Mercurial > hg > graal-compiler
annotate src/share/vm/opto/memnode.cpp @ 6231:7e5976e66c62
7182543: NMT ON: Aggregate a few NMT related bugs
Summary: 1) Fixed MemTrackWorker::generations_in_used() calculation 2) Ensured NMT not to leak memory recorders after shutdown 3) Used ThreadCritical to block safepoint safe threads
Reviewed-by: acorn, coleenp, dholmes, kvn
author | zgu |
---|---|
date | Thu, 19 Jul 2012 09:05:42 -0400 |
parents | 8c92982cbbc4 |
children | da91efe96a93 |
rev | line source |
---|---|
0 | 1 /* |
6179
8c92982cbbc4
7119644: Increase superword's vector size up to 256 bits
kvn
parents:
6139
diff
changeset
|
2 * Copyright (c) 1997, 2012, 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 | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
107 const TypeOopPtr *tinst = t_adr->isa_oopptr(); |
223 | 108 if (tinst == NULL || !tinst->is_known_instance_field()) |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
109 return mchain; // don't try to optimize non-instance types |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
110 uint instance_id = tinst->instance_id(); |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
111 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
|
112 Node *prev = NULL; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
113 Node *result = mchain; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
114 while (prev != result) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
115 prev = result; |
253
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
116 if (result == start_mem) |
605 | 117 break; // hit one of our sentinels |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
118 // 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
|
119 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
|
120 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
|
121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { |
605 | 122 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
|
123 } else if (proj_in->is_Call()) { |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
124 CallNode *call = proj_in->as_Call(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
125 if (!call->may_modify(t_adr, phase)) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
126 result = call->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
127 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
128 } else if (proj_in->is_Initialize()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
129 AllocateNode* alloc = proj_in->as_Initialize()->allocation(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
130 // 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
|
131 // which contains this memory slice, otherwise skip over it. |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
132 if (alloc != NULL && alloc->_idx != instance_id) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
133 result = proj_in->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
134 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
135 } else if (proj_in->is_MemBar()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
136 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
|
137 } else { |
b0fe4deeb9fb
6726999: nsk/stress/jck12a/jck12a010 assert(n != null,"Bad immediate dominator info.")
kvn
parents:
247
diff
changeset
|
138 assert(false, "unexpected projection"); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
139 } |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
140 } else if (result->is_ClearArray()) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
142 // Can not bypass initialization of the instance |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
143 // we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
144 break; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
145 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
146 // 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
|
147 } else if (result->is_MergeMem()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
149 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
150 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
151 return result; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
152 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
153 |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
155 const TypeOopPtr *t_oop = t_adr->isa_oopptr(); |
223 | 156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
157 PhaseIterGVN *igvn = phase->is_IterGVN(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
158 Node *result = mchain; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
159 result = optimize_simple_memory_chain(result, t_adr, phase); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
160 if (is_instance && igvn != NULL && result->is_Phi()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
161 PhiNode *mphi = result->as_Phi(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
162 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
|
163 const TypePtr *t = mphi->adr_type(); |
163 | 164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || |
223 | 165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && |
247 | 166 t->is_oopptr()->cast_to_exactness(true) |
167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) | |
168 ->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
|
169 // clone the Phi with our address type |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
170 result = mphi->split_out_instance(t_adr, igvn); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
171 } else { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
172 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
|
173 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
174 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
175 return result; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
176 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
177 |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
178 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
|
179 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
|
180 Node *mem = mmem; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
181 #ifdef ASSERT |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
182 { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
183 // 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
|
184 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
|
185 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
|
186 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
|
187 // 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
|
188 if( !consistent && adr_check != NULL && !adr_check->empty() && |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
189 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
|
190 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
|
191 ( 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
|
192 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
|
193 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
|
194 // 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
|
195 consistent = true; |
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 if( !consistent ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
198 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
|
199 if( adr_check == NULL ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
200 st->print("NULL"); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
201 } else { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
202 adr_check->dump(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
203 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
204 st->cr(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
205 print_alias_types(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
206 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
|
207 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
208 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
209 #endif |
1815 | 210 // 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
|
211 // 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
|
212 // alias stuff with it any time soon. |
1815 | 213 const TypeOopPtr *toop = tp->isa_oopptr(); |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
214 if( tp->base() != Type::AnyPtr && |
1815 | 215 !(toop && |
216 toop->klass() != NULL && | |
217 toop->klass()->is_java_lang_Object() && | |
218 toop->offset() == Type::OffsetBot) ) { | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
219 // 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
|
220 // 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
|
221 // Equivalent code in PhiNode::Ideal |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
222 Node* m = phase->transform(mmem); |
605 | 223 // 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
|
224 // 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
|
225 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
|
226 // 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
|
227 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
228 return mem; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
229 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
230 |
0 | 231 //--------------------------Ideal_common--------------------------------------- |
232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs. | |
233 // Unhook non-raw memories from complete (macro-expanded) initializations. | |
234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { | |
235 // If our control input is a dead region, kill all below the region | |
236 Node *ctl = in(MemNode::Control); | |
237 if (ctl && remove_dead_region(phase, can_reshape)) | |
238 return this; | |
305 | 239 ctl = in(MemNode::Control); |
240 // Don't bother trying to transform a dead node | |
241 if( ctl && ctl->is_top() ) return NodeSentinel; | |
0 | 242 |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
243 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
|
244 // 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
|
245 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
|
246 Node* bol = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
247 Node* cmp = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
248 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
|
249 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
|
250 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
|
251 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
|
252 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
|
253 } |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
254 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
|
255 (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
|
256 (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
|
257 // 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
|
258 // 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
|
259 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
|
260 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
|
261 } |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
262 } |
0 | 263 // Ignore if memory is dead, or self-loop |
264 Node *mem = in(MemNode::Memory); | |
265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL | |
266 assert( mem != this, "dead loop in MemNode::Ideal" ); | |
267 | |
4124
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
268 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
|
269 // This memory slice may be dead. |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
270 // 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
|
271 phase->is_IterGVN()->_worklist.push(this); |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
272 return NodeSentinel; // caller will return NULL |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
273 } |
f745b2be3737
7117282: assert(base == NULL || t_adr->isa_rawptr() || !phase->type(base)
kvn
parents:
4115
diff
changeset
|
274 |
0 | 275 Node *address = in(MemNode::Address); |
276 const Type *t_adr = phase->type( address ); | |
277 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL | |
278 | |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
279 if( can_reshape && igvn != NULL && |
1826
56601ef83436
6916062: assert(_inserts <= _insert_limit,"hash table overflow") in NodeHash::hash_insert
kvn
parents:
1815
diff
changeset
|
280 (igvn->_worklist.member(address) || |
56601ef83436
6916062: assert(_inserts <= _insert_limit,"hash table overflow") in NodeHash::hash_insert
kvn
parents:
1815
diff
changeset
|
281 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) { |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
282 // 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
|
283 // 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
|
284 phase->is_IterGVN()->_worklist.push(this); |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
285 return NodeSentinel; // caller will return NULL |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
286 } |
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
366
diff
changeset
|
287 |
1062
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
288 // 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
|
289 // 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
|
290 // Note: C++ will not remove it since the call has side effect. |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
291 if ( t_adr->isa_oopptr() ) { |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
292 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
|
293 } |
dcdcc8c16e20
6896352: CTW fails hotspot/src/share/vm/opto/escape.cpp:1155
kvn
parents:
986
diff
changeset
|
294 |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
295 #ifdef ASSERT |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
296 Node* base = NULL; |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
297 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
|
298 base = address->in(AddPNode::Base); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
299 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
|
300 !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
|
301 #endif |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
302 |
0 | 303 // Avoid independent memory operations |
304 Node* old_mem = mem; | |
305 | |
36 | 306 // The code which unhooks non-raw memories from complete (macro-expanded) |
307 // initializations was removed. After macro-expansion all stores catched | |
308 // by Initialize node became raw stores and there is no information | |
309 // which memory slices they modify. So it is unsafe to move any memory | |
310 // operation above these stores. Also in most cases hooked non-raw memories | |
311 // were already unhooked by using information from detect_ptr_independence() | |
312 // and find_previous_store(). | |
0 | 313 |
314 if (mem->is_MergeMem()) { | |
315 MergeMemNode* mmem = mem->as_MergeMem(); | |
316 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
|
317 |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); |
0 | 319 } |
320 | |
321 if (mem != old_mem) { | |
322 set_req(MemNode::Memory, mem); | |
305 | 323 if (phase->type( mem ) == Type::TOP) return NodeSentinel; |
0 | 324 return this; |
325 } | |
326 | |
327 // let the subclass continue analyzing... | |
328 return NULL; | |
329 } | |
330 | |
331 // Helper function for proving some simple control dominations. | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
332 // Attempt to prove that all control inputs of 'dom' dominate 'sub'. |
0 | 333 // Already assumes that 'dom' is available at 'sub', and that 'sub' |
334 // is not a constant (dominated by the method's StartNode). | |
335 // Used by MemNode::find_previous_store to prove that the | |
336 // control input of a memory operation predates (dominates) | |
337 // an allocation it wants to look past. | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
338 bool MemNode::all_controls_dominate(Node* dom, Node* sub) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
339 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
|
340 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
341 |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
342 // Check 'dom'. Skip Proj and CatchProj nodes. |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
343 dom = dom->find_exact_control(dom); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
344 if (dom == NULL || dom->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
345 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
346 |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
347 if (dom == sub) { |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
348 // 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
|
349 // 'dom' is Proj node of the 'sub'. |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
350 return false; |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
351 } |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
352 |
155
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
353 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
|
354 return true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
355 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
356 // '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
|
357 // 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
|
358 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
359 // Currently 'sub' is either Allocate, Initialize or Start nodes. |
163 | 360 // Or Region for the check in LoadNode::Ideal(); |
361 // 'sub' should have sub->in(0) != NULL. | |
362 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || | |
363 sub->is_Region(), "expecting only these nodes"); | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
364 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
365 // Get control edge of 'sub'. |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
366 Node* orig_sub = sub; |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
367 sub = sub->find_exact_control(sub->in(0)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
368 if (sub == NULL || sub->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
369 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
370 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
371 assert(sub->is_CFG(), "expecting control"); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
372 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
373 if (sub == dom) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
374 return true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
375 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
376 if (sub->is_Start() || sub->is_Root()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
377 return false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
378 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
379 { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
380 // Check all control edges of 'dom'. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
381 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
382 ResourceMark rm; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
383 Arena* arena = Thread::current()->resource_area(); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
384 Node_List nlist(arena); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
385 Unique_Node_List dom_list(arena); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
386 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
387 dom_list.push(dom); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
388 bool only_dominating_controls = false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
389 |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
390 for (uint next = 0; next < dom_list.size(); next++) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
391 Node* n = dom_list.at(next); |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
392 if (n == orig_sub) |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
393 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
|
394 if (!n->is_CFG() && n->pinned()) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
395 // 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
|
396 n = n->find_exact_control(n->in(0)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
397 if (n == NULL || n->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
398 return false; // Conservative answer for dead code |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
399 assert(n->is_CFG(), "expecting control"); |
193
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
400 dom_list.push(n); |
44a553b2809d
6714406: Node::dominates() does not always check for TOP
kvn
parents:
169
diff
changeset
|
401 } 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
|
402 only_dominating_controls = true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
403 } else if (n->is_CFG()) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
404 if (n->dominates(sub, nlist)) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
405 only_dominating_controls = true; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
406 else |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
407 return false; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
408 } else { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
409 // First, own control edge. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
410 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
|
411 if (m != NULL) { |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
412 if (m->is_top()) |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
413 return false; // Conservative answer for dead code |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
414 dom_list.push(m); |
723be81c1212
6701887: JDK7 server VM in endless loop in Node::dominates
kvn
parents:
128
diff
changeset
|
415 } |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
416 // Now, the rest of edges. |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
417 uint cnt = n->req(); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
418 for (uint i = 1; i < cnt; i++) { |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
419 m = n->find_exact_control(n->in(i)); |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
420 if (m == NULL || m->is_top()) |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
421 continue; |
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
422 dom_list.push(m); |
0 | 423 } |
424 } | |
425 } | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
426 return only_dominating_controls; |
0 | 427 } |
428 } | |
429 | |
430 //---------------------detect_ptr_independence--------------------------------- | |
431 // Used by MemNode::find_previous_store to prove that two base | |
432 // pointers are never equal. | |
433 // The pointers are accompanied by their associated allocations, | |
434 // if any, which have been previously discovered by the caller. | |
435 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, | |
436 Node* p2, AllocateNode* a2, | |
437 PhaseTransform* phase) { | |
438 // Attempt to prove that these two pointers cannot be aliased. | |
439 // They may both manifestly be allocations, and they should differ. | |
440 // Or, if they are not both allocations, they can be distinct constants. | |
441 // Otherwise, one is an allocation and the other a pre-existing value. | |
442 if (a1 == NULL && a2 == NULL) { // neither an allocation | |
443 return (p1 != p2) && p1->is_Con() && p2->is_Con(); | |
444 } else if (a1 != NULL && a2 != NULL) { // both allocations | |
445 return (a1 != a2); | |
446 } else if (a1 != NULL) { // one allocation a1 | |
447 // (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
|
448 return all_controls_dominate(p2, a1); |
0 | 449 } else { //(a2 != NULL) // one allocation a2 |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
450 return all_controls_dominate(p1, a2); |
0 | 451 } |
452 return false; | |
453 } | |
454 | |
455 | |
456 // The logic for reordering loads and stores uses four steps: | |
457 // (a) Walk carefully past stores and initializations which we | |
458 // can prove are independent of this load. | |
459 // (b) Observe that the next memory state makes an exact match | |
460 // with self (load or store), and locate the relevant store. | |
461 // (c) Ensure that, if we were to wire self directly to the store, | |
462 // the optimizer would fold it up somehow. | |
463 // (d) Do the rewiring, and return, depending on some other part of | |
464 // the optimizer to fold up the load. | |
465 // This routine handles steps (a) and (b). Steps (c) and (d) are | |
466 // specific to loads and stores, so they are handled by the callers. | |
467 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) | |
468 // | |
469 Node* MemNode::find_previous_store(PhaseTransform* phase) { | |
470 Node* ctrl = in(MemNode::Control); | |
471 Node* adr = in(MemNode::Address); | |
472 intptr_t offset = 0; | |
473 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
474 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); | |
475 | |
476 if (offset == Type::OffsetBot) | |
477 return NULL; // cannot unalias unless there are precise offsets | |
478 | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
479 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
|
480 |
0 | 481 intptr_t size_in_bytes = memory_size(); |
482 | |
483 Node* mem = in(MemNode::Memory); // start searching here... | |
484 | |
485 int cnt = 50; // Cycle limiter | |
486 for (;;) { // While we can dance past unrelated stores... | |
487 if (--cnt < 0) break; // Caught in cycle or a complicated dance? | |
488 | |
489 if (mem->is_Store()) { | |
490 Node* st_adr = mem->in(MemNode::Address); | |
491 intptr_t st_offset = 0; | |
492 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); | |
493 if (st_base == NULL) | |
494 break; // inscrutable pointer | |
495 if (st_offset != offset && st_offset != Type::OffsetBot) { | |
496 const int MAX_STORE = BytesPerLong; | |
497 if (st_offset >= offset + size_in_bytes || | |
498 st_offset <= offset - MAX_STORE || | |
499 st_offset <= offset - mem->as_Store()->memory_size()) { | |
500 // Success: The offsets are provably independent. | |
501 // (You may ask, why not just test st_offset != offset and be done? | |
502 // The answer is that stores of different sizes can co-exist | |
503 // in the same sequence of RawMem effects. We sometimes initialize | |
504 // a whole 'tile' of array elements with a single jint or jlong.) | |
505 mem = mem->in(MemNode::Memory); | |
506 continue; // (a) advance through independent store memory | |
507 } | |
508 } | |
509 if (st_base != base && | |
510 detect_ptr_independence(base, alloc, | |
511 st_base, | |
512 AllocateNode::Ideal_allocation(st_base, phase), | |
513 phase)) { | |
514 // Success: The bases are provably independent. | |
515 mem = mem->in(MemNode::Memory); | |
516 continue; // (a) advance through independent store memory | |
517 } | |
518 | |
519 // (b) At this point, if the bases or offsets do not agree, we lose, | |
520 // since we have not managed to prove 'this' and 'mem' independent. | |
521 if (st_base == base && st_offset == offset) { | |
522 return mem; // let caller handle steps (c), (d) | |
523 } | |
524 | |
525 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { | |
526 InitializeNode* st_init = mem->in(0)->as_Initialize(); | |
527 AllocateNode* st_alloc = st_init->allocation(); | |
528 if (st_alloc == NULL) | |
529 break; // something degenerated | |
530 bool known_identical = false; | |
531 bool known_independent = false; | |
532 if (alloc == st_alloc) | |
533 known_identical = true; | |
534 else if (alloc != NULL) | |
