
1865 juint* cases = _deoptimization_hist[reason][1+action]; | |
1866 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { | |
1867 juint counter = cases[bc_case]; | |
1868 if (counter != 0) { | |
1869 char name[1*K]; | |
1870 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); | |
1871 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) | |
1872 bc = Bytecodes::_illegal; | |
1873 sprintf(name, "%s/%s/%s", | |
1874 trap_reason_name(reason), | |
1875 trap_action_name(action), | |
1876 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); | |
1877 juint r = counter >> LSB_BITS; | |
1878 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); | |
1879 account -= r; | |
1880 } | |
1881 } | |
1882 } | |
1883 } | |
1884 if (account != 0) { | |
1885 PRINT_STAT_LINE("unaccounted", account); | |
1886 } | |
1887 #undef PRINT_STAT_LINE | |
1888 if (xtty != NULL) xtty->tail("statistics"); | |
1889 } | |
1890 } | |
1692 | 1891 #else // COMPILER2 || SHARK |
0 | 1892 |
1893 | |
1894 // Stubs for C1 only system. | |
1895 bool Deoptimization::trap_state_is_recompiled(int trap_state) { | |
1896 return false; | |
1897 } | |
1898 | |
1899 const char* Deoptimization::trap_reason_name(int reason) { | |
1900 return "unknown"; | |
1901 } | |
1902 | |
1903 void Deoptimization::print_statistics() { | |
1904 // no output | |
1905 } | |
1906 | |
1907 void | |
1908 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { | |
1909 // no udpate | |
1910 } | |
1911 | |
1912 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { | |
1913 return 0; | |
1914 } | |
1915 | |
1916 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, | |
1917 Bytecodes::Code bc) { | |
1918 // no update | |
1919 } | |
1920 | |
1921 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, | |
1922 int trap_state) { | |
1923 jio_snprintf(buf, buflen, "#%d", trap_state); | |
1924 return buf; | |
1925 } | |
1926 | |
1692 | 1927 #endif // COMPILER2 || SHARK |