Mercurial > hg > truffle
annotate src/share/vm/opto/memnode.cpp @ 14702:d4dd5204c0aa
Merge
author | ehelin |
---|---|
date | Fri, 14 Mar 2014 13:27:18 +0100 |
parents | bd2220947c1e |
children | 92aa6797d639 |
rev | line source |
---|---|
0 | 1 /* |
14223
de6a9e811145
8029233: Update copyright year to match last edit in jdk8 hotspot repository for 2013
mikael
parents:
14213
diff
changeset
|
2 * Copyright (c) 1997, 2013, 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:
1100
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1100
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:
1100
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "classfile/systemDictionary.hpp" | |
27 #include "compiler/compileLog.hpp" | |
28 #include "memory/allocation.inline.hpp" | |
29 #include "oops/objArrayKlass.hpp" | |
30 #include "opto/addnode.hpp" | |
31 #include "opto/cfgnode.hpp" | |
32 #include "opto/compile.hpp" | |
33 #include "opto/connode.hpp" | |
34 #include "opto/loopnode.hpp" | |
35 #include "opto/machnode.hpp" | |
36 #include "opto/matcher.hpp" | |
37 #include "opto/memnode.hpp" | |
38 #include "opto/mulnode.hpp" | |
39 #include "opto/phaseX.hpp" | |
40 #include "opto/regmask.hpp" | |
41 | |
0 | 42 // Portions of code courtesy of Clifford Click |
43 | |
44 // Optimization - Graph Style | |
45 | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
47 |
0 | 48 //============================================================================= |
49 uint MemNode::size_of() const { return sizeof(*this); } | |
50 | |
51 const TypePtr *MemNode::adr_type() const { | |
52 Node* adr = in(Address); | |
53 const TypePtr* cross_check = NULL; | |
54 DEBUG_ONLY(cross_check = _adr_type); | |
55 return calculate_adr_type(adr->bottom_type(), cross_check); | |
56 } | |
57 | |
58 #ifndef PRODUCT | |
59 void MemNode::dump_spec(outputStream *st) const { | |
60 if (in(Address) == NULL) return; // node is dead | |
61 #ifndef ASSERT | |
62 // fake the missing field | |
63 const TypePtr* _adr_type = NULL; | |
64 if (in(Address) != NULL) | |
65 _adr_type = in(Address)->bottom_type()->isa_ptr(); | |
66 #endif | |
67 dump_adr_type(this, _adr_type, st); | |
68 | |
69 Compile* C = Compile::current(); | |
70 if( C->alias_type(_adr_type)->is_volatile() ) | |
71 st->print(" Volatile!"); | |
72 } | |
73 | |
74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { | |
75 st->print(" @"); | |
76 if (adr_type == NULL) { | |
77 st->print("NULL"); | |
78 } else { | |
79 adr_type->dump_on(st); | |
80 Compile* C = Compile::current(); | |
81 Compile::AliasType* atp = NULL; | |
82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); | |
83 if (atp == NULL) | |
84 st->print(", idx=?\?;"); | |
85 else if (atp->index() == Compile::AliasIdxBot) | |
86 st->print(", idx=Bot;"); | |
87 else if (atp->index() == Compile::AliasIdxTop) | |
88 st->print(", idx=Top;"); | |
89 else if (atp->index() == Compile::AliasIdxRaw) | |
90 st->print(", idx=Raw;"); | |
91 else { | |
92 ciField* field = atp->field(); | |
93 if (field) { | |
94 st->print(", name="); | |
95 field->print_name_on(st); | |
96 } | |
97 st->print(", idx=%d;", atp->index()); | |
98 } | |
99 } | |
100 } | |
101 | |
102 extern void print_alias_types(); | |
103 | |
104 #endif | |
105 | |
10278 | 106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) { |
107 assert((t_oop != NULL), "sanity"); | |
108 bool is_instance = t_oop->is_known_instance_field(); | |
109 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() && | |
110 (load != NULL) && load->is_Load() && | |
111 (phase->is_IterGVN() != NULL); | |
112 if (!(is_instance || is_boxed_value_load)) | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
113 return mchain; // don't try to optimize non-instance types |
10278 | 114 uint instance_id = t_oop->instance_id(); |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
115 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
116 Node *prev = NULL; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
117 Node *result = mchain; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
118 while (prev != result) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
119 prev = result; |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
120 if (result == start_mem) |
605 | 121 break; // hit one of our sentinels |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
122 // skip over a call which does not affect this memory slice |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
123 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
124 Node *proj_in = result->in(0); |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
125 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { |
605 | 126 break; // hit one of our sentinels |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
127 } else if (proj_in->is_Call()) { |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
128 CallNode *call = proj_in->as_Call(); |
10278 | 129 if (!call->may_modify(t_oop, phase)) { // returns false for instances |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
130 result = call->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
131 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
132 } else if (proj_in->is_Initialize()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
133 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
134 // Stop if this is the initialization for the object instance which |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
135 // which contains this memory slice, otherwise skip over it. |
10278 | 136 if ((alloc == NULL) || (alloc->_idx == instance_id)) { |
137 break; | |
138 } | |
139 if (is_instance) { | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
140 result = proj_in->in(TypeFunc::Memory); |
10278 | 141 } else if (is_boxed_value_load) { |
142 Node* klass = alloc->in(AllocateNode::KlassNode); | |
143 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr(); | |
144 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) { | |
145 result = proj_in->in(TypeFunc::Memory); // not related allocation | |
146 } | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
147 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
148 } else if (proj_in->is_MemBar()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
149 result = proj_in->in(TypeFunc::Memory); |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
150 } else { |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
151 assert(false, "unexpected projection"); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
152 } |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
153 } else if (result->is_ClearArray()) { |
10278 | 154 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) { |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
155 // Can not bypass initialization of the instance |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
156 // we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
157 break; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
158 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
159 // Otherwise skip it (the call updated 'result' value). |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
160 } else if (result->is_MergeMem()) { |
10278 | 161 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
162 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
163 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
164 return result; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
165 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
166 |
10278 | 167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) { |
168 const TypeOopPtr* t_oop = t_adr->isa_oopptr(); | |
169 if (t_oop == NULL) | |
170 return mchain; // don't try to optimize non-oop types | |
171 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase); | |
172 bool is_instance = t_oop->is_known_instance_field(); | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
173 PhaseIterGVN *igvn = phase->is_IterGVN(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
174 if (is_instance && igvn != NULL && result->is_Phi()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
175 PhiNode *mphi = result->as_Phi(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
176 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
177 const TypePtr *t = mphi->adr_type(); |
163 | 178 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || |
223 | 179 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && |
247 | 180 t->is_oopptr()->cast_to_exactness(true) |
181 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) | |
182 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
183 // clone the Phi with our address type |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
184 result = mphi->split_out_instance(t_adr, igvn); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
185 } else { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
186 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
187 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
188 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
189 return result; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
190 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
191 |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
193 uint alias_idx = phase->C->get_alias_index(tp); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
194 Node *mem = mmem; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
195 #ifdef ASSERT |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
196 { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
197 // Check that current type is consistent with the alias index used during graph construction |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
198 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
199 bool consistent = adr_check == NULL || adr_check->empty() || |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
200 phase->C->must_alias(adr_check, alias_idx ); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
201 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
202 if( !consistent && adr_check != NULL && !adr_check->empty() && |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
203 tp->isa_aryptr() && tp->offset() == Type::OffsetBot && |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
204 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
205 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
206 adr_check->offset() == oopDesc::klass_offset_in_bytes() || |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
207 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
208 // don't assert if it is dead code. |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
209 consistent = true; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
210 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
211 if( !consistent ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
212 st->print("alias_idx==%d, adr_check==", alias_idx); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
213 if( adr_check == NULL ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
214 st->print("NULL"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
215 } else { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
216 adr_check->dump(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
217 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
218 st->cr(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
219 print_alias_types(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
220 assert(consistent, "adr_check must match alias idx"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
221 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
222 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
223 #endif |
1815 | 224 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
225 // means an array I have not precisely typed yet. Do not do any |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
226 // alias stuff with it any time soon. |
1815 | 227 const TypeOopPtr *toop = tp->isa_oopptr(); |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
228 if( tp->base() != Type::AnyPtr && |
1815 | 229 !(toop && |
230 toop->klass() != NULL && | |
231 toop->klass()->is_java_lang_Object() && | |
232 toop->offset() == Type::OffsetBot) ) { | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
233 // compress paths and change unreachable cycles to TOP |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
234 // If not, we can update the input infinitely along a MergeMem cycle |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
235 // Equivalent code in PhiNode::Ideal |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
236 Node* m = phase->transform(mmem); |
605 | 237 // If transformed to a MergeMem, get the desired slice |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
238 // Otherwise the returned node represents memory for every slice |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
239 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
240 // Update input if it is progress over what we have now |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
241 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
242 return mem; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
243 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
244 |
0 | 245 //--------------------------Ideal_common--------------------------------------- |
246 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. | |
247 // Unhook non-raw memories from complete (macro-expanded) initializations. | |
248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { | |
249 // If our control input is a dead region, kill all below the region | |
250 Node *ctl = in(MemNode::Control); | |
251 if (ctl && remove_dead_region(phase, can_reshape)) | |
252 return this; | |
305 | 253 ctl = in(MemNode::Control); |
254 // Don't bother trying to transform a dead node | |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
255 if (ctl && ctl->is_top()) return NodeSentinel; |
0 | 256 |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
257 PhaseIterGVN *igvn = phase->is_IterGVN(); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
258 // Wait if control on the worklist. |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
259 if (ctl && can_reshape && igvn != NULL) { |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
260 Node* bol = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
261 Node* cmp = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
262 if (ctl->in(0)->is_If()) { |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
263 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity"); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
264 bol = ctl->in(0)->in(1); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
265 if (bol->is_Bool()) |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
266 cmp = ctl->in(0)->in(1)->in(1); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
267 } |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
268 if (igvn->_worklist.member(ctl) || |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
269 (bol != NULL && igvn->_worklist.member(bol)) || |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
270 (cmp != NULL && igvn->_worklist.member(cmp)) ) { |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
271 // This control path may be dead. |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
272 // Delay this memory node transformation until the control is processed. |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
273 phase->is_IterGVN()->_worklist.push(this); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
274 return NodeSentinel; // caller will return NULL |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
275 } |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
276 } |
0 | 277 // Ignore if memory is dead, or self-loop |
278 Node *mem = in(MemNode::Memory); | |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
279 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
280 assert(mem != this, "dead loop in MemNode::Ideal"); |
0 | 281 |
4124
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
282 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) { |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
283 // This memory slice may be dead. |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
284 // Delay this mem node transformation until the memory is processed. |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
285 phase->is_IterGVN()->_worklist.push(this); |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
286 return NodeSentinel; // caller will return NULL |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
287 } |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
288 |
0 | 289 Node *address = in(MemNode::Address); |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
290 const Type *t_adr = phase->type(address); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
291 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
292 |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
293 if (can_reshape && igvn != NULL && |
1826
56601ef83436
6916062: assert(_inserts <= _insert_limit,"hash table overflow") in NodeHash::hash_insert
kvn
parents:
1815
diff
changeset
|
294 (igvn->_worklist.member(address) || |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
295 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) { |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
296 // The address's base and type may change when the address is processed. |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
297 // Delay this mem node transformation until the address is processed. |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
298 phase->is_IterGVN()->_worklist.push(this); |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
299 return NodeSentinel; // caller will return NULL |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
300 } |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
301 |
1062
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
302 // Do NOT remove or optimize the next lines: ensure a new alias index |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
303 // is allocated for an oop pointer type before Escape Analysis. |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
304 // Note: C++ will not remove it since the call has side effect. |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
305 if (t_adr->isa_oopptr()) { |
1062
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
306 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr()); |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
307 } |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
308 |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
309 #ifdef ASSERT |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
310 Node* base = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
311 if (address->is_AddP()) |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
312 base = address->in(AddPNode::Base); |
8695
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
313 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) && |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
314 !t_adr->isa_rawptr()) { |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
315 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true. |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
316 Compile* C = phase->C; |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
317 tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
318 tty->print_cr("===== NULL+offs not RAW address ====="); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
319 if (C->is_dead_node(this->_idx)) tty->print_cr("'this' is dead"); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
320 if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead"); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
321 if (C->is_dead_node(mem->_idx)) tty->print_cr("'mem' is dead"); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
322 if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead"); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
323 if (C->is_dead_node(base->_idx)) tty->print_cr("'base' is dead"); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
324 tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
325 base->dump(1); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
326 tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
327 this->dump(2); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
328 tty->print("this->adr_type(): "); adr_type()->dump(); tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
329 tty->print("phase->type(address): "); t_adr->dump(); tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
330 tty->print("phase->type(base): "); phase->type(address)->dump(); tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
331 tty->cr(); |
ff55877839bc
8009472: Print additional information for 8004640 failure
kvn
parents:
8116
diff
changeset
|
332 } |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
333 assert(base == NULL || t_adr->isa_rawptr() || |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
334 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?"); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
335 #endif |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
336 |
0 | 337 // Avoid independent memory operations |
338 Node* old_mem = mem; | |
339 | |
36 | 340 // The code which unhooks non-raw memories from complete (macro-expanded) |
341 // initializations was removed. After macro-expansion all stores catched | |
342 // by Initialize node became raw stores and there is no information | |
343 // which memory slices they modify. So it is unsafe to move any memory | |
344 // operation above these stores. Also in most cases hooked non-raw memories | |
345 // were already unhooked by using information from detect_ptr_independence() | |
346 // and find_previous_store(). | |
0 | 347 |
348 if (mem->is_MergeMem()) { | |
349 MergeMemNode* mmem = mem->as_MergeMem(); | |
350 const TypePtr *tp = t_adr->is_ptr(); | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
351 |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
352 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); |
0 | 353 } |
354 | |
355 if (mem != old_mem) { | |
356 set_req(MemNode::Memory, mem); | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
357 if (can_reshape && old_mem->outcnt() == 0) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
358 igvn->_worklist.push(old_mem); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
359 } |
305 | 360 if (phase->type( mem ) == Type::TOP) return NodeSentinel; |
0 | 361 return this; |
362 } | |
363 | |
364 // let the subclass continue analyzing... | |
365 return NULL; | |
366 } | |
367 | |
368 // Helper function for proving some simple control dominations. | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
369 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. |
0 | 370 // Already assumes that 'dom' is available at 'sub', and that 'sub' |
371 // is not a constant (dominated by the method's StartNode). | |
372 // Used by MemNode::find_previous_store to prove that the | |
373 // control input of a memory operation predates (dominates) | |
374 // an allocation it wants to look past. | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
375 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
376 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
377 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
378 |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
379 // Check 'dom'. Skip Proj and CatchProj nodes. |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
380 dom = dom->find_exact_control(dom); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
381 if (dom == NULL || dom->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
382 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
383 |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
384 if (dom == sub) { |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
385 // For the case when, for example, 'sub' is Initialize and the original |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
386 // 'dom' is Proj node of the 'sub'. |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
387 return false; |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
388 } |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
389 |
155
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
390 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
391 return true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
392 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
393 // 'dom' dominates 'sub' if its control edge and control edges |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
394 // of all its inputs dominate or equal to sub's control edge. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
395 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
396 // Currently 'sub' is either Allocate, Initialize or Start nodes. |
163 | 397 // Or Region for the check in LoadNode::Ideal(); |
398 // 'sub' should have sub->in(0) != NULL. | |
399 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || | |
10278 | 400 sub->is_Region() || sub->is_Call(), "expecting only these nodes"); |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
401 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
402 // Get control edge of 'sub'. |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
403 Node* orig_sub = sub; |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
404 sub = sub->find_exact_control(sub->in(0)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
405 if (sub == NULL || sub->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
406 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
407 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
408 assert(sub->is_CFG(), "expecting control"); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
409 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
410 if (sub == dom) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
411 return true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
412 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
413 if (sub->is_Start() || sub->is_Root()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
414 return false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
415 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
416 { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
417 // Check all control edges of 'dom'. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
418 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
419 ResourceMark rm; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
420 Arena* arena = Thread::current()->resource_area(); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
421 Node_List nlist(arena); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
422 Unique_Node_List dom_list(arena); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
423 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
424 dom_list.push(dom); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
425 bool only_dominating_controls = false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
426 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
427 for (uint next = 0; next < dom_list.size(); next++) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
428 Node* n = dom_list.at(next); |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
429 if (n == orig_sub) |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
430 return false; // One of dom's inputs dominated by sub. |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
431 if (!n->is_CFG() && n->pinned()) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
432 // Check only own control edge for pinned non-control nodes. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
433 n = n->find_exact_control(n->in(0)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
434 if (n == NULL || n->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
435 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
436 assert(n->is_CFG(), "expecting control"); |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
437 dom_list.push(n); |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
438 } else if (n->is_Con() || n->is_Start() || n->is_Root()) { |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
439 only_dominating_controls = true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
440 } else if (n->is_CFG()) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
441 if (n->dominates(sub, nlist)) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
442 only_dominating_controls = true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
443 else |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
444 return false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
445 } else { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
446 // First, own control edge. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
447 Node* m = n->find_exact_control(n->in(0)); |
155
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
448 if (m != NULL) { |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
449 if (m->is_top()) |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
450 return false; // Conservative answer for dead code |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
451 dom_list.push(m); |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
452 } |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
453 // Now, the rest of edges. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
454 uint cnt = n->req(); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
455 for (uint i = 1; i < cnt; i++) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
456 m = n->find_exact_control(n->in(i)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
457 if (m == NULL || m->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
458 continue; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
459 dom_list.push(m); |
0 | 460 } |
461 } | |
462 } | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
463 return only_dominating_controls; |
0 | 464 } |
465 } | |
466 | |
467 //---------------------detect_ptr_independence--------------------------------- | |
468 // Used by MemNode::find_previous_store to prove that two base | |
469 // pointers are never equal. | |
470 // The pointers are accompanied by their associated allocations, | |
471 // if any, which have been previously discovered by the caller. | |
472 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, | |
473 Node* p2, AllocateNode* a2, | |
474 PhaseTransform* phase) { | |
475 // Attempt to prove that these two pointers cannot be aliased. | |
476 // They may both manifestly be allocations, and they should differ. | |
477 // Or, if they are not both allocations, they can be distinct constants. | |
478 // Otherwise, one is an allocation and the other a pre-existing value. | |
479 if (a1 == NULL && a2 == NULL) { // neither an allocation | |
480 return (p1 != p2) && p1->is_Con() && p2->is_Con(); | |
481 } else if (a1 != NULL && a2 != NULL) { // both allocations | |
482 return (a1 != a2); | |
483 } else if (a1 != NULL) { // one allocation a1 | |
484 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
485 return all_controls_dominate(p2, a1); |