535 known_independent = true; | |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
536 else if (all_controls_dominate(this, st_alloc)) |
0 | 537 known_independent = true; |
538 | |
539 if (known_independent) { | |
540 // The bases are provably independent: Either they are | |
541 // manifestly distinct allocations, or else the control | |
542 // of this load dominates the store's allocation. | |
543 int alias_idx = phase->C->get_alias_index(adr_type()); | |
544 if (alias_idx == Compile::AliasIdxRaw) { | |
545 mem = st_alloc->in(TypeFunc::Memory); | |
546 } else { | |
547 mem = st_init->memory(alias_idx); | |
548 } | |
549 continue; // (a) advance through independent store memory | |
550 } | |
551 | |
552 // (b) at this point, if we are not looking at a store initializing | |
553 // the same allocation we are loading from, we lose. | |
554 if (known_identical) { | |
555 // From caller, can_see_stored_value will consult find_captured_store. | |
556 return mem; // let caller handle steps (c), (d) | |
557 } | |
558 | |
223 | 559 } 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
|
560 // 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
|
561 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
|
562 CallNode *call = mem->in(0)->as_Call(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
563 if (!call->may_modify(addr_t, phase)) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
564 mem = call->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
565 continue; // (a) advance through independent call memory |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
566 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
567 } 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
|
568 mem = mem->in(0)->in(TypeFunc::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
569 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
|
570 } else if (mem->is_ClearArray()) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
571 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
|
572 // (the call updated 'mem' value) |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
573 continue; // (a) advance through independent allocation memory |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
574 } else { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
575 // Can not bypass initialization of the instance |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
576 // we are looking for. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
577 return mem; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
578 } |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
579 } else if (mem->is_MergeMem()) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
580 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
|
581 mem = mem->as_MergeMem()->memory_at(alias_idx); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
582 continue; // (a) advance through independent MergeMem memory |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
583 } |
0 | 584 } |
585 | |
586 // Unless there is an explicit 'continue', we must bail out here, | |
587 // because 'mem' is an inscrutable memory state (e.g., a call). | |
588 break; | |
589 } | |
590 | |
591 return NULL; // bail out | |
592 } | |
593 | |
594 //----------------------calculate_adr_type------------------------------------- | |
595 // Helper function. Notices when the given type of address hits top or bottom. | |
596 // Also, asserts a cross-check of the type against the expected address type. | |
597 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { | |
598 if (t == Type::TOP) return NULL; // does not touch memory any more? | |
599 #ifdef PRODUCT | |
600 cross_check = NULL; | |
601 #else | |
602 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; | |
603 #endif | |
604 const TypePtr* tp = t->isa_ptr(); | |
605 if (tp == NULL) { | |
606 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); | |
607 return TypePtr::BOTTOM; // touches lots of memory | |
608 } else { | |
609 #ifdef ASSERT | |
610 // %%%% [phh] We don't check the alias index if cross_check is | |
611 // TypeRawPtr::BOTTOM. Needs to be investigated. | |
612 if (cross_check != NULL && | |
613 cross_check != TypePtr::BOTTOM && | |
614 cross_check != TypeRawPtr::BOTTOM) { | |
615 // Recheck the alias index, to see if it has changed (due to a bug). | |
616 Compile* C = Compile::current(); | |
617 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), | |
618 "must stay in the original alias category"); | |
619 // The type of the address must be contained in the adr_type, | |
620 // disregarding "null"-ness. | |
621 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) | |
622 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); | |
623 assert(cross_check->meet(tp_notnull) == cross_check, | |
624 "real address must not escape from expected memory type"); | |
625 } | |
626 #endif | |
627 return tp; | |
628 } | |
629 } | |
630 | |
631 //------------------------adr_phi_is_loop_invariant---------------------------- | |
632 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted | |
633 // loop is loop invariant. Make a quick traversal of Phi and associated | |
634 // CastPP nodes, looking to see if they are a closed group within the loop. | |
635 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { | |
636 // The idea is that the phi-nest must boil down to only CastPP nodes | |
637 // with the same data. This implies that any path into the loop already | |
638 // includes such a CastPP, and so the original cast, whatever its input, | |
639 // must be covered by an equivalent cast, with an earlier control input. | |
640 ResourceMark rm; | |
641 | |
642 // The loop entry input of the phi should be the unique dominating | |
643 // node for every Phi/CastPP in the loop. | |
644 Unique_Node_List closure; | |
645 closure.push(adr_phi->in(LoopNode::EntryControl)); | |
646 | |
647 // Add the phi node and the cast to the worklist. | |
648 Unique_Node_List worklist; | |
649 worklist.push(adr_phi); | |
650 if( cast != NULL ){ | |
651 if( !cast->is_ConstraintCast() ) return false; | |
652 worklist.push(cast); | |
653 } | |
654 | |
655 // Begin recursive walk of phi nodes. | |
656 while( worklist.size() ){ | |
657 // Take a node off the worklist | |
658 Node *n = worklist.pop(); | |
659 if( !closure.member(n) ){ | |
660 // Add it to the closure. | |
661 closure.push(n); | |
662 // Make a sanity check to ensure we don't waste too much time here. | |
663 if( closure.size() > 20) return false; | |
664 // This node is OK if: | |
665 // - it is a cast of an identical value | |
666 // - or it is a phi node (then we add its inputs to the worklist) | |
667 // Otherwise, the node is not OK, and we presume the cast is not invariant | |
668 if( n->is_ConstraintCast() ){ | |
669 worklist.push(n->in(1)); | |
670 } else if( n->is_Phi() ) { | |
671 for( uint i = 1; i < n->req(); i++ ) { | |
672 worklist.push(n->in(i)); | |
673 } | |
674 } else { | |
675 return false; | |
676 } | |
677 } | |
678 } | |
679 | |
680 // Quit when the worklist is empty, and we've found no offending nodes. | |
681 return true; | |
682 } | |
683 | |
684 //------------------------------Ideal_DU_postCCP------------------------------- | |
685 // Find any cast-away of null-ness and keep its control. Null cast-aways are | |
686 // going away in this pass and we need to make this memory op depend on the | |
687 // gating null check. | |
163 | 688 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { |
689 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); | |
690 } | |
0 | 691 |
692 // I tried to leave the CastPP's in. This makes the graph more accurate in | |
693 // some sense; we get to keep around the knowledge that an oop is not-null | |
694 // after some test. Alas, the CastPP's interfere with GVN (some values are | |
695 // the regular oop, some are the CastPP of the oop, all merge at Phi's which | |
696 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed | |
697 // some of the more trivial cases in the optimizer. Removing more useless | |
698 // Phi's started allowing Loads to illegally float above null checks. I gave | |
699 // up on this approach. CNC 10/20/2000 | |
163 | 700 // This static method may be called not from MemNode (EncodePNode calls it). |
701 // Only the control edge of the node 'n' might be updated. | |
702 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { | |
0 | 703 Node *skipped_cast = NULL; |
704 // Need a null check? Regular static accesses do not because they are | |
705 // from constant addresses. Array ops are gated by the range check (which | |
706 // always includes a NULL check). Just check field ops. | |
163 | 707 if( n->in(MemNode::Control) == NULL ) { |
0 | 708 // Scan upwards for the highest location we can place this memory op. |
709 while( true ) { | |
710 switch( adr->Opcode() ) { | |
711 | |
712 case Op_AddP: // No change to NULL-ness, so peek thru AddP's | |
713 adr = adr->in(AddPNode::Base); | |
714 continue; | |
715 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
716 case Op_DecodeN: // No change to NULL-ness, so peek thru |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
717 adr = adr->in(1); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
718 continue; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
719 |
6139
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
720 case Op_EncodeP: |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
721 // 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
|
722 // 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
|
723 // |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
724 // 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
|
725 // 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
|
726 if (adr->in(0) == NULL) { |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
727 // Keep looking for cast nodes. |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
728 adr = adr->in(1); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
729 continue; |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
730 } |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
731 ccp->hash_delete(n); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
732 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
|
733 ccp->hash_insert(n); |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
734 return n; |
0919b2e7895d
7174510: 19 JCK compiler tests fail with C2 error: memNode.cpp:812 - ShouldNotReachHere
kvn
parents:
4821
diff
changeset
|
735 |
0 | 736 case Op_CastPP: |
737 // If the CastPP is useless, just peek on through it. | |
738 if( ccp->type(adr) == ccp->type(adr->in(1)) ) { | |
739 // Remember the cast that we've peeked though. If we peek | |
740 // through more than one, then we end up remembering the highest | |
741 // one, that is, if in a loop, the one closest to the top. | |
742 skipped_cast = adr; | |
743 adr = adr->in(1); | |
744 continue; | |
745 } | |
746 // CastPP is going away in this pass! We need this memory op to be | |
747 // control-dependent on the test that is guarding the CastPP. | |
163 | 748 ccp->hash_delete(n); |
749 n->set_req(MemNode::Control, adr->in(0)); | |
750 ccp->hash_insert(n); | |
751 return n; | |
0 | 752 |
753 case Op_Phi: | |
754 // Attempt to float above a Phi to some dominating point. | |
755 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { | |
756 // If we've already peeked through a Cast (which could have set the | |
757 // control), we can't float above a Phi, because the skipped Cast | |
758 // may not be loop invariant. | |
759 if (adr_phi_is_loop_invariant(adr, skipped_cast)) { | |
760 adr = adr->in(1); | |
761 continue; | |
762 } | |
763 } | |
764 | |
765 // Intentional fallthrough! | |
766 | |
767 // No obvious dominating point. The mem op is pinned below the Phi | |
768 // by the Phi itself. If the Phi goes away (no true value is merged) | |
769 // then the mem op can float, but not indefinitely. It must be pinned | |
770 // behind the controls leading to the Phi. | |
771 case Op_CheckCastPP: | |
772 // These usually stick around to change address type, however a | |
773 // useless one can be elided and we still need to pick up a control edge | |
774 if (adr->in(0) == NULL) { | |
775 // This CheckCastPP node has NO control and is likely useless. But we | |
776 // need check further up the ancestor chain for a control input to keep | |
777 // the node in place. 4959717. | |
778 skipped_cast = adr; | |
779 adr = adr->in(1); | |
780 continue; | |
781 } | |
163 | 782 ccp->hash_delete(n); |
783 n->set_req(MemNode::Control, adr->in(0)); | |
784 ccp->hash_insert(n); | |
785 return n; | |
0 | 786 |
787 // List of "safe" opcodes; those that implicitly block the memory | |
788 // op below any null check. | |
789 case Op_CastX2P: // no null checks on native pointers | |
790 case Op_Parm: // 'this' pointer is not null | |
791 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
|
792 case Op_LoadN: // Loading from within a klass |
0 | 793 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
|
794 case Op_LoadNKlass: // Loading from within a klass |
0 | 795 case Op_ConP: // Loading from a klass |
163 | 796 case Op_ConN: // Loading from a klass |
0 | 797 case Op_CreateEx: // Sucking up the guts of an exception oop |
798 case Op_Con: // Reading from TLS | |
799 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
|
800 case Op_CMoveN: // CMoveN is pinned |
0 | 801 break; // No progress |
802 | |
803 case Op_Proj: // Direct call to an allocation routine | |
804 case Op_SCMemProj: // Memory state from store conditional ops | |
805 #ifdef ASSERT | |
806 { | |
807 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); | |
808 const Node* call = adr->in(0); | |
163 | 809 if (call->is_CallJava()) { |
810 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
|
811 const TypeTuple *r = call_java->tf()->range(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
812 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
|
813 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
|
814 assert(ret_type && ret_type->isa_ptr(), "must return pointer"); |
0 | 815 // We further presume that this is one of |
816 // new_instance_Java, new_array_Java, or | |
817 // the like, but do not assert for this. | |
818 } else if (call->is_Allocate()) { | |
819 // similar case to new_instance_Java, etc. | |
820 } else if (!call->is_CallLeaf()) { | |
821 // Projections from fetch_oop (OSR) are allowed as well. | |
822 ShouldNotReachHere(); | |
823 } | |
824 } | |
825 #endif | |
826 break; | |
827 default: | |
828 ShouldNotReachHere(); | |
829 } | |
830 break; | |
831 } | |
832 } | |
833 | |
834 return NULL; // No progress | |
835 } | |
836 | |
837 | |
838 //============================================================================= | |
839 uint LoadNode::size_of() const { return sizeof(*this); } | |
840 uint LoadNode::cmp( const Node &n ) const | |
841 { return !Type::cmp( _type, ((LoadNode&)n)._type ); } | |
842 const Type *LoadNode::bottom_type() const { return _type; } | |
843 uint LoadNode::ideal_reg() const { | |
844 return Matcher::base2reg[_type->base()]; | |
845 } | |
846 | |
847 #ifndef PRODUCT | |
848 void LoadNode::dump_spec(outputStream *st) const { | |
849 MemNode::dump_spec(st); | |
850 if( !Verbose && !WizardMode ) { | |
851 // standard dump does this in Verbose and WizardMode | |
852 st->print(" #"); _type->dump_on(st); | |
853 } | |
854 } | |
855 #endif | |
856 | |
1609 | 857 #ifdef ASSERT |
858 //----------------------------is_immutable_value------------------------------- | |
859 // Helper function to allow a raw load without control edge for some cases | |
860 bool LoadNode::is_immutable_value(Node* adr) { | |
861 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() && | |
862 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && | |
863 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) == | |
864 in_bytes(JavaThread::osthread_offset()))); | |
865 } | |
866 #endif | |
0 | 867 |
868 //----------------------------LoadNode::make----------------------------------- | |
869 // Polymorphic factory method: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
870 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
871 Compile* C = gvn.C; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
872 |
0 | 873 // sanity check the alias category against the created node type |
874 assert(!(adr_type->isa_oopptr() && | |
875 adr_type->offset() == oopDesc::klass_offset_in_bytes()), | |
876 "use LoadKlassNode instead"); | |
877 assert(!(adr_type->isa_aryptr() && | |
878 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), | |
879 "use LoadRangeNode instead"); | |
1609 | 880 // Check control edge of raw loads |
881 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw || | |
882 // oop will be recorded in oop map if load crosses safepoint | |
883 rt->isa_oopptr() || is_immutable_value(adr), | |
884 "raw memory operations should have control edge"); | |
0 | 885 switch (bt) { |
624 | 886 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() ); |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
887 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
888 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
889 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
890 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
891 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
892 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
893 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); |
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
894 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
895 case T_OBJECT: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
896 #ifdef _LP64 |
163 | 897 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
898 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
899 return new (C, 2) 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
|
900 } else |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
901 #endif |
163 | 902 { |
903 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); | |
904 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); | |
905 } | |
0 | 906 } |
907 ShouldNotReachHere(); | |
908 return (LoadNode*)NULL; | |
909 } | |
910 | |
911 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { | |
912 bool require_atomic = true; | |
913 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); | |
914 } | |
915 | |
916 | |
917 | |
918 | |
919 //------------------------------hash------------------------------------------- | |
920 uint LoadNode::hash() const { | |
921 // unroll addition of interesting fields | |
922 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); | |
923 } | |
924 | |
925 //---------------------------can_see_stored_value------------------------------ | |
926 // This routine exists to make sure this set of tests is done the same | |
927 // everywhere. We need to make a coordinated change: first LoadNode::Ideal | |
928 // will change the graph shape in a way which makes memory alive twice at the | |
929 // same time (uses the Oracle model of aliasing), then some | |
930 // LoadXNode::Identity will fold things back to the equivalence-class model | |
931 // of aliasing. | |
932 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { | |
933 Node* ld_adr = in(MemNode::Address); | |
934 | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
935 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
936 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
937 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
938 atp->field() != NULL && !atp->field()->is_volatile()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
939 uint alias_idx = atp->index(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
940 bool final = atp->field()->is_final(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
941 Node* result = NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
942 Node* current = st; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
943 // 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
|
944 // 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
|
945 // 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
|
946 // 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
|
947 // 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
|
948 // a synchronized region. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
949 while (current->is_Proj()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
950 int opc = current->in(0)->Opcode(); |
3849
f1c12354c3f7
7074017: Introduce MemBarAcquireLock/MemBarReleaseLock nodes for monitor enter/exit code paths
roland
parents:
3284
diff
changeset
|
951 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) || |
f1c12354c3f7
7074017: Introduce MemBarAcquireLock/MemBarReleaseLock nodes for monitor enter/exit code paths
roland
parents:
3284
diff
changeset
|
952 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder || |
f1c12354c3f7
7074017: Introduce MemBarAcquireLock/MemBarReleaseLock nodes for monitor enter/exit code paths
roland
parents:
3284
diff
changeset
|
953 opc == Op_MemBarReleaseLock) { |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
954 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
|
955 if (mem->is_MergeMem()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
956 MergeMemNode* merge = mem->as_MergeMem(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
957 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
|
958 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
|
959 // Keep searching |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
960 current = merge->base_memory(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
961 continue; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
962 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
963 // 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
|
964 // to exit. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
965 result = new_st; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
966 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
967 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
968 break; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
969 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
970 if (result != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
971 st = result; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
972 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
973 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
974 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
975 |
0 | 976 // Loop around twice in the case Load -> Initialize -> Store. |
977 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) | |
978 for (int trip = 0; trip <= 1; trip++) { | |
979 | |
980 if (st->is_Store()) { | |
981 Node* st_adr = st->in(MemNode::Address); | |
982 if (!phase->eqv(st_adr, ld_adr)) { | |
983 // Try harder before giving up... Match raw and non-raw pointers. | |
984 intptr_t st_off = 0; | |
985 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); | |
986 if (alloc == NULL) return NULL; | |
987 intptr_t ld_off = 0; | |
988 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); | |
989 if (alloc != allo2) return NULL; | |
990 if (ld_off != st_off) return NULL; | |
991 // At this point we have proven something like this setup: | |
992 // A = Allocate(...) | |
993 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) | |
994 // S = StoreQ(, AddP(, A.Parm , #Off), V) | |
995 // (Actually, we haven't yet proven the Q's are the same.) | |
996 // In other words, we are loading from a casted version of | |
997 // the same pointer-and-offset that we stored to. | |
998 // Thus, we are able to replace L by V. | |
999 } | |
1000 // Now prove that we have a LoadQ matched to a StoreQ, for some Q. | |
1001 if (store_Opcode() != st->Opcode()) | |
1002 return NULL; | |
1003 return st->in(MemNode::ValueIn); | |
1004 } | |
1005 | |
1006 intptr_t offset = 0; // scratch | |
1007 | |
1008 // A load from a freshly-created object always returns zero. | |
1009 // (This can happen after LoadNode::Ideal resets the load's memory input | |
1010 // to find_captured_store, which returned InitializeNode::zero_memory.) | |
1011 if (st->is_Proj() && st->in(0)->is_Allocate() && | |
1012 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && | |
1013 offset >= st->in(0)->as_Allocate()->minimum_header_size()) { | |
1014 // return a zero value for the load's basic type | |
1015 // (This is one of the few places where a generic PhaseTransform | |
1016 // can create new nodes. Think of it as lazily manifesting | |
1017 // virtually pre-existing constants.) | |
1018 return phase->zerocon(memory_type()); | |
1019 } | |
1020 | |
1021 // A load from an initialization barrier can match a captured store. | |
1022 if (st->is_Proj() && st->in(0)->is_Initialize()) { | |
1023 InitializeNode* init = st->in(0)->as_Initialize(); | |
1024 AllocateNode* alloc = init->allocation(); | |
1025 if (alloc != NULL && | |
1026 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { | |
1027 // examine a captured store value | |
1028 st = init->find_captured_store(offset, memory_size(), phase); | |
1029 if (st != NULL) | |
1030 continue; // take one more trip around | |
1031 } | |
1032 } | |
1033 | |
1034 break; | |
1035 } | |
1036 | |
1037 return NULL; | |
1038 } | |
1039 | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1040 //----------------------is_instance_field_load_with_local_phi------------------ |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1041 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1042 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1043 in(MemNode::Address)->is_AddP() ) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1044 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1045 // Only instances. |
223 | 1046 if( t_oop != NULL && t_oop->is_known_instance_field() && |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1047 t_oop->offset() != Type::OffsetBot && |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1048 t_oop->offset() != Type::OffsetTop) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1049 return true; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1050 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1051 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1052 return false; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1053 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1054 |
0 | 1055 //------------------------------Identity--------------------------------------- |
1056 // Loads are identity if previous store is to same address | |
1057 Node *LoadNode::Identity( PhaseTransform *phase ) { | |
1058 // If the previous store-maker is the right kind of Store, and the store is | |
1059 // to the same address, then we are equal to the value stored. | |
1060 Node* mem = in(MemNode::Memory); | |
1061 Node* value = can_see_stored_value(mem, phase); | |
1062 if( value ) { | |
1063 // byte, short & char stores truncate naturally. | |
1064 // A load has to load the truncated value which requires | |
1065 // some sort of masking operation and that requires an | |
1066 // Ideal call instead of an Identity call. | |
1067 if (memory_size() < BytesPerInt) { | |
1068 // If the input to the store does not fit with the load's result type, | |
1069 // it must be truncated via an Ideal call. | |
1070 if (!phase->type(value)->higher_equal(phase->type(this))) | |
1071 return this; | |
1072 } | |
1073 // (This works even when value is a Con, but LoadNode::Value | |
1074 // usually runs first, producing the singleton type of the Con.) | |
1075 return value; | |