0 | 486 } else { //(a2 != NULL) // one allocation a2 |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
487 return all_controls_dominate(p1, a2); |
0 | 488 } |
489 return false; | |
490 } | |
491 | |
492 | |
493 // The logic for reordering loads and stores uses four steps: | |
494 // (a) Walk carefully past stores and initializations which we | |
495 // can prove are independent of this load. | |
496 // (b) Observe that the next memory state makes an exact match | |
497 // with self (load or store), and locate the relevant store. | |
498 // (c) Ensure that, if we were to wire self directly to the store, | |
499 // the optimizer would fold it up somehow. | |
500 // (d) Do the rewiring, and return, depending on some other part of | |
501 // the optimizer to fold up the load. | |
502 // This routine handles steps (a) and (b). Steps (c) and (d) are | |
503 // specific to loads and stores, so they are handled by the callers. | |
504 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) | |
505 // | |
506 Node* MemNode::find_previous_store(PhaseTransform* phase) { | |
507 Node* ctrl = in(MemNode::Control); | |
508 Node* adr = in(MemNode::Address); | |
509 intptr_t offset = 0; | |
510 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
511 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); | |
512 | |
513 if (offset == Type::OffsetBot) | |
514 return NULL; // cannot unalias unless there are precise offsets | |
515 | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
516 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
517 |
0 | 518 intptr_t size_in_bytes = memory_size(); |
519 | |
520 Node* mem = in(MemNode::Memory); // start searching here... | |
521 | |
522 int cnt = 50; // Cycle limiter | |
523 for (;;) { // While we can dance past unrelated stores... | |
524 if (--cnt < 0) break; // Caught in cycle or a complicated dance? | |
525 | |
526 if (mem->is_Store()) { | |
527 Node* st_adr = mem->in(MemNode::Address); | |
528 intptr_t st_offset = 0; | |
529 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); | |
530 if (st_base == NULL) | |
531 break; // inscrutable pointer | |
532 if (st_offset != offset && st_offset != Type::OffsetBot) { | |
533 const int MAX_STORE = BytesPerLong; | |
534 if (st_offset >= offset + size_in_bytes || | |
535 st_offset <= offset - MAX_STORE || | |
536 st_offset <= offset - mem->as_Store()->memory_size()) { | |
537 // Success: The offsets are provably independent. | |
538 // (You may ask, why not just test st_offset != offset and be done? | |
539 // The answer is that stores of different sizes can co-exist | |
540 // in the same sequence of RawMem effects. We sometimes initialize | |
541 // a whole 'tile' of array elements with a single jint or jlong.) | |
542 mem = mem->in(MemNode::Memory); | |
543 continue; // (a) advance through independent store memory | |
544 } | |
545 } | |
546 if (st_base != base && | |
547 detect_ptr_independence(base, alloc, | |
548 st_base, | |
549 AllocateNode::Ideal_allocation(st_base, phase), | |
550 phase)) { | |
551 // Success: The bases are provably independent. | |
552 mem = mem->in(MemNode::Memory); | |
553 continue; // (a) advance through independent store memory | |
554 } | |
555 | |
556 // (b) At this point, if the bases or offsets do not agree, we lose, | |
557 // since we have not managed to prove 'this' and 'mem' independent. | |
558 if (st_base == base && st_offset == offset) { | |
559 return mem; // let caller handle steps (c), (d) | |
560 } | |
561 | |
562 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { | |
563 InitializeNode* st_init = mem->in(0)->as_Initialize(); | |
564 AllocateNode* st_alloc = st_init->allocation(); | |
565 if (st_alloc == NULL) | |
566 break; // something degenerated | |
567 bool known_identical = false; | |
568 bool known_independent = false; | |
569 if (alloc == st_alloc) | |
570 known_identical = true; | |
571 else if (alloc != NULL) | |
572 known_independent = true; | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
573 else if (all_controls_dominate(this, st_alloc)) |
0 | 574 known_independent = true; |
575 | |
576 if (known_independent) { | |
577 // The bases are provably independent: Either they are | |
578 // manifestly distinct allocations, or else the control | |
579 // of this load dominates the store's allocation. | |
580 int alias_idx = phase->C->get_alias_index(adr_type()); | |
581 if (alias_idx == Compile::AliasIdxRaw) { | |
582 mem = st_alloc->in(TypeFunc::Memory); | |
583 } else { | |
584 mem = st_init->memory(alias_idx); | |
585 } | |
586 continue; // (a) advance through independent store memory | |
587 } | |
588 | |
589 // (b) at this point, if we are not looking at a store initializing | |
590 // the same allocation we are loading from, we lose. | |
591 if (known_identical) { | |
592 // From caller, can_see_stored_value will consult find_captured_store. | |
593 return mem; // let caller handle steps (c), (d) | |
594 } | |
595 | |
223 | 596 } else if (addr_t != NULL && addr_t->is_known_instance_field()) { |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
597 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
598 if (mem->is_Proj() && mem->in(0)->is_Call()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
599 CallNode *call = mem->in(0)->as_Call(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
600 if (!call->may_modify(addr_t, phase)) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
601 mem = call->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
602 continue; // (a) advance through independent call memory |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
603 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
604 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
605 mem = mem->in(0)->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
606 continue; // (a) advance through independent MemBar memory |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
607 } else if (mem->is_ClearArray()) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
608 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
609 // (the call updated 'mem' value) |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
610 continue; // (a) advance through independent allocation memory |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
611 } else { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
612 // Can not bypass initialization of the instance |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
613 // we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
614 return mem; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
615 } |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
616 } else if (mem->is_MergeMem()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
617 int alias_idx = phase->C->get_alias_index(adr_type()); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
618 mem = mem->as_MergeMem()->memory_at(alias_idx); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
619 continue; // (a) advance through independent MergeMem memory |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
620 } |
0 | 621 } |
622 | |
623 // Unless there is an explicit 'continue', we must bail out here, | |
624 // because 'mem' is an inscrutable memory state (e.g., a call). | |
625 break; | |
626 } | |
627 | |
628 return NULL; // bail out | |
629 } | |
630 | |
631 //----------------------calculate_adr_type------------------------------------- | |
632 // Helper function. Notices when the given type of address hits top or bottom. | |
633 // Also, asserts a cross-check of the type against the expected address type. | |
634 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { | |
635 if (t == Type::TOP) return NULL; // does not touch memory any more? | |
636 #ifdef PRODUCT | |
637 cross_check = NULL; | |
638 #else | |
639 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; | |
640 #endif | |
641 const TypePtr* tp = t->isa_ptr(); | |
642 if (tp == NULL) { | |
643 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); | |
644 return TypePtr::BOTTOM; // touches lots of memory | |
645 } else { | |
646 #ifdef ASSERT | |
647 // %%%% [phh] We don't check the alias index if cross_check is | |
648 // TypeRawPtr::BOTTOM. Needs to be investigated. | |
649 if (cross_check != NULL && | |
650 cross_check != TypePtr::BOTTOM && | |
651 cross_check != TypeRawPtr::BOTTOM) { | |
652 // Recheck the alias index, to see if it has changed (due to a bug). | |
653 Compile* C = Compile::current(); | |
654 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), | |
655 "must stay in the original alias category"); | |
656 // The type of the address must be contained in the adr_type, | |
657 // disregarding "null"-ness. | |
658 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) | |
659 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); | |
14383
5ec7dace41a6
8027422: assert(_gvn.type(obj)->higher_equal(tjp)) failed: cast_up is no longer needed
roland
parents:
14223
diff
changeset
|
660 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(), |
0 | 661 "real address must not escape from expected memory type"); |
662 } | |
663 #endif | |
664 return tp; | |
665 } | |
666 } | |
667 | |
668 //------------------------adr_phi_is_loop_invariant---------------------------- | |
669 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted | |
670 // loop is loop invariant. Make a quick traversal of Phi and associated | |
671 // CastPP nodes, looking to see if they are a closed group within the loop. | |
672 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { | |
673 // The idea is that the phi-nest must boil down to only CastPP nodes | |
674 // with the same data. This implies that any path into the loop already | |
675 // includes such a CastPP, and so the original cast, whatever its input, | |
676 // must be covered by an equivalent cast, with an earlier control input. | |
677 ResourceMark rm; | |
678 | |
679 // The loop entry input of the phi should be the unique dominating | |
680 // node for every Phi/CastPP in the loop. | |
681 Unique_Node_List closure; | |
682 closure.push(adr_phi->in(LoopNode::EntryControl)); | |
683 | |
684 // Add the phi node and the cast to the worklist. | |
685 Unique_Node_List worklist; | |
686 worklist.push(adr_phi); | |
687 if( cast != NULL ){ | |
688 if( !cast->is_ConstraintCast() ) return false; | |
689 worklist.push(cast); | |
690 } | |
691 | |
692 // Begin recursive walk of phi nodes. | |
693 while( worklist.size() ){ | |
694 // Take a node off the worklist | |
695 Node *n = worklist.pop(); | |
696 if( !closure.member(n) ){ | |
697 // Add it to the closure. | |
698 closure.push(n); | |
699 // Make a sanity check to ensure we don't waste too much time here. | |
700 if( closure.size() > 20) return false; | |
701 // This node is OK if: | |
702 // - it is a cast of an identical value | |
703 // - or it is a phi node (then we add its inputs to the worklist) | |
704 // Otherwise, the node is not OK, and we presume the cast is not invariant | |
705 if( n->is_ConstraintCast() ){ | |
706 worklist.push(n->in(1)); | |
707 } else if( n->is_Phi() ) { | |
708 for( uint i = 1; i < n->req(); i++ ) { | |
709 worklist.push(n->in(i)); | |
710 } | |
711 } else { | |
712 return false; | |
713 } | |
714 } | |
715 } | |
716 | |
717 // Quit when the worklist is empty, and we've found no offending nodes. | |
718 return true; | |
719 } | |
720 | |
721 //------------------------------Ideal_DU_postCCP------------------------------- | |
722 // Find any cast-away of null-ness and keep its control. Null cast-aways are | |
723 // going away in this pass and we need to make this memory op depend on the | |
724 // gating null check. | |
163 | 725 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { |
726 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); | |
727 } | |
0 | 728 |
729 // I tried to leave the CastPP's in. This makes the graph more accurate in | |
730 // some sense; we get to keep around the knowledge that an oop is not-null | |
731 // after some test. Alas, the CastPP's interfere with GVN (some values are | |
732 // the regular oop, some are the CastPP of the oop, all merge at Phi's which | |
733 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed | |
734 // some of the more trivial cases in the optimizer. Removing more useless | |
735 // Phi's started allowing Loads to illegally float above null checks. I gave | |
736 // up on this approach. CNC 10/20/2000 | |
163 | 737 // This static method may be called not from MemNode (EncodePNode calls it). |
738 // Only the control edge of the node 'n' might be updated. | |
739 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { | |
0 | 740 Node *skipped_cast = NULL; |
741 // Need a null check? Regular static accesses do not because they are | |
742 // from constant addresses. Array ops are gated by the range check (which | |
743 // always includes a NULL check). Just check field ops. | |
163 | 744 if( n->in(MemNode::Control) == NULL ) { |
0 | 745 // Scan upwards for the highest location we can place this memory op. |
746 while( true ) { | |
747 switch( adr->Opcode() ) { | |
748 | |
749 case Op_AddP: // No change to NULL-ness, so peek thru AddP's | |
750 adr = adr->in(AddPNode::Base); | |
751 continue; | |
752 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
753 case Op_DecodeN: // No change to NULL-ness, so peek thru |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
754 case Op_DecodeNKlass: |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
755 adr = adr->in(1); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
756 continue; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
757 |
6139
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
758 case Op_EncodeP: |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
759 case Op_EncodePKlass: |
6139
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
760 // EncodeP node's control edge could be set by this method |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
761 // when EncodeP node depends on CastPP node. |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
762 // |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
763 // Use its control edge for memory op because EncodeP may go away |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
764 // later when it is folded with following or preceding DecodeN node. |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
765 if (adr->in(0) == NULL) { |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
766 // Keep looking for cast nodes. |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
767 adr = adr->in(1); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
768 continue; |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
769 } |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
770 ccp->hash_delete(n); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
771 n->set_req(MemNode::Control, adr->in(0)); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
772 ccp->hash_insert(n); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
773 return n; |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
774 |
0 | 775 case Op_CastPP: |
776 // If the CastPP is useless, just peek on through it. | |
777 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { | |
778 // Remember the cast that we've peeked though. If we peek | |
779 // through more than one, then we end up remembering the highest | |
780 // one, that is, if in a loop, the one closest to the top. | |
781 skipped_cast = adr; | |
782 adr = adr->in(1); | |
783 continue; | |
784 } | |
785 // CastPP is going away in this pass! We need this memory op to be | |
786 // control-dependent on the test that is guarding the CastPP. | |
163 | 787 ccp->hash_delete(n); |
788 n->set_req(MemNode::Control, adr->in(0)); | |
789 ccp->hash_insert(n); | |
790 return n; | |
0 | 791 |
792 case Op_Phi: | |
793 // Attempt to float above a Phi to some dominating point. | |
794 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { | |
795 // If we've already peeked through a Cast (which could have set the | |
796 // control), we can't float above a Phi, because the skipped Cast | |
797 // may not be loop invariant. | |
798 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { | |
799 adr = adr->in(1); | |
800 continue; | |
801 } | |
802 } | |
803 | |
804 // Intentional fallthrough! | |
805 | |
806 // No obvious dominating point. The mem op is pinned below the Phi | |
807 // by the Phi itself. If the Phi goes away (no true value is merged) | |
808 // then the mem op can float, but not indefinitely. It must be pinned | |
809 // behind the controls leading to the Phi. | |
810 case Op_CheckCastPP: | |
811 // These usually stick around to change address type, however a | |
812 // useless one can be elided and we still need to pick up a control edge | |
813 if (adr->in(0) == NULL) { | |
814 // This CheckCastPP node has NO control and is likely useless. But we | |
815 // need check further up the ancestor chain for a control input to keep | |
816 // the node in place. 4959717. | |
817 skipped_cast = adr; | |
818 adr = adr->in(1); | |
819 continue; | |
820 } | |
163 | 821 ccp->hash_delete(n); |
822 n->set_req(MemNode::Control, adr->in(0)); | |
823 ccp->hash_insert(n); | |
824 return n; | |
0 | 825 |
826 // List of "safe" opcodes; those that implicitly block the memory | |
827 // op below any null check. | |
828 case Op_CastX2P: // no null checks on native pointers | |
829 case Op_Parm: // 'this' pointer is not null | |
830 case Op_LoadP: // Loading from within a klass | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
831 case Op_LoadN: // Loading from within a klass |
0 | 832 case Op_LoadKlass: // Loading from within a klass |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
833 case Op_LoadNKlass: // Loading from within a klass |
0 | 834 case Op_ConP: // Loading from a klass |
163 | 835 case Op_ConN: // Loading from a klass |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
836 case Op_ConNKlass: // Loading from a klass |
0 | 837 case Op_CreateEx: // Sucking up the guts of an exception oop |
838 case Op_Con: // Reading from TLS | |
839 case Op_CMoveP: // CMoveP is pinned | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
840 case Op_CMoveN: // CMoveN is pinned |
0 | 841 break; // No progress |
842 | |
843 case Op_Proj: // Direct call to an allocation routine | |
844 case Op_SCMemProj: // Memory state from store conditional ops | |
845 #ifdef ASSERT | |
846 { | |
847 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); | |
848 const Node* call = adr->in(0); | |
163 | 849 if (call->is_CallJava()) { |
850 const CallJavaNode* call_java = call->as_CallJava(); | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
851 const TypeTuple *r = call_java->tf()->range(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
852 assert(r->cnt() > TypeFunc::Parms, "must return value"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
853 const Type* ret_type = r->field_at(TypeFunc::Parms); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
854 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); |
0 | 855 // We further presume that this is one of |
856 // new_instance_Java, new_array_Java, or | |
857 // the like, but do not assert for this. | |
858 } else if (call->is_Allocate()) { | |
859 // similar case to new_instance_Java, etc. | |
860 } else if (!call->is_CallLeaf()) { | |
861 // Projections from fetch_oop (OSR) are allowed as well. | |
862 ShouldNotReachHere(); | |
863 } | |
864 } | |
865 #endif | |
866 break; | |
867 default: | |
868 ShouldNotReachHere(); | |
869 } | |
870 break; | |
871 } | |
872 } | |
873 | |
874 return NULL; // No progress | |
875 } | |
876 | |
877 | |
878 //============================================================================= | |
879 uint LoadNode::size_of() const { return sizeof(*this); } | |
880 uint LoadNode::cmp( const Node &n ) const | |
881 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } | |
882 const Type *LoadNode::bottom_type() const { return _type; } | |
883 uint LoadNode::ideal_reg() const { | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
884 return _type->ideal_reg(); |
0 | 885 } |
886 | |
887 #ifndef PRODUCT | |
888 void LoadNode::dump_spec(outputStream *st) const { | |
889 MemNode::dump_spec(st); | |
890 if( !Verbose && !WizardMode ) { | |
891 // standard dump does this in Verbose and WizardMode | |
892 st->print(" #"); _type->dump_on(st); | |
893 } | |
894 } | |
895 #endif | |
896 | |
1609 | 897 #ifdef ASSERT |
898 //----------------------------is_immutable_value------------------------------- | |
899 // Helper function to allow a raw load without control edge for some cases | |
900 bool LoadNode::is_immutable_value(Node* adr) { | |
901 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && | |
902 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && | |
903 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == | |
904 in_bytes(JavaThread::osthread_offset()))); | |
905 } | |
906 #endif | |
0 | 907 |
908 //----------------------------LoadNode::make----------------------------------- | |
909 // Polymorphic factory method: | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
910 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) { |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
911 Compile* C = gvn.C; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
912 |
0 | 913 // sanity check the alias category against the created node type |
914 assert(!(adr_type->isa_oopptr() && | |
915 adr_type->offset() == oopDesc::klass_offset_in_bytes()), | |
916 "use LoadKlassNode instead"); | |
917 assert(!(adr_type->isa_aryptr() && | |
918 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), | |
919 "use LoadRangeNode instead"); | |
1609 | 920 // Check control edge of raw loads |
921 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || | |
922 // oop will be recorded in oop map if load crosses safepoint | |
923 rt->isa_oopptr() || is_immutable_value(adr), | |
924 "raw memory operations should have control edge"); | |
0 | 925 switch (bt) { |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
926 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
927 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
928 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
929 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
930 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
931 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
932 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
933 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
934 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
935 case T_OBJECT: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
936 #ifdef _LP64 |
163 | 937 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
938 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo)); |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
939 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr()); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
940 } else |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
941 #endif |
163 | 942 { |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
943 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop"); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
944 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo); |
163 | 945 } |
0 | 946 } |
947 ShouldNotReachHere(); | |
948 return (LoadNode*)NULL; | |
949 } | |
950 | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
951 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) { |
0 | 952 bool require_atomic = true; |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
953 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic); |
0 | 954 } |
955 | |
956 | |
957 | |
958 | |
959 //------------------------------hash------------------------------------------- | |
960 uint LoadNode::hash() const { | |
961 // unroll addition of interesting fields | |
962 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); | |
963 } | |
964 | |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
965 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) { |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
966 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) { |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
967 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile(); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
968 bool is_stable_ary = FoldStableValues && |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
969 (tp != NULL) && (tp->isa_aryptr() != NULL) && |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
970 tp->isa_aryptr()->is_stable(); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
971 |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
972 return (eliminate_boxing && non_volatile) || is_stable_ary; |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
973 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
974 |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
975 return false; |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
976 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
977 |
0 | 978 //---------------------------can_see_stored_value------------------------------ |
979 // This routine exists to make sure this set of tests is done the same | |
980 // everywhere. We need to make a coordinated change: first LoadNode::Ideal | |
981 // will change the graph shape in a way which makes memory alive twice at the | |
982 // same time (uses the Oracle model of aliasing), then some | |
983 // LoadXNode::Identity will fold things back to the equivalence-class model | |
984 // of aliasing. | |
985 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { | |
986 Node* ld_adr = in(MemNode::Address); | |
10278 | 987 intptr_t ld_off = 0; |
988 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
989 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); |
10278 | 990 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL; |
991 // This is more general than load from boxing objects. | |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
992 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) { |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
993 uint alias_idx = atp->index(); |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
994 bool final = !atp->is_rewritable(); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
995 Node* result = NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
996 Node* current = st; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
997 // Skip through chains of MemBarNodes checking the MergeMems for |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
998 // new states for the slice of this load. Stop once any other |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
999 // kind of node is encountered. Loads from final memory can skip |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1000 // through any kind of MemBar but normal loads shouldn't skip |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1001 // through MemBarAcquire since the could allow them to move out of |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1002 // a synchronized region. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1003 while (current->is_Proj()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1004 int opc = current->in(0)->Opcode(); |
14439
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1005 if ((final && (opc == Op_MemBarAcquire || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1006 opc == Op_MemBarAcquireLock || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1007 opc == Op_LoadFence)) || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1008 opc == Op_MemBarRelease || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1009 opc == Op_StoreFence || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1010 opc == Op_MemBarReleaseLock || |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
1011 opc == Op_MemBarCPUOrder) { |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1012 Node* mem = current->in(0)->in(TypeFunc::Memory); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1013 if (mem->is_MergeMem()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1014 MergeMemNode* merge = mem->as_MergeMem(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1015 Node* new_st = merge->memory_at(alias_idx); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1016 if (new_st == merge->base_memory()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1017 // Keep searching |
10278 | 1018 current = new_st; |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1019 continue; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1020 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1021 // Save the new memory state for the slice and fall through |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1022 // to exit. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1023 result = new_st; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1024 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1025 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1026 break; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1027 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1028 if (result != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1029 st = result; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1030 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1031 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1032 |
0 | 1033 // Loop around twice in the case Load -> Initialize -> Store. |
1034 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) | |
1035 for (int trip = 0; trip <= 1; trip++) { | |
1036 | |
1037 if (st->is_Store()) { | |
1038 Node* st_adr = st->in(MemNode::Address); | |
1039 if (!phase->eqv(st_adr, ld_adr)) { | |
1040 // Try harder before giving up... Match raw and non-raw pointers. | |
1041 intptr_t st_off = 0; | |
1042 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); | |
1043 if (alloc == NULL) return NULL; | |
10278 | 1044 if (alloc != ld_alloc) return NULL; |
0 | 1045 if (ld_off != st_off) return NULL; |
1046 // At this point we have proven something like this setup: | |
1047 // A = Allocate(...) | |
1048 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) | |
1049 // S = StoreQ(, AddP(, A.Parm , #Off), V) | |
1050 // (Actually, we haven't yet proven the Q's are the same.) | |
1051 // In other words, we are loading from a casted version of | |
1052 // the same pointer-and-offset that we stored to. | |
1053 // Thus, we are able to replace L by V. | |
1054 } | |
1055 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. | |
1056 if (store_Opcode() != st->Opcode()) | |
1057 return NULL; | |
1058 return st->in(MemNode::ValueIn); | |
1059 } | |
1060 | |
1061 // A load from a freshly-created object always returns zero. | |
1062 // (This can happen after LoadNode::Ideal resets the load's memory input | |
1063 // to find_captured_store, which returned InitializeNode::zero_memory.) | |
1064 if (st->is_Proj() && st->in(0)->is_Allocate() && | |
10278 | 1065 (st->in(0) == ld_alloc) && |
1066 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) { | |
0 | 1067 // return a zero value for the load's basic type |
1068 // (This is one of the few places where a generic PhaseTransform | |
1069 // can create new nodes. Think of it as lazily manifesting | |
1070 // virtually pre-existing constants.) | |