1076 } | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1077 |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1078 // Search for an existing data phi which was generated before for the same |
605 | 1079 // 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
|
1080 Node *region = mem->in(0); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1081 if (is_instance_field_load_with_local_phi(region)) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1082 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1083 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
|
1084 int this_offset = addr_t->offset(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1085 int this_id = addr_t->is_oopptr()->instance_id(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1086 const Type* this_type = bottom_type(); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1087 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
|
1088 Node* phi = region->fast_out(i); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1089 if (phi->is_Phi() && phi != mem && |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1090 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1091 return phi; |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1092 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1093 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1094 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1095 |
0 | 1096 return this; |
1097 } | |
1098 | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1099 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1100 // Returns true if the AliasType refers to the field that holds the |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1101 // cached box array. Currently only handles the IntegerCache case. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1102 static bool is_autobox_cache(Compile::AliasType* atp) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1103 if (atp != NULL && atp->field() != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1104 ciField* field = atp->field(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1105 ciSymbol* klass = field->holder()->name(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1106 if (field->name() == ciSymbol::cache_field_name() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1107 field->holder()->uses_default_loader() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1108 klass == ciSymbol::java_lang_Integer_IntegerCache()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1109 return true; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1110 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1111 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1112 return false; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1113 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1114 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1115 // Fetch the base value in the autobox array |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1116 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1117 if (atp != NULL && atp->field() != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1118 ciField* field = atp->field(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1119 ciSymbol* klass = field->holder()->name(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1120 if (field->name() == ciSymbol::cache_field_name() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1121 field->holder()->uses_default_loader() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1122 klass == ciSymbol::java_lang_Integer_IntegerCache()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1123 assert(field->is_constant(), "what?"); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1124 ciObjArray* array = field->constant_value().as_object()->as_obj_array(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1125 // Fetch the box object at the base of the array and get its value |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1126 ciInstance* box = array->obj_at(0)->as_instance(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1127 ciInstanceKlass* ik = box->klass()->as_instance_klass(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1128 if (ik->nof_nonstatic_fields() == 1) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1129 // This should be true nonstatic_field_at requires calling |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1130 // nof_nonstatic_fields so check it anyway |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1131 ciConstant c = box->field_value(ik->nonstatic_field_at(0)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1132 cache_offset = c.as_int(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1133 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1134 return true; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1135 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1136 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1137 return false; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1138 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1139 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1140 // Returns true if the AliasType refers to the value field of an |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1141 // autobox object. Currently only handles Integer. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1142 static bool is_autobox_object(Compile::AliasType* atp) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1143 if (atp != NULL && atp->field() != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1144 ciField* field = atp->field(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1145 ciSymbol* klass = field->holder()->name(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1146 if (field->name() == ciSymbol::value_name() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1147 field->holder()->uses_default_loader() && |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1148 klass == ciSymbol::java_lang_Integer()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1149 return true; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1150 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1151 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1152 return false; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1153 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1154 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1155 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1156 // 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
|
1157 // 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
|
1158 // 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
|
1159 // 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
|
1160 // argument to the valueOf call. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1161 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1162 Node* base = in(Address)->in(AddPNode::Base); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1163 if (base->is_Phi() && base->req() == 3) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1164 AllocateNode* allocation = NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1165 int allocation_index = -1; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1166 int load_index = -1; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1167 for (uint i = 1; i < base->req(); i++) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1168 allocation = AllocateNode::Ideal_allocation(base->in(i), phase); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1169 if (allocation != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1170 allocation_index = i; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1171 load_index = 3 - allocation_index; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1172 break; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1173 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1174 } |
583
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1175 bool has_load = ( allocation != NULL && |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1176 (base->in(load_index)->is_Load() || |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1177 base->in(load_index)->is_DecodeN() && |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1178 base->in(load_index)->in(1)->is_Load()) ); |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1179 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1180 // Push the loads from the phi that comes from valueOf up |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1181 // through it to allow elimination of the loads and the recovery |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1182 // of the original value. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1183 Node* mem_phi = in(Memory); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1184 Node* offset = in(Address)->in(AddPNode::Offset); |
553 | 1185 Node* region = base->in(0); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1186 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1187 Node* in1 = clone(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1188 Node* in1_addr = in1->in(Address)->clone(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1189 in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1190 in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1191 in1_addr->set_req(AddPNode::Offset, offset); |
553 | 1192 in1->set_req(0, region->in(allocation_index)); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1193 in1->set_req(Address, in1_addr); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1194 in1->set_req(Memory, mem_phi->in(allocation_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1195 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1196 Node* in2 = clone(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1197 Node* in2_addr = in2->in(Address)->clone(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1198 in2_addr->set_req(AddPNode::Base, base->in(load_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1199 in2_addr->set_req(AddPNode::Address, base->in(load_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1200 in2_addr->set_req(AddPNode::Offset, offset); |
553 | 1201 in2->set_req(0, region->in(load_index)); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1202 in2->set_req(Address, in2_addr); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1203 in2->set_req(Memory, mem_phi->in(load_index)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1204 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1205 in1_addr = phase->transform(in1_addr); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1206 in1 = phase->transform(in1); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1207 in2_addr = phase->transform(in2_addr); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1208 in2 = phase->transform(in2); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1209 |
553 | 1210 PhiNode* result = PhiNode::make_blank(region, this); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1211 result->set_req(allocation_index, in1); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1212 result->set_req(load_index, in2); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1213 return result; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1214 } |
583
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1215 } else if (base->is_Load() || |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1216 base->is_DecodeN() && base->in(1)->is_Load()) { |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1217 if (base->is_DecodeN()) { |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1218 // Get LoadN node which loads cached Integer object |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1219 base = base->in(1); |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1220 } |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1221 // Eliminate the load of Integer.value for integers from the cache |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1222 // 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
|
1223 // Capture the offset of the load and then reverse the computation. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1224 Node* load_base = base->in(Address)->in(AddPNode::Base); |
583
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1225 if (load_base->is_DecodeN()) { |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1226 // Get LoadN node which loads IntegerCache.cache field |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1227 load_base = load_base->in(1); |
5d75ab5f6698
6807084: AutoBox elimination is broken with compressed oops
kvn
parents:
570
diff
changeset
|
1228 } |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1229 if (load_base != NULL) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1230 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1231 intptr_t cache_offset; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1232 int shift = -1; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1233 Node* cache = NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1234 if (is_autobox_cache(atp)) { |
29
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
1235 shift = exact_log2(type2aelembytes(T_OBJECT)); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1236 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1237 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1238 if (cache != NULL && base->in(Address)->is_AddP()) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1239 Node* elements[4]; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1240 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1241 int cache_low; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1242 if (count > 0 && fetch_autobox_base(atp, cache_low)) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1243 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1244 // Add up all the offsets making of the address of the load |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1245 Node* result = elements[0]; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1246 for (int i = 1; i < count; i++) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1247 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1248 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1249 // Remove the constant offset from the address and then |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1250 // remove the scaling of the offset to recover the original index. |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1251 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1252 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1253 // Peel the shift off directly but wrap it in a dummy node |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1254 // since Ideal can't return existing nodes |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1255 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1256 } else { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1257 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1258 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1259 #ifdef _LP64 |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1260 result = new (phase->C, 2) ConvL2INode(phase->transform(result)); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1261 #endif |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1262 return result; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1263 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1264 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1265 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1266 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1267 return NULL; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1268 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1269 |
163 | 1270 //------------------------------split_through_phi------------------------------ |
1271 // Split instance field load through Phi. | |
1272 Node *LoadNode::split_through_phi(PhaseGVN *phase) { | |
1273 Node* mem = in(MemNode::Memory); | |
1274 Node* address = in(MemNode::Address); | |
1275 const TypePtr *addr_t = phase->type(address)->isa_ptr(); | |
1276 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); | |
1277 | |
1278 assert(mem->is_Phi() && (t_oop != NULL) && | |
223 | 1279 t_oop->is_known_instance_field(), "invalide conditions"); |
163 | 1280 |
1281 Node *region = mem->in(0); | |
1282 if (region == NULL) { | |
1283 return NULL; // Wait stable graph | |
1284 } | |
1285 uint cnt = mem->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) |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1289 return NULL; // Wait stable graph |
163 | 1290 Node *in = mem->in(i); |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1291 if (in == NULL) { |
163 | 1292 return NULL; // Wait stable graph |
1293 } | |
1294 } | |
1295 // Check for loop invariant. | |
1296 if (cnt == 3) { | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1297 for (uint i = 1; i < cnt; i++) { |
163 | 1298 Node *in = mem->in(i); |
1299 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); | |
1300 if (m == mem) { | |
1301 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. | |
1302 return this; | |
1303 } | |
1304 } | |
1305 } | |
1306 // Split through Phi (see original code in loopopts.cpp). | |
1307 assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); | |
1308 | |
1309 // Do nothing here if Identity will find a value | |
1310 // (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
|
1311 if (!phase->eqv(this, this->Identity(phase))) |
163 | 1312 return NULL; |
1313 | |
1314 // Skip the split if the region dominates some control edge of the address. | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1315 if (!MemNode::all_controls_dominate(address, region)) |
163 | 1316 return NULL; |
1317 | |
1318 const Type* this_type = this->bottom_type(); | |
1319 int this_index = phase->C->get_alias_index(addr_t); | |
1320 int this_offset = addr_t->offset(); | |
1321 int this_iid = addr_t->is_oopptr()->instance_id(); | |
1322 PhaseIterGVN *igvn = phase->is_IterGVN(); | |
1323 Node *phi = new (igvn->C, region->req()) 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
|
1324 for (uint i = 1; i < region->req(); i++) { |
163 | 1325 Node *x; |
1326 Node* the_clone = NULL; | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1327 if (region->in(i) == phase->C->top()) { |
163 | 1328 x = phase->C->top(); // Dead path? Use a dead data op |
1329 } else { | |
1330 x = this->clone(); // Else clone up the data op | |
1331 the_clone = x; // Remember for possible deletion. | |
1332 // 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
|
1333 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
|
1334 x->set_req(0, region->in(i)); |
163 | 1335 } else { |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1336 x->set_req(0, NULL); |
163 | 1337 } |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1338 for (uint j = 1; j < this->req(); j++) { |
163 | 1339 Node *in = this->in(j); |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1340 if (in->is_Phi() && in->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
|
1341 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone |
163 | 1342 } |
1343 } | |
1344 // Check for a 'win' on some paths | |
1345 const Type *t = x->Value(igvn); | |
1346 | |
1347 bool singleton = t->singleton(); | |
1348 | |
1349 // 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
|
1350 if (singleton && t == Type::TOP) { |
163 | 1351 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
1352 } | |
1353 | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1354 if (singleton) { |
163 | 1355 x = igvn->makecon(t); |
1356 } else { | |
1357 // We now call Identity to try to simplify the cloned node. | |
1358 // Note that some Identity methods call phase->type(this). | |
1359 // Make sure that the type array is big enough for | |
1360 // our new node, even though we may throw the node away. | |
1361 // (This tweaking with igvn only works because x is a new node.) | |
1362 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
|
1363 // If x is a TypeNode, capture any more-precise type permanently into Node |
605 | 1364 // 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
|
1365 // 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
|
1366 x->raise_bottom_type(t); |
163 | 1367 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
|
1368 if (y != x) { |
163 | 1369 x = y; |
1370 } else { | |
1371 y = igvn->hash_find(x); | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1372 if (y) { |
163 | 1373 x = y; |
1374 } else { | |
1375 // Else x is a new node we are keeping | |
1376 // We do not need register_new_node_with_optimizer | |
1377 // because set_type has already been called. | |
1378 igvn->_worklist.push(x); | |
1379 } | |
1380 } | |
1381 } | |
1382 if (x != the_clone && the_clone != NULL) | |
1383 igvn->remove_dead_node(the_clone); | |
1384 phi->set_req(i, x); | |
1385 } | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1386 // Record Phi |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1387 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
|
1388 return phi; |
163 | 1389 } |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1390 |
0 | 1391 //------------------------------Ideal------------------------------------------ |
1392 // If the load is from Field memory and the pointer is non-null, we can | |
1393 // zero out the control input. | |
1394 // If the offset is constant and the base is an object allocation, | |
1395 // try to hook me up to the exact initializing store. | |
1396 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1397 Node* p = MemNode::Ideal_common(phase, can_reshape); | |
1398 if (p) return (p == NodeSentinel) ? NULL : p; | |
1399 | |
1400 Node* ctrl = in(MemNode::Control); | |
1401 Node* address = in(MemNode::Address); | |
1402 | |
1403 // Skip up past a SafePoint control. Cannot do this for Stores because | |
1404 // pointer stores & cardmarks must stay on the same side of a SafePoint. | |
1405 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && | |
1406 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { | |
1407 ctrl = ctrl->in(0); | |
1408 set_req(MemNode::Control,ctrl); | |
1409 } | |
1410 | |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1411 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
|
1412 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
|
1413 if (base != NULL |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1414 && 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
|
1415 // 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
|
1416 if (in(MemNode::Control) != NULL |
0 | 1417 && phase->type(base)->higher_equal(TypePtr::NOTNULL) |
119
d1a5218d7eaf
6686791: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
113
diff
changeset
|
1418 && all_controls_dominate(base, phase->C->start())) { |
0 | 1419 // A method-invariant, non-null address (constant or 'this' argument). |
1420 set_req(MemNode::Control, NULL); | |
1421 } | |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1422 |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1423 if (EliminateAutoBox && can_reshape) { |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1424 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null"); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1425 Compile::AliasType* atp = phase->C->alias_type(adr_type()); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1426 if (is_autobox_object(atp)) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1427 Node* result = eliminate_autobox(phase); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1428 if (result != NULL) return result; |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1429 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1430 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1431 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1432 |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1433 Node* mem = in(MemNode::Memory); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1434 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
|
1435 |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1436 if (addr_t != NULL) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1437 // try to optimize our memory input |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1438 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1439 if (opt_mem != mem) { |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1440 set_req(MemNode::Memory, opt_mem); |
305 | 1441 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
|
1442 return this; |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1443 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1444 const TypeOopPtr *t_oop = addr_t->isa_oopptr(); |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1445 if (can_reshape && opt_mem->is_Phi() && |
223 | 1446 (t_oop != NULL) && t_oop->is_known_instance_field()) { |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1447 PhaseIterGVN *igvn = phase->is_IterGVN(); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1448 if (igvn != NULL && igvn->_worklist.member(opt_mem)) { |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1449 // Delay this transformation until memory Phi is processed. |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1450 phase->is_IterGVN()->_worklist.push(this); |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1451 return NULL; |
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3961
diff
changeset
|
1452 } |
163 | 1453 // Split instance field load through Phi. |
1454 Node* result = split_through_phi(phase); | |
1455 if (result != NULL) return result; | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1456 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1457 } |
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
68
diff
changeset
|
1458 |
0 | 1459 // Check for prior store with a different base or offset; make Load |
1460 // independent. Skip through any number of them. Bail out if the stores | |
1461 // are in an endless dead cycle and report no progress. This is a key | |
1462 // transform for Reflection. However, if after skipping through the Stores | |
1463 // we can't then fold up against a prior store do NOT do the transform as | |
1464 // this amounts to using the 'Oracle' model of aliasing. It leaves the same | |
1465 // array memory alive twice: once for the hoisted Load and again after the | |
1466 // bypassed Store. This situation only works if EVERYBODY who does | |
1467 // anti-dependence work knows how to bypass. I.e. we need all | |
1468 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is | |
1469 // the alias index stuff. So instead, peek through Stores and IFF we can | |
1470 // fold up, do so. | |
1471 Node* prev_mem = find_previous_store(phase); | |
1472 // Steps (a), (b): Walk past independent stores to find an exact match. | |
1473 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { | |
1474 // (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
|
1475 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or |
0 | 1476 // just return a prior value, which is done by Identity calls. |
1477 if (can_see_stored_value(prev_mem, phase)) { | |
1478 // Make ready for step (d): | |
1479 set_req(MemNode::Memory, prev_mem); | |
1480 return this; | |
1481 } | |
1482 } | |
1483 | |
1484 return NULL; // No further progress | |
1485 } | |
1486 | |
1487 // Helper to recognize certain Klass fields which are invariant across | |
1488 // some group of array types (e.g., int[] or all T[] where T < Object). | |
1489 const Type* | |
1490 LoadNode::load_array_final_field(const TypeKlassPtr *tkls, | |