1071 return phase->zerocon(memory_type()); | |
1072 } | |
1073 | |
1074 // A load from an initialization barrier can match a captured store. | |
1075 if (st->is_Proj() && st->in(0)->is_Initialize()) { | |
1076 InitializeNode* init = st->in(0)->as_Initialize(); | |
1077 AllocateNode* alloc = init->allocation(); | |
10278 | 1078 if ((alloc != NULL) && (alloc == ld_alloc)) { |
0 | 1079 // examine a captured store value |
10278 | 1080 st = init->find_captured_store(ld_off, memory_size(), phase); |
0 | 1081 if (st != NULL) |
1082 continue; // take one more trip around | |
1083 } | |
1084 } | |
1085 | |
10278 | 1086 // Load boxed value from result of valueOf() call is input parameter. |
1087 if (this->is_Load() && ld_adr->is_AddP() && | |
1088 (tp != NULL) && tp->is_ptr_to_boxed_value()) { | |
1089 intptr_t ignore = 0; | |
1090 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore); | |
1091 if (base != NULL && base->is_Proj() && | |
1092 base->as_Proj()->_con == TypeFunc::Parms && | |
1093 base->in(0)->is_CallStaticJava() && | |
1094 base->in(0)->as_CallStaticJava()->is_boxing_method()) { | |
1095 return base->in(0)->in(TypeFunc::Parms); | |
1096 } | |
1097 } | |
1098 | |
0 | 1099 break; |
1100 } | |
1101 | |
1102 return NULL; | |
1103 } | |
1104 | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1105 //----------------------is_instance_field_load_with_local_phi------------------ |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1106 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { |
10278 | 1107 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl && |
1108 in(Address)->is_AddP() ) { | |
1109 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr(); | |
1110 // Only instances and boxed values. | |
1111 if( t_oop != NULL && | |
1112 (t_oop->is_ptr_to_boxed_value() || | |
1113 t_oop->is_known_instance_field()) && | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1114 t_oop->offset() != Type::OffsetBot && |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1115 t_oop->offset() != Type::OffsetTop) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1116 return true; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1117 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1118 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1119 return false; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1120 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1121 |
0 | 1122 //------------------------------Identity--------------------------------------- |
1123 // Loads are identity if previous store is to same address | |
1124 Node *LoadNode::Identity( PhaseTransform *phase ) { | |
1125 // If the previous store-maker is the right kind of Store, and the store is | |
1126 // to the same address, then we are equal to the value stored. | |
10278 | 1127 Node* mem = in(Memory); |
0 | 1128 Node* value = can_see_stored_value(mem, phase); |
1129 if( value ) { | |
1130 // byte, short & char stores truncate naturally. | |
1131 // A load has to load the truncated value which requires | |
1132 // some sort of masking operation and that requires an | |
1133 // Ideal call instead of an Identity call. | |
1134 if (memory_size() < BytesPerInt) { | |
1135 // If the input to the store does not fit with the load's result type, | |
1136 // it must be truncated via an Ideal call. | |
1137 if (!phase->type(value)->higher_equal(phase->type(this))) | |
1138 return this; | |
1139 } | |
1140 // (This works even when value is a Con, but LoadNode::Value | |
1141 // usually runs first, producing the singleton type of the Con.) | |
1142 return value; | |
1143 } | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1144 |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1145 // Search for an existing data phi which was generated before for the same |
605 | 1146 // instance's field to avoid infinite generation of phis in a loop. |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1147 Node *region = mem->in(0); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1148 if (is_instance_field_load_with_local_phi(region)) { |
10278 | 1149 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr(); |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1150 int this_index = phase->C->get_alias_index(addr_t); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1151 int this_offset = addr_t->offset(); |
10278 | 1152 int this_iid = addr_t->instance_id(); |
1153 if (!addr_t->is_known_instance() && | |
1154 addr_t->is_ptr_to_boxed_value()) { | |
1155 // Use _idx of address base (could be Phi node) for boxed values. | |
1156 intptr_t ignore = 0; | |
1157 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); | |
1158 this_iid = base->_idx; | |
1159 } | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1160 const Type* this_type = bottom_type(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1161 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1162 Node* phi = region->fast_out(i); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1163 if (phi->is_Phi() && phi != mem && |
10278 | 1164 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) { |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1165 return phi; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1166 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1167 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1168 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1169 |
0 | 1170 return this; |
1171 } | |
1172 | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1173 // We're loading from an object which has autobox behaviour. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1174 // If this object is result of a valueOf call we'll have a phi |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1175 // merging a newly allocated object and a load from the cache. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1176 // We want to replace this load with the original incoming |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1177 // argument to the valueOf call. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1178 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { |
10278 | 1179 assert(phase->C->eliminate_boxing(), "sanity"); |
1180 intptr_t ignore = 0; | |
1181 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore); | |
1182 if ((base == NULL) || base->is_Phi()) { | |
1183 // Push the loads from the phi that comes from valueOf up | |
1184 // through it to allow elimination of the loads and the recovery | |
1185 // of the original value. It is done in split_through_phi(). | |
1186 return NULL; | |
583
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1187 } else if (base->is_Load() || |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1188 base->is_DecodeN() && base->in(1)->is_Load()) { |
10278 | 1189 // Eliminate the load of boxed value for integer types from the cache |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1190 // array by deriving the value from the index into the array. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1191 // Capture the offset of the load and then reverse the computation. |
10278 | 1192 |
1193 // Get LoadN node which loads a boxing object from 'cache' array. | |
1194 if (base->is_DecodeN()) { | |
1195 base = base->in(1); | |
1196 } | |
1197 if (!base->in(Address)->is_AddP()) { | |
1198 return NULL; // Complex address | |
583
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1199 } |
10278 | 1200 AddPNode* address = base->in(Address)->as_AddP(); |
1201 Node* cache_base = address->in(AddPNode::Base); | |
1202 if ((cache_base != NULL) && cache_base->is_DecodeN()) { | |
1203 // Get ConP node which is static 'cache' field. | |
1204 cache_base = cache_base->in(1); | |
1205 } | |
1206 if ((cache_base != NULL) && cache_base->is_Con()) { | |
1207 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr(); | |
1208 if ((base_type != NULL) && base_type->is_autobox_cache()) { | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1209 Node* elements[4]; |
10278 | 1210 int shift = exact_log2(type2aelembytes(T_OBJECT)); |
1211 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements)); | |
1212 if ((count > 0) && elements[0]->is_Con() && | |
1213 ((count == 1) || | |
1214 (count == 2) && elements[1]->Opcode() == Op_LShiftX && | |
1215 elements[1]->in(2) == phase->intcon(shift))) { | |
1216 ciObjArray* array = base_type->const_oop()->as_obj_array(); | |
1217 // Fetch the box object cache[0] at the base of the array and get its value | |
1218 ciInstance* box = array->obj_at(0)->as_instance(); | |
1219 ciInstanceKlass* ik = box->klass()->as_instance_klass(); | |
1220 assert(ik->is_box_klass(), "sanity"); | |
1221 assert(ik->nof_nonstatic_fields() == 1, "change following code"); | |
1222 if (ik->nof_nonstatic_fields() == 1) { | |
1223 // This should be true nonstatic_field_at requires calling | |
1224 // nof_nonstatic_fields so check it anyway | |
1225 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); | |
1226 BasicType bt = c.basic_type(); | |
1227 // Only integer types have boxing cache. | |
1228 assert(bt == T_BOOLEAN || bt == T_CHAR || | |
1229 bt == T_BYTE || bt == T_SHORT || | |
1230 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt))); | |
1231 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int(); | |
1232 if (cache_low != (int)cache_low) { | |
1233 return NULL; // should not happen since cache is array indexed by value | |
1234 } | |
1235 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift); | |
1236 if (offset != (int)offset) { | |
1237 return NULL; // should not happen since cache is array indexed by value | |
1238 } | |
1239 // Add up all the offsets making of the address of the load | |
1240 Node* result = elements[0]; | |
1241 for (int i = 1; i < count; i++) { | |
1242 result = phase->transform(new (phase->C) AddXNode(result, elements[i])); | |
1243 } | |
1244 // Remove the constant offset from the address and then | |
1245 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset))); | |
1246 // remove the scaling of the offset to recover the original index. | |
1247 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { | |
1248 // Peel the shift off directly but wrap it in a dummy node | |
1249 // since Ideal can't return existing nodes | |
1250 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0)); | |
1251 } else if (result->is_Add() && result->in(2)->is_Con() && | |
1252 result->in(1)->Opcode() == Op_LShiftX && | |
1253 result->in(1)->in(2) == phase->intcon(shift)) { | |
1254 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z) | |
1255 // but for boxing cache access we know that X<<Z will not overflow | |
1256 // (there is range check) so we do this optimizatrion by hand here. | |
1257 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift)); | |
1258 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con)); | |
1259 } else { | |
1260 result = new (phase->C) RShiftXNode(result, phase->intcon(shift)); | |
1261 } | |
1262 #ifdef _LP64 | |
1263 if (bt != T_LONG) { | |
1264 result = new (phase->C) ConvL2INode(phase->transform(result)); | |
1265 } | |
1266 #else | |
1267 if (bt == T_LONG) { | |
1268 result = new (phase->C) ConvI2LNode(phase->transform(result)); | |
1269 } | |
1270 #endif | |
1271 return result; | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1272 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1273 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1274 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1275 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1276 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1277 return NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1278 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1279 |
10278 | 1280 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) { |
1281 Node* region = phi->in(0); | |
163 | 1282 if (region == NULL) { |
10278 | 1283 return false; // Wait stable graph |
163 | 1284 } |
10278 | 1285 uint cnt = phi->req(); |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1286 for (uint i = 1; i < cnt; i++) { |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1287 Node* rc = region->in(i); |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1288 if (rc == NULL || phase->type(rc) == Type::TOP) |
10278 | 1289 return false; // Wait stable graph |
1290 Node* in = phi->in(i); | |
1291 if (in == NULL || phase->type(in) == Type::TOP) | |
1292 return false; // Wait stable graph | |
1293 } | |
1294 return true; | |
1295 } | |
1296 //------------------------------split_through_phi------------------------------ | |
1297 // Split instance or boxed field load through Phi. | |
1298 Node *LoadNode::split_through_phi(PhaseGVN *phase) { | |
1299 Node* mem = in(Memory); | |
1300 Node* address = in(Address); | |
1301 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr(); | |
1302 | |
1303 assert((t_oop != NULL) && | |
1304 (t_oop->is_known_instance_field() || | |
1305 t_oop->is_ptr_to_boxed_value()), "invalide conditions"); | |
1306 | |
1307 Compile* C = phase->C; | |
1308 intptr_t ignore = 0; | |
1309 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); | |
1310 bool base_is_phi = (base != NULL) && base->is_Phi(); | |
1311 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() && | |
1312 (base != NULL) && (base == address->in(AddPNode::Base)) && | |
1313 phase->type(base)->higher_equal(TypePtr::NOTNULL); | |
1314 | |
1315 if (!((mem->is_Phi() || base_is_phi) && | |
1316 (load_boxed_values || t_oop->is_known_instance_field()))) { | |
1317 return NULL; // memory is not Phi | |
1318 } | |
1319 | |
1320 if (mem->is_Phi()) { | |
1321 if (!stable_phi(mem->as_Phi(), phase)) { | |
163 | 1322 return NULL; // Wait stable graph |
1323 } | |
10278 | 1324 uint cnt = mem->req(); |
1325 // Check for loop invariant memory. | |
1326 if (cnt == 3) { | |
1327 for (uint i = 1; i < cnt; i++) { | |
1328 Node* in = mem->in(i); | |
1329 Node* m = optimize_memory_chain(in, t_oop, this, phase); | |
1330 if (m == mem) { | |
1331 set_req(Memory, mem->in(cnt - i)); | |
1332 return this; // made change | |
1333 } | |
163 | 1334 } |
1335 } | |
1336 } | |
10278 | 1337 if (base_is_phi) { |
1338 if (!stable_phi(base->as_Phi(), phase)) { | |
1339 return NULL; // Wait stable graph | |
1340 } | |
1341 uint cnt = base->req(); | |
1342 // Check for loop invariant memory. | |
1343 if (cnt == 3) { | |
1344 for (uint i = 1; i < cnt; i++) { | |
1345 if (base->in(i) == base) { | |
1346 return NULL; // Wait stable graph | |
1347 } | |
1348 } | |
1349 } | |
1350 } | |
1351 | |
1352 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0)); | |
1353 | |
163 | 1354 // Split through Phi (see original code in loopopts.cpp). |
10278 | 1355 assert(C->have_alias_type(t_oop), "instance should have alias type"); |
163 | 1356 |
1357 // Do nothing here if Identity will find a value | |
1358 // (to avoid infinite chain of value phis generation). | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1359 if (!phase->eqv(this, this->Identity(phase))) |
163 | 1360 return NULL; |
1361 | |
10278 | 1362 // Select Region to split through. |
1363 Node* region; | |
1364 if (!base_is_phi) { | |
1365 assert(mem->is_Phi(), "sanity"); | |
1366 region = mem->in(0); | |
1367 // Skip if the region dominates some control edge of the address. | |
1368 if (!MemNode::all_controls_dominate(address, region)) | |
1369 return NULL; | |
1370 } else if (!mem->is_Phi()) { | |
1371 assert(base_is_phi, "sanity"); | |
1372 region = base->in(0); | |
1373 // Skip if the region dominates some control edge of the memory. | |
1374 if (!MemNode::all_controls_dominate(mem, region)) | |
1375 return NULL; | |
1376 } else if (base->in(0) != mem->in(0)) { | |
1377 assert(base_is_phi && mem->is_Phi(), "sanity"); | |
1378 if (MemNode::all_controls_dominate(mem, base->in(0))) { | |
1379 region = base->in(0); | |
1380 } else if (MemNode::all_controls_dominate(address, mem->in(0))) { | |
1381 region = mem->in(0); | |
1382 } else { | |
1383 return NULL; // complex graph | |
1384 } | |
1385 } else { | |
1386 assert(base->in(0) == mem->in(0), "sanity"); | |
1387 region = mem->in(0); | |
1388 } | |
163 | 1389 |
1390 const Type* this_type = this->bottom_type(); | |
10278 | 1391 int this_index = C->get_alias_index(t_oop); |
1392 int this_offset = t_oop->offset(); | |
1393 int this_iid = t_oop->instance_id(); | |
1394 if (!t_oop->is_known_instance() && load_boxed_values) { | |
1395 // Use _idx of address base for boxed values. | |
1396 this_iid = base->_idx; | |
1397 } | |
1398 PhaseIterGVN* igvn = phase->is_IterGVN(); | |
1399 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1400 for (uint i = 1; i < region->req(); i++) { |
10278 | 1401 Node* x; |
163 | 1402 Node* the_clone = NULL; |
10278 | 1403 if (region->in(i) == C->top()) { |
1404 x = C->top(); // Dead path? Use a dead data op | |
163 | 1405 } else { |
1406 x = this->clone(); // Else clone up the data op | |
1407 the_clone = x; // Remember for possible deletion. | |
1408 // Alter data node to use pre-phi inputs | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1409 if (this->in(0) == region) { |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1410 x->set_req(0, region->in(i)); |
163 | 1411 } else { |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1412 x->set_req(0, NULL); |
163 | 1413 } |
10278 | 1414 if (mem->is_Phi() && (mem->in(0) == region)) { |
1415 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone. | |
1416 } | |
1417 if (address->is_Phi() && address->in(0) == region) { | |
1418 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone | |
1419 } | |
1420 if (base_is_phi && (base->in(0) == region)) { | |
1421 Node* base_x = base->in(i); // Clone address for loads from boxed objects. | |
1422 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset))); | |
1423 x->set_req(Address, adr_x); | |
163 | 1424 } |
1425 } | |
1426 // Check for a 'win' on some paths | |
1427 const Type *t = x->Value(igvn); | |
1428 | |
1429 bool singleton = t->singleton(); | |
1430 | |
1431 // See comments in PhaseIdealLoop::split_thru_phi(). | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1432 if (singleton && t == Type::TOP) { |
163 | 1433 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
1434 } | |
1435 | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1436 if (singleton) { |
163 | 1437 x = igvn->makecon(t); |
1438 } else { | |
1439 // We now call Identity to try to simplify the cloned node. | |
1440 // Note that some Identity methods call phase->type(this). | |
1441 // Make sure that the type array is big enough for | |
1442 // our new node, even though we may throw the node away. | |
1443 // (This tweaking with igvn only works because x is a new node.) | |
1444 igvn->set_type(x, t); | |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
1445 // If x is a TypeNode, capture any more-precise type permanently into Node |
605 | 1446 // otherwise it will be not updated during igvn->transform since |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
1447 // igvn->type(x) is set to x->Value() already. |
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
1448 x->raise_bottom_type(t); |
163 | 1449 Node *y = x->Identity(igvn); |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1450 if (y != x) { |
163 | 1451 x = y; |
1452 } else { | |
10278 | 1453 y = igvn->hash_find_insert(x); |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1454 if (y) { |
163 | 1455 x = y; |
1456 } else { | |
1457 // Else x is a new node we are keeping | |
1458 // We do not need register_new_node_with_optimizer | |
1459 // because set_type has already been called. | |
1460 igvn->_worklist.push(x); | |
1461 } | |
1462 } | |
1463 } | |
10278 | 1464 if (x != the_clone && the_clone != NULL) { |
163 | 1465 igvn->remove_dead_node(the_clone); |
10278 | 1466 } |
163 | 1467 phi->set_req(i, x); |
1468 } | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1469 // Record Phi |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1470 igvn->register_new_node_with_optimizer(phi); |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1471 return phi; |
163 | 1472 } |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1473 |
0 | 1474 //------------------------------Ideal------------------------------------------ |
1475 // If the load is from Field memory and the pointer is non-null, we can | |
1476 // zero out the control input. | |
1477 // If the offset is constant and the base is an object allocation, | |
1478 // try to hook me up to the exact initializing store. | |
1479 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1480 Node* p = MemNode::Ideal_common(phase, can_reshape); | |
1481 if (p) return (p == NodeSentinel) ? NULL : p; | |
1482 | |
1483 Node* ctrl = in(MemNode::Control); | |
1484 Node* address = in(MemNode::Address); | |
1485 | |
1486 // Skip up past a SafePoint control. Cannot do this for Stores because | |
1487 // pointer stores & cardmarks must stay on the same side of a SafePoint. | |
1488 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && | |
1489 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { | |
1490 ctrl = ctrl->in(0); | |
1491 set_req(MemNode::Control,ctrl); | |
1492 } | |
1493 | |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1494 intptr_t ignore = 0; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1495 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1496 if (base != NULL |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1497 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) { |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1498 // Check for useless control edge in some common special cases |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1499 if (in(MemNode::Control) != NULL |
0 | 1500 && phase->type(base)->higher_equal(TypePtr::NOTNULL) |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
1501 && all_controls_dominate(base, phase->C->start())) { |
0 | 1502 // A method-invariant, non-null address (constant or 'this' argument). |
1503 set_req(MemNode::Control, NULL); | |
1504 } | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1505 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1506 |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1507 Node* mem = in(MemNode::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1508 const TypePtr *addr_t = phase->type(address)->isa_ptr(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1509 |
10278 | 1510 if (can_reshape && (addr_t != NULL)) { |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1511 // try to optimize our memory input |
10278 | 1512 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1513 if (opt_mem != mem) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1514 set_req(MemNode::Memory, opt_mem); |
305 | 1515 if (phase->type( opt_mem ) == Type::TOP) return NULL; |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1516 return this; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1517 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1518 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); |
10278 | 1519 if ((t_oop != NULL) && |
1520 (t_oop->is_known_instance_field() || | |
1521 t_oop->is_ptr_to_boxed_value())) { | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1522 PhaseIterGVN *igvn = phase->is_IterGVN(); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1523 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1524 // Delay this transformation until memory Phi is processed. |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1525 phase->is_IterGVN()->_worklist.push(this); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1526 return NULL; |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1527 } |
163 | 1528 // Split instance field load through Phi. |
1529 Node* result = split_through_phi(phase); | |
1530 if (result != NULL) return result; | |
10278 | 1531 |
1532 if (t_oop->is_ptr_to_boxed_value()) { | |
1533 Node* result = eliminate_autobox(phase); | |
1534 if (result != NULL) return result; | |
1535 } | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1536 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1537 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1538 |
0 | 1539 // Check for prior store with a different base or offset; make Load |
1540 // independent. Skip through any number of them. Bail out if the stores | |
1541 // are in an endless dead cycle and report no progress. This is a key | |
1542 // transform for Reflection. However, if after skipping through the Stores | |
1543 // we can't then fold up against a prior store do NOT do the transform as | |
1544 // this amounts to using the 'Oracle' model of aliasing. It leaves the same | |
1545 // array memory alive twice: once for the hoisted Load and again after the | |
1546 // bypassed Store. This situation only works if EVERYBODY who does | |
1547 // anti-dependence work knows how to bypass. I.e. we need all | |
1548 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is | |
1549 // the alias index stuff. So instead, peek through Stores and IFF we can | |
1550 // fold up, do so. | |
1551 Node* prev_mem = find_previous_store(phase); | |
1552 // Steps (a), (b): Walk past independent stores to find an exact match. | |
1553 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { | |
1554 // (c) See if we can fold up on the spot, but don't fold up here. | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
1555 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or |
0 | 1556 // just return a prior value, which is done by Identity calls. |
1557 if (can_see_stored_value(prev_mem, phase)) { | |
1558 // Make ready for step (d): | |
1559 set_req(MemNode::Memory, prev_mem); | |
1560 return this; | |
1561 } | |
1562 } | |
1563 | |
1564 return NULL; // No further progress | |
1565 } | |
1566 | |
1567 // Helper to recognize certain Klass fields which are invariant across | |
1568 // some group of array types (e.g., int[] or all T[] where T < Object). | |
1569 const Type* | |
1570 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, | |
1571 ciKlass* klass) const { | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1572 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { |
0 | 1573 // The field is Klass::_modifier_flags. Return its (constant) value. |
1574 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) | |
1575 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); | |
1576 return TypeInt::make(klass->modifier_flags()); | |
1577 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1578 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { |
0 | 1579 // The field is Klass::_access_flags. Return its (constant) value. |
1580 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) | |
1581 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); | |
1582 return TypeInt::make(klass->access_flags()); | |
1583 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1584 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { |
0 | 1585 // The field is Klass::_layout_helper. Return its constant value if known. |
1586 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); | |
1587 return TypeInt::make(klass->layout_helper()); | |
1588 } | |
1589 | |
1590 // No match. | |
1591 return NULL; | |
1592 } | |
1593 | |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1594 // Try to constant-fold a stable array element. |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1595 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) { |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1596 assert(ary->const_oop(), "array should be constant"); |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1597 assert(ary->is_stable(), "array should be stable"); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1598 |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1599 // Decode the results of GraphKit::array_element_address. |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1600 ciArray* aobj = ary->const_oop()->as_array(); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1601 ciConstant con = aobj->element_value_by_offset(off); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1602 |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1603 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1604 const Type* con_type = Type::make_from_constant(con); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1605 if (con_type != NULL) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1606 if (con_type->isa_aryptr()) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1607 // Join with the array element type, in case it is also stable. |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1608 int dim = ary->stable_dimension(); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1609 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1610 } |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1611 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1612 con_type = con_type->make_narrowoop(); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1613 } |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1614 #ifndef PRODUCT |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1615 if (TraceIterativeGVN) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1616 tty->print("FoldStableValues: array element [off=%d]: con_type=", off); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1617 con_type->dump(); tty->cr(); |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1618 } |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1619 #endif //PRODUCT |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1620 return con_type; |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1621 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1622 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1623 return NULL; |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1624 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1625 |
0 | 1626 //------------------------------Value----------------------------------------- |
1627 const Type *LoadNode::Value( PhaseTransform *phase ) const { | |
1628 // Either input is TOP ==> the result is TOP | |
1629 Node* mem = in(MemNode::Memory); | |
1630 const Type *t1 = phase->type(mem); | |
1631 if (t1 == Type::TOP) return Type::TOP; | |
1632 Node* adr = in(MemNode::Address); | |
1633 const TypePtr* tp = phase->type(adr)->isa_ptr(); | |
1634 if (tp == NULL || tp->empty()) return Type::TOP; | |
1635 int off = tp->offset(); | |
1636 assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1637 Compile* C = phase->C; |
0 | 1638 |
1639 // Try to guess loaded type from pointer type | |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1640 if (tp->isa_aryptr()) { |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1641 const TypeAryPtr* ary = tp->is_aryptr(); |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1642 const Type* t = ary->elem(); |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1643 |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1644 // Determine whether the reference is beyond the header or not, by comparing |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1645 // the offset against the offset of the start of the array's data. |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1646 // Different array types begin at slightly different offsets (12 vs. 16). |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1647 // We choose T_BYTE as an example base type that is least restrictive |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1648 // as to alignment, which will therefore produce the smallest |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1649 // possible base offset. |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1650 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1651 const bool off_beyond_header = ((uint)off >= (uint)min_base_off); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1652 |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1653 // Try to constant-fold a stable array element. |