1491 ciKlass* klass) const { | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1492 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) { |
0 | 1493 // The field is Klass::_modifier_flags. Return its (constant) value. |
1494 // (Folds up the 2nd indirection in aClassConstant.getModifiers().) | |
1495 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); | |
1496 return TypeInt::make(klass->modifier_flags()); | |
1497 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1498 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) { |
0 | 1499 // The field is Klass::_access_flags. Return its (constant) value. |
1500 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) | |
1501 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); | |
1502 return TypeInt::make(klass->access_flags()); | |
1503 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1504 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) { |
0 | 1505 // The field is Klass::_layout_helper. Return its constant value if known. |
1506 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); | |
1507 return TypeInt::make(klass->layout_helper()); | |
1508 } | |
1509 | |
1510 // No match. | |
1511 return NULL; | |
1512 } | |
1513 | |
1514 //------------------------------Value----------------------------------------- | |
1515 const Type *LoadNode::Value( PhaseTransform *phase ) const { | |
1516 // Either input is TOP ==> the result is TOP | |
1517 Node* mem = in(MemNode::Memory); | |
1518 const Type *t1 = phase->type(mem); | |
1519 if (t1 == Type::TOP) return Type::TOP; | |
1520 Node* adr = in(MemNode::Address); | |
1521 const TypePtr* tp = phase->type(adr)->isa_ptr(); | |
1522 if (tp == NULL || tp->empty()) return Type::TOP; | |
1523 int off = tp->offset(); | |
1524 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
|
1525 Compile* C = phase->C; |
0 | 1526 |
1527 // Try to guess loaded type from pointer type | |
1528 if (tp->base() == Type::AryPtr) { | |
1529 const Type *t = tp->is_aryptr()->elem(); | |
1530 // Don't do this for integer types. There is only potential profit if | |
1531 // the element type t is lower than _type; that is, for int types, if _type is | |
1532 // more restrictive than t. This only happens here if one is short and the other | |
1533 // char (both 16 bits), and in those cases we've made an intentional decision | |
1534 // to use one kind of load over the other. See AndINode::Ideal and 4965907. | |
1535 // Also, do not try to narrow the type for a LoadKlass, regardless of offset. | |
1536 // | |
1537 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) | |
1538 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier | |
1539 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been | |
1540 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, | |
1541 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. | |
1542 // In fact, that could have been the original type of p1, and p1 could have | |
1543 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the | |
1544 // expression (LShiftL quux 3) independently optimized to the constant 8. | |
1545 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
|
1546 && (_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
|
1547 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { |
0 | 1548 // t might actually be lower than _type, if _type is a unique |
1549 // concrete subclass of abstract class t. | |
1550 // Make sure the reference is not into the header, by comparing | |
1551 // the offset against the offset of the start of the array's data. | |
1552 // Different array types begin at slightly different offsets (12 vs. 16). | |
1553 // We choose T_BYTE as an example base type that is least restrictive | |
1554 // as to alignment, which will therefore produce the smallest | |
1555 // possible base offset. | |
1556 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); | |
1557 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? | |
1558 const Type* jt = t->join(_type); | |
1559 // In any case, do not allow the join, per se, to empty out the type. | |
1560 if (jt->empty() && !t->empty()) { | |
1561 // This can happen if a interface-typed array narrows to a class type. | |
1562 jt = _type; | |
1563 } | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1564 |
708
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1565 if (EliminateAutoBox && adr->is_AddP()) { |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1566 // 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
|
1567 Node* base = adr->in(AddPNode::Base); |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1568 if (base != NULL && |
f2049ae95c3d
6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr")
kvn
parents:
681
diff
changeset
|
1569 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) { |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1570 Compile::AliasType* atp = C->alias_type(base->adr_type()); |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1571 if (is_autobox_cache(atp)) { |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1572 return jt->join(TypePtr::NOTNULL)->is_ptr(); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1573 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1574 } |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
1575 } |
0 | 1576 return jt; |
1577 } | |
1578 } | |
1579 } else if (tp->base() == Type::InstPtr) { | |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1580 ciEnv* env = C->env(); |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1581 const TypeInstPtr* tinst = tp->is_instptr(); |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1582 ciKlass* klass = tinst->klass(); |
0 | 1583 assert( off != Type::OffsetBot || |
1584 // arrays can be cast to Objects | |
1585 tp->is_oopptr()->klass()->is_java_lang_Object() || | |
1586 // 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
|
1587 C->has_unsafe_access(), |
0 | 1588 "Field accesses must be precise" ); |
1589 // 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
|
1590 if (klass == env->String_klass() && |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1591 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
|
1592 // 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
|
1593 Node* base = adr->in(AddPNode::Base); |
1766
02f0a9b6f654
6969586: OptimizeStringConcat: SIGSEGV in LoadNode::Value()
never
parents:
1609
diff
changeset
|
1594 const TypeOopPtr* t = phase->type(base)->isa_oopptr(); |
02f0a9b6f654
6969586: OptimizeStringConcat: SIGSEGV in LoadNode::Value()
never
parents:
1609
diff
changeset
|
1595 if (t != NULL && t->singleton()) { |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1596 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
|
1597 if (field != NULL && field->is_final()) { |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1598 ciObject* string = t->const_oop(); |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1599 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
|
1600 if (constant.basic_type() == T_INT) { |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1601 return TypeInt::make(constant.as_int()); |
41d4973cf100
6942326: x86 code in string_indexof() could read beyond reserved heap space
kvn
parents:
2317
diff
changeset
|
1602 } 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
|
1603 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
2379
b099aaf51bf8
6962931: move interned strings out of the perm gen
jcoomes
parents:
2320
diff
changeset
|
1604 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
|
1605 } else { |
2379
b099aaf51bf8
6962931: move interned strings out of the perm gen
jcoomes
parents:
2320
diff
changeset
|
1606 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
|
1607 } |
1080
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1608 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1609 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1610 } |
7c57aead6d3e
6892658: C2 should optimize some stringbuilder patterns
never
parents:
1062
diff
changeset
|
1611 } |
3902
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1612 // Optimizations for constant objects |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1613 ciObject* const_oop = tinst->const_oop(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1614 if (const_oop != NULL) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1615 // 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
|
1616 if (const_oop->is_call_site()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1617 ciCallSite* call_site = const_oop->as_call_site(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1618 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
|
1619 if (field != NULL && field->is_call_site_target()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1620 ciMethodHandle* target = call_site->get_target(); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1621 if (target != NULL) { // just in case |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1622 ciConstant constant(T_OBJECT, target); |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1623 const Type* t; |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1624 if (adr->bottom_type()->is_ptr_to_narrowoop()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1625 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
|
1626 } else { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1627 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
|
1628 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1629 // Add a dependence for invalidation of the optimization. |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1630 if (!call_site->is_constant_call_site()) { |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1631 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
|
1632 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1633 return t; |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1634 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1635 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1636 } |
11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
twisti
parents:
3849
diff
changeset
|
1637 } |
0 | 1638 } else if (tp->base() == Type::KlassPtr) { |
1639 assert( off != Type::OffsetBot || | |
1640 // arrays can be cast to Objects | |
1641 tp->is_klassptr()->klass()->is_java_lang_Object() || | |
1642 // also allow array-loading from the primary supertype | |
1643 // array during subtype checks | |
1644 Opcode() == Op_LoadKlass, | |
1645 "Field accesses must be precise" ); | |
1646 // For klass/static loads, we expect the _type to be precise | |
1647 } | |
1648 | |
1649 const TypeKlassPtr *tkls = tp->isa_klassptr(); | |
1650 if (tkls != NULL && !StressReflectiveCode) { | |
1651 ciKlass* klass = tkls->klass(); | |
1652 if (klass->is_loaded() && tkls->klass_is_exact()) { | |
1653 // We are loading a field from a Klass metaobject whose identity | |
1654 // is known at compile time (the type is "exact" or "precise"). | |
1655 // 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
|
1656 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) { |
0 | 1657 // The field is Klass::_super_check_offset. Return its (constant) value. |
1658 // (Folds up type checking code.) | |
1659 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); | |
1660 return TypeInt::make(klass->super_check_offset()); | |
1661 } | |
1662 // Compute index into primary_supers array | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1663 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop); |
0 | 1664 // Check for overflowing; use unsigned compare to handle the negative case. |
1665 if( depth < ciKlass::primary_super_limit() ) { | |
1666 // The field is an element of Klass::_primary_supers. Return its (constant) value. | |
1667 // (Folds up type checking code.) | |
1668 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); | |
1669 ciKlass *ss = klass->super_of_depth(depth); | |
1670 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; | |
1671 } | |
1672 const Type* aift = load_array_final_field(tkls, klass); | |
1673 if (aift != NULL) return aift; | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1674 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) |
0 | 1675 && klass->is_array_klass()) { |
1676 // The field is arrayKlass::_component_mirror. Return its (constant) value. | |
1677 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) | |
1678 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); | |
1679 return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); | |
1680 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1681 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) { |
0 | 1682 // The field is Klass::_java_mirror. Return its (constant) value. |
1683 // (Folds up the 2nd indirection in anObjConstant.getClass().) | |
1684 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); | |
1685 return TypeInstPtr::make(klass->java_mirror()); | |
1686 } | |
1687 } | |
1688 | |
1689 // We can still check if we are loading from the primary_supers array at a | |
1690 // shallow enough depth. Even though the klass is not exact, entries less | |
1691 // than or equal to its super depth are correct. | |
1692 if (klass->is_loaded() ) { | |
1693 ciType *inner = klass->klass(); | |
1694 while( inner->is_obj_array_klass() ) | |
1695 inner = inner->as_obj_array_klass()->base_element_type(); | |
1696 if( inner->is_instance_klass() && | |
1697 !inner->as_instance_klass()->flags().is_interface() ) { | |
1698 // Compute index into primary_supers array | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
1699 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop); |
0 | 1700 // Check for overflowing; use unsigned compare to handle the negative case. |
1701 if( depth < ciKlass::primary_super_limit() && | |
1702 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case | |
1703 // The field is an element of Klass::_primary_supers. Return its (constant) value. | |
1704 // (Folds up type checking code.) | |
1705 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); | |
1706 ciKlass *ss = klass->super_of_depth(depth); | |
1707 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; | |
1708 } | |
1709 } | |
1710 } | |
1711 | |
1712 // If the type is enough to determine that the thing is not an array, | |
1713 // we can give the layout_helper a positive interval type. | |
1714 // 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
|
1715 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) |
0 | 1716 && !klass->is_array_klass() // not directly typed as an array |
1717 && !klass->is_interface() // specifically not Serializable & Cloneable | |
1718 && !klass->is_java_lang_Object() // not the supertype of all T[] | |
1719 ) { | |
1720 // Note: When interfaces are reliable, we can narrow the interface | |
1721 // test to (klass != Serializable && klass != Cloneable). | |
1722 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); | |
1723 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); | |
1724 // The key property of this type is that it folds up tests | |
1725 // for array-ness, since it proves that the layout_helper is positive. | |
1726 // Thus, a generic value like the basic object layout helper works fine. | |
1727 return TypeInt::make(min_size, max_jint, Type::WidenMin); | |
1728 } | |
1729 } | |
1730 | |
1731 // If we are loading from a freshly-allocated object, produce a zero, | |
1732 // if the load is provably beyond the header of the object. | |
1733 // (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
|
1734 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
|
1735 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field(); |
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1736 if (ReduceFieldZeroing || is_instance) { |
0 | 1737 Node* value = can_see_stored_value(mem,phase); |
4815 | 1738 if (value != NULL && value->is_Con()) { |
1739 assert(value->bottom_type()->higher_equal(_type),"sanity"); | |
0 | 1740 return value->bottom_type(); |
4815 | 1741 } |
0 | 1742 } |
1743 | |
3278
66b0e2371912
7026700: regression in 6u24-rev-b23: Crash in C2 compiler in PhaseIdealLoop::build_loop_late_post
kvn
parents:
2417
diff
changeset
|
1744 if (is_instance) { |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1745 // 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
|
1746 // 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
|
1747 // 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
|
1748 Node *mem = in(MemNode::Memory); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1749 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
|
1750 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
|
1751 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
|
1752 } |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
1753 } |
0 | 1754 return _type; |
1755 } | |
1756 | |
1757 //------------------------------match_edge------------------------------------- | |
1758 // Do we Match on this edge index or not? Match only the address. | |
1759 uint LoadNode::match_edge(uint idx) const { | |
1760 return idx == MemNode::Address; | |
1761 } | |
1762 | |
1763 //--------------------------LoadBNode::Ideal-------------------------------------- | |
1764 // | |
1765 // If the previous store is to the same address as this load, | |
1766 // and the value stored was larger than a byte, replace this load | |
1767 // with the value stored truncated to a byte. If no truncation is | |
1768 // needed, the replacement is done in LoadNode::Identity(). | |
1769 // | |
1770 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1771 Node* mem = in(MemNode::Memory); | |
1772 Node* value = can_see_stored_value(mem,phase); | |
1773 if( value && !phase->type(value)->higher_equal( _type ) ) { | |
1774 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); | |
1775 return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); | |
1776 } | |
1777 // Identity call will handle the case where truncation is not needed. | |
1778 return LoadNode::Ideal(phase, can_reshape); | |
1779 } | |
1780 | |
4815 | 1781 const Type* LoadBNode::Value(PhaseTransform *phase) const { |
1782 Node* mem = in(MemNode::Memory); | |
1783 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1784 if (value != NULL && value->is_Con() && |
1785 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1786 // If the input to the store does not fit with the load's result type, |
1787 // it must be truncated. We can't delay until Ideal call since | |
1788 // a singleton Value is needed for split_thru_phi optimization. | |
1789 int con = value->get_int(); | |
1790 return TypeInt::make((con << 24) >> 24); | |
1791 } | |
1792 return LoadNode::Value(phase); | |
1793 } | |
1794 | |
624 | 1795 //--------------------------LoadUBNode::Ideal------------------------------------- |
1796 // | |
1797 // If the previous store is to the same address as this load, | |
1798 // and the value stored was larger than a byte, replace this load | |
1799 // with the value stored truncated to a byte. If no truncation is | |
1800 // needed, the replacement is done in LoadNode::Identity(). | |
1801 // | |
1802 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) { | |
1803 Node* mem = in(MemNode::Memory); | |
1804 Node* value = can_see_stored_value(mem, phase); | |
1805 if (value && !phase->type(value)->higher_equal(_type)) | |
1806 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF)); | |
1807 // Identity call will handle the case where truncation is not needed. | |
1808 return LoadNode::Ideal(phase, can_reshape); | |
1809 } | |
1810 | |
4815 | 1811 const Type* LoadUBNode::Value(PhaseTransform *phase) const { |
1812 Node* mem = in(MemNode::Memory); | |
1813 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1814 if (value != NULL && value->is_Con() && |
1815 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1816 // If the input to the store does not fit with the load's result type, |
1817 // it must be truncated. We can't delay until Ideal call since | |
1818 // a singleton Value is needed for split_thru_phi optimization. | |
1819 int con = value->get_int(); | |
1820 return TypeInt::make(con & 0xFF); | |
1821 } | |
1822 return LoadNode::Value(phase); | |
1823 } | |
1824 | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
1825 //--------------------------LoadUSNode::Ideal------------------------------------- |
0 | 1826 // |
1827 // If the previous store is to the same address as this load, | |
1828 // and the value stored was larger than a char, replace this load | |
1829 // with the value stored truncated to a char. If no truncation is | |
1830 // needed, the replacement is done in LoadNode::Identity(). | |
1831 // | |
558
3b5ac9e7e6ea
6796746: rename LoadC (char) opcode class to LoadUS (unsigned short)
twisti
parents:
553
diff
changeset
|
1832 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
0 | 1833 Node* mem = in(MemNode::Memory); |
1834 Node* value = can_see_stored_value(mem,phase); | |
1835 if( value && !phase->type(value)->higher_equal( _type ) ) | |
1836 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); | |
1837 // Identity call will handle the case where truncation is not needed. | |
1838 return LoadNode::Ideal(phase, can_reshape); | |
1839 } | |
1840 | |
4815 | 1841 const Type* LoadUSNode::Value(PhaseTransform *phase) const { |
1842 Node* mem = in(MemNode::Memory); | |
1843 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1844 if (value != NULL && value->is_Con() && |
1845 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1846 // If the input to the store does not fit with the load's result type, |
1847 // it must be truncated. We can't delay until Ideal call since | |
1848 // a singleton Value is needed for split_thru_phi optimization. | |
1849 int con = value->get_int(); | |
1850 return TypeInt::make(con & 0xFFFF); | |
1851 } | |
1852 return LoadNode::Value(phase); | |
1853 } | |
1854 | |
0 | 1855 //--------------------------LoadSNode::Ideal-------------------------------------- |
1856 // | |
1857 // If the previous store is to the same address as this load, | |
1858 // and the value stored was larger than a short, replace this load | |
1859 // with the value stored truncated to a short. If no truncation is | |
1860 // needed, the replacement is done in LoadNode::Identity(). | |
1861 // | |
1862 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
1863 Node* mem = in(MemNode::Memory); | |
1864 Node* value = can_see_stored_value(mem,phase); | |
1865 if( value && !phase->type(value)->higher_equal( _type ) ) { | |
1866 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); | |
1867 return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); | |
1868 } | |
1869 // Identity call will handle the case where truncation is not needed. | |
1870 return LoadNode::Ideal(phase, can_reshape); | |
1871 } | |
1872 | |
4815 | 1873 const Type* LoadSNode::Value(PhaseTransform *phase) const { |
1874 Node* mem = in(MemNode::Memory); | |
1875 Node* value = can_see_stored_value(mem,phase); | |
4821 | 1876 if (value != NULL && value->is_Con() && |
1877 !value->bottom_type()->higher_equal(_type)) { | |
4815 | 1878 // If the input to the store does not fit with the load's result type, |
1879 // it must be truncated. We can't delay until Ideal call since | |
1880 // a singleton Value is needed for split_thru_phi optimization. | |
1881 int con = value->get_int(); | |
1882 return TypeInt::make((con << 16) >> 16); | |
1883 } | |
1884 return LoadNode::Value(phase); | |
1885 } | |
1886 | |
0 | 1887 //============================================================================= |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1888 //----------------------------LoadKlassNode::make------------------------------ |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1889 // Polymorphic factory method: |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1890 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
|
1891 Compile* C = gvn.C; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1892 Node *ctl = NULL; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1893 // sanity check the alias category against the created node type |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1894 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1895 assert(adr_type != NULL, "expecting TypeOopPtr"); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1896 #ifdef _LP64 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1897 if (adr_type->is_ptr_to_narrowoop()) { |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
1898 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
1899 return new (C, 2) DecodeNNode(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
|
1900 } |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1901 #endif |
168
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1902 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); |
7793bd37a336
6705887: Compressed Oops: generate x64 addressing and implicit null checks with narrow oops
kvn
parents:
164
diff
changeset
|
1903 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1904 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1905 |
0 | 1906 //------------------------------Value------------------------------------------ |
1907 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
|
1908 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
|
1909 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1910 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
1911 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { |
0 | 1912 // Either input is TOP ==> the result is TOP |
1913 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
1914 if (t1 == Type::TOP) return Type::TOP; | |
1915 Node *adr = in(MemNode::Address); | |
1916 const Type *t2 = phase->type( adr ); | |
1917 if (t2 == Type::TOP) return Type::TOP; | |
1918 const TypePtr *tp = t2->is_ptr(); | |
1919 if (TypePtr::above_centerline(tp->ptr()) || | |
1920 tp->ptr() == TypePtr::Null) return Type::TOP; | |
1921 | |
1922 // Return a more precise klass, if possible | |
1923 const TypeInstPtr *tinst = tp->isa_instptr(); | |
1924 if (tinst != NULL) { | |
1925 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); | |
1926 int offset = tinst->offset(); | |
1927 if (ik == phase->C->env()->Class_klass() | |
1928 && (offset == java_lang_Class::klass_offset_in_bytes() || | |
1929 offset == java_lang_Class::array_klass_offset_in_bytes())) { | |
1930 // We are loading a special hidden field from a Class mirror object, | |
1931 // the field which points to the VM's Klass metaobject. | |
1932 ciType* t = tinst->java_mirror_type(); | |
1933 // java_mirror_type returns non-null for compile-time Class constants. | |
1934 if (t != NULL) { | |
1935 // constant oop => constant klass | |
1936 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { | |
1937 return TypeKlassPtr::make(ciArrayKlass::make(t)); | |
1938 } | |
1939 if (!t->is_klass()) { | |
1940 // a primitive Class (e.g., int.class) has NULL for a klass field | |
1941 return TypePtr::NULL_PTR; | |
1942 } | |
1943 // (Folds up the 1st indirection in aClassConstant.getModifiers().) | |
1944 return TypeKlassPtr::make(t->as_klass()); | |
1945 } | |
1946 // non-constant mirror, so we can't tell what's going on | |
1947 } | |
1948 if( !ik->is_loaded() ) | |
1949 return _type; // Bail out if not loaded | |
1950 if (offset == oopDesc::klass_offset_in_bytes()) { | |
1951 if (tinst->klass_is_exact()) { | |
1952 return TypeKlassPtr::make(ik); | |
1953 } | |
1954 // See if we can become precise: no subklasses and no interface | |
1955 // (Note: We need to support verified interfaces.) | |
1956 if (!ik->is_interface() && !ik->has_subklass()) { | |
1957 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
1958 // Add a dependence; if any subclass added we need to recompile | |
1959 if (!ik->is_final()) { | |
1960 // %%% should use stronger assert_unique_concrete_subtype instead | |
1961 phase->C->dependencies()->assert_leaf_type(ik); | |
1962 } | |
1963 // Return precise klass | |
1964 return TypeKlassPtr::make(ik); | |
1965 } | |
1966 | |
1967 // Return root of possible klass | |
1968 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); | |
1969 } | |
1970 } | |
1971 | |
1972 // Check for loading klass from an array | |
1973 const TypeAryPtr *tary = tp->isa_aryptr(); | |
1974 if( tary != NULL ) { | |
1975 ciKlass *tary_klass = tary->klass(); | |
1976 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP | |
1977 && tary->offset() == oopDesc::klass_offset_in_bytes()) { | |
1978 if (tary->klass_is_exact()) { | |
1979 return TypeKlassPtr::make(tary_klass); | |
1980 } | |
1981 ciArrayKlass *ak = tary->klass()->as_array_klass(); | |
1982 // If the klass is an object array, we defer the question to the | |
1983 // array component klass. | |
1984 if( ak->is_obj_array_klass() ) { | |
1985 assert( ak->is_loaded(), "" ); | |
1986 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); | |
1987 if( base_k->is_loaded() && base_k->is_instance_klass() ) { | |
1988 ciInstanceKlass* ik = base_k->as_instance_klass(); | |
1989 // See if we can become precise: no subklasses and no interface | |
1990 if (!ik->is_interface() && !ik->has_subklass()) { | |
1991 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
1992 // Add a dependence; if any subclass added we need to recompile | |
1993 if (!ik->is_final()) { | |
1994 phase->C->dependencies()->assert_leaf_type(ik); | |
1995 } | |
1996 // Return precise array klass | |
1997 return TypeKlassPtr::make(ak); | |
1998 } | |
1999 } | |
2000 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); | |
2001 } else { // Found a type-array? | |
2002 //assert(!UseExactTypes, "this code should be useless with exact types"); | |
2003 assert( ak->is_type_array_klass(), "" ); | |
2004 return TypeKlassPtr::make(ak); // These are always precise | |
2005 } | |
2006 } | |
2007 } | |
2008 | |
2009 // Check for loading klass from an array klass | |
2010 const TypeKlassPtr *tkls = tp->isa_klassptr(); | |
2011 if (tkls != NULL && !StressReflectiveCode) { | |
2012 ciKlass* klass = tkls->klass(); | |
2013 if( !klass->is_loaded() ) | |
2014 return _type; // Bail out if not loaded | |
2015 if( klass->is_obj_array_klass() && | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2016 tkls->offset() == in_bytes(objArrayKlass::element_klass_offset())) { |
0 | 2017 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); |
2018 // // Always returning precise element type is incorrect, | |
2019 // // e.g., element type could be object and array may contain strings | |
2020 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); | |
2021 | |
2022 // The array's TypeKlassPtr was declared 'precise' or 'not precise' | |
2023 // according to the element type's subclassing. | |
2024 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); | |
2025 } | |
2026 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
|
2027 tkls->offset() == in_bytes(Klass::super_offset())) { |
0 | 2028 ciKlass* sup = klass->as_instance_klass()->super(); |
2029 // The field is Klass::_super. Return its (constant) value. | |
2030 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) | |
2031 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; | |
2032 } | |
2033 } | |
2034 | |
2035 // Bailout case | |
2036 return LoadNode::Value(phase); | |
2037 } | |
2038 | |
2039 //------------------------------Identity--------------------------------------- | |
2040 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. | |
2041 // Also feed through the klass in Allocate(...klass...)._klass. | |
2042 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
|
2043 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
|
2044 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2045 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2046 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { |
0 | 2047 Node* x = LoadNode::Identity(phase); |
2048 if (x != this) return x; | |
2049 | |
2050 // Take apart the address into an oop and and offset. | |
2051 // Return 'this' if we cannot. | |
2052 Node* adr = in(MemNode::Address); | |
2053 intptr_t offset = 0; | |
2054 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
2055 if (base == NULL) return this; | |
2056 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); | |
2057 if (toop == NULL) return this; | |
2058 | |
2059 // We can fetch the klass directly through an AllocateNode. | |
2060 // This works even if the klass is not constant (clone or newArray). | |
2061 if (offset == oopDesc::klass_offset_in_bytes()) { | |
2062 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); | |
2063 if (allocated_klass != NULL) { | |
2064 return allocated_klass; | |
2065 } | |
2066 } | |
2067 | |
2068 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. | |
2069 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. | |
2070 // See inline_native_Class_query for occurrences of these patterns. | |
2071 // Java Example: x.getClass().isAssignableFrom(y) | |
2072 // Java Example: Array.newInstance(x.getClass().getComponentType(), n) | |
2073 // | |
2074 // This improves reflective code, often making the Class | |
2075 // mirror go completely dead. (Current exception: Class | |
2076 // mirrors may appear in debug info, but we could clean them out by | |
2077 // introducing a new debug info operator for klassOop.java_mirror). | |
2078 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() | |
2079 && (offset == java_lang_Class::klass_offset_in_bytes() || | |
2080 offset == java_lang_Class::array_klass_offset_in_bytes())) { | |
2081 // We are loading a special hidden field from a Class mirror, | |
2082 // the field which points to its Klass or arrayKlass metaobject. | |
2083 if (base->is_Load()) { | |
2084 Node* adr2 = base->in(MemNode::Address); | |
2085 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); | |
2086 if (tkls != NULL && !tkls->empty() | |
2087 && (tkls->klass()->is_instance_klass() || | |
2088 tkls->klass()->is_array_klass()) | |
2089 && adr2->is_AddP() | |
2090 ) { | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2091 int mirror_field = in_bytes(Klass::java_mirror_offset()); |
0 | 2092 if (offset == java_lang_Class::array_klass_offset_in_bytes()) { |
2093 mirror_field = in_bytes(arrayKlass::component_mirror_offset()); | |
2094 } | |
4762
069ab3f976d3
7118863: Move sizeof(klassOopDesc) into the *Klass::*_offset_in_bytes() functions
stefank
parents:
4124
diff
changeset
|
2095 if (tkls->offset() == mirror_field) { |
0 | 2096 return adr2->in(AddPNode::Base); |
2097 } | |
2098 } | |
2099 } | |
2100 } | |
2101 | |
2102 return this; | |
2103 } | |
2104 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2105 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2106 //------------------------------Value------------------------------------------ |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2107 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
|
2108 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
|
2109 if (t == Type::TOP) |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2110 return t; |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2111 |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2112 return t->make_narrowoop(); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2113 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2114 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2115 //------------------------------Identity--------------------------------------- |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2116 // 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
|
2117 // 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
|
2118 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
|
2119 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
|
2120 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2121 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
|
2122 if( t == Type::TOP ) return x; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2123 if( t->isa_narrowoop()) return x; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2124 |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2125 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2126 } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
2127 |
0 | 2128 //------------------------------Value----------------------------------------- |
2129 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { | |
2130 // Either input is TOP ==> the result is TOP | |
2131 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
2132 if( t1 == Type::TOP ) return Type::TOP; | |
2133 Node *adr = in(MemNode::Address); | |
2134 const Type *t2 = phase->type( adr ); | |
2135 if( t2 == Type::TOP ) return Type::TOP; | |
2136 const TypePtr *tp = t2->is_ptr(); | |
2137 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; | |
2138 const TypeAryPtr *tap = tp->isa_aryptr(); | |
2139 if( !tap ) return _type; | |
2140 return tap->size(); | |
2141 } | |
2142 | |
366
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2143 //-------------------------------Ideal--------------------------------------- |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2144 // 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
|
2145 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
|
2146 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
|
2147 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
|
2148 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2149 // 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
|
2150 // Return 'this' if we cannot. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2151 Node* adr = in(MemNode::Address); |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2152 intptr_t offset = 0; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2153 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
|
2154 if (base == NULL) return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2155 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
|
2156 if (tary == NULL) return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2157 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2158 // 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
|
2159 // 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
|
2160 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
|
2161 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
|
2162 if (alloc != NULL) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2163 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
|
2164 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
|
2165 if (allocated_length != len) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2166 // New CastII improves on this. |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2167 return len; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2168 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2169 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2170 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2171 |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2172 return NULL; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2173 } |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2174 |
0 | 2175 //------------------------------Identity--------------------------------------- |
2176 // Feed through the length in AllocateArray(...length...)._length. | |
2177 Node* LoadRangeNode::Identity( PhaseTransform *phase ) { | |
2178 Node* x = LoadINode::Identity(phase); | |
2179 if (x != this) return x; | |
2180 | |
2181 // Take apart the address into an oop and and offset. | |
2182 // Return 'this' if we cannot. | |
2183 Node* adr = in(MemNode::Address); | |
2184 intptr_t offset = 0; | |
2185 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); | |
2186 if (base == NULL) return this; | |
2187 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); | |
2188 if (tary == NULL) return this; | |
2189 | |
2190 // We can fetch the length directly through an AllocateArrayNode. | |
2191 // This works even if the length is not constant (clone or newArray). | |
2192 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
|
2193 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
|
2194 if (alloc != NULL) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2195 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
|
2196 // 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
|
2197 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
|
2198 if (allocated_length == len) { |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2199 // 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
|
2200 return allocated_length; |
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2201 } |
0 | 2202 } |
2203 } | |
2204 | |
2205 return this; | |
2206 | |
2207 } | |
366
8261ee795323
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
rasbold
parents:
305
diff
changeset
|
2208 |
0 | 2209 //============================================================================= |
2210 //---------------------------StoreNode::make----------------------------------- | |
2211 // Polymorphic factory method: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2212 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2213 Compile* C = gvn.C; |
1609 | 2214 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw || |
2215 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
|
2216 |
0 | 2217 switch (bt) { |
2218 case T_BOOLEAN: | |
2219 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); | |
2220 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); | |
2221 case T_CHAR: | |
2222 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); | |
2223 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); | |
2224 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); | |
2225 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); | |
2226 case T_ADDRESS: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2227 case T_OBJECT: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2228 #ifdef _LP64 |
163 | 2229 if (adr->bottom_type()->is_ptr_to_narrowoop() || |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2230 (UseCompressedOops && val->bottom_type()->isa_klassptr() && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2231 adr->bottom_type()->isa_rawptr())) { |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2232 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2233 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2234 } else |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2235 #endif |
221
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2236 { |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2237 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); |
1e026f8da827
6710487: More than half of JDI Regression tests hang with COOPs in -Xcomp mode
kvn
parents:
193
diff
changeset
|
2238 } |
0 | 2239 } |
2240 ShouldNotReachHere(); | |
2241 return (StoreNode*)NULL; | |
2242 } | |
2243 | |
2244 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { | |
2245 bool require_atomic = true; | |
2246 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); | |
2247 } | |
2248 | |
2249 | |
2250 //--------------------------bottom_type---------------------------------------- | |
2251 const Type *StoreNode::bottom_type() const { | |
2252 return Type::MEMORY; | |
2253 } | |
2254 | |
2255 //------------------------------hash------------------------------------------- | |
2256 uint StoreNode::hash() const { | |
2257 // unroll addition of interesting fields | |
2258 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); | |
2259 | |
2260 // Since they are not commoned, do not hash them: | |
2261 return NO_HASH; | |
2262 } | |
2263 | |
2264 //------------------------------Ideal------------------------------------------ | |
2265 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). | |
2266 // When a store immediately follows a relevant allocation/initialization, | |
2267 // try to capture it into the initialization, or hoist it above. | |
2268 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
2269 Node* p = MemNode::Ideal_common(phase, can_reshape); | |
2270 if (p) return (p == NodeSentinel) ? NULL : p; | |
2271 | |
2272 Node* mem = in(MemNode::Memory); | |
2273 Node* address = in(MemNode::Address); | |
2274 | |
3248
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2275 // 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
|
2276 // unsafe if I have intervening uses... Also disallowed for StoreCM |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2277 // since they must follow each StoreP operation. Redundant StoreCMs |
e6beb62de02d
7032963: StoreCM shouldn't participate in store elimination
never
parents:
2379
diff
changeset
|
2278 // are eliminated just before matching in final_graph_reshape. |
4778 | 2279 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
|
2280 mem->Opcode() != Op_StoreCM) { |
0 | 2281 // Looking at a dead closed cycle of memory? |
2282 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); | |
2283 | |
2284 assert(Opcode() == mem->Opcode() || | |
2285 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, | |
2286 "no mismatched stores, except on raw memory"); | |
2287 | |
2288 if (mem->outcnt() == 1 && // check for intervening uses | |
2289 mem->as_Store()->memory_size() <= this->memory_size()) { | |
2290 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. | |
2291 // For example, 'mem' might be the final state at a conditional return. | |
2292 // Or, 'mem' might be used by some node which is live at the same time | |
2293 // 'this' is live, which might be unschedulable. So, require exactly | |
2294 // ONE user, the 'this' store, until such time as we clone 'mem' for | |
2295 // each of 'mem's uses (thus making the exactly-1-user-rule hold true). | |
2296 if (can_reshape) { // (%%% is this an anachronism?) | |
2297 set_req_X(MemNode::Memory, mem->in(MemNode::Memory), | |
2298 phase->is_IterGVN()); | |
2299 } else { | |
2300 // It's OK to do this in the parser, since DU info is always accurate, | |
2301 // and the parser always refers to nodes via SafePointNode maps. | |
2302 set_req(MemNode::Memory, mem->in(MemNode::Memory)); | |
2303 } | |
2304 return this; | |
2305 } | |
2306 } | |
2307 | |
2308 // Capture an unaliased, unconditional, simple store into an initializer. | |
2309 // Or, if it is independent of the allocation, hoist it above the allocation. | |
2310 if (ReduceFieldZeroing && /*can_reshape &&*/ | |
2311 mem->is_Proj() && mem->in(0)->is_Initialize()) { | |
2312 InitializeNode* init = mem->in(0)->as_Initialize(); | |
2313 intptr_t offset = init->can_capture_store(this, phase); | |
2314 if (offset > 0) { | |
2315 Node* moved = init->capture_store(this, offset, phase); | |
2316 // If the InitializeNode captured me, it made a raw copy of me, | |
2317 // and I need to disappear. | |
2318 if (moved != NULL) { | |
2319 // %%% hack to ensure that Ideal returns a new node: | |
2320 mem = MergeMemNode::make(phase->C, mem); | |
2321 return mem; // fold me away | |
2322 } | |
2323 } | |
2324 } | |
2325 | |
2326 return NULL; // No further progress | |
2327 } | |
2328 | |
2329 //------------------------------Value----------------------------------------- | |
2330 const Type *StoreNode::Value( PhaseTransform *phase ) const { | |
2331 // Either input is TOP ==> the result is TOP | |
2332 const Type *t1 = phase->type( in(MemNode::Memory) ); | |
2333 if( t1 == Type::TOP ) return Type::TOP; | |
2334 const Type *t2 = phase->type( in(MemNode::Address) ); | |
2335 if( t2 == Type::TOP ) return Type::TOP; | |
2336 const Type *t3 = phase->type( in(MemNode::ValueIn) ); | |
2337 if( t3 == Type::TOP ) return Type::TOP; | |
2338 return Type::MEMORY; | |
2339 } | |
2340 | |
2341 //------------------------------Identity--------------------------------------- | |
2342 // Remove redundant stores: | |
2343 // Store(m, p, Load(m, p)) changes to m. | |
2344 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). | |
2345 Node *StoreNode::Identity( PhaseTransform *phase ) { | |
2346 Node* mem = in(MemNode::Memory); | |
2347 Node* adr = in(MemNode::Address); | |
2348 Node* val = in(MemNode::ValueIn); | |
2349 | |
2350 // Load then Store? Then the Store is useless | |
2351 if (val->is_Load() && | |
4778 | 2352 val->in(MemNode::Address)->eqv_uncast(adr) && |
2353 val->in(MemNode::Memory )->eqv_uncast(mem) && | |
0 | 2354 val->as_Load()->store_Opcode() == Opcode()) { |
2355 return mem; | |
2356 } | |
2357 | |
2358 // Two stores in a row of the same value? | |
2359 if (mem->is_Store() && | |
4778 | 2360 mem->in(MemNode::Address)->eqv_uncast(adr) && |
2361 mem->in(MemNode::ValueIn)->eqv_uncast(val) && | |
0 | 2362 mem->Opcode() == Opcode()) { |
2363 return mem; | |
2364 } | |
2365 | |
2366 // Store of zero anywhere into a freshly-allocated object? | |
2367 // Then the store is useless. | |
2368 // (It must already have been captured by the InitializeNode.) | |
2369 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { | |
2370 // a newly allocated object is already all-zeroes everywhere | |
2371 if (mem->is_Proj() && mem->in(0)->is_Allocate()) { | |
2372 return mem; | |
2373 } | |
2374 | |
2375 // the store may also apply to zero-bits in an earlier object | |
2376 Node* prev_mem = find_previous_store(phase); | |
2377 // Steps (a), (b): Walk past independent stores to find an exact match. | |
2378 if (prev_mem != NULL) { | |
2379 Node* prev_val = can_see_stored_value(prev_mem, phase); | |
2380 if (prev_val != NULL && phase->eqv(prev_val, val)) { | |
2381 // prev_val and val might differ by a cast; it would be good | |
2382 // to keep the more informative of the two. | |
2383 return mem; | |
2384 } | |
2385 } | |
2386 } | |
2387 | |
2388 return this; | |
2389 } | |
2390 | |
2391 //------------------------------match_edge------------------------------------- | |
2392 // Do we Match on this edge index or not? Match only memory & value | |
2393 uint StoreNode::match_edge(uint idx) const { | |
2394 return idx == MemNode::Address || idx == MemNode::ValueIn; | |
2395 } | |
2396 | |
2397 //------------------------------cmp-------------------------------------------- | |
2398 // Do not common stores up together. They generally have to be split | |
2399 // back up anyways, so do not bother. | |
2400 uint StoreNode::cmp( const Node &n ) const { | |
2401 return (&n == this); // Always fail except on self | |
2402 } | |
2403 | |
2404 //------------------------------Ideal_masked_input----------------------------- | |
2405 // Check for a useless mask before a partial-word store | |
2406 // (StoreB ... (AndI valIn conIa) ) | |
2407 // If (conIa & mask == mask) this simplifies to | |
2408 // (StoreB ... (valIn) ) | |
2409 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { | |
2410 Node *val = in(MemNode::ValueIn); | |
2411 if( val->Opcode() == Op_AndI ) { | |
2412 const TypeInt *t = phase->type( val->in(2) )->isa_int(); | |
2413 if( t && t->is_con() && (t->get_con() & mask) == mask ) { | |
2414 set_req(MemNode::ValueIn, val->in(1)); | |
2415 return this; | |
2416 } | |
2417 } | |
2418 return NULL; | |
2419 } | |
2420 | |
2421 | |
2422 //------------------------------Ideal_sign_extended_input---------------------- | |
2423 // Check for useless sign-extension before a partial-word store | |
2424 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) | |
2425 // If (conIL == conIR && conIR <= num_bits) this simplifies to | |
2426 // (StoreB ... (valIn) ) | |
2427 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { | |
2428 Node *val = in(MemNode::ValueIn); | |
2429 if( val->Opcode() == Op_RShiftI ) { | |
2430 const TypeInt *t = phase->type( val->in(2) )->isa_int(); | |
2431 if( t && t->is_con() && (t->get_con() <= num_bits) ) { | |
2432 Node *shl = val->in(1); | |
2433 if( shl->Opcode() == Op_LShiftI ) { | |
2434 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); | |
2435 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { | |
2436 set_req(MemNode::ValueIn, shl->in(1)); | |
2437 return this; | |
2438 } | |
2439 } | |
2440 } | |
2441 } | |
2442 return NULL; | |
2443 } | |
2444 | |
2445 //------------------------------value_never_loaded----------------------------------- | |
2446 // Determine whether there are any possible loads of the value stored. | |
2447 // For simplicity, we actually check if there are any loads from the | |
2448 // address stored to, not just for loads of the value stored by this node. | |
2449 // | |
2450 bool StoreNode::value_never_loaded( PhaseTransform *phase) const { | |
2451 Node *adr = in(Address); | |
2452 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); | |
2453 if (adr_oop == NULL) | |
2454 return false; | |
223 | 2455 if (!adr_oop->is_known_instance_field()) |
0 | 2456 return false; // if not a distinct instance, there may be aliases of the address |
2457 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { | |
2458 Node *use = adr->fast_out(i); | |
2459 int opc = use->Opcode(); | |
2460 if (use->is_Load() || use->is_LoadStore()) { | |
2461 return false; | |
2462 } | |
2463 } | |
2464 return true; | |
2465 } | |
2466 | |
2467 //============================================================================= | |
2468 //------------------------------Ideal------------------------------------------ | |
2469 // If the store is from an AND mask that leaves the low bits untouched, then | |
2470 // we can skip the AND operation. If the store is from a sign-extension | |
2471 // (a left shift, then right shift) we can skip both. | |
2472 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2473 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); | |
2474 if( progress != NULL ) return progress; | |
2475 | |
2476 progress = StoreNode::Ideal_sign_extended_input(phase, 24); | |
2477 if( progress != NULL ) return progress; | |
2478 | |
2479 // Finally check the default case | |
2480 return StoreNode::Ideal(phase, can_reshape); | |
2481 } | |
2482 | |
2483 //============================================================================= | |
2484 //------------------------------Ideal------------------------------------------ | |
2485 // If the store is from an AND mask that leaves the low bits untouched, then | |
2486 // we can skip the AND operation | |
2487 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2488 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); | |
2489 if( progress != NULL ) return progress; | |
2490 | |
2491 progress = StoreNode::Ideal_sign_extended_input(phase, 16); | |
2492 if( progress != NULL ) return progress; | |
2493 | |
2494 // Finally check the default case | |
2495 return StoreNode::Ideal(phase, can_reshape); | |
2496 } | |
2497 | |
2498 //============================================================================= | |
2499 //------------------------------Identity--------------------------------------- | |
2500 Node *StoreCMNode::Identity( PhaseTransform *phase ) { | |
2501 // No need to card mark when storing a null ptr | |
2502 Node* my_store = in(MemNode::OopStore); | |
2503 if (my_store->is_Store()) { | |
2504 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); | |
2505 if( t1 == TypePtr::NULL_PTR ) { | |
2506 return in(MemNode::Memory); | |
2507 } | |
2508 } | |
2509 return this; | |
2510 } | |
2511 | |
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
|