14683
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1654 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) { |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1655 // Make sure the reference is not into the header and the offset is constant |
bd2220947c1e
8036667: "assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con()) failed: offset is a constant" with FoldStableValues on
vlivanov
parents:
14457
diff
changeset
|
1656 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) { |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1657 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type()); |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1658 if (con_type != NULL) { |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1659 return con_type; |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1660 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1661 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1662 } |
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1663 |
0 | 1664 // Don't do this for integer types. There is only potential profit if |
1665 // the element type t is lower than _type; that is, for int types, if _type is | |
1666 // more restrictive than t. This only happens here if one is short and the other | |
1667 // char (both 16 bits), and in those cases we've made an intentional decision | |
1668 // to use one kind of load over the other. See AndINode::Ideal and 4965907. | |
1669 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. | |
1670 // | |
1671 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) | |
1672 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier | |
1673 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been | |
1674 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, | |
1675 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. | |
1676 // In fact, that could have been the original type of p1, and p1 could have | |
1677 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the | |
1678 // expression (LShiftL quux 3) independently optimized to the constant 8. | |
1679 if ((t->isa_int() == NULL) && (t->isa_long() == NULL) | |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
6139
diff
changeset
|
1680 && (_type->isa_vect() == NULL) |
293
c3e045194476
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
kvn
parents:
253
diff
changeset
|
1681 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { |
0 | 1682 // t might actually be lower than _type, if _type is a unique |
1683 // concrete subclass of abstract class t. | |
12190
edb5ab0f3fe5
8001107: @Stable annotation for constant folding of lazily evaluated variables
vlivanov
parents:
11044
diff
changeset
|
1684 if (off_beyond_header) { // is the offset beyond the header? |
14383
5ec7dace41a6
8027422: assert(_gvn.type(obj)->higher_equal(tjp)) failed: cast_up is no longer needed
roland
parents:
14223
diff
changeset
|
1685 const Type* jt = t->join_speculative(_type); |
0 | 1686 // In any case, do not allow the join, per se, to empty out the type. |
1687 if (jt->empty() && !t->empty()) { | |
1688 // This can happen if a interface-typed array narrows to a class type. | |
1689 jt = _type; | |
1690 } | |
10278 | 1691 #ifdef ASSERT |
1692 if (phase->C->eliminate_boxing() && adr->is_AddP()) { | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1693 // The pointers in the autobox arrays are always non-null |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1694 Node* base = adr->in(AddPNode::Base); |
10278 | 1695 if ((base != NULL) && base->is_DecodeN()) { |
1696 // Get LoadN node which loads IntegerCache.cache field | |
1697 base = base->in(1); | |
1698 } | |
1699 if ((base != NULL) && base->is_Con()) { | |
1700 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr(); | |
1701 if ((base_type != NULL) && base_type->is_autobox_cache()) { | |
1702 // It could be narrow oop | |
1703 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity"); | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1704 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1705 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1706 } |
10278 | 1707 #endif |
0 | 1708 return jt; |
1709 } | |
1710 } | |
1711 } else if (tp->base() == Type::InstPtr) { | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1712 ciEnv* env = C->env(); |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1713 const TypeInstPtr* tinst = tp->is_instptr(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1714 ciKlass* klass = tinst->klass(); |
0 | 1715 assert( off != Type::OffsetBot || |
1716 // arrays can be cast to Objects | |
1717 tp->is_oopptr()->klass()->is_java_lang_Object() || | |
1718 // unsafe field access may not have a constant offset | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1719 C->has_unsafe_access(), |
0 | 1720 "Field accesses must be precise" ); |
1721 // For oop loads, we expect the _type to be precise | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1722 if (klass == env->String_klass() && |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1723 adr->is_AddP() && off != Type::OffsetBot) { |
2320
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1724 // For constant Strings treat the final fields as compile time constants. |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1725 Node* base = adr->in(AddPNode::Base); |
1766
02f0a9b6f654
6969586: OptimizeStringConcat: SIGSEGV in LoadNode::Value()
never
parents:
1609
diff
changeset
|
1726 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); |
02f0a9b6f654
6969586: OptimizeStringConcat: SIGSEGV in LoadNode::Value()
never
parents:
1609
diff
changeset
|
1727 if (t != NULL && t->singleton()) { |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1728 ciField* field = env->String_klass()->get_field_by_offset(off, false); |
2320
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1729 if (field != NULL && field->is_final()) { |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1730 ciObject* string = t->const_oop(); |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1731 ciConstant constant = string->as_instance()->field_value(field); |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1732 if (constant.basic_type() == T_INT) { |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1733 return TypeInt::make(constant.as_int()); |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1734 } else if (constant.basic_type() == T_ARRAY) { |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1735 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
2379
b099aaf51bf8
6962931: move interned strings out of the perm gen
jcoomes
parents:
2320
diff
changeset
|
1736 return TypeNarrowOop::make_from_constant(constant.as_object(), true); |
2320
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1737 } else { |
2379
b099aaf51bf8
6962931: move interned strings out of the perm gen
jcoomes
parents:
2320
diff
changeset
|
1738 return TypeOopPtr::make_from_constant(constant.as_object(), true); |
2320
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1739 } |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1740 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1741 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1742 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1743 } |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1744 // Optimizations for constant objects |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1745 ciObject* const_oop = tinst->const_oop(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1746 if (const_oop != NULL) { |
10278 | 1747 // For constant Boxed value treat the target field as a compile time constant. |
1748 if (tinst->is_ptr_to_boxed_value()) { | |
1749 return tinst->get_const_boxed_value(); | |
1750 } else | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1751 // For constant CallSites treat the target field as a compile time constant. |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1752 if (const_oop->is_call_site()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1753 ciCallSite* call_site = const_oop->as_call_site(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1754 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1755 if (field != NULL && field->is_call_site_target()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1756 ciMethodHandle* target = call_site->get_target(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1757 if (target != NULL) { // just in case |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1758 ciConstant constant(T_OBJECT, target); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1759 const Type* t; |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1760 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1761 t = TypeNarrowOop::make_from_constant(constant.as_object(), true); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1762 } else { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1763 t = TypeOopPtr::make_from_constant(constant.as_object(), true); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1764 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1765 // Add a dependence for invalidation of the optimization. |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1766 if (!call_site->is_constant_call_site()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1767 C->dependencies()->assert_call_site_target_value(call_site, target); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1768 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1769 return t; |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1770 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1771 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1772 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1773 } |
0 | 1774 } else if (tp->base() == Type::KlassPtr) { |
1775 assert( off != Type::OffsetBot || | |
1776 // arrays can be cast to Objects | |
1777 tp->is_klassptr()->klass()->is_java_lang_Object() || | |
1778 // also allow array-loading from the primary supertype | |
1779 // array during subtype checks | |
1780 Opcode() == Op_LoadKlass, | |
1781 "Field accesses must be precise" ); | |
1782 // For klass/static loads, we expect the _type to be precise | |
1783 } | |
1784 | |
1785 const TypeKlassPtr *tkls = tp->isa_klassptr(); | |
1786 if (tkls != NULL && !StressReflectiveCode) { | |
1787 ciKlass* klass = tkls->klass(); | |
1788 if (klass->is_loaded() && tkls->klass_is_exact()) { | |
1789 // We are loading a field from a Klass metaobject whose identity | |
1790 // is known at compile time (the type is "exact" or "precise"). | |
1791 // Check for fields we know are maintained as constants by the VM. | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1792 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { |
0 | 1793 // The field is Klass::_super_check_offset. Return its (constant) value. |
1794 // (Folds up type checking code.) | |
1795 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); | |
1796 return TypeInt::make(klass->super_check_offset()); | |
1797 } | |
1798 // Compute index into primary_supers array | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
1799 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
0 | 1800 // Check for overflowing; use unsigned compare to handle the negative case. |
1801 if( depth < ciKlass::primary_super_limit() ) { | |
1802 // The field is an element of Klass::_primary_supers. Return its (constant) value. | |
1803 // (Folds up type checking code.) | |
1804 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); | |
1805 ciKlass *ss = klass->super_of_depth(depth); | |
1806 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; | |
1807 } | |
1808 const Type* aift = load_array_final_field(tkls, klass); | |
1809 if (aift != NULL) return aift; | |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
1810 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset()) |
0 | 1811 && klass->is_array_klass()) { |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
1812 // The field is ArrayKlass::_component_mirror. Return its (constant) value. |
0 | 1813 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) |
1814 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); | |
1815 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); | |
1816 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1817 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { |
0 | 1818 // The field is Klass::_java_mirror. Return its (constant) value. |
1819 // (Folds up the 2nd indirection in anObjConstant.getClass().) | |
1820 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); | |
1821 return TypeInstPtr::make(klass->java_mirror()); | |
1822 } | |
1823 } | |
1824 | |
1825 // We can still check if we are loading from the primary_supers array at a | |
1826 // shallow enough depth. Even though the klass is not exact, entries less | |
1827 // than or equal to its super depth are correct. | |
1828 if (klass->is_loaded() ) { | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
1829 ciType *inner = klass; |
0 | 1830 while( inner->is_obj_array_klass() ) |
1831 inner = inner->as_obj_array_klass()->base_element_type(); | |
1832 if( inner->is_instance_klass() && | |
1833 !inner->as_instance_klass()->flags().is_interface() ) { | |
1834 // Compute index into primary_supers array | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
1835 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*); |
0 | 1836 // Check for overflowing; use unsigned compare to handle the negative case. |
1837 if( depth < ciKlass::primary_super_limit() && | |
1838 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case | |
1839 // The field is an element of Klass::_primary_supers. Return its (constant) value. | |
1840 // (Folds up type checking code.) | |
1841 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); | |
1842 ciKlass *ss = klass->super_of_depth(depth); | |
1843 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; | |
1844 } | |
1845 } | |
1846 } | |
1847 | |
1848 // If the type is enough to determine that the thing is not an array, | |
1849 // we can give the layout_helper a positive interval type. | |
1850 // This will help short-circuit some reflective code. | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1851 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) |
0 | 1852 && !klass->is_array_klass() // not directly typed as an array |
1853 && !klass->is_interface() // specifically not Serializable & Cloneable | |
1854 && !klass->is_java_lang_Object() // not the supertype of all T[] | |
1855 ) { | |
1856 // Note: When interfaces are reliable, we can narrow the interface | |
1857 // test to (klass != Serializable && klass != Cloneable). | |
1858 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); | |
1859 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); | |
1860 // The key property of this type is that it folds up tests | |
1861 // for array-ness, since it proves that the layout_helper is positive. | |
1862 // Thus, a generic value like the basic object layout helper works fine. | |
1863 return TypeInt::make(min_size, max_jint, Type::WidenMin); | |
1864 } | |
1865 } | |
1866 | |
1867 // If we are loading from a freshly-allocated object, produce a zero, | |
1868 // if the load is provably beyond the header of the object. | |
1869 // (Also allow a variable load from a fresh array to produce zero.) | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1870 const TypeOopPtr *tinst = tp->isa_oopptr(); |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1871 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); |
10278 | 1872 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value(); |
1873 if (ReduceFieldZeroing || is_instance || is_boxed_value) { | |
0 | 1874 Node* value = can_see_stored_value(mem,phase); |
4815 | 1875 if (value != NULL && value->is_Con()) { |
1876 assert(value->bottom_type()->higher_equal(_type),"sanity"); | |
0 | 1877 return value->bottom_type(); |
4815 | 1878 } |
0 | 1879 } |
1880 | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1881 if (is_instance) { |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1882 // If we have an instance type and our memory input is the |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1883 // programs's initial memory state, there is no matching store, |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1884 // so just return a zero of the appropriate type |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1885 Node *mem = in(MemNode::Memory); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1886 if (mem->is_Parm() && mem->in(0)->is_Start()) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1887 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1888 return Type::get_zero_type(_type->basic_type()); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1889 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1890 } |
0 | 1891 return _type; |
1892 } | |
1893 | |
1894 //------------------------------match_edge------------------------------------- | |
1895 // Do we Match on this edge index or not? Match only the address. | |
1896 uint LoadNode::match_edge(uint idx) const { | |
1897 return idx == MemNode::Address; | |
1898 } | |
1899 | |
1900 //--------------------------LoadBNode::Ideal-------------------------------------- | |
1901 // | |
1902 // If the previous store is to the same address as this load, | |
1903 // and the value stored was larger than a byte, replace this load | |
1904 // with the value stored truncated to a byte. If no truncation is | |
1905 // needed, the replacement is done in LoadNode::Identity(). | |
1906 // | |
1907 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1908 Node* mem = in(MemNode::Memory); | |
1909 Node* value = can_see_stored_value(mem,phase); | |
1910 if( value && !phase->type(value)->higher_equal( _type ) ) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1911 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1912 return new (phase->C) RShiftINode(result, phase->intcon(24)); |
0 | 1913 } |
1914 // Identity call will handle the case where truncation is not needed. | |
1915 return LoadNode::Ideal(phase, can_reshape); | |
1916 } | |
1917 | |
4815 | 1918 const Type* LoadBNode::Value(PhaseTransform *phase) const { |
1919 Node* mem = in(MemNode::Memory); | |
1920 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1921 if (value != NULL && value->is_Con() && |
1922 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1923 // If the input to the store does not fit with the load's result type, |
1924 // it must be truncated. We can't delay until Ideal call since | |
1925 // a singleton Value is needed for split_thru_phi optimization. | |
1926 int con = value->get_int(); | |
1927 return TypeInt::make((con << 24) >> 24); | |
1928 } | |
1929 return LoadNode::Value(phase); | |
1930 } | |
1931 | |
624 | 1932 //--------------------------LoadUBNode::Ideal------------------------------------- |
1933 // | |
1934 // If the previous store is to the same address as this load, | |
1935 // and the value stored was larger than a byte, replace this load | |
1936 // with the value stored truncated to a byte. If no truncation is | |
1937 // needed, the replacement is done in LoadNode::Identity(). | |
1938 // | |
1939 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { | |
1940 Node* mem = in(MemNode::Memory); | |
1941 Node* value = can_see_stored_value(mem, phase); | |
1942 if (value && !phase->type(value)->higher_equal(_type)) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1943 return new (phase->C) AndINode(value, phase->intcon(0xFF)); |
624 | 1944 // Identity call will handle the case where truncation is not needed. |
1945 return LoadNode::Ideal(phase, can_reshape); | |
1946 } | |
1947 | |
4815 | 1948 const Type* LoadUBNode::Value(PhaseTransform *phase) const { |
1949 Node* mem = in(MemNode::Memory); | |
1950 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1951 if (value != NULL && value->is_Con() && |
1952 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1953 // If the input to the store does not fit with the load's result type, |
1954 // it must be truncated. We can't delay until Ideal call since | |
1955 // a singleton Value is needed for split_thru_phi optimization. | |
1956 int con = value->get_int(); | |
1957 return TypeInt::make(con & 0xFF); | |
1958 } | |
1959 return LoadNode::Value(phase); | |
1960 } | |
1961 | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
1962 //--------------------------LoadUSNode::Ideal------------------------------------- |
0 | 1963 // |
1964 // If the previous store is to the same address as this load, | |
1965 // and the value stored was larger than a char, replace this load | |
1966 // with the value stored truncated to a char. If no truncation is | |
1967 // needed, the replacement is done in LoadNode::Identity(). | |
1968 // | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
1969 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
0 | 1970 Node* mem = in(MemNode::Memory); |
1971 Node* value = can_see_stored_value(mem,phase); | |
1972 if( value && !phase->type(value)->higher_equal( _type ) ) | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
1973 return new (phase->C) AndINode(value,phase->intcon(0xFFFF)); |
0 | 1974 // Identity call will handle the case where truncation is not needed. |
1975 return LoadNode::Ideal(phase, can_reshape); | |
1976 } | |
1977 | |
4815 | 1978 const Type* LoadUSNode::Value(PhaseTransform *phase) const { |
1979 Node* mem = in(MemNode::Memory); | |
1980 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1981 if (value != NULL && value->is_Con() && |
1982 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1983 // If the input to the store does not fit with the load's result type, |
1984 // it must be truncated. We can't delay until Ideal call since | |
1985 // a singleton Value is needed for split_thru_phi optimization. | |
1986 int con = value->get_int(); | |
1987 return TypeInt::make(con & 0xFFFF); | |
1988 } | |
1989 return LoadNode::Value(phase); | |
1990 } | |
1991 | |
0 | 1992 //--------------------------LoadSNode::Ideal-------------------------------------- |
1993 // | |
1994 // If the previous store is to the same address as this load, | |
1995 // and the value stored was larger than a short, replace this load | |
1996 // with the value stored truncated to a short. If no truncation is | |
1997 // needed, the replacement is done in LoadNode::Identity(). | |
1998 // | |
1999 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
2000 Node* mem = in(MemNode::Memory); | |
2001 Node* value = can_see_stored_value(mem,phase); | |
2002 if( value && !phase->type(value)->higher_equal( _type ) ) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2003 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2004 return new (phase->C) RShiftINode(result, phase->intcon(16)); |
0 | 2005 } |
2006 // Identity call will handle the case where truncation is not needed. | |
2007 return LoadNode::Ideal(phase, can_reshape); | |
2008 } | |
2009 | |
4815 | 2010 const Type* LoadSNode::Value(PhaseTransform *phase) const { |
2011 Node* mem = in(MemNode::Memory); | |
2012 Node* value = can_see_stored_value(mem,phase); | |
4821 | 2013 if (value != NULL && value->is_Con() && |
2014 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 2015 // If the input to the store does not fit with the load's result type, |
2016 // it must be truncated. We can't delay until Ideal call since | |
2017 // a singleton Value is needed for split_thru_phi optimization. | |
2018 int con = value->get_int(); | |
2019 return TypeInt::make((con << 16) >> 16); | |
2020 } | |
2021 return LoadNode::Value(phase); | |
2022 } | |
2023 | |
0 | 2024 //============================================================================= |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2025 //----------------------------LoadKlassNode::make------------------------------ |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2026 // Polymorphic factory method: |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2027 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2028 Compile* C = gvn.C; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2029 Node *ctl = NULL; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2030 // sanity check the alias category against the created node type |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
2031 const TypePtr *adr_type = adr->bottom_type()->isa_ptr(); |
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
2032 assert(adr_type != NULL, "expecting TypeKlassPtr"); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2033 #ifdef _LP64 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2034 if (adr_type->is_ptr_to_narrowklass()) { |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
11044
diff
changeset
|
2035 assert(UseCompressedClassPointers, "no compressed klasses"); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2036 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered)); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2037 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr()); |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
2038 } |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2039 #endif |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2040 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2041 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2042 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2043 |
0 | 2044 //------------------------------Value------------------------------------------ |
2045 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2046 return klass_value_common(phase); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2047 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2048 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2049 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { |
0 | 2050 // Either input is TOP ==> the result is TOP |
2051 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
2052 if (t1 == Type::TOP) return Type::TOP; | |
2053 Node *adr = in(MemNode::Address); | |
2054 const Type *t2 = phase->type( adr ); | |
2055 if (t2 == Type::TOP) return Type::TOP; | |
2056 const TypePtr *tp = t2->is_ptr(); | |
2057 if (TypePtr::above_centerline(tp->ptr()) || | |
2058 tp->ptr() == TypePtr::Null) return Type::TOP; | |
2059 | |
2060 // Return a more precise klass, if possible | |
2061 const TypeInstPtr *tinst = tp->isa_instptr(); | |
2062 if (tinst != NULL) { | |
2063 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); | |
2064 int offset = tinst->offset(); | |
2065 if (ik == phase->C->env()->Class_klass() | |
2066 && (offset == java_lang_Class::klass_offset_in_bytes() || | |
2067 offset == java_lang_Class::array_klass_offset_in_bytes())) { | |
2068 // We are loading a special hidden field from a Class mirror object, | |
2069 // the field which points to the VM's Klass metaobject. | |
2070 ciType* t = tinst->java_mirror_type(); | |
2071 // java_mirror_type returns non-null for compile-time Class constants. | |
2072 if (t != NULL) { | |
2073 // constant oop => constant klass | |
2074 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { | |
14213
6c4c27c5cc9a
8029366: ShouldNotReachHere error when creating an array with component type of void
twisti
parents:
12242
diff
changeset
|
2075 if (t->is_void()) { |
6c4c27c5cc9a
8029366: ShouldNotReachHere error when creating an array with component type of void
twisti
parents:
12242
diff
changeset
|
2076 // We cannot create a void array. Since void is a primitive type return null |
6c4c27c5cc9a
8029366: ShouldNotReachHere error when creating an array with component type of void
twisti
parents:
12242
diff
changeset
|
2077 // klass. Users of this result need to do a null check on the returned klass. |
6c4c27c5cc9a
8029366: ShouldNotReachHere error when creating an array with component type of void
twisti
parents:
12242
diff
changeset
|
2078 return TypePtr::NULL_PTR; |
6c4c27c5cc9a
8029366: ShouldNotReachHere error when creating an array with component type of void
twisti
parents:
12242
diff
changeset
|
2079 } |
0 | 2080 return TypeKlassPtr::make(ciArrayKlass::make(t)); |
2081 } | |
2082 if (!t->is_klass()) { | |
2083 // a primitive Class (e.g., int.class) has NULL for a klass field | |
2084 return TypePtr::NULL_PTR; | |
2085 } | |
2086 // (Folds up the 1st indirection in aClassConstant.getModifiers().) | |
2087 return TypeKlassPtr::make(t->as_klass()); | |
2088 } | |
2089 // non-constant mirror, so we can't tell what's going on | |
2090 } | |
2091 if( !ik->is_loaded() ) | |
2092 return _type; // Bail out if not loaded | |
2093 if (offset == oopDesc::klass_offset_in_bytes()) { | |
2094 if (tinst->klass_is_exact()) { | |
2095 return TypeKlassPtr::make(ik); | |
2096 } | |
2097 // See if we can become precise: no subklasses and no interface | |
2098 // (Note: We need to support verified interfaces.) | |
2099 if (!ik->is_interface() && !ik->has_subklass()) { | |
2100 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
2101 // Add a dependence; if any subclass added we need to recompile | |
2102 if (!ik->is_final()) { | |
2103 // %%% should use stronger assert_unique_concrete_subtype instead | |
2104 phase->C->dependencies()->assert_leaf_type(ik); | |
2105 } | |
2106 // Return precise klass | |
2107 return TypeKlassPtr::make(ik); | |
2108 } | |
2109 | |
2110 // Return root of possible klass | |
2111 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); | |
2112 } | |
2113 } | |
2114 | |
2115 // Check for loading klass from an array | |
2116 const TypeAryPtr *tary = tp->isa_aryptr(); | |
2117 if( tary != NULL ) { | |
2118 ciKlass *tary_klass = tary->klass(); | |
2119 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP | |
2120 && tary->offset() == oopDesc::klass_offset_in_bytes()) { | |
2121 if (tary->klass_is_exact()) { | |
2122 return TypeKlassPtr::make(tary_klass); | |
2123 } | |
2124 ciArrayKlass *ak = tary->klass()->as_array_klass(); | |
2125 // If the klass is an object array, we defer the question to the | |
2126 // array component klass. | |
2127 if( ak->is_obj_array_klass() ) { | |
2128 assert( ak->is_loaded(), "" ); | |
2129 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); | |
2130 if( base_k->is_loaded() && base_k->is_instance_klass() ) { | |
2131 ciInstanceKlass* ik = base_k->as_instance_klass(); | |
2132 // See if we can become precise: no subklasses and no interface | |
2133 if (!ik->is_interface() && !ik->has_subklass()) { | |
2134 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
2135 // Add a dependence; if any subclass added we need to recompile | |
2136 if (!ik->is_final()) { | |
2137 phase->C->dependencies()->assert_leaf_type(ik); | |
2138 } | |
2139 // Return precise array klass | |
2140 return TypeKlassPtr::make(ak); | |
2141 } | |
2142 } | |
2143 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); | |
2144 } else { // Found a type-array? | |
2145 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
2146 assert( ak->is_type_array_klass(), "" ); | |
2147 return TypeKlassPtr::make(ak); // These are always precise | |
2148 } | |
2149 } | |
2150 } | |
2151 | |
2152 // Check for loading klass from an array klass | |
2153 const TypeKlassPtr *tkls = tp->isa_klassptr(); | |
2154 if (tkls != NULL && !StressReflectiveCode) { | |
2155 ciKlass* klass = tkls->klass(); | |
2156 if( !klass->is_loaded() ) | |
2157 return _type; // Bail out if not loaded | |
2158 if( klass->is_obj_array_klass() && | |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
2159 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) { |
0 | 2160 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); |
2161 // // Always returning precise element type is incorrect, | |
2162 // // e.g., element type could be object and array may contain strings | |
2163 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); | |
2164 | |
2165 // The array's TypeKlassPtr was declared 'precise' or 'not precise' | |
2166 // according to the element type's subclassing. | |
2167 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); | |
2168 } | |
2169 if( klass->is_instance_klass() && tkls->klass_is_exact() && | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2170 tkls->offset() == in_bytes(Klass::super_offset())) { |
0 | 2171 ciKlass* sup = klass->as_instance_klass()->super(); |
2172 // The field is Klass::_super. Return its (constant) value. | |
2173 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) | |
2174 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; | |
2175 } | |
2176 } | |
2177 | |
2178 // Bailout case | |
2179 return LoadNode::Value(phase); | |
2180 } | |
2181 | |
2182 //------------------------------Identity--------------------------------------- | |
2183 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. | |
2184 // Also feed through the klass in Allocate(...klass...)._klass. | |
2185 Node* LoadKlassNode::Identity( PhaseTransform *phase ) { | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2186 return klass_identity_common(phase); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2187 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2188 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2189 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { |
0 | 2190 Node* x = LoadNode::Identity(phase); |
2191 if (x != this) return x; | |
2192 | |
2193 // Take apart the address into an oop and and offset. | |
2194 // Return 'this' if we cannot. | |
2195 Node* adr = in(MemNode::Address); | |
2196 intptr_t offset = 0; | |
2197 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
2198 if (base == NULL) return this; | |
2199 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); | |
2200 if (toop == NULL) return this; | |
2201 | |
2202 // We can fetch the klass directly through an AllocateNode. | |
2203 // This works even if the klass is not constant (clone or newArray). | |
2204 if (offset == oopDesc::klass_offset_in_bytes()) { | |
2205 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); | |
2206 if (allocated_klass != NULL) { | |
2207 return allocated_klass; | |
2208 } | |
2209 } | |
2210 | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
2211 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*. |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
2212 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass. |
0 | 2213 // See inline_native_Class_query for occurrences of these patterns. |
2214 // Java Example: x.getClass().isAssignableFrom(y) | |
2215 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) | |
2216 // | |
2217 // This improves reflective code, often making the Class | |
2218 // mirror go completely dead. (Current exception: Class | |
2219 // mirrors may appear in debug info, but we could clean them out by | |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
2220 // introducing a new debug info operator for Klass*.java_mirror). |
0 | 2221 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() |
2222 && (offset == java_lang_Class::klass_offset_in_bytes() || | |
2223 offset == java_lang_Class::array_klass_offset_in_bytes())) { | |
2224 // We are loading a special hidden field from a Class mirror, | |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
2225 // the field which points to its Klass or ArrayKlass metaobject. |
0 | 2226 if (base->is_Load()) { |
2227 Node* adr2 = base->in(MemNode::Address); | |
2228 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); | |
2229 if (tkls != NULL && !tkls->empty() | |
2230 && (tkls->klass()->is_instance_klass() || | |
2231 tkls->klass()->is_array_klass()) | |
2232 && adr2->is_AddP() | |
2233 ) { | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2234 int mirror_field = in_bytes(Klass::java_mirror_offset()); |
0 | 2235 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
6831
d8ce2825b193
8000213: NPG: Should have renamed arrayKlass and typeArrayKlass
coleenp
parents:
6804
diff
changeset
|
2236 mirror_field = in_bytes(ArrayKlass::component_mirror_offset()); |
0 | 2237 } |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2238 if (tkls->offset() == mirror_field) { |
0 | 2239 return adr2->in(AddPNode::Base); |
2240 } | |
2241 } | |
2242 } | |
2243 } | |
2244 | |
2245 return this; | |
2246 } | |
2247 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2248 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2249 //------------------------------Value------------------------------------------ |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2250 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2251 const Type *t = klass_value_common(phase); |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2252 if (t == Type::TOP) |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2253 return t; |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2254 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2255 return t->make_narrowklass(); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2256 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2257 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2258 //------------------------------Identity--------------------------------------- |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2259 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2260 // Also feed through the klass in Allocate(...klass...)._klass. |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2261 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2262 Node *x = klass_identity_common(phase); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2263 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2264 const Type *t = phase->type( x ); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2265 if( t == Type::TOP ) return x; |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2266 if( t->isa_narrowklass()) return x; |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2267 assert (!t->isa_narrowoop(), "no narrow oop here"); |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2268 |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2269 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass())); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2270 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2271 |
0 | 2272 //------------------------------Value----------------------------------------- |
2273 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { | |
2274 // Either input is TOP ==> the result is TOP | |
2275 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
2276 if( t1 == Type::TOP ) return Type::TOP; | |
2277 Node *adr = in(MemNode::Address); | |
2278 const Type *t2 = phase->type( adr ); | |
2279 if( t2 == Type::TOP ) return Type::TOP; | |
2280 const TypePtr *tp = t2->is_ptr(); | |
2281 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; | |
2282 const TypeAryPtr *tap = tp->isa_aryptr(); | |
2283 if( !tap ) return _type; | |
2284 return tap->size(); | |
2285 } | |
2286 | |
366
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2287 //-------------------------------Ideal--------------------------------------- |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2288 // Feed through the length in AllocateArray(...length...)._length. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2289 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2290 Node* p = MemNode::Ideal_common(phase, can_reshape); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2291 if (p) return (p == NodeSentinel) ? NULL : p; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2292 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2293 // Take apart the address into an oop and and offset. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2294 // Return 'this' if we cannot. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2295 Node* adr = in(MemNode::Address); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2296 intptr_t offset = 0; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2297 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2298 if (base == NULL) return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2299 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2300 if (tary == NULL) return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2301 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2302 // We can fetch the length directly through an AllocateArrayNode. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2303 // This works even if the length is not constant (clone or newArray). |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2304 if (offset == arrayOopDesc::length_offset_in_bytes()) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2305 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2306 if (alloc != NULL) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2307 Node* allocated_length = alloc->Ideal_length(); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2308 Node* len = alloc->make_ideal_length(tary, phase); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2309 if (allocated_length != len) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2310 // New CastII improves on this. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2311 return len; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2312 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2313 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2314 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2315 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2316 return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2317 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2318 |
0 | 2319 //------------------------------Identity--------------------------------------- |
2320 // Feed through the length in AllocateArray(...length...)._length. | |
2321 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { | |
2322 Node* x = LoadINode::Identity(phase); | |
2323 if (x != this) return x; | |
2324 | |
2325 // Take apart the address into an oop and and offset. | |
2326 // Return 'this' if we cannot. | |
2327 Node* adr = in(MemNode::Address); | |
2328 intptr_t offset = 0; | |
2329 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
2330 if (base == NULL) return this; | |
2331 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); | |
2332 if (tary == NULL) return this; | |
2333 | |
2334 // We can fetch the length directly through an AllocateArrayNode. | |
2335 // This works even if the length is not constant (clone or newArray). | |
2336 if (offset == arrayOopDesc::length_offset_in_bytes()) { | |
366
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2337 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2338 if (alloc != NULL) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2339 Node* allocated_length = alloc->Ideal_length(); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2340 // Do not allow make_ideal_length to allocate a CastII node. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2341 Node* len = alloc->make_ideal_length(tary, phase, false); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2342 if (allocated_length == len) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2343 // Return allocated_length only if it would not be improved by a CastII. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2344 return allocated_length; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2345 } |
0 | 2346 } |
2347 } | |
2348 | |
2349 return this; | |
2350 | |
2351 } | |
366
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2352 |
0 | 2353 //============================================================================= |
2354 //---------------------------StoreNode::make----------------------------------- | |
2355 // Polymorphic factory method: | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2356 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) { |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2357 assert((mo == unordered || mo == release), "unexpected"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2358 Compile* C = gvn.C; |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2359 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2360 ctl != NULL, "raw memory operations should have control edge"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2361 |
0 | 2362 switch (bt) { |
2363 case T_BOOLEAN: | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2364 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2365 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo); |
0 | 2366 case T_CHAR: |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2367 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2368 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2369 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo); |
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2370 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo); |
6725
da91efe96a93
6964458: Reimplement class meta-data storage to use native memory
coleenp
parents:
6179
diff
changeset
|
2371 case T_METADATA: |
0 | 2372 case T_ADDRESS: |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2373 case T_OBJECT: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2374 #ifdef _LP64 |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2375 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2376 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop())); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2377 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2378 } else if (adr->bottom_type()->is_ptr_to_narrowklass() || |
12226
7944aba7ba41
8015107: NPG: Use consistent naming for metaspace concepts
ehelin
parents:
11044
diff
changeset
|
2379 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() && |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2380 adr->bottom_type()->isa_rawptr())) { |
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2381 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass())); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2382 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo); |
6848
8e47bac5643a
7054512: Compress class pointers after perm gen removal
roland
parents:
6804
diff
changeset
|
2383 } |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2384 #endif |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2385 { |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2386 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo); |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2387 } |
0 | 2388 } |
2389 ShouldNotReachHere(); | |
2390 return (StoreNode*)NULL; | |
2391 } | |
2392 | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2393 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) { |
0 | 2394 bool require_atomic = true; |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2395 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic); |
0 | 2396 } |
2397 | |
2398 | |
2399 //--------------------------bottom_type---------------------------------------- | |
2400 const Type *StoreNode::bottom_type() const { | |
2401 return Type::MEMORY; | |
2402 } | |
2403 | |
2404 //------------------------------hash------------------------------------------- | |
2405 uint StoreNode::hash() const { | |
2406 // unroll addition of interesting fields | |
2407 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); | |
2408 | |
2409 // Since they are not commoned, do not hash them: | |
2410 return NO_HASH; | |
2411 } | |
2412 | |
2413 //------------------------------Ideal------------------------------------------ | |
2414 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). | |
2415 // When a store immediately follows a relevant allocation/initialization, | |
2416 // try to capture it into the initialization, or hoist it above. | |
2417 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
2418 Node* p = MemNode::Ideal_common(phase, can_reshape); | |
2419 if (p) return (p == NodeSentinel) ? NULL : p; | |
2420 | |
2421 Node* mem = in(MemNode::Memory); | |
2422 Node* address = in(MemNode::Address); | |
2423 | |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2424 // Back-to-back stores to same address? Fold em up. Generally |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2425 // unsafe if I have intervening uses... Also disallowed for StoreCM |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2426 // since they must follow each StoreP operation. Redundant StoreCMs |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2427 // are eliminated just before matching in final_graph_reshape. |
4778 | 2428 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) && |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2429 mem->Opcode() != Op_StoreCM) { |
0 | 2430 // Looking at a dead closed cycle of memory? |
2431 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); | |
2432 | |
2433 assert(Opcode() == mem->Opcode() || | |
2434 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, | |
2435 "no mismatched stores, except on raw memory"); | |
2436 | |
2437 if (mem->outcnt() == 1 && // check for intervening uses | |
2438 mem->as_Store()->memory_size() <= this->memory_size()) { | |
2439 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. | |
2440 // For example, 'mem' might be the final state at a conditional return. | |
2441 // Or, 'mem' might be used by some node which is live at the same time | |
2442 // 'this' is live, which might be unschedulable. So, require exactly | |
2443 // ONE user, the 'this' store, until such time as we clone 'mem' for | |
2444 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). | |
2445 if (can_reshape) { // (%%% is this an anachronism?) | |
2446 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), | |
2447 phase->is_IterGVN()); | |
2448 } else { | |
2449 // It's OK to do this in the parser, since DU info is always accurate, | |
2450 // and the parser always refers to nodes via SafePointNode maps. | |
2451 set_req(MemNode::Memory, mem->in(MemNode::Memory)); | |
2452 } | |
2453 return this; | |
2454 } | |
2455 } | |
2456 | |
2457 // Capture an unaliased, unconditional, simple store into an initializer. | |
2458 // Or, if it is independent of the allocation, hoist it above the allocation. | |
2459 if (ReduceFieldZeroing && /*can_reshape &&*/ | |
2460 mem->is_Proj() && mem->in(0)->is_Initialize()) { | |
2461 InitializeNode* init = mem->in(0)->as_Initialize(); | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
2462 intptr_t offset = init->can_capture_store(this, phase, can_reshape); |
0 | 2463 if (offset > 0) { |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
2464 Node* moved = init->capture_store(this, offset, phase, can_reshape); |
0 | 2465 // If the InitializeNode captured me, it made a raw copy of me, |
2466 // and I need to disappear. | |
2467 if (moved != NULL) { | |
2468 // %%% hack to ensure that Ideal returns a new node: | |
2469 mem = MergeMemNode::make(phase->C, mem); | |
2470 return mem; // fold me away | |
2471 } | |
2472 } | |
2473 } | |
2474 | |
2475 return NULL; // No further progress | |
2476 } | |
2477 | |
2478 //------------------------------Value----------------------------------------- | |
2479 const Type *StoreNode::Value( PhaseTransform *phase ) const { | |
2480 // Either input is TOP ==> the result is TOP | |
2481 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
2482 if( t1 == Type::TOP ) return Type::TOP; | |
2483 const Type *t2 = phase->type( in(MemNode::Address) ); | |
2484 if( t2 == Type::TOP ) return Type::TOP; | |
2485 const Type *t3 = phase->type( in(MemNode::ValueIn) ); | |
2486 if( t3 == Type::TOP ) return Type::TOP; | |
2487 return Type::MEMORY; | |
2488 } | |
2489 | |
2490 //------------------------------Identity--------------------------------------- | |
2491 // Remove redundant stores: | |
2492 // Store(m, p, Load(m, p)) changes to m. | |
2493 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). | |
2494 Node *StoreNode::Identity( PhaseTransform *phase ) { | |
2495 Node* mem = in(MemNode::Memory); | |
2496 Node* adr = in(MemNode::Address); | |
2497 Node* val = in(MemNode::ValueIn); | |
2498 | |
2499 // Load then Store? Then the Store is useless | |
2500 if (val->is_Load() && | |
4778 | 2501 val->in(MemNode::Address)->eqv_uncast(adr) && |
2502 val->in(MemNode::Memory )->eqv_uncast(mem) && | |
0 | 2503 val->as_Load()->store_Opcode() == Opcode()) { |
2504 return mem; | |
2505 } | |
2506 | |
2507 // Two stores in a row of the same value? | |
2508 if (mem->is_Store() && | |
4778 | 2509 mem->in(MemNode::Address)->eqv_uncast(adr) && |
2510 mem->in(MemNode::ValueIn)->eqv_uncast(val) && | |
0 | 2511 mem->Opcode() == Opcode()) { |
2512 return mem; | |
2513 } | |
2514 | |
2515 // Store of zero anywhere into a freshly-allocated object? | |
2516 // Then the store is useless. | |
2517 // (It must already have been captured by the InitializeNode.) | |
2518 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { | |
2519 // a newly allocated object is already all-zeroes everywhere | |
2520 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { | |
2521 return mem; | |
2522 } | |
2523 | |
2524 // the store may also apply to zero-bits in an earlier object | |
2525 Node* prev_mem = find_previous_store(phase); | |
2526 // Steps (a), (b): Walk past independent stores to find an exact match. | |
2527 if (prev_mem != NULL) { | |
2528 Node* prev_val = can_see_stored_value(prev_mem, phase); | |
2529 if (prev_val != NULL && phase->eqv(prev_val, val)) { | |
2530 // prev_val and val might differ by a cast; it would be good | |
2531 // to keep the more informative of the two. | |
2532 return mem; | |
2533 } | |
2534 } | |
2535 } | |
2536 | |
2537 return this; | |
2538 } | |
2539 | |
2540 //------------------------------match_edge------------------------------------- | |
2541 // Do we Match on this edge index or not? Match only memory & value | |
2542 uint StoreNode::match_edge(uint idx) const { | |
2543 return idx == MemNode::Address || idx == MemNode::ValueIn; | |
2544 } | |
2545 | |
2546 //------------------------------cmp-------------------------------------------- | |
2547 // Do not common stores up together. They generally have to be split | |
2548 // back up anyways, so do not bother. | |
2549 uint StoreNode::cmp( const Node &n ) const { | |
2550 return (&n == this); // Always fail except on self | |
2551 } | |
2552 | |
2553 //------------------------------Ideal_masked_input----------------------------- | |
2554 // Check for a useless mask before a partial-word store | |
2555 // (StoreB ... (AndI valIn conIa) ) | |
2556 // If (conIa & mask == mask) this simplifies to | |
2557 // (StoreB ... (valIn) ) | |
2558 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { | |
2559 Node *val = in(MemNode::ValueIn); | |
2560 if( val->Opcode() == Op_AndI ) { | |
2561 const TypeInt *t = phase->type( val->in(2) )->isa_int(); | |
2562 if( t && t->is_con() && (t->get_con() & mask) == mask ) { | |
2563 set_req(MemNode::ValueIn, val->in(1)); | |
2564 return this; | |
2565 } | |
2566 } | |
2567 return NULL; | |
2568 } | |
2569 | |
2570 | |
2571 //------------------------------Ideal_sign_extended_input---------------------- | |
2572 // Check for useless sign-extension before a partial-word store | |
2573 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) | |
2574 // If (conIL == conIR && conIR <= num_bits) this simplifies to | |
2575 // (StoreB ... (valIn) ) | |
2576 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { | |
2577 Node *val = in(MemNode::ValueIn); | |
2578 if( val->Opcode() == Op_RShiftI ) { | |
2579 const TypeInt *t = phase->type( val->in(2) )->isa_int(); | |
2580 if( t && t->is_con() && (t->get_con() <= num_bits) ) { | |
2581 Node *shl = val->in(1); | |
2582 if( shl->Opcode() == Op_LShiftI ) { | |
2583 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); | |
2584 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { | |
2585 set_req(MemNode::ValueIn, shl->in(1)); | |
2586 return this; | |
2587 } | |
2588 } | |
2589 } | |
2590 } | |
2591 return NULL; | |
2592 } | |
2593 | |
2594 //------------------------------value_never_loaded----------------------------------- | |
2595 // Determine whether there are any possible loads of the value stored. | |
2596 // For simplicity, we actually check if there are any loads from the | |
2597 // address stored to, not just for loads of the value stored by this node. | |
2598 // | |
2599 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { | |
2600 Node *adr = in(Address); | |
2601 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); | |
2602 if (adr_oop == NULL) | |
2603 return false; | |
223 | 2604 if (!adr_oop->is_known_instance_field()) |
0 | 2605 return false; // if not a distinct instance, there may be aliases of the address |
2606 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { | |
2607 Node *use = adr->fast_out(i); | |
2608 int opc = use->Opcode(); | |
2609 if (use->is_Load() || use->is_LoadStore()) { | |
2610 return false; | |
2611 } | |
2612 } | |
2613 return true; | |
2614 } | |
2615 | |
2616 //============================================================================= | |
2617 //------------------------------Ideal------------------------------------------ | |
2618 // If the store is from an AND mask that leaves the low bits untouched, then | |
2619 // we can skip the AND operation. If the store is from a sign-extension | |
2620 // (a left shift, then right shift) we can skip both. | |
2621 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2622 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); | |
2623 if( progress != NULL ) return progress; | |
2624 | |
2625 progress = StoreNode::Ideal_sign_extended_input(phase, 24); | |
2626 if( progress != NULL ) return progress; | |
2627 | |
2628 // Finally check the default case | |
2629 return StoreNode::Ideal(phase, can_reshape); | |
2630 } | |
2631 | |
2632 //============================================================================= | |
2633 //------------------------------Ideal------------------------------------------ | |
2634 // If the store is from an AND mask that leaves the low bits untouched, then | |
2635 // we can skip the AND operation | |
2636 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2637 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); | |
2638 if( progress != NULL ) return progress; | |
2639 | |
2640 progress = StoreNode::Ideal_sign_extended_input(phase, 16); | |
2641 if( progress != NULL ) return progress; | |
2642 | |
2643 // Finally check the default case | |
2644 return StoreNode::Ideal(phase, can_reshape); | |
2645 } | |
2646 | |
2647 //============================================================================= | |
2648 //------------------------------Identity--------------------------------------- | |
2649 Node *StoreCMNode::Identity( PhaseTransform *phase ) { | |
2650 // No need to card mark when storing a null ptr | |
2651 Node* my_store = in(MemNode::OopStore); | |
2652 if (my_store->is_Store()) { | |
2653 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); | |
2654 if( t1 == TypePtr::NULL_PTR ) { | |
2655 return in(MemNode::Memory); | |
2656 } | |
2657 } | |
2658 return this; | |
2659 } | |
2660 | |
985
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2661 //============================================================================= |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2662 //------------------------------Ideal--------------------------------------- |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2663 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){ |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2664 Node* progress = StoreNode::Ideal(phase, can_reshape); |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2665 if (progress != NULL) return progress; |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2666 |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2667 Node* my_store = in(MemNode::OopStore); |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2668 if (my_store->is_MergeMem()) { |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2669 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx()); |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2670 set_req(MemNode::OopStore, mem); |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2671 return this; |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2672 } |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2673 |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2674 return NULL; |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2675 } |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2676 |
0 | 2677 //------------------------------Value----------------------------------------- |
2678 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { | |
43 | 2679 // Either input is TOP ==> the result is TOP |
2680 const Type *t = phase->type( in(MemNode::Memory) ); | |
2681 if( t == Type::TOP ) return Type::TOP; | |
2682 t = phase->type( in(MemNode::Address) ); | |
2683 if( t == Type::TOP ) return Type::TOP; | |
2684 t = phase->type( in(MemNode::ValueIn) ); | |
2685 if( t == Type::TOP ) return Type::TOP; | |
0 | 2686 // If extra input is TOP ==> the result is TOP |
43 | 2687 t = phase->type( in(MemNode::OopStore) ); |
2688 if( t == Type::TOP ) return Type::TOP; | |
0 | 2689 |
2690 return StoreNode::Value( phase ); | |
2691 } | |
2692 | |
2693 | |
2694 //============================================================================= | |
2695 //----------------------------------SCMemProjNode------------------------------ | |
2696 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const | |
2697 { | |
2698 return bottom_type(); | |
2699 } | |
2700 | |
2701 //============================================================================= | |
6795
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2702 //----------------------------------LoadStoreNode------------------------------ |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2703 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required ) |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2704 : Node(required), |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2705 _type(rt), |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2706 _adr_type(at) |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2707 { |