2512 //============================================================================= |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2513 //------------------------------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
|
2514 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
|
2515 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
|
2516 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
|
2517 |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2518 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
|
2519 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
|
2520 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
|
2521 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
|
2522 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
|
2523 } |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2524 |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2525 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
|
2526 } |
685e959d09ea
6877254: Server vm crashes with no branches off of store slice" when run with CMS and UseSuperWord(default)
cfang
parents:
708
diff
changeset
|
2527 |
0 | 2528 //------------------------------Value----------------------------------------- |
2529 const Type *StoreCMNode::Value( PhaseTransform *phase ) const { | |
43 | 2530 // Either input is TOP ==> the result is TOP |
2531 const Type *t = phase->type( in(MemNode::Memory) ); | |
2532 if( t == Type::TOP ) return Type::TOP; | |
2533 t = phase->type( in(MemNode::Address) ); | |
2534 if( t == Type::TOP ) return Type::TOP; | |
2535 t = phase->type( in(MemNode::ValueIn) ); | |
2536 if( t == Type::TOP ) return Type::TOP; | |
0 | 2537 // If extra input is TOP ==> the result is TOP |
43 | 2538 t = phase->type( in(MemNode::OopStore) ); |
2539 if( t == Type::TOP ) return Type::TOP; | |
0 | 2540 |
2541 return StoreNode::Value( phase ); | |
2542 } | |
2543 | |
2544 | |
2545 //============================================================================= | |
2546 //----------------------------------SCMemProjNode------------------------------ | |
2547 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const | |
2548 { | |
2549 return bottom_type(); | |
2550 } | |
2551 | |
2552 //============================================================================= | |
2553 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { | |
2554 init_req(MemNode::Control, c ); | |
2555 init_req(MemNode::Memory , mem); | |
2556 init_req(MemNode::Address, adr); | |
2557 init_req(MemNode::ValueIn, val); | |
2558 init_req( ExpectedIn, ex ); | |
2559 init_class_id(Class_LoadStore); | |
2560 | |
2561 } | |
2562 | |
2563 //============================================================================= | |
2564 //-------------------------------adr_type-------------------------------------- | |
2565 // Do we Match on this edge index or not? Do not match memory | |
2566 const TypePtr* ClearArrayNode::adr_type() const { | |
2567 Node *adr = in(3); | |
2568 return MemNode::calculate_adr_type(adr->bottom_type()); | |
2569 } | |
2570 | |
2571 //------------------------------match_edge------------------------------------- | |
2572 // Do we Match on this edge index or not? Do not match memory | |
2573 uint ClearArrayNode::match_edge(uint idx) const { | |
2574 return idx > 1; | |
2575 } | |
2576 | |
2577 //------------------------------Identity--------------------------------------- | |
2578 // Clearing a zero length array does nothing | |
2579 Node *ClearArrayNode::Identity( PhaseTransform *phase ) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2580 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; |
0 | 2581 } |
2582 | |
2583 //------------------------------Idealize--------------------------------------- | |
2584 // Clearing a short array is faster with stores | |
2585 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ | |
2586 const int unit = BytesPerLong; | |
2587 const TypeX* t = phase->type(in(2))->isa_intptr_t(); | |
2588 if (!t) return NULL; | |
2589 if (!t->is_con()) return NULL; | |
2590 intptr_t raw_count = t->get_con(); | |
2591 intptr_t size = raw_count; | |
2592 if (!Matcher::init_array_count_is_in_bytes) size *= unit; | |
2593 // Clearing nothing uses the Identity call. | |
2594 // Negative clears are possible on dead ClearArrays | |
2595 // (see jck test stmt114.stmt11402.val). | |
2596 if (size <= 0 || size % unit != 0) return NULL; | |
2597 intptr_t count = size / unit; | |
2598 // Length too long; use fast hardware clear | |
2599 if (size > Matcher::init_array_short_size) return NULL; | |
2600 Node *mem = in(1); | |
2601 if( phase->type(mem)==Type::TOP ) return NULL; | |
2602 Node *adr = in(3); | |
2603 const Type* at = phase->type(adr); | |
2604 if( at==Type::TOP ) return NULL; | |
2605 const TypePtr* atp = at->isa_ptr(); | |
2606 // adjust atp to be the correct array element address type | |
2607 if (atp == NULL) atp = TypePtr::BOTTOM; | |
2608 else atp = atp->add_offset(Type::OffsetBot); | |
2609 // Get base for derived pointer purposes | |
2610 if( adr->Opcode() != Op_AddP ) Unimplemented(); | |
2611 Node *base = adr->in(1); | |
2612 | |
2613 Node *zero = phase->makecon(TypeLong::ZERO); | |
2614 Node *off = phase->MakeConX(BytesPerLong); | |
2615 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); | |
2616 count--; | |
2617 while( count-- ) { | |
2618 mem = phase->transform(mem); | |
2619 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); | |
2620 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); | |
2621 } | |
2622 return mem; | |
2623 } | |
2624 | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2625 //----------------------------step_through---------------------------------- |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2626 // 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
|
2627 // 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
|
2628 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
|
2629 Node* n = *np; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2630 assert(n->is_ClearArray(), "sanity"); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2631 intptr_t offset; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2632 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
|
2633 // 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
|
2634 // 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
|
2635 // in LibraryCallKit::generate_arraycopy() which follows allocations. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2636 assert(alloc != NULL, "should have allocation"); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2637 if (alloc->_idx == instance_id) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2638 // 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
|
2639 return false; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2640 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2641 // Otherwise skip it. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2642 InitializeNode* init = alloc->initialization(); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2643 if (init != NULL) |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2644 *np = init->in(TypeFunc::Memory); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2645 else |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2646 *np = alloc->in(TypeFunc::Memory); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2647 return true; |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2648 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2649 |
0 | 2650 //----------------------------clear_memory------------------------------------- |
2651 // Generate code to initialize object storage to zero. | |
2652 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2653 intptr_t start_offset, | |
2654 Node* end_offset, | |
2655 PhaseGVN* phase) { | |
2656 Compile* C = phase->C; | |
2657 intptr_t offset = start_offset; | |
2658 | |
2659 int unit = BytesPerLong; | |
2660 if ((offset % unit) != 0) { | |
2661 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); | |
2662 adr = phase->transform(adr); | |
2663 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2664 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
0 | 2665 mem = phase->transform(mem); |
2666 offset += BytesPerInt; | |
2667 } | |
2668 assert((offset % unit) == 0, ""); | |
2669 | |
2670 // Initialize the remaining stuff, if any, with a ClearArray. | |
2671 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); | |
2672 } | |
2673 | |
2674 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2675 Node* start_offset, | |
2676 Node* end_offset, | |
2677 PhaseGVN* phase) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2678 if (start_offset == end_offset) { |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2679 // nothing to do |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2680 return mem; |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2681 } |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2682 |
0 | 2683 Compile* C = phase->C; |
2684 int unit = BytesPerLong; | |
2685 Node* zbase = start_offset; | |
2686 Node* zend = end_offset; | |
2687 | |
2688 // Scale to the unit required by the CPU: | |
2689 if (!Matcher::init_array_count_is_in_bytes) { | |
2690 Node* shift = phase->intcon(exact_log2(unit)); | |
2691 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); | |
2692 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); | |
2693 } | |
2694 | |
2695 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); | |
2696 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); | |
2697 | |
2698 // Bulk clear double-words | |
2699 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); | |
2700 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); | |
2701 return phase->transform(mem); | |
2702 } | |
2703 | |
2704 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, | |
2705 intptr_t start_offset, | |
2706 intptr_t end_offset, | |
2707 PhaseGVN* phase) { | |
68
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2708 if (start_offset == end_offset) { |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2709 // nothing to do |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2710 return mem; |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2711 } |
daf38130e60d
6676841: ClearArrayNode::Identity is incorrect for 64-bit
never
parents:
64
diff
changeset
|
2712 |
0 | 2713 Compile* C = phase->C; |
2714 assert((end_offset % BytesPerInt) == 0, "odd end offset"); | |
2715 intptr_t done_offset = end_offset; | |
2716 if ((done_offset % BytesPerLong) != 0) { | |
2717 done_offset -= BytesPerInt; | |
2718 } | |
2719 if (done_offset > start_offset) { | |
2720 mem = clear_memory(ctl, mem, dest, | |
2721 start_offset, phase->MakeConX(done_offset), phase); | |
2722 } | |
2723 if (done_offset < end_offset) { // emit the final 32-bit store | |
2724 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); | |
2725 adr = phase->transform(adr); | |
2726 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
2727 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); |
0 | 2728 mem = phase->transform(mem); |
2729 done_offset += BytesPerInt; | |
2730 } | |
2731 assert(done_offset == end_offset, ""); | |
2732 return mem; | |
2733 } | |
2734 | |
2735 //============================================================================= | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2736 // Do not match memory edge. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2737 uint StrIntrinsicNode::match_edge(uint idx) const { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2738 return idx == 2 || idx == 3; |
681 | 2739 } |
2740 | |
2741 //------------------------------Ideal------------------------------------------ | |
2742 // Return a node which is more "ideal" than the current node. Strip out | |
2743 // control copies | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2744 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2745 if (remove_dead_region(phase, can_reshape)) return this; |
4115 | 2746 // Don't bother trying to transform a dead node |
2747 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
|
2748 |
2417 | 2749 if (can_reshape) { |
2750 Node* mem = phase->transform(in(MemNode::Memory)); | |
2751 // If transformed to a MergeMem, get the desired slice | |
2752 uint alias_idx = phase->C->get_alias_index(adr_type()); | |
2753 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem; | |
2754 if (mem != in(MemNode::Memory)) { | |
2755 set_req(MemNode::Memory, mem); | |
2756 return this; | |
2757 } | |
2758 } | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2379
diff
changeset
|
2759 return NULL; |
169
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
2760 } |
9148c65abefc
6695049: (coll) Create an x86 intrinsic for Arrays.equals
rasbold
parents:
168
diff
changeset
|
2761 |
4115 | 2762 //------------------------------Value------------------------------------------ |
2763 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const { | |
2764 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP; | |
2765 return bottom_type(); | |
2766 } | |
2767 | |
0 | 2768 //============================================================================= |
2769 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) | |
2770 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), | |
2771 _adr_type(C->get_adr_type(alias_idx)) | |
2772 { | |
2773 init_class_id(Class_MemBar); | |
2774 Node* top = C->top(); | |
2775 init_req(TypeFunc::I_O,top); | |
2776 init_req(TypeFunc::FramePtr,top); | |
2777 init_req(TypeFunc::ReturnAdr,top); | |
2778 if (precedent != NULL) | |
2779 init_req(TypeFunc::Parms, precedent); | |
2780 } | |
2781 | |
2782 //------------------------------cmp-------------------------------------------- | |
2783 uint MemBarNode::hash() const { return NO_HASH; } | |
2784 uint MemBarNode::cmp( const Node &n ) const { | |
2785 return (&n == this); // Always fail except on self | |
2786 } | |
2787 | |
2788 //------------------------------make------------------------------------------- | |
2789 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { | |
2790 int len = Precedent + (pn == NULL? 0: 1); | |
2791 switch (opcode) { | |
2792 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); | |
2793 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); | |
3849
f1c12354c3f7
7074017: Introduce MemBarAcquireLock/MemBarReleaseLock nodes for monitor enter/exit code paths
roland
parents:
3284
diff
changeset
|
2794 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn); |
f1c12354c3f7
7074017: Introduce MemBarAcquireLock/MemBarReleaseLock nodes for monitor enter/exit code paths
roland
parents:
3284
diff
changeset
|
2795 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn); |
0 | 2796 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); |
2797 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); | |
2798 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); | |
4763
1dc233a8c7fe
7121140: Allocation paths require explicit memory synchronization operations for RMO systems
roland
parents:
4762
diff
changeset
|
2799 case Op_MemBarStoreStore: return new(C, len) MemBarStoreStoreNode(C, atp, pn); |
0 | 2800 default: ShouldNotReachHere(); return NULL; |
2801 } | |
2802 } | |
2803 | |
2804 //------------------------------Ideal------------------------------------------ | |
2805 // Return a node which is more "ideal" than the current node. Strip out | |
2806 // control copies | |
2807 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
|
2808 if (remove_dead_region(phase, can_reshape)) return this; |
4115 | 2809 // Don't bother trying to transform a dead node |
2810 if (in(0) && in(0)->is_top()) return NULL; | |
1100
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2811 |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2812 // Eliminate volatile MemBars for scalar replaced objects. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2813 if (can_reshape && req() == (Precedent+1) && |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2814 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2815 // Volatile field loads and stores. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2816 Node* my_mem = in(MemBarNode::Precedent); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2817 if (my_mem != NULL && my_mem->is_Mem()) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2818 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr(); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2819 // Check for scalar replaced object reference. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2820 if( t_oop != NULL && t_oop->is_known_instance_field() && |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2821 t_oop->offset() != Type::OffsetBot && |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2822 t_oop->offset() != Type::OffsetTop) { |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2823 // Replace MemBar projections by its inputs. |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2824 PhaseIterGVN* igvn = phase->is_IterGVN(); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2825 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory)); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2826 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control)); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2827 // Must return either the original node (now dead) or a new node |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2828 // (Do not return a top here, since that would break the uniqueness of top.) |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2829 return new (phase->C, 1) ConINode(TypeInt::ZERO); |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2830 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2831 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2832 } |
f96a1a986f7b
6895383: JCK test throws NPE for method compiled with Escape Analysis
kvn
parents:
1080
diff
changeset
|
2833 return NULL; |
0 | 2834 } |
2835 | |
2836 //------------------------------Value------------------------------------------ | |
2837 const Type *MemBarNode::Value( PhaseTransform *phase ) const { | |
2838 if( !in(0) ) return Type::TOP; | |
2839 if( phase->type(in(0)) == Type::TOP ) | |
2840 return Type::TOP; | |
2841 return TypeTuple::MEMBAR; | |
2842 } | |
2843 | |
2844 //------------------------------match------------------------------------------ | |
2845 // Construct projections for memory. | |
2846 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { | |
2847 switch (proj->_con) { | |
2848 case TypeFunc::Control: | |
2849 case TypeFunc::Memory: | |
2850 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); | |
2851 } | |
2852 ShouldNotReachHere(); | |
2853 return NULL; | |
2854 } | |
2855 | |
2856 //===========================InitializeNode==================================== | |
2857 // SUMMARY: | |
2858 // This node acts as a memory barrier on raw memory, after some raw stores. | |
2859 // The 'cooked' oop value feeds from the Initialize, not the Allocation. | |
2860 // The Initialize can 'capture' suitably constrained stores as raw inits. | |
2861 // It can coalesce related raw stores into larger units (called 'tiles'). | |
2862 // It can avoid zeroing new storage for memory units which have raw inits. | |
2863 // At macro-expansion, it is marked 'complete', and does not optimize further. | |
2864 // | |
2865 // EXAMPLE: | |
2866 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. | |
2867 // ctl = incoming control; mem* = incoming memory | |
2868 // (Note: A star * on a memory edge denotes I/O and other standard edges.) | |
2869 // First allocate uninitialized memory and fill in the header: | |
2870 // alloc = (Allocate ctl mem* 16 #short[].klass ...) | |
2871 // ctl := alloc.Control; mem* := alloc.Memory* | |
2872 // rawmem = alloc.Memory; rawoop = alloc.RawAddress | |
2873 // Then initialize to zero the non-header parts of the raw memory block: | |
2874 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) | |
2875 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory | |
2876 // After the initialize node executes, the object is ready for service: | |
2877 // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) | |
2878 // Suppose its body is immediately initialized as {1,2}: | |
2879 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) | |
2880 // store2 = (StoreC init.Control store1 (+ oop 14) 2) | |
2881 // mem.SLICE(#short[*]) := store2 | |
2882 // | |
2883 // DETAILS: | |
2884 // An InitializeNode collects and isolates object initialization after | |
2885 // an AllocateNode and before the next possible safepoint. As a | |
2886 // memory barrier (MemBarNode), it keeps critical stores from drifting | |
2887 // down past any safepoint or any publication of the allocation. | |
2888 // Before this barrier, a newly-allocated object may have uninitialized bits. | |
2889 // After this barrier, it may be treated as a real oop, and GC is allowed. | |
2890 // | |
2891 // The semantics of the InitializeNode include an implicit zeroing of | |
2892 // the new object from object header to the end of the object. | |
2893 // (The object header and end are determined by the AllocateNode.) | |
2894 // | |
2895 // Certain stores may be added as direct inputs to the InitializeNode. | |
2896 // These stores must update raw memory, and they must be to addresses | |
2897 // derived from the raw address produced by AllocateNode, and with | |
2898 // a constant offset. They must be ordered by increasing offset. | |
2899 // The first one is at in(RawStores), the last at in(req()-1). | |
2900 // Unlike most memory operations, they are not linked in a chain, | |
2901 // but are displayed in parallel as users of the rawmem output of | |
2902 // the allocation. | |
2903 // | |
2904 // (See comments in InitializeNode::capture_store, which continue | |
2905 // the example given above.) | |
2906 // | |
2907 // When the associated Allocate is macro-expanded, the InitializeNode | |
2908 // may be rewritten to optimize collected stores. A ClearArrayNode | |
2909 // may also be created at that point to represent any required zeroing. | |
2910 // The InitializeNode is then marked 'complete', prohibiting further | |
2911 // capturing of nearby memory operations. | |
2912 // | |
2913 // During macro-expansion, all captured initializations which store | |
605 | 2914 // constant values of 32 bits or smaller are coalesced (if advantageous) |
0 | 2915 // into larger 'tiles' 32 or 64 bits. This allows an object to be |
2916 // initialized in fewer memory operations. Memory words which are | |
2917 // covered by neither tiles nor non-constant stores are pre-zeroed | |
2918 // by explicit stores of zero. (The code shape happens to do all | |
2919 // zeroing first, then all other stores, with both sequences occurring | |
2920 // in order of ascending offsets.) | |
2921 // | |
2922 // Alternatively, code may be inserted between an AllocateNode and its | |
2923 // InitializeNode, to perform arbitrary initialization of the new object. | |
2924 // E.g., the object copying intrinsics insert complex data transfers here. | |
2925 // The initialization must then be marked as 'complete' disable the | |
2926 // built-in zeroing semantics and the collection of initializing stores. | |
2927 // | |
2928 // While an InitializeNode is incomplete, reads from the memory state | |
2929 // produced by it are optimizable if they match the control edge and | |
2930 // new oop address associated with the allocation/initialization. | |
2931 // They return a stored value (if the offset matches) or else zero. | |
2932 // A write to the memory state, if it matches control and address, | |
2933 // and if it is to a constant offset, may be 'captured' by the | |
2934 // InitializeNode. It is cloned as a raw memory operation and rewired | |
2935 // inside the initialization, to the raw oop produced by the allocation. | |
2936 // Operations on addresses which are provably distinct (e.g., to | |
2937 // other AllocateNodes) are allowed to bypass the initialization. | |
2938 // | |
2939 // The effect of all this is to consolidate object initialization | |
2940 // (both arrays and non-arrays, both piecewise and bulk) into a | |
2941 // single location, where it can be optimized as a unit. | |
2942 // | |
2943 // Only stores with an offset less than TrackedInitializationLimit words | |
2944 // will be considered for capture by an InitializeNode. This puts a | |
2945 // reasonable limit on the complexity of optimized initializations. | |
2946 | |
2947 //---------------------------InitializeNode------------------------------------ | |
2948 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
|
2949 : _is_complete(Incomplete), _does_not_escape(false), |
0 | 2950 MemBarNode(C, adr_type, rawoop) |
2951 { | |
2952 init_class_id(Class_Initialize); | |
2953 | |
2954 assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); | |
2955 assert(in(RawAddress) == rawoop, "proper init"); | |
2956 // Note: allocation() can be NULL, for secondary initialization barriers | |
2957 } | |
2958 | |
2959 // Since this node is not matched, it will be processed by the | |
2960 // register allocator. Declare that there are no constraints | |
2961 // on the allocation of the RawAddress edge. | |
2962 const RegMask &InitializeNode::in_RegMask(uint idx) const { | |
2963 // This edge should be set to top, by the set_complete. But be conservative. | |
2964 if (idx == InitializeNode::RawAddress) | |
2965 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); | |
2966 return RegMask::Empty; | |
2967 } | |
2968 | |
2969 Node* InitializeNode::memory(uint alias_idx) { | |
2970 Node* mem = in(Memory); | |
2971 if (mem->is_MergeMem()) { | |
2972 return mem->as_MergeMem()->memory_at(alias_idx); | |
2973 } else { | |
2974 // incoming raw memory is not split | |
2975 return mem; | |
2976 } | |
2977 } | |
2978 | |
2979 bool InitializeNode::is_non_zero() { | |
2980 if (is_complete()) return false; | |
2981 remove_extra_zeroes(); | |
2982 return (req() > RawStores); | |
2983 } | |
2984 | |
2985 void InitializeNode::set_complete(PhaseGVN* phase) { | |
2986 assert(!is_complete(), "caller responsibility"); | |
3961
a92cdbac8b9e
7081933: Use zeroing elimination optimization for large array
kvn
parents:
3902
diff
changeset
|
2987 _is_complete = Complete; |
0 | 2988 |
2989 // After this node is complete, it contains a bunch of | |
2990 // raw-memory initializations. There is no need for | |
2991 // it to have anything to do with non-raw memory effects. | |
2992 // Therefore, tell all non-raw users to re-optimize themselves, | |
2993 // after skipping the memory effects of this initialization. | |
2994 PhaseIterGVN* igvn = phase->is_IterGVN(); | |
2995 if (igvn) igvn->add_users_to_worklist(this); | |
2996 } | |
2997 | |
2998 // convenience function | |
2999 // return false if the init contains any stores already | |
3000 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { | |
3001 InitializeNode* init = initialization(); | |
3002 if (init == NULL || init->is_complete()) return false; | |
3003 init->remove_extra_zeroes(); | |
3004 // for now, if this allocation has already collected any inits, bail: | |
3005 if (init->is_non_zero()) return false; | |
3006 init->set_complete(phase); | |
3007 return true; | |
3008 } | |
3009 | |
3010 void InitializeNode::remove_extra_zeroes() { | |
3011 if (req() == RawStores) return; | |
3012 Node* zmem = zero_memory(); | |
3013 uint fill = RawStores; | |
3014 for (uint i = fill; i < req(); i++) { | |
3015 Node* n = in(i); | |
3016 if (n->is_top() || n == zmem) continue; // skip | |
3017 if (fill < i) set_req(fill, n); // compact | |
3018 ++fill; | |
3019 } | |
3020 // delete any empty spaces created: | |
3021 while (fill < req()) { | |
3022 del_req(fill); | |
3023 } | |
3024 } | |
3025 | |
3026 // Helper for remembering which stores go with which offsets. | |
3027 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { | |
3028 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node | |
3029 intptr_t offset = -1; | |
3030 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), | |
3031 phase, offset); | |
3032 if (base == NULL) return -1; // something is dead, | |
3033 if (offset < 0) return -1; // dead, dead | |
3034 return offset; | |
3035 } | |
3036 | |
3037 // Helper for proving that an initialization expression is | |
3038 // "simple enough" to be folded into an object initialization. | |
3039 // Attempts to prove that a store's initial value 'n' can be captured | |
3040 // within the initialization without creating a vicious cycle, such as: | |
3041 // { Foo p = new Foo(); p.next = p; } | |
3042 // True for constants and parameters and small combinations thereof. | |
3043 bool InitializeNode::detect_init_independence(Node* n, | |
3044 bool st_is_pinned, | |
3045 int& count) { | |
3046 if (n == NULL) return true; // (can this really happen?) | |
3047 if (n->is_Proj()) n = n->in(0); | |
3048 if (n == this) return false; // found a cycle | |
3049 if (n->is_Con()) return true; | |
3050 if (n->is_Start()) return true; // params, etc., are OK | |
3051 if (n->is_Root()) return true; // even better | |
3052 | |
3053 Node* ctl = n->in(0); | |
3054 if (ctl != NULL && !ctl->is_top()) { | |
3055 if (ctl->is_Proj()) ctl = ctl->in(0); | |
3056 if (ctl == this) return false; | |
3057 | |
3058 // If we already know that the enclosing memory op is pinned right after | |
3059 // the init, then any control flow that the store has picked up | |
3060 // must have preceded the init, or else be equal to the init. | |
3061 // Even after loop optimizations (which might change control edges) | |
3062 // 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
|
3063 if (!MemNode::all_controls_dominate(n, this)) |
0 | 3064 return false; // failed to prove a good control |
3065 | |
3066 } | |
3067 | |
3068 // Check data edges for possible dependencies on 'this'. | |
3069 if ((count += 1) > 20) return false; // complexity limit | |
3070 for (uint i = 1; i < n->req(); i++) { | |
3071 Node* m = n->in(i); | |
3072 if (m == NULL || m == n || m->is_top()) continue; | |
3073 uint first_i = n->find_edge(m); | |
3074 if (i != first_i) continue; // process duplicate edge just once | |
3075 if (!detect_init_independence(m, st_is_pinned, count)) { | |
3076 return false; | |
3077 } | |
3078 } | |
3079 | |
3080 return true; | |
3081 } | |
3082 | |
3083 // Here are all the checks a Store must pass before it can be moved into | |
3084 // an initialization. Returns zero if a check fails. | |
3085 // On success, returns the (constant) offset to which the store applies, | |
3086 // within the initialized memory. | |
3087 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { | |
3088 const int FAIL = 0; | |
3089 if (st->req() != MemNode::ValueIn + 1) | |
3090 return FAIL; // an inscrutable StoreNode (card mark?) | |
3091 Node* ctl = st->in(MemNode::Control); | |
3092 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) | |
3093 return FAIL; // must be unconditional after the initialization | |
3094 Node* mem = st->in(MemNode::Memory); | |
3095 if (!(mem->is_Proj() && mem->in(0) == this)) | |
3096 return FAIL; // must not be preceded by other stores | |
3097 Node* adr = st->in(MemNode::Address); | |
3098 intptr_t offset; | |
3099 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); | |
3100 if (alloc == NULL) | |
3101 return FAIL; // inscrutable address | |
3102 if (alloc != allocation()) | |
3103 return FAIL; // wrong allocation! (store needs to float up) | |
3104 Node* val = st->in(MemNode::ValueIn); | |
3105 int complexity_count = 0; | |
3106 if (!detect_init_independence(val, true, complexity_count)) | |
3107 return FAIL; // stored value must be 'simple enough' | |
3108 | |
3109 return offset; // success | |
3110 } | |
3111 | |
3112 // Find the captured store in(i) which corresponds to the range | |
3113 // [start..start+size) in the initialized object. | |
3114 // If there is one, return its index i. If there isn't, return the | |
3115 // negative of the index where it should be inserted. | |
3116 // Return 0 if the queried range overlaps an initialization boundary | |
3117 // or if dead code is encountered. | |
3118 // If size_in_bytes is zero, do not bother with overlap checks. | |
3119 int InitializeNode::captured_store_insertion_point(intptr_t start, | |
3120 int size_in_bytes, | |
3121 PhaseTransform* phase) { | |
3122 const int FAIL = 0, MAX_STORE = BytesPerLong; | |
3123 | |
3124 if (is_complete()) | |
3125 return FAIL; // arraycopy got here first; punt | |
3126 | |
3127 assert(allocation() != NULL, "must be present"); | |
3128 | |
3129 // 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
|
3130 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; |
0 | 3131 |
3132 // after a certain size, we bail out on tracking all the stores: | |
3133 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); | |
3134 if (start >= ti_limit) return FAIL; | |
3135 | |
3136 for (uint i = InitializeNode::RawStores, limit = req(); ; ) { | |
3137 if (i >= limit) return -(int)i; // not found; here is where to put it | |
3138 | |
3139 Node* st = in(i); | |
3140 intptr_t st_off = get_store_offset(st, phase); | |
3141 if (st_off < 0) { | |
3142 if (st != zero_memory()) { | |
3143 return FAIL; // bail out if there is dead garbage | |
3144 } | |
3145 } else if (st_off > start) { | |
3146 // ...we are done, since stores are ordered | |
3147 if (st_off < start + size_in_bytes) { | |
3148 return FAIL; // the next store overlaps | |
3149 } | |
3150 return -(int)i; // not found; here is where to put it | |
3151 } else if (st_off < start) { | |
3152 if (size_in_bytes != 0 && | |
3153 start < st_off + MAX_STORE && | |
3154 start < st_off + st->as_Store()->memory_size()) { | |
3155 return FAIL; // the previous store overlaps | |
3156 } | |
3157 } else { | |
3158 if (size_in_bytes != 0 && | |
3159 st->as_Store()->memory_size() != size_in_bytes) { | |
3160 return FAIL; // mismatched store size | |
3161 } | |
3162 return i; | |
3163 } | |
3164 | |
3165 ++i; | |
3166 } | |
3167 } | |
3168 | |
3169 // Look for a captured store which initializes at the offset 'start' | |
3170 // with the given size. If there is no such store, and no other | |
3171 // initialization interferes, then return zero_memory (the memory | |
3172 // projection of the AllocateNode). | |
3173 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, | |
3174 PhaseTransform* phase) { | |
3175 assert(stores_are_sane(phase), ""); | |
3176 int i = captured_store_insertion_point(start, size_in_bytes, phase); | |
3177 if (i == 0) { | |
3178 return NULL; // something is dead | |
3179 } else if (i < 0) { | |
3180 return zero_memory(); // just primordial zero bits here | |
3181 } else { | |
3182 Node* st = in(i); // here is the store at this position | |
3183 assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); | |
3184 return st; | |
3185 } | |
3186 } | |
3187 | |
3188 // Create, as a raw pointer, an address within my new object at 'offset'. | |
3189 Node* InitializeNode::make_raw_address(intptr_t offset, | |
3190 PhaseTransform* phase) { | |
3191 Node* addr = in(RawAddress); | |
3192 if (offset != 0) { | |
3193 Compile* C = phase->C; | |
3194 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, | |
3195 phase->MakeConX(offset)) ); | |
3196 } | |
3197 return addr; | |
3198 } | |
3199 | |
3200 // Clone the given store, converting it into a raw store | |
3201 // initializing a field or element of my new object. | |
3202 // Caller is responsible for retiring the original store, | |
3203 // with subsume_node or the like. | |
3204 // | |
3205 // From the example above InitializeNode::InitializeNode, | |
3206 // here are the old stores to be captured: | |
3207 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) | |
3208 // store2 = (StoreC init.Control store1 (+ oop 14) 2) | |
3209 // | |
3210 // Here is the changed code; note the extra edges on init: | |
3211 // alloc = (Allocate ...) | |
3212 // rawoop = alloc.RawAddress | |
3213 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) | |
3214 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) | |
3215 // init = (Initialize alloc.Control alloc.Memory rawoop | |
3216 // rawstore1 rawstore2) | |
3217 // | |
3218 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, | |
3219 PhaseTransform* phase) { | |
3220 assert(stores_are_sane(phase), ""); | |
3221 | |
3222 if (start < 0) return NULL; | |
3223 assert(can_capture_store(st, phase) == start, "sanity"); | |
3224 | |
3225 Compile* C = phase->C; | |
3226 int size_in_bytes = st->memory_size(); | |
3227 int i = captured_store_insertion_point(start, size_in_bytes, phase); | |
3228 if (i == 0) return NULL; // bail out | |
3229 Node* prev_mem = NULL; // raw memory for the captured store | |
3230 if (i > 0) { | |
3231 prev_mem = in(i); // there is a pre-existing store under this one | |
3232 set_req(i, C->top()); // temporarily disconnect it | |
3233 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. | |
3234 } else { | |
3235 i = -i; // no pre-existing store | |
3236 prev_mem = zero_memory(); // a slice of the newly allocated object | |
3237 if (i > InitializeNode::RawStores && in(i-1) == prev_mem) | |
3238 set_req(--i, C->top()); // reuse this edge; it has been folded away | |
3239 else | |
3240 ins_req(i, C->top()); // build a new edge | |
3241 } | |
3242 Node* new_st = st->clone(); | |
3243 new_st->set_req(MemNode::Control, in(Control)); | |
3244 new_st->set_req(MemNode::Memory, prev_mem); | |
3245 new_st->set_req(MemNode::Address, make_raw_address(start, phase)); | |
3246 new_st = phase->transform(new_st); | |
3247 | |
3248 // At this point, new_st might have swallowed a pre-existing store | |
3249 // at the same offset, or perhaps new_st might have disappeared, | |
3250 // if it redundantly stored the same value (or zero to fresh memory). | |
3251 | |
3252 // In any case, wire it in: | |
3253 set_req(i, new_st); | |
3254 | |
3255 // The caller may now kill the old guy. | |
3256 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); | |
3257 assert(check_st == new_st || check_st == NULL, "must be findable"); | |
3258 assert(!is_complete(), ""); | |
3259 return new_st; | |
3260 } | |
3261 | |
3262 static bool store_constant(jlong* tiles, int num_tiles, | |
3263 intptr_t st_off, int st_size, | |
3264 jlong con) { | |
3265 if ((st_off & (st_size-1)) != 0) | |
3266 return false; // strange store offset (assume size==2**N) | |
3267 address addr = (address)tiles + st_off; | |
3268 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); | |
3269 switch (st_size) { | |
3270 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; | |
3271 case sizeof(jchar): *(jchar*) addr = (jchar) con; break; | |
3272 case sizeof(jint): *(jint*) addr = (jint) con; break; | |
3273 case sizeof(jlong): *(jlong*) addr = (jlong) con; break; | |
3274 default: return false; // strange store size (detect size!=2**N here) | |
3275 } | |
3276 return true; // return success to caller | |
3277 } | |
3278 | |
3279 // Coalesce subword constants into int constants and possibly | |
3280 // into long constants. The goal, if the CPU permits, | |
3281 // is to initialize the object with a small number of 64-bit tiles. | |
3282 // Also, convert floating-point constants to bit patterns. | |
3283 // Non-constants are not relevant to this pass. | |
3284 // | |
3285 // In terms of the running example on InitializeNode::InitializeNode | |
3286 // and InitializeNode::capture_store, here is the transformation | |
3287 // of rawstore1 and rawstore2 into rawstore12: | |
3288 // alloc = (Allocate ...) | |
3289 // rawoop = alloc.RawAddress | |
3290 // tile12 = 0x00010002 | |
3291 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) | |
3292 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) | |
3293 // | |
3294 void | |
3295 InitializeNode::coalesce_subword_stores(intptr_t header_size, | |
3296 Node* size_in_bytes, | |
3297 PhaseGVN* phase) { | |
3298 Compile* C = phase->C; | |
3299 | |
3300 assert(stores_are_sane(phase), ""); | |
3301 // Note: After this pass, they are not completely sane, | |
3302 // since there may be some overlaps. | |
3303 | |
3304 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; | |
3305 | |
3306 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); | |
3307 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); | |
3308 size_limit = MIN2(size_limit, ti_limit); | |
3309 size_limit = align_size_up(size_limit, BytesPerLong); | |
3310 int num_tiles = size_limit / BytesPerLong; | |
3311 | |
3312 // allocate space for the tile map: | |
3313 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small | |
3314 jlong tiles_buf[small_len]; | |
3315 Node* nodes_buf[small_len]; | |
3316 jlong inits_buf[small_len]; | |
3317 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] | |
3318 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); | |
3319 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] | |
3320 : NEW_RESOURCE_ARRAY(Node*, num_tiles)); | |
3321 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] | |
3322 : NEW_RESOURCE_ARRAY(jlong, num_tiles)); | |
3323 // tiles: exact bitwise model of all primitive constants | |
3324 // nodes: last constant-storing node subsumed into the tiles model | |
3325 // inits: which bytes (in each tile) are touched by any initializations | |
3326 | |
3327 //// Pass A: Fill in the tile model with any relevant stores. | |
3328 | |
3329 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); | |
3330 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); | |
3331 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); | |
3332 Node* zmem = zero_memory(); // initially zero memory state | |
3333 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { | |
3334 Node* st = in(i); | |
3335 intptr_t st_off = get_store_offset(st, phase); | |
3336 | |
3337 // Figure out the store's offset and constant value: | |
3338 if (st_off < header_size) continue; //skip (ignore header) | |
3339 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) | |
3340 int st_size = st->as_Store()->memory_size(); | |
3341 if (st_off + st_size > size_limit) break; | |
3342 | |
3343 // Record which bytes are touched, whether by constant or not. | |
3344 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) | |
3345 continue; // skip (strange store size) | |
3346 | |
3347 const Type* val = phase->type(st->in(MemNode::ValueIn)); | |
3348 if (!val->singleton()) continue; //skip (non-con store) | |
3349 BasicType type = val->basic_type(); | |
3350 | |
3351 jlong con = 0; | |
3352 switch (type) { | |
3353 case T_INT: con = val->is_int()->get_con(); break; | |
3354 case T_LONG: con = val->is_long()->get_con(); break; | |
3355 case T_FLOAT: con = jint_cast(val->getf()); break; | |
3356 case T_DOUBLE: con = jlong_cast(val->getd()); break; | |
3357 default: continue; //skip (odd store type) | |
3358 } | |
3359 | |
3360 if (type == T_LONG && Matcher::isSimpleConstant64(con) && | |
3361 st->Opcode() == Op_StoreL) { | |
3362 continue; // This StoreL is already optimal. | |
3363 } | |
3364 | |
3365 // Store down the constant. | |
3366 store_constant(tiles, num_tiles, st_off, st_size, con); | |
3367 | |
3368 intptr_t j = st_off >> LogBytesPerLong; | |
3369 | |
3370 if (type == T_INT && st_size == BytesPerInt | |
3371 && (st_off & BytesPerInt) == BytesPerInt) { | |
3372 jlong lcon = tiles[j]; | |
3373 if (!Matcher::isSimpleConstant64(lcon) && | |
3374 st->Opcode() == Op_StoreI) { | |
3375 // This StoreI is already optimal by itself. | |
3376 jint* intcon = (jint*) &tiles[j]; | |
3377 intcon[1] = 0; // undo the store_constant() | |
3378 | |
3379 // If the previous store is also optimal by itself, back up and | |
3380 // undo the action of the previous loop iteration... if we can. | |
3381 // But if we can't, just let the previous half take care of itself. | |
3382 st = nodes[j]; | |
3383 st_off -= BytesPerInt; | |
3384 con = intcon[0]; | |
3385 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { | |
3386 assert(st_off >= header_size, "still ignoring header"); | |
3387 assert(get_store_offset(st, phase) == st_off, "must be"); | |
3388 assert(in(i-1) == zmem, "must be"); | |
3389 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); | |
3390 assert(con == tcon->is_int()->get_con(), "must be"); | |
3391 // Undo the effects of the previous loop trip, which swallowed st: | |
3392 intcon[0] = 0; // undo store_constant() | |
3393 set_req(i-1, st); // undo set_req(i, zmem) | |
3394 nodes[j] = NULL; // undo nodes[j] = st | |
3395 --old_subword; // undo ++old_subword | |
3396 } | |
3397 continue; // This StoreI is already optimal. | |
3398 } | |
3399 } | |
3400 | |
3401 // This store is not needed. | |
3402 set_req(i, zmem); | |
3403 nodes[j] = st; // record for the moment | |
3404 if (st_size < BytesPerLong) // something has changed | |
3405 ++old_subword; // includes int/float, but who's counting... | |
3406 else ++old_long; | |
3407 } | |
3408 | |
3409 if ((old_subword + old_long) == 0) | |
3410 return; // nothing more to do | |
3411 | |
3412 //// Pass B: Convert any non-zero tiles into optimal constant stores. | |
3413 // Be sure to insert them before overlapping non-constant stores. | |
3414 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) | |
3415 for (int j = 0; j < num_tiles; j++) { | |
3416 jlong con = tiles[j]; | |
3417 jlong init = inits[j]; | |
3418 if (con == 0) continue; | |
3419 jint con0, con1; // split the constant, address-wise | |
3420 jint init0, init1; // split the init map, address-wise | |
3421 { union { jlong con; jint intcon[2]; } u; | |
3422 u.con = con; | |
3423 con0 = u.intcon[0]; | |
3424 con1 = u.intcon[1]; | |
3425 u.con = init; | |
3426 init0 = u.intcon[0]; | |
3427 init1 = u.intcon[1]; | |
3428 } | |
3429 | |
3430 Node* old = nodes[j]; | |
3431 assert(old != NULL, "need the prior store"); | |
3432 intptr_t offset = (j * BytesPerLong); | |
3433 | |
3434 bool split = !Matcher::isSimpleConstant64(con); | |
3435 | |
3436 if (offset < header_size) { | |
3437 assert(offset + BytesPerInt >= header_size, "second int counts"); | |
3438 assert(*(jint*)&tiles[j] == 0, "junk in header"); | |
3439 split = true; // only the second word counts | |
3440 // Example: int a[] = { 42 ... } | |
3441 } else if (con0 == 0 && init0 == -1) { | |
3442 split = true; // first word is covered by full inits | |
3443 // Example: int a[] = { ... foo(), 42 ... } | |
3444 } else if (con1 == 0 && init1 == -1) { | |
3445 split = true; // second word is covered by full inits | |
3446 // Example: int a[] = { ... 42, foo() ... } | |
3447 } | |
3448 | |
3449 // Here's a case where init0 is neither 0 nor -1: | |
3450 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } | |
3451 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. | |
3452 // In this case the tile is not split; it is (jlong)42. | |
3453 // The big tile is stored down, and then the foo() value is inserted. | |
3454 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) | |
3455 | |
3456 Node* ctl = old->in(MemNode::Control); | |
3457 Node* adr = make_raw_address(offset, phase); | |
3458 const TypePtr* atp = TypeRawPtr::BOTTOM; | |
3459 | |
3460 // One or two coalesced stores to plop down. | |
3461 Node* st[2]; | |
3462 intptr_t off[2]; | |
3463 int nst = 0; | |
3464 if (!split) { | |
3465 ++new_long; | |
3466 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3467 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
0 | 3468 phase->longcon(con), T_LONG); |
3469 } else { | |
3470 // Omit either if it is a zero. | |
3471 if (con0 != 0) { | |
3472 ++new_int; | |
3473 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3474 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
0 | 3475 phase->intcon(con0), T_INT); |
3476 } | |
3477 if (con1 != 0) { | |
3478 ++new_int; | |
3479 offset += BytesPerInt; | |
3480 adr = make_raw_address(offset, phase); | |
3481 off[nst] = offset; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3482 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, |
0 | 3483 phase->intcon(con1), T_INT); |
3484 } | |
3485 } | |
3486 | |
3487 // Insert second store first, then the first before the second. | |
3488 // Insert each one just before any overlapping non-constant stores. | |
3489 while (nst > 0) { | |
3490 Node* st1 = st[--nst]; | |
3491 C->copy_node_notes_to(st1, old); | |
3492 st1 = phase->transform(st1); | |
3493 offset = off[nst]; | |
3494 assert(offset >= header_size, "do not smash header"); | |
3495 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); | |
3496 guarantee(ins_idx != 0, "must re-insert constant store"); | |
3497 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap | |
3498 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) | |
3499 set_req(--ins_idx, st1); | |
3500 else | |
3501 ins_req(ins_idx, st1); | |
3502 } | |
3503 } | |
3504 | |
3505 if (PrintCompilation && WizardMode) | |
3506 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", | |
3507 old_subword, old_long, new_int, new_long); | |
3508 if (C->log() != NULL) | |
3509 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", | |
3510 old_subword, old_long, new_int, new_long); | |
3511 | |
3512 // Clean up any remaining occurrences of zmem: | |
3513 remove_extra_zeroes(); | |
3514 } | |
3515 | |
3516 // Explore forward from in(start) to find the first fully initialized | |
3517 // word, and return its offset. Skip groups of subword stores which | |
3518 // together initialize full words. If in(start) is itself part of a | |
3519 // fully initialized word, return the offset of in(start). If there | |
3520 // are no following full-word stores, or if something is fishy, return | |
3521 // a negative value. | |
3522 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { | |
3523 int int_map = 0; | |
3524 intptr_t int_map_off = 0; | |
3525 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for | |
3526 | |
3527 for (uint i = start, limit = req(); i < limit; i++) { | |
3528 Node* st = in(i); | |
3529 | |
3530 intptr_t st_off = get_store_offset(st, phase); | |
3531 if (st_off < 0) break; // return conservative answer | |
3532 | |
3533 int st_size = st->as_Store()->memory_size(); | |
3534 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { | |
3535 return st_off; // we found a complete word init | |
3536 } | |
3537 | |
3538 // update the map: | |
3539 | |
3540 intptr_t this_int_off = align_size_down(st_off, BytesPerInt); | |
3541 if (this_int_off != int_map_off) { | |
3542 // reset the map: | |
3543 int_map = 0; | |
3544 int_map_off = this_int_off; | |
3545 } | |
3546 | |
3547 int subword_off = st_off - this_int_off; | |
3548 int_map |= right_n_bits(st_size) << subword_off; | |
3549 if ((int_map & FULL_MAP) == FULL_MAP) { | |
3550 return this_int_off; // we found a complete word init | |
3551 } | |
3552 | |
3553 // Did this store hit or cross the word boundary? | |
3554 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); | |
3555 if (next_int_off == this_int_off + BytesPerInt) { | |
3556 // We passed the current int, without fully initializing it. | |
3557 int_map_off = next_int_off; | |
3558 int_map >>= BytesPerInt; | |
3559 } else if (next_int_off > this_int_off + BytesPerInt) { | |
3560 // We passed the current and next int. | |
3561 return this_int_off + BytesPerInt; | |
3562 } | |
3563 } | |
3564 | |
3565 return -1; | |
3566 } | |
3567 | |
3568 | |
3569 // Called when the associated AllocateNode is expanded into CFG. | |
3570 // At this point, we may perform additional optimizations. | |
3571 // Linearize the stores by ascending offset, to make memory | |
3572 // activity as coherent as possible. | |
3573 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, | |
3574 intptr_t header_size, | |
3575 Node* size_in_bytes, | |
3576 PhaseGVN* phase) { | |
3577 assert(!is_complete(), "not already complete"); | |
3578 assert(stores_are_sane(phase), ""); | |
3579 assert(allocation() != NULL, "must be present"); | |
3580 | |
3581 remove_extra_zeroes(); | |
3582 | |
3583 if (ReduceFieldZeroing || ReduceBulkZeroing) | |
3584 // reduce instruction count for common initialization patterns | |
3585 coalesce_subword_stores(header_size, size_in_bytes, phase); | |
3586 | |
3587 Node* zmem = zero_memory(); // initially zero memory state | |
3588 Node* inits = zmem; // accumulating a linearized chain of inits | |
3589 #ifdef ASSERT | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
82
diff
changeset
|
3590 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
|
3591 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
|
3592 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
|
3593 intptr_t last_tile_end = first_offset; // previous tile offset+size |
0 | 3594 #endif |
3595 intptr_t zeroes_done = header_size; | |
3596 | |
3597 bool do_zeroing = true; // we might give up if inits are very sparse | |
3598 int big_init_gaps = 0; // how many large gaps have we seen? | |
3599 | |
3600 if (ZeroTLAB) do_zeroing = false; | |
3601 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; | |
3602 | |
3603 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { | |
3604 Node* st = in(i); | |
3605 intptr_t st_off = get_store_offset(st, phase); | |
3606 if (st_off < 0) | |
3607 break; // unknown junk in the inits | |
3608 if (st->in(MemNode::Memory) != zmem) | |
3609 break; // complicated store chains somehow in list | |
3610 | |
3611 int st_size = st->as_Store()->memory_size(); | |
3612 intptr_t next_init_off = st_off + st_size; | |
3613 | |
3614 if (do_zeroing && zeroes_done < next_init_off) { | |
3615 // See if this store needs a zero before it or under it. | |
3616 intptr_t zeroes_needed = st_off; | |
3617 | |
3618 if (st_size < BytesPerInt) { | |
3619 // Look for subword stores which only partially initialize words. | |
3620 // If we find some, we must lay down some word-level zeroes first, | |
3621 // underneath the subword stores. | |
3622 // | |
3623 // Examples: | |
3624 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s | |
3625 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y | |
3626 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z | |
3627 // | |
3628 // Note: coalesce_subword_stores may have already done this, | |
3629 // if it was prompted by constant non-zero subword initializers. | |
3630 // But this case can still arise with non-constant stores. | |
3631 | |
3632 intptr_t next_full_store = find_next_fullword_store(i, phase); | |
3633 | |
3634 // In the examples above: | |
3635 // in(i) p q r s x y z | |
3636 // st_off 12 13 14 15 12 13 14 | |
3637 // st_size 1 1 1 1 1 1 1 | |
3638 // next_full_s. 12 16 16 16 16 16 16 | |
3639 // z's_done 12 16 16 16 12 16 12 | |
3640 // z's_needed 12 16 16 16 16 16 16 | |
3641 // zsize 0 0 0 0 4 0 4 | |
3642 if (next_full_store < 0) { | |
3643 // Conservative tack: Zero to end of current word. | |
3644 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); | |
3645 } else { | |
3646 // Zero to beginning of next fully initialized word. | |
3647 // Or, don't zero at all, if we are already in that word. | |
3648 assert(next_full_store >= zeroes_needed, "must go forward"); | |
3649 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); | |
3650 zeroes_needed = next_full_store; | |
3651 } | |
3652 } | |
3653 | |
3654 if (zeroes_needed > zeroes_done) { | |
3655 intptr_t zsize = zeroes_needed - zeroes_done; | |
3656 // Do some incremental zeroing on rawmem, in parallel with inits. | |
3657 zeroes_done = align_size_down(zeroes_done, BytesPerInt); | |