0 | 2708 init_req(MemNode::Control, c ); |
2709 init_req(MemNode::Memory , mem); | |
2710 init_req(MemNode::Address, adr); | |
2711 init_req(MemNode::ValueIn, val); | |
2712 init_class_id(Class_LoadStore); | |
6795
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2713 } |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2714 |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2715 uint LoadStoreNode::ideal_reg() const { |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2716 return _type->ideal_reg(); |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2717 } |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2718 |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2719 bool LoadStoreNode::result_not_used() const { |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2720 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) { |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2721 Node *x = fast_out(i); |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2722 if (x->Opcode() == Op_SCMemProj) continue; |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2723 return false; |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2724 } |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2725 return true; |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2726 } |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2727 |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2728 uint LoadStoreNode::size_of() const { return sizeof(*this); } |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2729 |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2730 //============================================================================= |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2731 //----------------------------------LoadStoreConditionalNode-------------------- |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2732 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) { |
7eca5de9e0b6
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
roland
parents:
6725
diff
changeset
|
2733 init_req(ExpectedIn, ex ); |
0 | 2734 } |
2735 | |
2736 //============================================================================= | |
2737 //-------------------------------adr_type-------------------------------------- | |
2738 // Do we Match on this edge index or not? Do not match memory | |
2739 const TypePtr* ClearArrayNode::adr_type() const { | |
2740 Node *adr = in(3); | |
2741 return MemNode::calculate_adr_type(adr->bottom_type()); | |
2742 } | |
2743 | |
2744 //------------------------------match_edge------------------------------------- | |
2745 // Do we Match on this edge index or not? Do not match memory | |
2746 uint ClearArrayNode::match_edge(uint idx) const { | |
2747 return idx > 1; | |
2748 } | |
2749 | |
2750 //------------------------------Identity--------------------------------------- | |
2751 // Clearing a zero length array does nothing | |
2752 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2753 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; |
0 | 2754 } |
2755 | |
2756 //------------------------------Idealize--------------------------------------- | |
2757 // Clearing a short array is faster with stores | |
2758 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2759 const int unit = BytesPerLong; | |
2760 const TypeX* t = phase->type(in(2))->isa_intptr_t(); | |
2761 if (!t) return NULL; | |
2762 if (!t->is_con()) return NULL; | |
2763 intptr_t raw_count = t->get_con(); | |
2764 intptr_t size = raw_count; | |
2765 if (!Matcher::init_array_count_is_in_bytes) size *= unit; | |
2766 // Clearing nothing uses the Identity call. | |
2767 // Negative clears are possible on dead ClearArrays | |
2768 // (see jck test stmt114.stmt11402.val). | |
2769 if (size <= 0 || size % unit != 0) return NULL; | |
2770 intptr_t count = size / unit; | |
2771 // Length too long; use fast hardware clear | |
2772 if (size > Matcher::init_array_short_size) return NULL; | |
2773 Node *mem = in(1); | |
2774 if( phase->type(mem)==Type::TOP ) return NULL; | |
2775 Node *adr = in(3); | |
2776 const Type* at = phase->type(adr); | |
2777 if( at==Type::TOP ) return NULL; | |
2778 const TypePtr* atp = at->isa_ptr(); | |
2779 // adjust atp to be the correct array element address type | |
2780 if (atp == NULL) atp = TypePtr::BOTTOM; | |
2781 else atp = atp->add_offset(Type::OffsetBot); | |
2782 // Get base for derived pointer purposes | |
2783 if( adr->Opcode() != Op_AddP ) Unimplemented(); | |
2784 Node *base = adr->in(1); | |
2785 | |
2786 Node *zero = phase->makecon(TypeLong::ZERO); | |
2787 Node *off = phase->MakeConX(BytesPerLong); | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2788 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); |
0 | 2789 count--; |
2790 while( count-- ) { | |
2791 mem = phase->transform(mem); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2792 adr = phase->transform(new (phase->C) AddPNode(base,adr,off)); |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2793 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false); |
0 | 2794 } |
2795 return mem; | |
2796 } | |
2797 | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2798 //----------------------------step_through---------------------------------- |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2799 // Return allocation input memory edge if it is different instance |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2800 // or itself if it is the one we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2801 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2802 Node* n = *np; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2803 assert(n->is_ClearArray(), "sanity"); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2804 intptr_t offset; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2805 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2806 // This method is called only before Allocate nodes are expanded during |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2807 // macro nodes expansion. Before that ClearArray nodes are only generated |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2808 // in LibraryCallKit::generate_arraycopy() which follows allocations. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2809 assert(alloc != NULL, "should have allocation"); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2810 if (alloc->_idx == instance_id) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2811 // Can not bypass initialization of the instance we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2812 return false; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2813 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2814 // Otherwise skip it. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2815 InitializeNode* init = alloc->initialization(); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2816 if (init != NULL) |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2817 *np = init->in(TypeFunc::Memory); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2818 else |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2819 *np = alloc->in(TypeFunc::Memory); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2820 return true; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2821 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2822 |
0 | 2823 //----------------------------clear_memory------------------------------------- |
2824 // Generate code to initialize object storage to zero. | |
2825 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2826 intptr_t start_offset, | |
2827 Node* end_offset, | |
2828 PhaseGVN* phase) { | |
2829 Compile* C = phase->C; | |
2830 intptr_t offset = start_offset; | |
2831 | |
2832 int unit = BytesPerLong; | |
2833 if ((offset % unit) != 0) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2834 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset)); |
0 | 2835 adr = phase->transform(adr); |
2836 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2837 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); |
0 | 2838 mem = phase->transform(mem); |
2839 offset += BytesPerInt; | |
2840 } | |
2841 assert((offset % unit) == 0, ""); | |
2842 | |
2843 // Initialize the remaining stuff, if any, with a ClearArray. | |
2844 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); | |
2845 } | |
2846 | |
2847 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2848 Node* start_offset, | |
2849 Node* end_offset, | |
2850 PhaseGVN* phase) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2851 if (start_offset == end_offset) { |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2852 // nothing to do |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2853 return mem; |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2854 } |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2855 |
0 | 2856 Compile* C = phase->C; |
2857 int unit = BytesPerLong; | |
2858 Node* zbase = start_offset; | |
2859 Node* zend = end_offset; | |
2860 | |
2861 // Scale to the unit required by the CPU: | |
2862 if (!Matcher::init_array_count_is_in_bytes) { | |
2863 Node* shift = phase->intcon(exact_log2(unit)); | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2864 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2865 zend = phase->transform( new(C) URShiftXNode(zend, shift) ); |
0 | 2866 } |
2867 | |
7474
00af3a3a8df4
8005522: use fast-string instructions on x86 for zeroing
kvn
parents:
6853
diff
changeset
|
2868 // Bulk clear double-words |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2869 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2870 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) ); |
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2871 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr); |
0 | 2872 return phase->transform(mem); |
2873 } | |
2874 | |
2875 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2876 intptr_t start_offset, | |
2877 intptr_t end_offset, | |
2878 PhaseGVN* phase) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2879 if (start_offset == end_offset) { |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2880 // nothing to do |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2881 return mem; |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2882 } |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2883 |
0 | 2884 Compile* C = phase->C; |
2885 assert((end_offset % BytesPerInt) == 0, "odd end offset"); | |
2886 intptr_t done_offset = end_offset; | |
2887 if ((done_offset % BytesPerLong) != 0) { | |
2888 done_offset -= BytesPerInt; | |
2889 } | |
2890 if (done_offset > start_offset) { | |
2891 mem = clear_memory(ctl, mem, dest, | |
2892 start_offset, phase->MakeConX(done_offset), phase); | |
2893 } | |
2894 if (done_offset < end_offset) { // emit the final 32-bit store | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
2895 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset)); |
0 | 2896 adr = phase->transform(adr); |
2897 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
2898 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered); |
0 | 2899 mem = phase->transform(mem); |
2900 done_offset += BytesPerInt; | |
2901 } | |
2902 assert(done_offset == end_offset, ""); | |
2903 return mem; | |
2904 } | |
2905 | |
2906 //============================================================================= | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2907 // Do not match memory edge. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2908 uint StrIntrinsicNode::match_edge(uint idx) const { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2909 return idx == 2 || idx == 3; |
681 | 2910 } |
2911 | |
2912 //------------------------------Ideal------------------------------------------ | |
2913 // Return a node which is more "ideal" than the current node. Strip out | |
2914 // control copies | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2915 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2916 if (remove_dead_region(phase, can_reshape)) return this; |
4115 | 2917 // Don't bother trying to transform a dead node |
2918 if (in(0) && in(0)->is_top()) return NULL; | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2919 |
2417 | 2920 if (can_reshape) { |
2921 Node* mem = phase->transform(in(MemNode::Memory)); | |
2922 // If transformed to a MergeMem, get the desired slice | |
2923 uint alias_idx = phase->C->get_alias_index(adr_type()); | |
2924 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; | |
2925 if (mem != in(MemNode::Memory)) { | |
2926 set_req(MemNode::Memory, mem); | |
2927 return this; | |
2928 } | |
2929 } | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2930 return NULL; |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
2931 } |
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
2932 |
4115 | 2933 //------------------------------Value------------------------------------------ |
2934 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { | |
2935 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; | |
2936 return bottom_type(); | |
2937 } | |
2938 | |
0 | 2939 //============================================================================= |
7637
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2940 //------------------------------match_edge------------------------------------- |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2941 // Do not match memory edge |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2942 uint EncodeISOArrayNode::match_edge(uint idx) const { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2943 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len) |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2944 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2945 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2946 //------------------------------Ideal------------------------------------------ |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2947 // Return a node which is more "ideal" than the current node. Strip out |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2948 // control copies |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2949 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2950 return remove_dead_region(phase, can_reshape) ? this : NULL; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2951 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2952 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2953 //------------------------------Value------------------------------------------ |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2954 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const { |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2955 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2956 return bottom_type(); |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2957 } |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2958 |
b30b3c2a0cf2
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
kvn
parents:
7474
diff
changeset
|
2959 //============================================================================= |
0 | 2960 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) |
2961 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), | |
2962 _adr_type(C->get_adr_type(alias_idx)) | |
2963 { | |
2964 init_class_id(Class_MemBar); | |
2965 Node* top = C->top(); | |
2966 init_req(TypeFunc::I_O,top); | |
2967 init_req(TypeFunc::FramePtr,top); | |
2968 init_req(TypeFunc::ReturnAdr,top); | |
2969 if (precedent != NULL) | |
2970 init_req(TypeFunc::Parms, precedent); | |
2971 } | |
2972 | |
2973 //------------------------------cmp-------------------------------------------- | |
2974 uint MemBarNode::hash() const { return NO_HASH; } | |
2975 uint MemBarNode::cmp( const Node &n ) const { | |
2976 return (&n == this); // Always fail except on self | |
2977 } | |
2978 | |
2979 //------------------------------make------------------------------------------- | |
2980 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { | |
2981 switch (opcode) { | |
14439
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2982 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2983 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2984 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2985 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2986 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2987 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2988 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2989 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2990 case Op_Initialize: return new(C) InitializeNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2991 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn); |
50fdb38839eb
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
goetz
parents:
14429
diff
changeset
|
2992 default: ShouldNotReachHere(); return NULL; |
0 | 2993 } |
2994 } | |
2995 | |
2996 //------------------------------Ideal------------------------------------------ | |
2997 // Return a node which is more "ideal" than the current node. Strip out | |
2998 // control copies | |
2999 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3000 if (remove_dead_region(phase, can_reshape)) return this; |
4115 | 3001 // Don't bother trying to transform a dead node |
11005
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3002 if (in(0) && in(0)->is_top()) { |
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3003 return NULL; |
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3004 } |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3005 |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3006 // Eliminate volatile MemBars for scalar replaced objects. |
10278 | 3007 if (can_reshape && req() == (Precedent+1)) { |
3008 bool eliminate = false; | |
3009 int opc = Opcode(); | |
3010 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) { | |
3011 // Volatile field loads and stores. | |
3012 Node* my_mem = in(MemBarNode::Precedent); | |
11005
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3013 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge |
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3014 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) { |
11044 | 3015 // if the Precedent is a decodeN and its input (a Load) is used at more than one place, |
3016 // replace this Precedent (decodeN) with the Load instead. | |
3017 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) { | |
3018 Node* load_node = my_mem->in(1); | |
3019 set_req(MemBarNode::Precedent, load_node); | |
3020 phase->is_IterGVN()->_worklist.push(my_mem); | |
3021 my_mem = load_node; | |
3022 } else { | |
3023 assert(my_mem->unique_out() == this, "sanity"); | |
3024 del_req(Precedent); | |
3025 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later | |
3026 my_mem = NULL; | |
3027 } | |
11005
08d35fd1b599
8001345: VM crashes with assert(n->outcnt() != 0 || C->top() == n || n->is_Proj()) failed: No dead instructions after post-alloc
adlertz
parents:
10278
diff
changeset
|
3028 } |
10278 | 3029 if (my_mem != NULL && my_mem->is_Mem()) { |
3030 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); | |
3031 // Check for scalar replaced object reference. | |
3032 if( t_oop != NULL && t_oop->is_known_instance_field() && | |
3033 t_oop->offset() != Type::OffsetBot && | |
3034 t_oop->offset() != Type::OffsetTop) { | |
3035 eliminate = true; | |
3036 } | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3037 } |
10278 | 3038 } else if (opc == Op_MemBarRelease) { |
3039 // Final field stores. | |
3040 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase); | |
3041 if ((alloc != NULL) && alloc->is_Allocate() && | |
3042 alloc->as_Allocate()->_is_non_escaping) { | |
3043 // The allocated object does not escape. | |
3044 eliminate = true; | |
3045 } | |
3046 } | |
3047 if (eliminate) { | |
3048 // Replace MemBar projections by its inputs. | |
3049 PhaseIterGVN* igvn = phase->is_IterGVN(); | |
3050 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); | |
3051 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); | |
3052 // Must return either the original node (now dead) or a new node | |
3053 // (Do not return a top here, since that would break the uniqueness of top.) | |
3054 return new (phase->C) ConINode(TypeInt::ZERO); | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3055 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3056 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
3057 return NULL; |
0 | 3058 } |
3059 | |
3060 //------------------------------Value------------------------------------------ | |
3061 const Type *MemBarNode::Value( PhaseTransform *phase ) const { | |
3062 if( !in(0) ) return Type::TOP; | |
3063 if( phase->type(in(0)) == Type::TOP ) | |
3064 return Type::TOP; | |
3065 return TypeTuple::MEMBAR; | |
3066 } | |
3067 | |
3068 //------------------------------match------------------------------------------ | |
3069 // Construct projections for memory. | |
3070 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { | |
3071 switch (proj->_con) { | |
3072 case TypeFunc::Control: | |
3073 case TypeFunc::Memory: | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
3074 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); |
0 | 3075 } |
3076 ShouldNotReachHere(); | |
3077 return NULL; | |
3078 } | |
3079 | |
3080 //===========================InitializeNode==================================== | |
3081 // SUMMARY: | |
3082 // This node acts as a memory barrier on raw memory, after some raw stores. | |
3083 // The 'cooked' oop value feeds from the Initialize, not the Allocation. | |
3084 // The Initialize can 'capture' suitably constrained stores as raw inits. | |
3085 // It can coalesce related raw stores into larger units (called 'tiles'). | |
3086 // It can avoid zeroing new storage for memory units which have raw inits. | |
3087 // At macro-expansion, it is marked 'complete', and does not optimize further. | |
3088 // | |
3089 // EXAMPLE: | |
3090 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. | |
3091 // ctl = incoming control; mem* = incoming memory | |
3092 // (Note: A star * on a memory edge denotes I/O and other standard edges.) | |
3093 // First allocate uninitialized memory and fill in the header: | |
3094 // alloc = (Allocate ctl mem* 16 #short[].klass ...) | |
3095 // ctl := alloc.Control; mem* := alloc.Memory* | |
3096 // rawmem = alloc.Memory; rawoop = alloc.RawAddress | |
3097 // Then initialize to zero the non-header parts of the raw memory block: | |
3098 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) | |
3099 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory | |
3100 // After the initialize node executes, the object is ready for service: | |
3101 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) | |
3102 // Suppose its body is immediately initialized as {1,2}: | |
3103 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) | |
3104 // store2 = (StoreC init.Control store1 (+ oop 14) 2) | |
3105 // mem.SLICE(#short[*]) := store2 | |
3106 // | |
3107 // DETAILS: | |
3108 // An InitializeNode collects and isolates object initialization after | |
3109 // an AllocateNode and before the next possible safepoint. As a | |
3110 // memory barrier (MemBarNode), it keeps critical stores from drifting | |
3111 // down past any safepoint or any publication of the allocation. | |
3112 // Before this barrier, a newly-allocated object may have uninitialized bits. | |
3113 // After this barrier, it may be treated as a real oop, and GC is allowed. | |
3114 // | |
3115 // The semantics of the InitializeNode include an implicit zeroing of | |
3116 // the new object from object header to the end of the object. | |
3117 // (The object header and end are determined by the AllocateNode.) | |
3118 // | |
3119 // Certain stores may be added as direct inputs to the InitializeNode. | |
3120 // These stores must update raw memory, and they must be to addresses | |
3121 // derived from the raw address produced by AllocateNode, and with | |
3122 // a constant offset. They must be ordered by increasing offset. | |
3123 // The first one is at in(RawStores), the last at in(req()-1). | |
3124 // Unlike most memory operations, they are not linked in a chain, | |
3125 // but are displayed in parallel as users of the rawmem output of | |
3126 // the allocation. | |
3127 // | |
3128 // (See comments in InitializeNode::capture_store, which continue | |
3129 // the example given above.) | |
3130 // | |
3131 // When the associated Allocate is macro-expanded, the InitializeNode | |
3132 // may be rewritten to optimize collected stores. A ClearArrayNode | |
3133 // may also be created at that point to represent any required zeroing. | |
3134 // The InitializeNode is then marked 'complete', prohibiting further | |
3135 // capturing of nearby memory operations. | |
3136 // | |
3137 // During macro-expansion, all captured initializations which store | |
605 | 3138 // constant values of 32 bits or smaller are coalesced (if advantageous) |
0 | 3139 // into larger 'tiles' 32 or 64 bits. This allows an object to be |
3140 // initialized in fewer memory operations. Memory words which are | |
3141 // covered by neither tiles nor non-constant stores are pre-zeroed | |
3142 // by explicit stores of zero. (The code shape happens to do all | |
3143 // zeroing first, then all other stores, with both sequences occurring | |
3144 // in order of ascending offsets.) | |
3145 // | |
3146 // Alternatively, code may be inserted between an AllocateNode and its | |
3147 // InitializeNode, to perform arbitrary initialization of the new object. | |
3148 // E.g., the object copying intrinsics insert complex data transfers here. | |
3149 // The initialization must then be marked as 'complete' disable the | |
3150 // built-in zeroing semantics and the collection of initializing stores. | |
3151 // | |
3152 // While an InitializeNode is incomplete, reads from the memory state | |
3153 // produced by it are optimizable if they match the control edge and | |
3154 // new oop address associated with the allocation/initialization. | |
3155 // They return a stored value (if the offset matches) or else zero. | |
3156 // A write to the memory state, if it matches control and address, | |
3157 // and if it is to a constant offset, may be 'captured' by the | |
3158 // InitializeNode. It is cloned as a raw memory operation and rewired | |
3159 // inside the initialization, to the raw oop produced by the allocation. | |
3160 // Operations on addresses which are provably distinct (e.g., to | |
3161 // other AllocateNodes) are allowed to bypass the initialization. | |
3162 // | |
3163 // The effect of all this is to consolidate object initialization | |
3164 // (both arrays and non-arrays, both piecewise and bulk) into a | |
3165 // single location, where it can be optimized as a unit. | |
3166 // | |
3167 // Only stores with an offset less than TrackedInitializationLimit words | |
3168 // will be considered for capture by an InitializeNode. This puts a | |
3169 // reasonable limit on the complexity of optimized initializations. | |
3170 | |
3171 //---------------------------InitializeNode------------------------------------ | |
3172 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) | |
4763
1dc233a8c7fe
7121140: Allocation paths require explicit memory synchronization operations for RMO systems
roland
parents:
4762
diff
changeset
|
3173 : _is_complete(Incomplete), _does_not_escape(false), |
0 | 3174 MemBarNode(C, adr_type, rawoop) |
3175 { | |
3176 init_class_id(Class_Initialize); | |
3177 | |
3178 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); | |
3179 assert(in(RawAddress) == rawoop, "proper init"); | |
3180 // Note: allocation() can be NULL, for secondary initialization barriers | |
3181 } | |
3182 | |
3183 // Since this node is not matched, it will be processed by the | |
3184 // register allocator. Declare that there are no constraints | |
3185 // on the allocation of the RawAddress edge. | |
3186 const RegMask &InitializeNode::in_RegMask(uint idx) const { | |
3187 // This edge should be set to top, by the set_complete. But be conservative. | |
3188 if (idx == InitializeNode::RawAddress) | |
3189 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); | |
3190 return RegMask::Empty; | |
3191 } | |
3192 | |
3193 Node* InitializeNode::memory(uint alias_idx) { | |
3194 Node* mem = in(Memory); | |
3195 if (mem->is_MergeMem()) { | |
3196 return mem->as_MergeMem()->memory_at(alias_idx); | |
3197 } else { | |
3198 // incoming raw memory is not split | |
3199 return mem; | |
3200 } | |
3201 } | |
3202 | |
3203 bool InitializeNode::is_non_zero() { | |
3204 if (is_complete()) return false; | |
3205 remove_extra_zeroes(); | |
3206 return (req() > RawStores); | |
3207 } | |
3208 | |
3209 void InitializeNode::set_complete(PhaseGVN* phase) { | |
3210 assert(!is_complete(), "caller responsibility"); | |
3961
a92cdbac8b9e
7081933: Use zeroing elimination optimization for large array
kvn
parents:
3902
diff
changeset
|
3211 _is_complete = Complete; |
0 | 3212 |
3213 // After this node is complete, it contains a bunch of | |
3214 // raw-memory initializations. There is no need for | |
3215 // it to have anything to do with non-raw memory effects. | |
3216 // Therefore, tell all non-raw users to re-optimize themselves, | |
3217 // after skipping the memory effects of this initialization. | |
3218 PhaseIterGVN* igvn = phase->is_IterGVN(); | |
3219 if (igvn) igvn->add_users_to_worklist(this); | |
3220 } | |
3221 | |
3222 // convenience function | |
3223 // return false if the init contains any stores already | |
3224 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { | |
3225 InitializeNode* init = initialization(); | |
3226 if (init == NULL || init->is_complete()) return false; | |
3227 init->remove_extra_zeroes(); | |
3228 // for now, if this allocation has already collected any inits, bail: | |
3229 if (init->is_non_zero()) return false; | |
3230 init->set_complete(phase); | |
3231 return true; | |
3232 } | |
3233 | |
3234 void InitializeNode::remove_extra_zeroes() { | |
3235 if (req() == RawStores) return; | |
3236 Node* zmem = zero_memory(); | |
3237 uint fill = RawStores; | |
3238 for (uint i = fill; i < req(); i++) { | |
3239 Node* n = in(i); | |
3240 if (n->is_top() || n == zmem) continue; // skip | |
3241 if (fill < i) set_req(fill, n); // compact | |
3242 ++fill; | |
3243 } | |
3244 // delete any empty spaces created: | |
3245 while (fill < req()) { | |
3246 del_req(fill); | |
3247 } | |
3248 } | |
3249 | |
3250 // Helper for remembering which stores go with which offsets. | |
3251 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { | |
3252 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node | |
3253 intptr_t offset = -1; | |
3254 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), | |
3255 phase, offset); | |
3256 if (base == NULL) return -1; // something is dead, | |
3257 if (offset < 0) return -1; // dead, dead | |
3258 return offset; | |
3259 } | |
3260 | |
3261 // Helper for proving that an initialization expression is | |
3262 // "simple enough" to be folded into an object initialization. | |
3263 // Attempts to prove that a store's initial value 'n' can be captured | |
3264 // within the initialization without creating a vicious cycle, such as: | |