3658 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, | |
3659 zeroes_done, zeroes_needed, | |
3660 phase); | |
3661 zeroes_done = zeroes_needed; | |
3662 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) | |
3663 do_zeroing = false; // leave the hole, next time | |
3664 } | |
3665 } | |
3666 | |
3667 // Collect the store and move on: | |
3668 st->set_req(MemNode::Memory, inits); | |
3669 inits = st; // put it on the linearized chain | |
3670 set_req(i, zmem); // unhook from previous position | |
3671 | |
3672 if (zeroes_done == st_off) | |
3673 zeroes_done = next_init_off; | |
3674 | |
3675 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); | |
3676 | |
3677 #ifdef ASSERT | |
3678 // Various order invariants. Weaker than stores_are_sane because | |
3679 // a large constant tile can be filled in by smaller non-constant stores. | |
3680 assert(st_off >= last_init_off, "inits do not reverse"); | |
3681 last_init_off = st_off; | |
3682 const Type* val = NULL; | |
3683 if (st_size >= BytesPerInt && | |
3684 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && | |
3685 (int)val->basic_type() < (int)T_OBJECT) { | |
3686 assert(st_off >= last_tile_end, "tiles do not overlap"); | |
3687 assert(st_off >= last_init_end, "tiles do not overwrite inits"); | |
3688 last_tile_end = MAX2(last_tile_end, next_init_off); | |
3689 } else { | |
3690 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); | |
3691 assert(st_tile_end >= last_tile_end, "inits stay with tiles"); | |
3692 assert(st_off >= last_init_end, "inits do not overlap"); | |
3693 last_init_end = next_init_off; // it's a non-tile | |
3694 } | |
3695 #endif //ASSERT | |
3696 } | |
3697 | |
3698 remove_extra_zeroes(); // clear out all the zmems left over | |
3699 add_req(inits); | |
3700 | |
3701 if (!ZeroTLAB) { | |
3702 // If anything remains to be zeroed, zero it all now. | |
3703 zeroes_done = align_size_down(zeroes_done, BytesPerInt); | |
3704 // if it is the last unused 4 bytes of an instance, forget about it | |
3705 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); | |
3706 if (zeroes_done + BytesPerLong >= size_limit) { | |
3707 assert(allocation() != NULL, ""); | |
2004
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3708 if (allocation()->Opcode() == Op_Allocate) { |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3709 Node* klass_node = allocation()->in(AllocateNode::KlassNode); |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3710 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3711 if (zeroes_done == k->layout_helper()) |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3712 zeroes_done = size_limit; |
4da76e32c0be
7002666: eclipse CDT projects crash with compressed oops
never
parents:
1972
diff
changeset
|
3713 } |
0 | 3714 } |
3715 if (zeroes_done < size_limit) { | |
3716 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, | |
3717 zeroes_done, size_in_bytes, phase); | |
3718 } | |
3719 } | |
3720 | |
3721 set_complete(phase); | |
3722 return rawmem; | |
3723 } | |
3724 | |
3725 | |
3726 #ifdef ASSERT | |
3727 bool InitializeNode::stores_are_sane(PhaseTransform* phase) { | |
3728 if (is_complete()) | |
3729 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
|
3730 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
|
3731 intptr_t last_off = allocation()->minimum_header_size(); |
0 | 3732 for (uint i = InitializeNode::RawStores; i < req(); i++) { |
3733 Node* st = in(i); | |
3734 intptr_t st_off = get_store_offset(st, phase); | |
3735 if (st_off < 0) continue; // ignore dead garbage | |
3736 if (last_off > st_off) { | |
3737 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); | |
3738 this->dump(2); | |
3739 assert(false, "ascending store offsets"); | |
3740 return false; | |
3741 } | |
3742 last_off = st_off + st->as_Store()->memory_size(); | |
3743 } | |
3744 return true; | |
3745 } | |
3746 #endif //ASSERT | |
3747 | |
3748 | |
3749 | |
3750 | |
3751 //============================MergeMemNode===================================== | |
3752 // | |
3753 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several | |
3754 // contributing store or call operations. Each contributor provides the memory | |
3755 // state for a particular "alias type" (see Compile::alias_type). For example, | |
3756 // if a MergeMem has an input X for alias category #6, then any memory reference | |
3757 // to alias category #6 may use X as its memory state input, as an exact equivalent | |
3758 // to using the MergeMem as a whole. | |
3759 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) | |
3760 // | |
3761 // (Here, the <N> notation gives the index of the relevant adr_type.) | |
3762 // | |
3763 // In one special case (and more cases in the future), alias categories overlap. | |
3764 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory | |
3765 // states. Therefore, if a MergeMem has only one contributing input W for Bot, | |
3766 // it is exactly equivalent to that state W: | |
3767 // MergeMem(<Bot>: W) <==> W | |
3768 // | |
3769 // Usually, the merge has more than one input. In that case, where inputs | |
3770 // overlap (i.e., one is Bot), the narrower alias type determines the memory | |
3771 // state for that type, and the wider alias type (Bot) fills in everywhere else: | |
3772 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p) | |
3773 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p) | |
3774 // | |
3775 // A merge can take a "wide" memory state as one of its narrow inputs. | |
3776 // This simply means that the merge observes out only the relevant parts of | |
3777 // the wide input. That is, wide memory states arriving at narrow merge inputs | |
3778 // are implicitly "filtered" or "sliced" as necessary. (This is rare.) | |
3779 // | |
3780 // These rules imply that MergeMem nodes may cascade (via their <Bot> links), | |
3781 // and that memory slices "leak through": | |
3782 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y) | |
3783 // | |
3784 // But, in such a cascade, repeated memory slices can "block the leak": | |
3785 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y') | |
3786 // | |
3787 // In the last example, Y is not part of the combined memory state of the | |
3788 // outermost MergeMem. The system must, of course, prevent unschedulable | |
3789 // memory states from arising, so you can be sure that the state Y is somehow | |
3790 // a precursor to state Y'. | |
3791 // | |
3792 // | |
3793 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array | |
3794 // of each MergeMemNode array are exactly the numerical alias indexes, including | |
3795 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions | |
3796 // Compile::alias_type (and kin) produce and manage these indexes. | |
3797 // | |
3798 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. | |
3799 // (Note that this provides quick access to the top node inside MergeMem methods, | |
3800 // without the need to reach out via TLS to Compile::current.) | |
3801 // | |
3802 // As a consequence of what was just described, a MergeMem that represents a full | |
3803 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, | |
3804 // containing all alias categories. | |
3805 // | |
3806 // MergeMem nodes never (?) have control inputs, so in(0) is NULL. | |
3807 // | |
3808 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either | |
3809 // a memory state for the alias type <N>, or else the top node, meaning that | |
3810 // there is no particular input for that alias type. Note that the length of | |
3811 // a MergeMem is variable, and may be extended at any time to accommodate new | |
3812 // memory states at larger alias indexes. When merges grow, they are of course | |
3813 // filled with "top" in the unused in() positions. | |
3814 // | |
3815 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. | |
3816 // (Top was chosen because it works smoothly with passes like GCM.) | |
3817 // | |
3818 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is | |
3819 // the type of random VM bits like TLS references.) Since it is always the | |
3820 // first non-Bot memory slice, some low-level loops use it to initialize an | |
3821 // index variable: for (i = AliasIdxRaw; i < req(); i++). | |
3822 // | |
3823 // | |
3824 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns | |
3825 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns | |
3826 // the memory state for alias type <N>, or (if there is no particular slice at <N>, | |
3827 // it returns the base memory. To prevent bugs, memory_at does not accept <Top> | |
3828 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over | |
3829 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. | |
3830 // | |
3831 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't | |
3832 // really that different from the other memory inputs. An abbreviation called | |
3833 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. | |
3834 // | |
3835 // | |
3836 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent | |
3837 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi | |
3838 // that "emerges though" the base memory will be marked as excluding the alias types | |
3839 // of the other (narrow-memory) copies which "emerged through" the narrow edges: | |
3840 // | |
3841 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y)) | |
3842 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y)) | |
3843 // | |
3844 // This strange "subtraction" effect is necessary to ensure IGVN convergence. | |
3845 // (It is currently unimplemented.) As you can see, the resulting merge is | |
3846 // actually a disjoint union of memory states, rather than an overlay. | |
3847 // | |
3848 | |
3849 //------------------------------MergeMemNode----------------------------------- | |
3850 Node* MergeMemNode::make_empty_memory() { | |
3851 Node* empty_memory = (Node*) Compile::current()->top(); | |
3852 assert(empty_memory->is_top(), "correct sentinel identity"); | |
3853 return empty_memory; | |
3854 } | |
3855 | |
3856 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { | |
3857 init_class_id(Class_MergeMem); | |
3858 // all inputs are nullified in Node::Node(int) | |
3859 // set_input(0, NULL); // no control input | |
3860 | |
3861 // Initialize the edges uniformly to top, for starters. | |
3862 Node* empty_mem = make_empty_memory(); | |
3863 for (uint i = Compile::AliasIdxTop; i < req(); i++) { | |
3864 init_req(i,empty_mem); | |
3865 } | |
3866 assert(empty_memory() == empty_mem, ""); | |
3867 | |
3868 if( new_base != NULL && new_base->is_MergeMem() ) { | |
3869 MergeMemNode* mdef = new_base->as_MergeMem(); | |
3870 assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); | |
3871 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { | |
3872 mms.set_memory(mms.memory2()); | |
3873 } | |
3874 assert(base_memory() == mdef->base_memory(), ""); | |
3875 } else { | |
3876 set_base_memory(new_base); | |
3877 } | |
3878 } | |
3879 | |
3880 // Make a new, untransformed MergeMem with the same base as 'mem'. | |
3881 // If mem is itself a MergeMem, populate the result with the same edges. | |
3882 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { | |
3883 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); | |
3884 } | |
3885 | |
3886 //------------------------------cmp-------------------------------------------- | |
3887 uint MergeMemNode::hash() const { return NO_HASH; } | |
3888 uint MergeMemNode::cmp( const Node &n ) const { | |
3889 return (&n == this); // Always fail except on self | |
3890 } | |
3891 | |
3892 //------------------------------Identity--------------------------------------- | |
3893 Node* MergeMemNode::Identity(PhaseTransform *phase) { | |
3894 // Identity if this merge point does not record any interesting memory | |
3895 // disambiguations. | |
3896 Node* base_mem = base_memory(); | |
3897 Node* empty_mem = empty_memory(); | |
3898 if (base_mem != empty_mem) { // Memory path is not dead? | |
3899 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
3900 Node* mem = in(i); | |
3901 if (mem != empty_mem && mem != base_mem) { | |
3902 return this; // Many memory splits; no change | |
3903 } | |
3904 } | |
3905 } | |
3906 return base_mem; // No memory splits; ID on the one true input | |
3907 } | |
3908 | |
3909 //------------------------------Ideal------------------------------------------ | |
3910 // This method is invoked recursively on chains of MergeMem nodes | |
3911 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { | |
3912 // Remove chain'd MergeMems | |
3913 // | |
3914 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted | |
3915 // relative to the "in(Bot)". Since we are patching both at the same time, | |
3916 // we have to be careful to read each "in(i)" relative to the old "in(Bot)", | |
3917 // but rewrite each "in(i)" relative to the new "in(Bot)". | |
3918 Node *progress = NULL; | |
3919 | |
3920 | |
3921 Node* old_base = base_memory(); | |
3922 Node* empty_mem = empty_memory(); | |
3923 if (old_base == empty_mem) | |
3924 return NULL; // Dead memory path. | |
3925 | |
3926 MergeMemNode* old_mbase; | |
3927 if (old_base != NULL && old_base->is_MergeMem()) | |
3928 old_mbase = old_base->as_MergeMem(); | |
3929 else | |
3930 old_mbase = NULL; | |
3931 Node* new_base = old_base; | |
3932 | |
3933 // simplify stacked MergeMems in base memory | |
3934 if (old_mbase) new_base = old_mbase->base_memory(); | |
3935 | |
3936 // the base memory might contribute new slices beyond my req() | |
3937 if (old_mbase) grow_to_match(old_mbase); | |
3938 | |
3939 // Look carefully at the base node if it is a phi. | |
3940 PhiNode* phi_base; | |
3941 if (new_base != NULL && new_base->is_Phi()) | |
3942 phi_base = new_base->as_Phi(); | |
3943 else | |
3944 phi_base = NULL; | |
3945 | |
3946 Node* phi_reg = NULL; | |
3947 uint phi_len = (uint)-1; | |
3948 if (phi_base != NULL && !phi_base->is_copy()) { | |
3949 // do not examine phi if degraded to a copy | |
3950 phi_reg = phi_base->region(); | |
3951 phi_len = phi_base->req(); | |
3952 // see if the phi is unfinished | |
3953 for (uint i = 1; i < phi_len; i++) { | |
3954 if (phi_base->in(i) == NULL) { | |
3955 // incomplete phi; do not look at it yet! | |
3956 phi_reg = NULL; | |
3957 phi_len = (uint)-1; | |
3958 break; | |
3959 } | |
3960 } | |
3961 } | |
3962 | |
3963 // Note: We do not call verify_sparse on entry, because inputs | |
3964 // can normalize to the base_memory via subsume_node or similar | |
3965 // mechanisms. This method repairs that damage. | |
3966 | |
3967 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); | |
3968 | |
3969 // Look at each slice. | |
3970 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
3971 Node* old_in = in(i); | |
3972 // calculate the old memory value | |
3973 Node* old_mem = old_in; | |
3974 if (old_mem == empty_mem) old_mem = old_base; | |
3975 assert(old_mem == memory_at(i), ""); | |
3976 | |
3977 // maybe update (reslice) the old memory value | |
3978 | |
3979 // simplify stacked MergeMems | |
3980 Node* new_mem = old_mem; | |
3981 MergeMemNode* old_mmem; | |
3982 if (old_mem != NULL && old_mem->is_MergeMem()) | |
3983 old_mmem = old_mem->as_MergeMem(); | |
3984 else | |
3985 old_mmem = NULL; | |
3986 if (old_mmem == this) { | |
3987 // This can happen if loops break up and safepoints disappear. | |
3988 // A merge of BotPtr (default) with a RawPtr memory derived from a | |
3989 // safepoint can be rewritten to a merge of the same BotPtr with | |
3990 // the BotPtr phi coming into the loop. If that phi disappears | |
3991 // also, we can end up with a self-loop of the mergemem. | |
3992 // In general, if loops degenerate and memory effects disappear, | |
3993 // a mergemem can be left looking at itself. This simply means | |
3994 // that the mergemem's default should be used, since there is | |
3995 // no longer any apparent effect on this slice. | |
3996 // Note: If a memory slice is a MergeMem cycle, it is unreachable | |
3997 // from start. Update the input to TOP. | |
3998 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; | |
3999 } | |
4000 else if (old_mmem != NULL) { | |
4001 new_mem = old_mmem->memory_at(i); | |
4002 } | |
605 | 4003 // else preceding memory was not a MergeMem |
0 | 4004 |
4005 // replace equivalent phis (unfortunately, they do not GVN together) | |
4006 if (new_mem != NULL && new_mem != new_base && | |
4007 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { | |
4008 if (new_mem->is_Phi()) { | |
4009 PhiNode* phi_mem = new_mem->as_Phi(); | |
4010 for (uint i = 1; i < phi_len; i++) { | |
4011 if (phi_base->in(i) != phi_mem->in(i)) { | |
4012 phi_mem = NULL; | |
4013 break; | |
4014 } | |
4015 } | |
4016 if (phi_mem != NULL) { | |
4017 // equivalent phi nodes; revert to the def | |
4018 new_mem = new_base; | |
4019 } | |
4020 } | |
4021 } | |
4022 | |
4023 // maybe store down a new value | |
4024 Node* new_in = new_mem; | |
4025 if (new_in == new_base) new_in = empty_mem; | |
4026 | |
4027 if (new_in != old_in) { | |
4028 // Warning: Do not combine this "if" with the previous "if" | |
4029 // A memory slice might have be be rewritten even if it is semantically | |
4030 // unchanged, if the base_memory value has changed. | |
4031 set_req(i, new_in); | |
4032 progress = this; // Report progress | |
4033 } | |
4034 } | |
4035 | |
4036 if (new_base != old_base) { | |
4037 set_req(Compile::AliasIdxBot, new_base); | |
4038 // Don't use set_base_memory(new_base), because we need to update du. | |
4039 assert(base_memory() == new_base, ""); | |
4040 progress = this; | |
4041 } | |
4042 | |
4043 if( base_memory() == this ) { | |
4044 // a self cycle indicates this memory path is dead | |
4045 set_req(Compile::AliasIdxBot, empty_mem); | |
4046 } | |
4047 | |
4048 // Resolve external cycles by calling Ideal on a MergeMem base_memory | |
4049 // Recursion must occur after the self cycle check above | |
4050 if( base_memory()->is_MergeMem() ) { | |
4051 MergeMemNode *new_mbase = base_memory()->as_MergeMem(); | |
4052 Node *m = phase->transform(new_mbase); // Rollup any cycles | |
4053 if( m != NULL && (m->is_top() || | |
4054 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { | |
4055 // propagate rollup of dead cycle to self | |
4056 set_req(Compile::AliasIdxBot, empty_mem); | |
4057 } | |
4058 } | |
4059 | |
4060 if( base_memory() == empty_mem ) { | |
4061 progress = this; | |
4062 // Cut inputs during Parse phase only. | |
4063 // During Optimize phase a dead MergeMem node will be subsumed by Top. | |
4064 if( !can_reshape ) { | |
4065 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4066 if( in(i) != empty_mem ) { set_req(i, empty_mem); } | |
4067 } | |
4068 } | |
4069 } | |
4070 | |
4071 if( !progress && base_memory()->is_Phi() && can_reshape ) { | |
4072 // Check if PhiNode::Ideal's "Split phis through memory merges" | |
4073 // transform should be attempted. Look for this->phi->this cycle. | |
4074 uint merge_width = req(); | |
4075 if (merge_width > Compile::AliasIdxRaw) { | |
4076 PhiNode* phi = base_memory()->as_Phi(); | |
4077 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in | |
4078 if (phi->in(i) == this) { | |
4079 phase->is_IterGVN()->_worklist.push(phi); | |
4080 break; | |
4081 } | |
4082 } | |
4083 } | |
4084 } | |
4085 | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
43
diff
changeset
|
4086 assert(progress || verify_sparse(), "please, no dups of base"); |
0 | 4087 return progress; |
4088 } | |
4089 | |
4090 //-------------------------set_base_memory------------------------------------- | |
4091 void MergeMemNode::set_base_memory(Node *new_base) { | |
4092 Node* empty_mem = empty_memory(); | |
4093 set_req(Compile::AliasIdxBot, new_base); | |
4094 assert(memory_at(req()) == new_base, "must set default memory"); | |
4095 // Clear out other occurrences of new_base: | |
4096 if (new_base != empty_mem) { | |
4097 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4098 if (in(i) == new_base) set_req(i, empty_mem); | |
4099 } | |
4100 } | |
4101 } | |
4102 | |
4103 //------------------------------out_RegMask------------------------------------ | |
4104 const RegMask &MergeMemNode::out_RegMask() const { | |
4105 return RegMask::Empty; | |
4106 } | |
4107 | |
4108 //------------------------------dump_spec-------------------------------------- | |
4109 #ifndef PRODUCT | |
4110 void MergeMemNode::dump_spec(outputStream *st) const { | |
4111 st->print(" {"); | |
4112 Node* base_mem = base_memory(); | |
4113 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { | |
4114 Node* mem = memory_at(i); | |
4115 if (mem == base_mem) { st->print(" -"); continue; } | |
4116 st->print( " N%d:", mem->_idx ); | |
4117 Compile::current()->get_adr_type(i)->dump_on(st); | |
4118 } | |
4119 st->print(" }"); | |
4120 } | |
4121 #endif // !PRODUCT | |
4122 | |
4123 | |
4124 #ifdef ASSERT | |
4125 static bool might_be_same(Node* a, Node* b) { | |
4126 if (a == b) return true; | |
4127 if (!(a->is_Phi() || b->is_Phi())) return false; | |
4128 // phis shift around during optimization | |
4129 return true; // pretty stupid... | |
4130 } | |
4131 | |
4132 // verify a narrow slice (either incoming or outgoing) | |
4133 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { | |
4134 if (!VerifyAliases) return; // don't bother to verify unless requested | |
4135 if (is_error_reported()) return; // muzzle asserts when debugging an error | |
4136 if (Node::in_dump()) return; // muzzle asserts when printing | |
4137 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); | |
4138 assert(n != NULL, ""); | |
4139 // Elide intervening MergeMem's | |
4140 while (n->is_MergeMem()) { | |
4141 n = n->as_MergeMem()->memory_at(alias_idx); | |
4142 } | |
4143 Compile* C = Compile::current(); | |
4144 const TypePtr* n_adr_type = n->adr_type(); | |
4145 if (n == m->empty_memory()) { | |
4146 // Implicit copy of base_memory() | |
4147 } else if (n_adr_type != TypePtr::BOTTOM) { | |
4148 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); | |
4149 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); | |
4150 } else { | |
4151 // A few places like make_runtime_call "know" that VM calls are narrow, | |
4152 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. | |
4153 bool expected_wide_mem = false; | |
4154 if (n == m->base_memory()) { | |
4155 expected_wide_mem = true; | |
4156 } else if (alias_idx == Compile::AliasIdxRaw || | |
4157 n == m->memory_at(Compile::AliasIdxRaw)) { | |
4158 expected_wide_mem = true; | |
4159 } else if (!C->alias_type(alias_idx)->is_rewritable()) { | |
4160 // memory can "leak through" calls on channels that | |
4161 // are write-once. Allow this also. | |
4162 expected_wide_mem = true; | |
4163 } | |
4164 assert(expected_wide_mem, "expected narrow slice replacement"); | |
4165 } | |
4166 } | |
4167 #else // !ASSERT | |
4168 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op | |
4169 #endif | |
4170 | |
4171 | |
4172 //-----------------------------memory_at--------------------------------------- | |
4173 Node* MergeMemNode::memory_at(uint alias_idx) const { | |
4174 assert(alias_idx >= Compile::AliasIdxRaw || | |
4175 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, | |
4176 "must avoid base_memory and AliasIdxTop"); | |
4177 | |
4178 // Otherwise, it is a narrow slice. | |
4179 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); | |
4180 Compile *C = Compile::current(); | |
4181 if (is_empty_memory(n)) { | |
4182 // the array is sparse; empty slots are the "top" node | |
4183 n = base_memory(); | |
4184 assert(Node::in_dump() | |
4185 || 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
|
4186 || n->adr_type() == NULL // address is TOP |
0 | 4187 || n->adr_type() == TypePtr::BOTTOM |
4188 || n->adr_type() == TypeRawPtr::BOTTOM | |
4189 || Compile::current()->AliasLevel() == 0, | |
4190 "must be a wide memory"); | |
4191 // AliasLevel == 0 if we are organizing the memory states manually. | |
4192 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. | |
4193 } else { | |
4194 // make sure the stored slice is sane | |
4195 #ifdef ASSERT | |
4196 if (is_error_reported() || Node::in_dump()) { | |
4197 } else if (might_be_same(n, base_memory())) { | |
4198 // Give it a pass: It is a mostly harmless repetition of the base. | |
4199 // This can arise normally from node subsumption during optimization. | |
4200 } else { | |
4201 verify_memory_slice(this, alias_idx, n); | |
4202 } | |
4203 #endif | |
4204 } | |
4205 return n; | |
4206 } | |
4207 | |
4208 //---------------------------set_memory_at------------------------------------- | |
4209 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { | |
4210 verify_memory_slice(this, alias_idx, n); | |
4211 Node* empty_mem = empty_memory(); | |
4212 if (n == base_memory()) n = empty_mem; // collapse default | |
4213 uint need_req = alias_idx+1; | |
4214 if (req() < need_req) { | |
4215 if (n == empty_mem) return; // already the default, so do not grow me | |
4216 // grow the sparse array | |
4217 do { | |
4218 add_req(empty_mem); | |
4219 } while (req() < need_req); | |
4220 } | |
4221 set_req( alias_idx, n ); | |
4222 } | |
4223 | |
4224 | |
4225 | |
4226 //--------------------------iteration_setup------------------------------------ | |
4227 void MergeMemNode::iteration_setup(const MergeMemNode* other) { | |
4228 if (other != NULL) { | |
4229 grow_to_match(other); | |
4230 // invariant: the finite support of mm2 is within mm->req() | |
4231 #ifdef ASSERT | |
4232 for (uint i = req(); i < other->req(); i++) { | |
4233 assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); | |
4234 } | |
4235 #endif | |
4236 } | |
4237 // Replace spurious copies of base_memory by top. | |
4238 Node* base_mem = base_memory(); | |
4239 if (base_mem != NULL && !base_mem->is_top()) { | |
4240 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { | |
4241 if (in(i) == base_mem) | |
4242 set_req(i, empty_memory()); | |
4243 } | |
4244 } | |
4245 } | |
4246 | |
4247 //---------------------------grow_to_match------------------------------------- | |
4248 void MergeMemNode::grow_to_match(const MergeMemNode* other) { | |
4249 Node* empty_mem = empty_memory(); | |
4250 assert(other->is_empty_memory(empty_mem), "consistent sentinels"); | |
4251 // look for the finite support of the other memory | |
4252 for (uint i = other->req(); --i >= req(); ) { | |
4253 if (other->in(i) != empty_mem) { | |
4254 uint new_len = i+1; | |
4255 while (req() < new_len) add_req(empty_mem); | |
4256 break; | |
4257 } | |
4258 } | |
4259 } | |
4260 | |
4261 //---------------------------verify_sparse------------------------------------- | |
4262 #ifndef PRODUCT | |
4263 bool MergeMemNode::verify_sparse() const { | |
4264 assert(is_empty_memory(make_empty_memory()), "sane sentinel"); | |
4265 Node* base_mem = base_memory(); | |
4266 // The following can happen in degenerate cases, since empty==top. | |
4267 if (is_empty_memory(base_mem)) return true; | |
4268 for (uint i = Compile::AliasIdxRaw; i < req(); i++) { | |
4269 assert(in(i) != NULL, "sane slice"); | |
4270 if (in(i) == base_mem) return false; // should have been the sentinel value! | |
4271 } | |
4272 return true; | |
4273 } | |
4274 | |
4275 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { | |
4276 Node* n; | |
4277 n = mm->in(idx); | |
4278 if (mem == n) return true; // might be empty_memory() | |
4279 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); | |
4280 if (mem == n) return true; | |
4281 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { | |
4282 if (mem == n) return true; | |
4283 if (n == NULL) break; | |
4284 } | |
4285 return false; | |
4286 } | |
4287 #endif // !PRODUCT |