3265 // { Foo p = new Foo(); p.next = p; } | |
3266 // True for constants and parameters and small combinations thereof. | |
10278 | 3267 bool InitializeNode::detect_init_independence(Node* n, int& count) { |
0 | 3268 if (n == NULL) return true; // (can this really happen?) |
3269 if (n->is_Proj()) n = n->in(0); | |
3270 if (n == this) return false; // found a cycle | |
3271 if (n->is_Con()) return true; | |
3272 if (n->is_Start()) return true; // params, etc., are OK | |
3273 if (n->is_Root()) return true; // even better | |
3274 | |
3275 Node* ctl = n->in(0); | |
3276 if (ctl != NULL && !ctl->is_top()) { | |
3277 if (ctl->is_Proj()) ctl = ctl->in(0); | |
3278 if (ctl == this) return false; | |
3279 | |
3280 // If we already know that the enclosing memory op is pinned right after | |
3281 // the init, then any control flow that the store has picked up | |
3282 // must have preceded the init, or else be equal to the init. | |
3283 // Even after loop optimizations (which might change control edges) | |
3284 // a store is never pinned *before* the availability of its inputs. | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
3285 if (!MemNode::all_controls_dominate(n, this)) |
0 | 3286 return false; // failed to prove a good control |
3287 } | |
3288 | |
3289 // Check data edges for possible dependencies on 'this'. | |
3290 if ((count += 1) > 20) return false; // complexity limit | |
3291 for (uint i = 1; i < n->req(); i++) { | |
3292 Node* m = n->in(i); | |
3293 if (m == NULL || m == n || m->is_top()) continue; | |
3294 uint first_i = n->find_edge(m); | |
3295 if (i != first_i) continue; // process duplicate edge just once | |
10278 | 3296 if (!detect_init_independence(m, count)) { |
0 | 3297 return false; |
3298 } | |
3299 } | |
3300 | |
3301 return true; | |
3302 } | |
3303 | |
3304 // Here are all the checks a Store must pass before it can be moved into | |
3305 // an initialization. Returns zero if a check fails. | |
3306 // On success, returns the (constant) offset to which the store applies, | |
3307 // within the initialized memory. | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3308 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) { |
0 | 3309 const int FAIL = 0; |
3310 if (st->req() != MemNode::ValueIn + 1) | |
3311 return FAIL; // an inscrutable StoreNode (card mark?) | |
3312 Node* ctl = st->in(MemNode::Control); | |
3313 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) | |
3314 return FAIL; // must be unconditional after the initialization | |
3315 Node* mem = st->in(MemNode::Memory); | |
3316 if (!(mem->is_Proj() && mem->in(0) == this)) | |
3317 return FAIL; // must not be preceded by other stores | |
3318 Node* adr = st->in(MemNode::Address); | |
3319 intptr_t offset; | |
3320 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); | |
3321 if (alloc == NULL) | |
3322 return FAIL; // inscrutable address | |
3323 if (alloc != allocation()) | |
3324 return FAIL; // wrong allocation! (store needs to float up) | |
3325 Node* val = st->in(MemNode::ValueIn); | |
3326 int complexity_count = 0; | |
10278 | 3327 if (!detect_init_independence(val, complexity_count)) |
0 | 3328 return FAIL; // stored value must be 'simple enough' |
3329 | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3330 // The Store can be captured only if nothing after the allocation |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3331 // and before the Store is using the memory location that the store |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3332 // overwrites. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3333 bool failed = false; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3334 // If is_complete_with_arraycopy() is true the shape of the graph is |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3335 // well defined and is safe so no need for extra checks. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3336 if (!is_complete_with_arraycopy()) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3337 // We are going to look at each use of the memory state following |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3338 // the allocation to make sure nothing reads the memory that the |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3339 // Store writes. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3340 const TypePtr* t_adr = phase->type(adr)->isa_ptr(); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3341 int alias_idx = phase->C->get_alias_index(t_adr); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3342 ResourceMark rm; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3343 Unique_Node_List mems; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3344 mems.push(mem); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3345 Node* unique_merge = NULL; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3346 for (uint next = 0; next < mems.size(); ++next) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3347 Node *m = mems.at(next); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3348 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3349 Node *n = m->fast_out(j); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3350 if (n->outcnt() == 0) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3351 continue; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3352 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3353 if (n == st) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3354 continue; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3355 } else if (n->in(0) != NULL && n->in(0) != ctl) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3356 // If the control of this use is different from the control |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3357 // of the Store which is right after the InitializeNode then |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3358 // this node cannot be between the InitializeNode and the |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3359 // Store. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3360 continue; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3361 } else if (n->is_MergeMem()) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3362 if (n->as_MergeMem()->memory_at(alias_idx) == m) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3363 // We can hit a MergeMemNode (that will likely go away |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3364 // later) that is a direct use of the memory state |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3365 // following the InitializeNode on the same slice as the |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3366 // store node that we'd like to capture. We need to check |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3367 // the uses of the MergeMemNode. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3368 mems.push(n); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3369 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3370 } else if (n->is_Mem()) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3371 Node* other_adr = n->in(MemNode::Address); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3372 if (other_adr == adr) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3373 failed = true; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3374 break; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3375 } else { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3376 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr(); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3377 if (other_t_adr != NULL) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3378 int other_alias_idx = phase->C->get_alias_index(other_t_adr); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3379 if (other_alias_idx == alias_idx) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3380 // A load from the same memory slice as the store right |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3381 // after the InitializeNode. We check the control of the |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3382 // object/array that is loaded from. If it's the same as |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3383 // the store control then we cannot capture the store. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3384 assert(!n->is_Store(), "2 stores to same slice on same control?"); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3385 Node* base = other_adr; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3386 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name())); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3387 base = base->in(AddPNode::Base); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3388 if (base != NULL) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3389 base = base->uncast(); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3390 if (base->is_Proj() && base->in(0) == alloc) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3391 failed = true; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3392 break; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3393 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3394 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3395 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3396 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3397 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3398 } else { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3399 failed = true; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3400 break; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3401 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3402 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3403 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3404 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3405 if (failed) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3406 if (!can_reshape) { |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3407 // We decided we couldn't capture the store during parsing. We |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3408 // should try again during the next IGVN once the graph is |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3409 // cleaner. |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3410 phase->C->record_for_igvn(st); |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3411 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3412 return FAIL; |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3413 } |
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3414 |
0 | 3415 return offset; // success |
3416 } | |
3417 | |
3418 // Find the captured store in(i) which corresponds to the range | |
3419 // [start..start+size) in the initialized object. | |
3420 // If there is one, return its index i. If there isn't, return the | |
3421 // negative of the index where it should be inserted. | |
3422 // Return 0 if the queried range overlaps an initialization boundary | |
3423 // or if dead code is encountered. | |
3424 // If size_in_bytes is zero, do not bother with overlap checks. | |
3425 int InitializeNode::captured_store_insertion_point(intptr_t start, | |
3426 int size_in_bytes, | |
3427 PhaseTransform* phase) { | |
3428 const int FAIL = 0, MAX_STORE = BytesPerLong; | |
3429 | |
3430 if (is_complete()) | |
3431 return FAIL; // arraycopy got here first; punt | |
3432 | |
3433 assert(allocation() != NULL, "must be present"); | |
3434 | |
3435 // no negatives, no header fields: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3436 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; |
0 | 3437 |
3438 // after a certain size, we bail out on tracking all the stores: | |
3439 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); | |
3440 if (start >= ti_limit) return FAIL; | |
3441 | |
3442 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { | |
3443 if (i >= limit) return -(int)i; // not found; here is where to put it | |
3444 | |
3445 Node* st = in(i); | |
3446 intptr_t st_off = get_store_offset(st, phase); | |
3447 if (st_off < 0) { | |
3448 if (st != zero_memory()) { | |
3449 return FAIL; // bail out if there is dead garbage | |
3450 } | |
3451 } else if (st_off > start) { | |
3452 // ...we are done, since stores are ordered | |
3453 if (st_off < start + size_in_bytes) { | |
3454 return FAIL; // the next store overlaps | |
3455 } | |
3456 return -(int)i; // not found; here is where to put it | |
3457 } else if (st_off < start) { | |
3458 if (size_in_bytes != 0 && | |
3459 start < st_off + MAX_STORE && | |
3460 start < st_off + st->as_Store()->memory_size()) { | |
3461 return FAIL; // the previous store overlaps | |
3462 } | |
3463 } else { | |
3464 if (size_in_bytes != 0 && | |
3465 st->as_Store()->memory_size() != size_in_bytes) { | |
3466 return FAIL; // mismatched store size | |
3467 } | |
3468 return i; | |
3469 } | |
3470 | |
3471 ++i; | |
3472 } | |
3473 } | |
3474 | |
3475 // Look for a captured store which initializes at the offset 'start' | |
3476 // with the given size. If there is no such store, and no other | |
3477 // initialization interferes, then return zero_memory (the memory | |
3478 // projection of the AllocateNode). | |
3479 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, | |
3480 PhaseTransform* phase) { | |
3481 assert(stores_are_sane(phase), ""); | |
3482 int i = captured_store_insertion_point(start, size_in_bytes, phase); | |
3483 if (i == 0) { | |
3484 return NULL; // something is dead | |
3485 } else if (i < 0) { | |
3486 return zero_memory(); // just primordial zero bits here | |
3487 } else { | |
3488 Node* st = in(i); // here is the store at this position | |
3489 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); | |
3490 return st; | |
3491 } | |
3492 } | |
3493 | |
3494 // Create, as a raw pointer, an address within my new object at 'offset'. | |
3495 Node* InitializeNode::make_raw_address(intptr_t offset, | |
3496 PhaseTransform* phase) { | |
3497 Node* addr = in(RawAddress); | |
3498 if (offset != 0) { | |
3499 Compile* C = phase->C; | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
3500 addr = phase->transform( new (C) AddPNode(C->top(), addr, |
0 | 3501 phase->MakeConX(offset)) ); |
3502 } | |
3503 return addr; | |
3504 } | |
3505 | |
3506 // Clone the given store, converting it into a raw store | |
3507 // initializing a field or element of my new object. | |
3508 // Caller is responsible for retiring the original store, | |
3509 // with subsume_node or the like. | |
3510 // | |
3511 // From the example above InitializeNode::InitializeNode, | |
3512 // here are the old stores to be captured: | |
3513 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) | |
3514 // store2 = (StoreC init.Control store1 (+ oop 14) 2) | |
3515 // | |
3516 // Here is the changed code; note the extra edges on init: | |
3517 // alloc = (Allocate ...) | |
3518 // rawoop = alloc.RawAddress | |
3519 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) | |
3520 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) | |
3521 // init = (Initialize alloc.Control alloc.Memory rawoop | |
3522 // rawstore1 rawstore2) | |
3523 // | |
3524 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3525 PhaseTransform* phase, bool can_reshape) { |
0 | 3526 assert(stores_are_sane(phase), ""); |
3527 | |
3528 if (start < 0) return NULL; | |
8116
6931f425c517
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
roland
parents:
7637
diff
changeset
|
3529 assert(can_capture_store(st, phase, can_reshape) == start, "sanity"); |
0 | 3530 |
3531 Compile* C = phase->C; | |
3532 int size_in_bytes = st->memory_size(); | |
3533 int i = captured_store_insertion_point(start, size_in_bytes, phase); | |
3534 if (i == 0) return NULL; // bail out | |
3535 Node* prev_mem = NULL; // raw memory for the captured store | |
3536 if (i > 0) { | |
3537 prev_mem = in(i); // there is a pre-existing store under this one | |
3538 set_req(i, C->top()); // temporarily disconnect it | |
3539 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. | |
3540 } else { | |
3541 i = -i; // no pre-existing store | |
3542 prev_mem = zero_memory(); // a slice of the newly allocated object | |
3543 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) | |
3544 set_req(--i, C->top()); // reuse this edge; it has been folded away | |
3545 else | |
3546 ins_req(i, C->top()); // build a new edge | |
3547 } | |
3548 Node* new_st = st->clone(); | |
3549 new_st->set_req(MemNode::Control, in(Control)); | |
3550 new_st->set_req(MemNode::Memory, prev_mem); | |
3551 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); | |
3552 new_st = phase->transform(new_st); | |
3553 | |
3554 // At this point, new_st might have swallowed a pre-existing store | |
3555 // at the same offset, or perhaps new_st might have disappeared, | |
3556 // if it redundantly stored the same value (or zero to fresh memory). | |
3557 | |
3558 // In any case, wire it in: | |
3559 set_req(i, new_st); | |
3560 | |
3561 // The caller may now kill the old guy. | |
3562 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); | |
3563 assert(check_st == new_st || check_st == NULL, "must be findable"); | |
3564 assert(!is_complete(), ""); | |
3565 return new_st; | |
3566 } | |
3567 | |
3568 static bool store_constant(jlong* tiles, int num_tiles, | |
3569 intptr_t st_off, int st_size, | |
3570 jlong con) { | |
3571 if ((st_off & (st_size-1)) != 0) | |
3572 return false; // strange store offset (assume size==2**N) | |
3573 address addr = (address)tiles + st_off; | |
3574 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); | |
3575 switch (st_size) { | |
3576 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; | |
3577 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; | |
3578 case sizeof(jint): *(jint*) addr = (jint) con; break; | |
3579 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; | |
3580 default: return false; // strange store size (detect size!=2**N here) | |
3581 } | |
3582 return true; // return success to caller | |
3583 } | |
3584 | |
3585 // Coalesce subword constants into int constants and possibly | |
3586 // into long constants. The goal, if the CPU permits, | |
3587 // is to initialize the object with a small number of 64-bit tiles. | |
3588 // Also, convert floating-point constants to bit patterns. | |
3589 // Non-constants are not relevant to this pass. | |
3590 // | |
3591 // In terms of the running example on InitializeNode::InitializeNode | |
3592 // and InitializeNode::capture_store, here is the transformation | |
3593 // of rawstore1 and rawstore2 into rawstore12: | |
3594 // alloc = (Allocate ...) | |
3595 // rawoop = alloc.RawAddress | |
3596 // tile12 = 0x00010002 | |
3597 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) | |
3598 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) | |
3599 // | |
3600 void | |
3601 InitializeNode::coalesce_subword_stores(intptr_t header_size, | |
3602 Node* size_in_bytes, | |
3603 PhaseGVN* phase) { | |
3604 Compile* C = phase->C; | |
3605 | |
3606 assert(stores_are_sane(phase), ""); | |
3607 // Note: After this pass, they are not completely sane, | |
3608 // since there may be some overlaps. | |
3609 | |
3610 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; | |
3611 | |
3612 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); | |
3613 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); | |
3614 size_limit = MIN2(size_limit, ti_limit); | |
3615 size_limit = align_size_up(size_limit, BytesPerLong); | |
3616 int num_tiles = size_limit / BytesPerLong; | |
3617 | |
3618 // allocate space for the tile map: | |
3619 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small | |
3620 jlong tiles_buf[small_len]; | |
3621 Node* nodes_buf[small_len]; | |
3622 jlong inits_buf[small_len]; | |
3623 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] | |
3624 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); | |
3625 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] | |
3626 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); | |
3627 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] | |
3628 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); | |
3629 // tiles: exact bitwise model of all primitive constants | |
3630 // nodes: last constant-storing node subsumed into the tiles model | |
3631 // inits: which bytes (in each tile) are touched by any initializations | |
3632 | |
3633 //// Pass A: Fill in the tile model with any relevant stores. | |
3634 | |
3635 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); | |
3636 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); | |
3637 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); | |
3638 Node* zmem = zero_memory(); // initially zero memory state | |
3639 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { | |
3640 Node* st = in(i); | |
3641 intptr_t st_off = get_store_offset(st, phase); | |
3642 | |
3643 // Figure out the store's offset and constant value: | |
3644 if (st_off < header_size) continue; //skip (ignore header) | |
3645 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) | |
3646 int st_size = st->as_Store()->memory_size(); | |
3647 if (st_off + st_size > size_limit) break; | |
3648 | |
3649 // Record which bytes are touched, whether by constant or not. | |
3650 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) | |
3651 continue; // skip (strange store size) | |
3652 | |
3653 const Type* val = phase->type(st->in(MemNode::ValueIn)); | |
3654 if (!val->singleton()) continue; //skip (non-con store) | |
3655 BasicType type = val->basic_type(); | |
3656 | |
3657 jlong con = 0; | |
3658 switch (type) { | |
3659 case T_INT: con = val->is_int()->get_con(); break; | |
3660 case T_LONG: con = val->is_long()->get_con(); break; | |
3661 case T_FLOAT: con = jint_cast(val->getf()); break; | |
3662 case T_DOUBLE: con = jlong_cast(val->getd()); break; | |
3663 default: continue; //skip (odd store type) | |
3664 } | |
3665 | |
3666 if (type == T_LONG && Matcher::isSimpleConstant64(con) && | |
3667 st->Opcode() == Op_StoreL) { | |
3668 continue; // This StoreL is already optimal. | |
3669 } | |
3670 | |
3671 // Store down the constant. | |
3672 store_constant(tiles, num_tiles, st_off, st_size, con); | |
3673 | |
3674 intptr_t j = st_off >> LogBytesPerLong; | |
3675 | |
3676 if (type == T_INT && st_size == BytesPerInt | |
3677 && (st_off & BytesPerInt) == BytesPerInt) { | |
3678 jlong lcon = tiles[j]; | |
3679 if (!Matcher::isSimpleConstant64(lcon) && | |
3680 st->Opcode() == Op_StoreI) { | |
3681 // This StoreI is already optimal by itself. | |
3682 jint* intcon = (jint*) &tiles[j]; | |
3683 intcon[1] = 0; // undo the store_constant() | |
3684 | |
3685 // If the previous store is also optimal by itself, back up and | |
3686 // undo the action of the previous loop iteration... if we can. | |
3687 // But if we can't, just let the previous half take care of itself. | |
3688 st = nodes[j]; | |
3689 st_off -= BytesPerInt; | |
3690 con = intcon[0]; | |
3691 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { | |
3692 assert(st_off >= header_size, "still ignoring header"); | |
3693 assert(get_store_offset(st, phase) == st_off, "must be"); | |
3694 assert(in(i-1) == zmem, "must be"); | |
3695 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); | |
3696 assert(con == tcon->is_int()->get_con(), "must be"); | |
3697 // Undo the effects of the previous loop trip, which swallowed st: | |
3698 intcon[0] = 0; // undo store_constant() | |
3699 set_req(i-1, st); // undo set_req(i, zmem) | |
3700 nodes[j] = NULL; // undo nodes[j] = st | |
3701 --old_subword; // undo ++old_subword | |
3702 } | |
3703 continue; // This StoreI is already optimal. | |
3704 } | |
3705 } | |
3706 | |
3707 // This store is not needed. | |
3708 set_req(i, zmem); | |
3709 nodes[j] = st; // record for the moment | |
3710 if (st_size < BytesPerLong) // something has changed | |
3711 ++old_subword; // includes int/float, but who's counting... | |
3712 else ++old_long; | |
3713 } | |
3714 | |
3715 if ((old_subword + old_long) == 0) | |
3716 return; // nothing more to do | |
3717 | |
3718 //// Pass B: Convert any non-zero tiles into optimal constant stores. | |
3719 // Be sure to insert them before overlapping non-constant stores. | |
3720 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) | |
3721 for (int j = 0; j < num_tiles; j++) { | |
3722 jlong con = tiles[j]; | |
3723 jlong init = inits[j]; | |
3724 if (con == 0) continue; | |
3725 jint con0, con1; // split the constant, address-wise | |
3726 jint init0, init1; // split the init map, address-wise | |
3727 { union { jlong con; jint intcon[2]; } u; | |
3728 u.con = con; | |
3729 con0 = u.intcon[0]; | |
3730 con1 = u.intcon[1]; | |
3731 u.con = init; | |
3732 init0 = u.intcon[0]; | |
3733 init1 = u.intcon[1]; | |
3734 } | |
3735 | |
3736 Node* old = nodes[j]; | |
3737 assert(old != NULL, "need the prior store"); | |
3738 intptr_t offset = (j * BytesPerLong); | |
3739 | |
3740 bool split = !Matcher::isSimpleConstant64(con); | |
3741 | |
3742 if (offset < header_size) { | |
3743 assert(offset + BytesPerInt >= header_size, "second int counts"); | |
3744 assert(*(jint*)&tiles[j] == 0, "junk in header"); | |
3745 split = true; // only the second word counts | |
3746 // Example: int a[] = { 42 ... } | |
3747 } else if (con0 == 0 && init0 == -1) { | |
3748 split = true; // first word is covered by full inits | |
3749 // Example: int a[] = { ... foo(), 42 ... } | |
3750 } else if (con1 == 0 && init1 == -1) { | |
3751 split = true; // second word is covered by full inits | |
3752 // Example: int a[] = { ... 42, foo() ... } | |
3753 } | |
3754 | |
3755 // Here's a case where init0 is neither 0 nor -1: | |
3756 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } | |
3757 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. | |
3758 // In this case the tile is not split; it is (jlong)42. | |
3759 // The big tile is stored down, and then the foo() value is inserted. | |
3760 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) | |
3761 | |
3762 Node* ctl = old->in(MemNode::Control); | |
3763 Node* adr = make_raw_address(offset, phase); | |
3764 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
3765 | |
3766 // One or two coalesced stores to plop down. | |
3767 Node* st[2]; | |
3768 intptr_t off[2]; | |
3769 int nst = 0; | |
3770 if (!split) { | |
3771 ++new_long; | |
3772 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3773 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
3774 phase->longcon(con), T_LONG, MemNode::unordered); |
0 | 3775 } else { |
3776 // Omit either if it is a zero. | |
3777 if (con0 != 0) { | |
3778 ++new_int; | |
3779 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3780 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
3781 phase->intcon(con0), T_INT, MemNode::unordered); |
0 | 3782 } |
3783 if (con1 != 0) { | |
3784 ++new_int; | |
3785 offset += BytesPerInt; | |
3786 adr = make_raw_address(offset, phase); | |
3787 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3788 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
14429
2113136690bc
8024921: PPC64 (part 113): Extend Load and Store nodes to know about memory ordering
goetz
parents:
12242
diff
changeset
|
3789 phase->intcon(con1), T_INT, MemNode::unordered); |
0 | 3790 } |
3791 } | |
3792 | |
3793 // Insert second store first, then the first before the second. | |
3794 // Insert each one just before any overlapping non-constant stores. | |
3795 while (nst > 0) { | |
3796 Node* st1 = st[--nst]; | |
3797 C->copy_node_notes_to(st1, old); | |
3798 st1 = phase->transform(st1); | |
3799 offset = off[nst]; | |
3800 assert(offset >= header_size, "do not smash header"); | |
3801 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); | |
3802 guarantee(ins_idx != 0, "must re-insert constant store"); | |
3803 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap | |
3804 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) | |
3805 set_req(--ins_idx, st1); | |
3806 else | |
3807 ins_req(ins_idx, st1); | |
3808 } | |
3809 } | |
3810 | |
3811 if (PrintCompilation && WizardMode) | |
3812 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", | |
3813 old_subword, old_long, new_int, new_long); | |
3814 if (C->log() != NULL) | |
3815 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", | |
3816 old_subword, old_long, new_int, new_long); | |
3817 | |
3818 // Clean up any remaining occurrences of zmem: | |
3819 remove_extra_zeroes(); | |
3820 } | |
3821 | |
3822 // Explore forward from in(start) to find the first fully initialized | |
3823 // word, and return its offset. Skip groups of subword stores which | |
3824 // together initialize full words. If in(start) is itself part of a | |
3825 // fully initialized word, return the offset of in(start). If there | |
3826 // are no following full-word stores, or if something is fishy, return | |
3827 // a negative value. | |
3828 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { | |
3829 int int_map = 0; | |
3830 intptr_t int_map_off = 0; | |
3831 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for | |
3832 | |
3833 for (uint i = start, limit = req(); i < limit; i++) { | |
3834 Node* st = in(i); | |
3835 | |
3836 intptr_t st_off = get_store_offset(st, phase); | |
3837 if (st_off < 0) break; // return conservative answer | |
3838 | |
3839 int st_size = st->as_Store()->memory_size(); | |
3840 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { | |
3841 return st_off; // we found a complete word init | |
3842 } | |
3843 | |
3844 // update the map: | |
3845 | |
3846 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); | |
3847 if (this_int_off != int_map_off) { | |
3848 // reset the map: | |
3849 int_map = 0; | |
3850 int_map_off = this_int_off; | |
3851 } | |
3852 | |
3853 int subword_off = st_off - this_int_off; | |
3854 int_map |= right_n_bits(st_size) << subword_off; | |
3855 if ((int_map & FULL_MAP) == FULL_MAP) { | |
3856 return this_int_off; // we found a complete word init | |
3857 } | |
3858 | |
3859 // Did this store hit or cross the word boundary? | |
3860 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); | |
3861 if (next_int_off == this_int_off + BytesPerInt) { | |
3862 // We passed the current int, without fully initializing it. | |
3863 int_map_off = next_int_off; | |
3864 int_map >>= BytesPerInt; | |
3865 } else if (next_int_off > this_int_off + BytesPerInt) { | |
3866 // We passed the current and next int. | |
3867 return this_int_off + BytesPerInt; | |
3868 } | |
3869 } | |
3870 | |
3871 return -1; | |
3872 } | |
3873 | |
3874 | |
3875 // Called when the associated AllocateNode is expanded into CFG. | |
3876 // At this point, we may perform additional optimizations. | |
3877 // Linearize the stores by ascending offset, to make memory | |
3878 // activity as coherent as possible. | |
3879 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, | |
3880 intptr_t header_size, | |
3881 Node* size_in_bytes, | |
3882 PhaseGVN* phase) { | |
3883 assert(!is_complete(), "not already complete"); | |
3884 assert(stores_are_sane(phase), ""); | |
3885 assert(allocation() != NULL, "must be present"); | |
3886 | |
3887 remove_extra_zeroes(); | |
3888 | |
3889 if (ReduceFieldZeroing || ReduceBulkZeroing) | |
3890 // reduce instruction count for common initialization patterns | |
3891 coalesce_subword_stores(header_size, size_in_bytes, phase); | |
3892 | |
3893 Node* zmem = zero_memory(); // initially zero memory state | |
3894 Node* inits = zmem; // accumulating a linearized chain of inits | |
3895 #ifdef ASSERT | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3896 intptr_t first_offset = allocation()->minimum_header_size(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3897 intptr_t last_init_off = first_offset; // previous init offset |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3898 intptr_t last_init_end = first_offset; // previous init offset+size |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3899 intptr_t last_tile_end = first_offset; // previous tile offset+size |
0 | 3900 #endif |
3901 intptr_t zeroes_done = header_size; | |
3902 | |
3903 bool do_zeroing = true; // we might give up if inits are very sparse | |
3904 int big_init_gaps = 0; // how many large gaps have we seen? | |
3905 | |
3906 if (ZeroTLAB) do_zeroing = false; | |
3907 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; | |
3908 | |
3909 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { | |
3910 Node* st = in(i); | |
3911 intptr_t st_off = get_store_offset(st, phase); | |
3912 if (st_off < 0) | |
3913 break; // unknown junk in the inits | |
3914 if (st->in(MemNode::Memory) != zmem) | |
3915 break; // complicated store chains somehow in list | |
3916 | |
3917 int st_size = st->as_Store()->memory_size(); | |
3918 intptr_t next_init_off = st_off + st_size; | |
3919 | |
3920 if (do_zeroing && zeroes_done < next_init_off) { | |
3921 // See if this store needs a zero before it or under it. | |
3922 intptr_t zeroes_needed = st_off; | |
3923 | |
3924 if (st_size < BytesPerInt) { | |
3925 // Look for subword stores which only partially initialize words. | |
3926 // If we find some, we must lay down some word-level zeroes first, | |
3927 // underneath the subword stores. | |
3928 // | |
3929 // Examples: | |
3930 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s | |
3931 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y | |
3932 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z | |
3933 // | |
3934 // Note: coalesce_subword_stores may have already done this, | |
3935 // if it was prompted by constant non-zero subword initializers. | |
3936 // But this case can still arise with non-constant stores. | |
3937 | |
3938 intptr_t next_full_store = find_next_fullword_store(i, phase); | |
3939 | |
3940 // In the examples above: | |
3941 // in(i) p q r s x y z | |
3942 // st_off 12 13 14 15 12 13 14 | |
3943 // st_size 1 1 1 1 1 1 1 | |
3944 // next_full_s. 12 16 16 16 16 16 16 | |
3945 // z's_done 12 16 16 16 12 16 12 | |
3946 // z's_needed 12 16 16 16 16 16 16 | |
3947 // zsize 0 0 0 0 4 0 4 | |
3948 if (next_full_store < 0) { | |
3949 // Conservative tack: Zero to end of current word. | |
3950 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); | |
3951 } else { | |
3952 // Zero to beginning of next fully initialized word. | |
3953 // Or, don't zero at all, if we are already in that word. | |
3954 assert(next_full_store >= zeroes_needed, "must go forward"); | |
3955 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); | |
3956 zeroes_needed = next_full_store; | |
3957 } | |
3958 } | |
3959 | |
3960 if (zeroes_needed > zeroes_done) { | |
3961 intptr_t zsize = zeroes_needed - zeroes_done; | |
3962 // Do some incremental zeroing on rawmem, in parallel with inits. | |
3963 zeroes_done = align_size_down(zeroes_done, BytesPerInt); | |
3964 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, | |
3965 zeroes_done, zeroes_needed, | |
3966 phase); | |
3967 zeroes_done = zeroes_needed; | |
3968 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) | |
3969 do_zeroing = false; // leave the hole, next time | |
3970 } | |
3971 } | |
3972 | |
3973 // Collect the store and move on: | |
3974 st->set_req(MemNode::Memory, inits); | |
3975 inits = st; // put it on the linearized chain | |
3976 set_req(i, zmem); // unhook from previous position | |
3977 | |
3978 if (zeroes_done == st_off) | |
3979 zeroes_done = next_init_off; | |
3980 | |
3981 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); | |
3982 | |
3983 #ifdef ASSERT | |
3984 // Various order invariants. Weaker than stores_are_sane because | |
3985 // a large constant tile can be filled in by smaller non-constant stores. | |
3986 assert(st_off >= last_init_off, "inits do not reverse"); | |
3987 last_init_off = st_off; | |
3988 const Type* val = NULL; | |
3989 if (st_size >= BytesPerInt && | |
3990 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && | |
3991 (int)val->basic_type() < (int)T_OBJECT) { | |
3992 assert(st_off >= last_tile_end, "tiles do not overlap"); | |
3993 assert(st_off >= last_init_end, "tiles do not overwrite inits"); | |
3994 last_tile_end = MAX2(last_tile_end, next_init_off); | |
3995 } else { | |
3996 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); | |
3997 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); | |
3998 assert(st_off >= last_init_end, "inits do not overlap"); | |
3999 last_init_end = next_init_off; // it's a non-tile | |
4000 } | |
4001 #endif //ASSERT | |
4002 } | |
4003 | |
4004 remove_extra_zeroes(); // clear out all the zmems left over | |
4005 add_req(inits); | |
4006 | |
4007 if (!ZeroTLAB) { | |
4008 // If anything remains to be zeroed, zero it all now. | |
4009 zeroes_done = align_size_down(zeroes_done, BytesPerInt); | |
4010 // if it is the last unused 4 bytes of an instance, forget about it | |
4011 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); | |
4012 if (zeroes_done + BytesPerLong >= size_limit) { | |
4013 assert(allocation() != NULL, ""); | |
2004
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4014 if (allocation()->Opcode() == Op_Allocate) { |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4015 Node* klass_node = allocation()->in(AllocateNode::KlassNode); |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4016 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4017 if (zeroes_done == k->layout_helper()) |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4018 zeroes_done = size_limit; |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
4019 } |
0 | 4020 } |
4021 if (zeroes_done < size_limit) { | |
4022 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, | |
4023 zeroes_done, size_in_bytes, phase); | |
4024 } | |
4025 } | |
4026 | |
4027 set_complete(phase); | |
4028 return rawmem; | |
4029 } | |
4030 | |
4031 | |
4032 #ifdef ASSERT | |
4033 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { | |
4034 if (is_complete()) | |
4035 return true; // stores could be anything at this point | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
4036 assert(allocation() != NULL, "must be present"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
4037 intptr_t last_off = allocation()->minimum_header_size(); |
0 | 4038 for (uint i = InitializeNode::RawStores; i < req(); i++) { |
4039 Node* st = in(i); | |
4040 intptr_t st_off = get_store_offset(st, phase); | |
4041 if (st_off < 0) continue; // ignore dead garbage | |
4042 if (last_off > st_off) { | |
4043 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); | |
4044 this->dump(2); | |
4045 assert(false, "ascending store offsets"); | |
4046 return false; | |
4047 } | |
4048 last_off = st_off + st->as_Store()->memory_size(); | |
4049 } | |
4050 return true; | |
4051 } | |
4052 #endif //ASSERT | |
4053 | |
4054 | |
4055 | |
4056 | |
4057 //============================MergeMemNode===================================== | |
4058 // | |
4059 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several | |
4060 // contributing store or call operations. Each contributor provides the memory | |
4061 // state for a particular "alias type" (see Compile::alias_type). For example, | |
4062 // if a MergeMem has an input X for alias category #6, then any memory reference | |
4063 // to alias category #6 may use X as its memory state input, as an exact equivalent | |
4064 // to using the MergeMem as a whole. | |
4065 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) | |
4066 // | |
4067 // (Here, the <N> notation gives the index of the relevant adr_type.) | |
4068 // | |
4069 // In one special case (and more cases in the future), alias categories overlap. | |
4070 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory | |
4071 // states. Therefore, if a MergeMem has only one contributing input W for Bot, | |
4072 // it is exactly equivalent to that state W: | |
4073 // MergeMem(<Bot>: W) <==> W | |
4074 // | |
4075 // Usually, the merge has more than one input. In that case, where inputs | |
4076 // overlap (i.e., one is Bot), the narrower alias type determines the memory | |
4077 // state for that type, and the wider alias type (Bot) fills in everywhere else: | |
4078 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) | |
4079 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) | |
4080 // | |
4081 // A merge can take a "wide" memory state as one of its narrow inputs. | |
4082 // This simply means that the merge observes out only the relevant parts of | |
4083 // the wide input. That is, wide memory states arriving at narrow merge inputs | |
4084 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) | |
4085 // | |
4086 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), | |
4087 // and that memory slices "leak through": | |
4088 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) | |
4089 // | |
4090 // But, in such a cascade, repeated memory slices can "block the leak": | |
4091 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') | |
4092 // | |
4093 // In the last example, Y is not part of the combined memory state of the | |
4094 // outermost MergeMem. The system must, of course, prevent unschedulable | |
4095 // memory states from arising, so you can be sure that the state Y is somehow | |
4096 // a precursor to state Y'. | |
4097 // | |
4098 // | |
4099 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array | |
4100 // of each MergeMemNode array are exactly the numerical alias indexes, including | |
4101 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions | |
4102 // Compile::alias_type (and kin) produce and manage these indexes. | |
4103 // | |
4104 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. | |
4105 // (Note that this provides quick access to the top node inside MergeMem methods, | |
4106 // without the need to reach out via TLS to Compile::current.) | |
4107 // | |
4108 // As a consequence of what was just described, a MergeMem that represents a full | |
4109 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, | |
4110 // containing all alias categories. | |
4111 // | |
4112 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. | |
4113 // | |
4114 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either | |
4115 // a memory state for the alias type <N>, or else the top node, meaning that | |
4116 // there is no particular input for that alias type. Note that the length of | |
4117 // a MergeMem is variable, and may be extended at any time to accommodate new | |
4118 // memory states at larger alias indexes. When merges grow, they are of course | |
4119 // filled with "top" in the unused in() positions. | |
4120 // | |
4121 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. | |
4122 // (Top was chosen because it works smoothly with passes like GCM.) | |
4123 // | |
4124 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is | |
4125 // the type of random VM bits like TLS references.) Since it is always the | |
4126 // first non-Bot memory slice, some low-level loops use it to initialize an | |
4127 // index variable: for (i = AliasIdxRaw; i < req(); i++). | |
4128 // | |
4129 // | |
4130 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns | |
4131 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns | |
4132 // the memory state for alias type <N>, or (if there is no particular slice at <N>, | |
4133 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> | |
4134 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over | |
4135 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. | |
4136 // | |
4137 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't | |
4138 // really that different from the other memory inputs. An abbreviation called | |
4139 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. | |
4140 // | |
4141 // | |
4142 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent | |
4143 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi | |
4144 // that "emerges though" the base memory will be marked as excluding the alias types | |
4145 // of the other (narrow-memory) copies which "emerged through" the narrow edges: | |
4146 // | |
4147 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) | |
4148 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) | |
4149 // | |
4150 // This strange "subtraction" effect is necessary to ensure IGVN convergence. | |
4151 // (It is currently unimplemented.) As you can see, the resulting merge is | |
4152 // actually a disjoint union of memory states, rather than an overlay. | |
4153 // | |
4154 | |
4155 //------------------------------MergeMemNode----------------------------------- | |
4156 Node* MergeMemNode::make_empty_memory() { | |
4157 Node* empty_memory = (Node*) Compile::current()->top(); | |
4158 assert(empty_memory->is_top(), "correct sentinel identity"); | |
4159 return empty_memory; | |
4160 } | |
4161 | |
4162 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { | |
4163 init_class_id(Class_MergeMem); | |
4164 // all inputs are nullified in Node::Node(int) | |
4165 // set_input(0, NULL); // no control input | |
4166 | |
4167 // Initialize the edges uniformly to top, for starters. | |
4168 Node* empty_mem = make_empty_memory(); | |
4169 for (uint i = Compile::AliasIdxTop; i < req(); i++) { | |
4170 init_req(i,empty_mem); | |
4171 } | |
4172 assert(empty_memory() == empty_mem, ""); | |
4173 | |
4174 if( new_base != NULL && new_base->is_MergeMem() ) { | |
4175 MergeMemNode* mdef = new_base->as_MergeMem(); | |
4176 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); | |
4177 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { | |
4178 mms.set_memory(mms.memory2()); | |
4179 } | |
4180 assert(base_memory() == mdef->base_memory(), ""); | |
4181 } else { | |
4182 set_base_memory(new_base); | |
4183 } | |
4184 } | |
4185 | |
4186 // Make a new, untransformed MergeMem with the same base as 'mem'. | |
4187 // If mem is itself a MergeMem, populate the result with the same edges. | |
4188 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { | |
6804
e626685e9f6c
7193318: C2: remove number of inputs requirement from Node's new operator
kvn
parents:
6795
diff
changeset
|
4189 return new(C) MergeMemNode(mem); |
0 | 4190 } |
4191 | |
4192 //------------------------------cmp-------------------------------------------- | |
4193 uint MergeMemNode::hash() const { return NO_HASH; } | |
4194 uint MergeMemNode::cmp( const Node &n ) const { | |
4195 return (&n == this); // Always fail except on self | |
4196 } | |
4197 | |
4198 //------------------------------Identity--------------------------------------- | |
4199 Node* MergeMemNode::Identity(PhaseTransform *phase) { | |
4200 // Identity if this merge point does not record any interesting memory | |
4201 // disambiguations. | |
4202 Node* base_mem = base_memory(); | |
4203 Node* empty_mem = empty_memory(); | |
4204 if (base_mem != empty_mem) { // Memory path is not dead? | |
4205 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4206 Node* mem = in(i); | |
4207 if (mem != empty_mem && mem != base_mem) { | |
4208 return this; // Many memory splits; no change | |
4209 } | |
4210 } | |
4211 } | |
4212 return base_mem; // No memory splits; ID on the one true input | |
4213 } | |
4214 | |
4215 //------------------------------Ideal------------------------------------------ | |
4216 // This method is invoked recursively on chains of MergeMem nodes | |
4217 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
4218 // Remove chain'd MergeMems | |
4219 // | |
4220 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted | |
4221 // relative to the "in(Bot)". Since we are patching both at the same time, | |
4222 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", | |
4223 // but rewrite each "in(i)" relative to the new "in(Bot)". | |
4224 Node *progress = NULL; | |
4225 | |
4226 | |
4227 Node* old_base = base_memory(); | |
4228 Node* empty_mem = empty_memory(); | |
4229 if (old_base == empty_mem) | |
4230 return NULL; // Dead memory path. | |
4231 | |
4232 MergeMemNode* old_mbase; | |
4233 if (old_base != NULL && old_base->is_MergeMem()) | |
4234 old_mbase = old_base->as_MergeMem(); | |
4235 else | |
4236 old_mbase = NULL; | |
4237 Node* new_base = old_base; | |
4238 | |
4239 // simplify stacked MergeMems in base memory | |
4240 if (old_mbase) new_base = old_mbase->base_memory(); | |
4241 | |
4242 // the base memory might contribute new slices beyond my req() | |
4243 if (old_mbase) grow_to_match(old_mbase); | |
4244 | |
4245 // Look carefully at the base node if it is a phi. | |
4246 PhiNode* phi_base; | |
4247 if (new_base != NULL && new_base->is_Phi()) | |
4248 phi_base = new_base->as_Phi(); | |
4249 else | |
4250 phi_base = NULL; | |
4251 | |
4252 Node* phi_reg = NULL; | |
4253 uint phi_len = (uint)-1; | |
4254 if (phi_base != NULL && !phi_base->is_copy()) { | |
4255 // do not examine phi if degraded to a copy | |
4256 phi_reg = phi_base->region(); | |
4257 phi_len = phi_base->req(); | |
4258 // see if the phi is unfinished | |
4259 for (uint i = 1; i < phi_len; i++) { | |
4260 if (phi_base->in(i) == NULL) { | |
4261 // incomplete phi; do not look at it yet! | |
4262 phi_reg = NULL; | |
4263 phi_len = (uint)-1; | |
4264 break; | |
4265 } | |
4266 } | |
4267 } | |
4268 | |
4269 // Note: We do not call verify_sparse on entry, because inputs | |
4270 // can normalize to the base_memory via subsume_node or similar | |
4271 // mechanisms. This method repairs that damage. | |
4272 | |
4273 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); | |
4274 | |
4275 // Look at each slice. | |
4276 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4277 Node* old_in = in(i); | |
4278 // calculate the old memory value | |
4279 Node* old_mem = old_in; | |
4280 if (old_mem == empty_mem) old_mem = old_base; | |
4281 assert(old_mem == memory_at(i), ""); | |
4282 | |
4283 // maybe update (reslice) the old memory value | |
4284 | |
4285 // simplify stacked MergeMems | |
4286 Node* new_mem = old_mem; | |
4287 MergeMemNode* old_mmem; | |
4288 if (old_mem != NULL && old_mem->is_MergeMem()) | |
4289 old_mmem = old_mem->as_MergeMem(); | |
4290 else | |
4291 old_mmem = NULL; | |
4292 if (old_mmem == this) { | |
4293 // This can happen if loops break up and safepoints disappear. | |
4294 // A merge of BotPtr (default) with a RawPtr memory derived from a | |
4295 // safepoint can be rewritten to a merge of the same BotPtr with | |
4296 // the BotPtr phi coming into the loop. If that phi disappears | |
4297 // also, we can end up with a self-loop of the mergemem. | |
4298 // In general, if loops degenerate and memory effects disappear, | |
4299 // a mergemem can be left looking at itself. This simply means | |
4300 // that the mergemem's default should be used, since there is | |
4301 // no longer any apparent effect on this slice. | |
4302 // Note: If a memory slice is a MergeMem cycle, it is unreachable | |
4303 // from start. Update the input to TOP. | |
4304 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; | |
4305 } | |
4306 else if (old_mmem != NULL) { | |
4307 new_mem = old_mmem->memory_at(i); | |
4308 } | |
605 | 4309 // else preceding memory was not a MergeMem |
0 | 4310 |
4311 // replace equivalent phis (unfortunately, they do not GVN together) | |
4312 if (new_mem != NULL && new_mem != new_base && | |
4313 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { | |
4314 if (new_mem->is_Phi()) { | |
4315 PhiNode* phi_mem = new_mem->as_Phi(); | |
4316 for (uint i = 1; i < phi_len; i++) { | |
4317 if (phi_base->in(i) != phi_mem->in(i)) { | |
4318 phi_mem = NULL; | |
4319 break; | |
4320 } | |
4321 } | |
4322 if (phi_mem != NULL) { | |
4323 // equivalent phi nodes; revert to the def | |
4324 new_mem = new_base; | |
4325 } | |
4326 } | |
4327 } | |
4328 | |
4329 // maybe store down a new value | |
4330 Node* new_in = new_mem; | |
4331 if (new_in == new_base) new_in = empty_mem; | |
4332 | |
4333 if (new_in != old_in) { | |
4334 // Warning: Do not combine this "if" with the previous "if" | |
4335 // A memory slice might have be be rewritten even if it is semantically | |
4336 // unchanged, if the base_memory value has changed. | |
4337 set_req(i, new_in); | |
4338 progress = this; // Report progress | |
4339 } | |
4340 } | |
4341 | |
4342 if (new_base != old_base) { | |
4343 set_req(Compile::AliasIdxBot, new_base); | |
4344 // Don't use set_base_memory(new_base), because we need to update du. | |
4345 assert(base_memory() == new_base, ""); | |
4346 progress = this; | |
4347 } | |
4348 | |
4349 if( base_memory() == this ) { | |
4350 // a self cycle indicates this memory path is dead | |
4351 set_req(Compile::AliasIdxBot, empty_mem); | |
4352 } | |
4353 | |
4354 // Resolve external cycles by calling Ideal on a MergeMem base_memory | |
4355 // Recursion must occur after the self cycle check above | |
4356 if( base_memory()->is_MergeMem() ) { | |
4357 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); | |
4358 Node *m = phase->transform(new_mbase); // Rollup any cycles | |
4359 if( m != NULL && (m->is_top() || | |
4360 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { | |
4361 // propagate rollup of dead cycle to self | |
4362 set_req(Compile::AliasIdxBot, empty_mem); | |
4363 } | |
4364 } | |
4365 | |
4366 if( base_memory() == empty_mem ) { | |
4367 progress = this; | |
4368 // Cut inputs during Parse phase only. | |
4369 // During Optimize phase a dead MergeMem node will be subsumed by Top. | |
4370 if( !can_reshape ) { | |
4371 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4372 if( in(i) != empty_mem ) { set_req(i, empty_mem); } | |
4373 } | |
4374 } | |
4375 } | |
4376 | |
4377 if( !progress && base_memory()->is_Phi() && can_reshape ) { | |
4378 // Check if PhiNode::Ideal's "Split phis through memory merges" | |
4379 // transform should be attempted. Look for this->phi->this cycle. | |
4380 uint merge_width = req(); | |
4381 if (merge_width > Compile::AliasIdxRaw) { | |
4382 PhiNode* phi = base_memory()->as_Phi(); | |
4383 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in | |
4384 if (phi->in(i) == this) { | |
4385 phase->is_IterGVN()->_worklist.push(phi); | |
4386 break; | |
4387 } | |
4388 } | |
4389 } | |
4390 } | |
4391 | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
4392 assert(progress || verify_sparse(), "please, no dups of base"); |
0 | 4393 return progress; |
4394 } | |
4395 | |
4396 //-------------------------set_base_memory------------------------------------- | |
4397 void MergeMemNode::set_base_memory(Node *new_base) { | |
4398 Node* empty_mem = empty_memory(); | |
4399 set_req(Compile::AliasIdxBot, new_base); | |
4400 assert(memory_at(req()) == new_base, "must set default memory"); | |
4401 // Clear out other occurrences of new_base: | |
4402 if (new_base != empty_mem) { | |
4403 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4404 if (in(i) == new_base) set_req(i, empty_mem); | |
4405 } | |
4406 } | |
4407 } | |
4408 | |
4409 //------------------------------out_RegMask------------------------------------ | |
4410 const RegMask &MergeMemNode::out_RegMask() const { | |
4411 return RegMask::Empty; | |
4412 } | |
4413 | |
4414 //------------------------------dump_spec-------------------------------------- | |
4415 #ifndef PRODUCT | |
4416 void MergeMemNode::dump_spec(outputStream *st) const { | |
4417 st->print(" {"); | |
4418 Node* base_mem = base_memory(); | |
4419 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { | |
4420 Node* mem = memory_at(i); | |
4421 if (mem == base_mem) { st->print(" -"); continue; } | |
4422 st->print( " N%d:", mem->_idx ); | |
4423 Compile::current()->get_adr_type(i)->dump_on(st); | |
4424 } | |
4425 st->print(" }"); | |
4426 } | |
4427 #endif // !PRODUCT | |
4428 | |
4429 | |
4430 #ifdef ASSERT | |
4431 static bool might_be_same(Node* a, Node* b) { | |
4432 if (a == b) return true; | |
4433 if (!(a->is_Phi() || b->is_Phi())) return false; | |
4434 // phis shift around during optimization | |
4435 return true; // pretty stupid... | |
4436 } | |
4437 | |
4438 // verify a narrow slice (either incoming or outgoing) | |
4439 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { | |
4440 if (!VerifyAliases) return; // don't bother to verify unless requested | |
4441 if (is_error_reported()) return; // muzzle asserts when debugging an error | |
4442 if (Node::in_dump()) return; // muzzle asserts when printing | |
4443 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); | |
4444 assert(n != NULL, ""); | |
4445 // Elide intervening MergeMem's | |
4446 while (n->is_MergeMem()) { | |
4447 n = n->as_MergeMem()->memory_at(alias_idx); | |
4448 } | |
4449 Compile* C = Compile::current(); | |
4450 const TypePtr* n_adr_type = n->adr_type(); | |
4451 if (n == m->empty_memory()) { | |
4452 // Implicit copy of base_memory() | |
4453 } else if (n_adr_type != TypePtr::BOTTOM) { | |
4454 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); | |
4455 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); | |
4456 } else { | |
4457 // A few places like make_runtime_call "know" that VM calls are narrow, | |
4458 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. | |
4459 bool expected_wide_mem = false; | |
4460 if (n == m->base_memory()) { | |
4461 expected_wide_mem = true; | |
4462 } else if (alias_idx == Compile::AliasIdxRaw || | |
4463 n == m->memory_at(Compile::AliasIdxRaw)) { | |
4464 expected_wide_mem = true; | |
4465 } else if (!C->alias_type(alias_idx)->is_rewritable()) { | |
4466 // memory can "leak through" calls on channels that | |
4467 // are write-once. Allow this also. | |
4468 expected_wide_mem = true; | |
4469 } | |
4470 assert(expected_wide_mem, "expected narrow slice replacement"); | |
4471 } | |
4472 } | |
4473 #else // !ASSERT | |
10973
ef57c43512d6
8014431: cleanup warnings indicated by the -Wunused-value compiler option on linux
ccheung
parents:
10278
diff
changeset
|
4474 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op |
0 | 4475 #endif |
4476 | |
4477 | |
4478 //-----------------------------memory_at--------------------------------------- | |
4479 Node* MergeMemNode::memory_at(uint alias_idx) const { | |
4480 assert(alias_idx >= Compile::AliasIdxRaw || | |
4481 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, | |
4482 "must avoid base_memory and AliasIdxTop"); | |
4483 | |
4484 // Otherwise, it is a narrow slice. | |
4485 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); | |
4486 Compile *C = Compile::current(); | |
4487 if (is_empty_memory(n)) { | |
4488 // the array is sparse; empty slots are the "top" node | |
4489 n = base_memory(); | |
4490 assert(Node::in_dump() | |
4491 || n == NULL || n->bottom_type() == Type::TOP | |
2317
5a41a201d08c
6812217: Base memory of MergeMem node violates assert during killing expanded AllocateArray node
kvn
parents:
2004
diff
changeset
|
4492 || n->adr_type() == NULL // address is TOP |
0 | 4493 || n->adr_type() == TypePtr::BOTTOM |
4494 || n->adr_type() == TypeRawPtr::BOTTOM | |
4495 || Compile::current()->AliasLevel() == 0, | |
4496 "must be a wide memory"); | |
4497 // AliasLevel == 0 if we are organizing the memory states manually. | |
4498 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. | |
4499 } else { | |
4500 // make sure the stored slice is sane | |
4501 #ifdef ASSERT | |
4502 if (is_error_reported() || Node::in_dump()) { | |
4503 } else if (might_be_same(n, base_memory())) { | |
4504 // Give it a pass: It is a mostly harmless repetition of the base. | |
4505 // This can arise normally from node subsumption during optimization. | |
4506 } else { | |
4507 verify_memory_slice(this, alias_idx, n); | |
4508 } | |
4509 #endif | |
4510 } | |
4511 return n; | |
4512 } | |
4513 | |
4514 //---------------------------set_memory_at------------------------------------- | |
4515 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { | |
4516 verify_memory_slice(this, alias_idx, n); | |
4517 Node* empty_mem = empty_memory(); | |
4518 if (n == base_memory()) n = empty_mem; // collapse default | |
4519 uint need_req = alias_idx+1; | |
4520 if (req() < need_req) { | |
4521 if (n == empty_mem) return; // already the default, so do not grow me | |
4522 // grow the sparse array | |
4523 do { | |
4524 add_req(empty_mem); | |
4525 } while (req() < need_req); | |
4526 } | |
4527 set_req( alias_idx, n ); | |
4528 } | |
4529 | |
4530 | |
4531 | |
4532 //--------------------------iteration_setup------------------------------------ | |
4533 void MergeMemNode::iteration_setup(const MergeMemNode* other) { | |
4534 if (other != NULL) { | |
4535 grow_to_match(other); | |
4536 // invariant: the finite support of mm2 is within mm->req() | |
4537 #ifdef ASSERT | |
4538 for (uint i = req(); i < other->req(); i++) { | |
4539 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); | |
4540 } | |
4541 #endif | |
4542 } | |
4543 // Replace spurious copies of base_memory by top. | |
4544 Node* base_mem = base_memory(); | |
4545 if (base_mem != NULL && !base_mem->is_top()) { | |
4546 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { | |
4547 if (in(i) == base_mem) | |
4548 set_req(i, empty_memory()); | |
4549 } | |
4550 } | |
4551 } | |
4552 | |
4553 //---------------------------grow_to_match------------------------------------- | |
4554 void MergeMemNode::grow_to_match(const MergeMemNode* other) { | |
4555 Node* empty_mem = empty_memory(); | |
4556 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); | |
4557 // look for the finite support of the other memory | |
4558 for (uint i = other->req(); --i >= req(); ) { | |
4559 if (other->in(i) != empty_mem) { | |
4560 uint new_len = i+1; | |
4561 while (req() < new_len) add_req(empty_mem); | |
4562 break; | |
4563 } | |
4564 } | |
4565 } | |
4566 | |
4567 //---------------------------verify_sparse------------------------------------- | |
4568 #ifndef PRODUCT | |
4569 bool MergeMemNode::verify_sparse() const { | |
4570 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); | |
4571 Node* base_mem = base_memory(); | |
4572 // The following can happen in degenerate cases, since empty==top. | |
4573 if (is_empty_memory(base_mem)) return true; | |
4574 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4575 assert(in(i) != NULL, "sane slice"); | |
4576 if (in(i) == base_mem) return false; // should have been the sentinel value! | |
4577 } | |
4578 return true; | |
4579 } | |
4580 | |
4581 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { | |
4582 Node* n; | |
4583 n = mm->in(idx); | |
4584 if (mem == n) return true; // might be empty_memory() | |
4585 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); | |
4586 if (mem == n) return true; | |
4587 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { | |
4588 if (mem == n) return true; | |
4589 if (n == NULL) break; | |
4590 } | |
4591 return false; | |
4592 } | |
4593 #endif // !PRODUCT |