Mercurial > hg > graal-compiler
annotate src/share/vm/opto/memnode.hpp @ 164:c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
Summary: Add LoadNKlass and CMoveN nodes, use CmpN and ConN nodes to generate narrow oops compare instructions.
Reviewed-by: never, rasbold
author | kvn |
---|---|
date | Wed, 21 May 2008 13:46:23 -0700 |
parents | 885ed790ecf0 |
children | 9148c65abefc |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 // Portions of code courtesy of Clifford Click | |
26 | |
27 class MultiNode; | |
28 class PhaseCCP; | |
29 class PhaseTransform; | |
30 | |
31 //------------------------------MemNode---------------------------------------- | |
32 // Load or Store, possibly throwing a NULL pointer exception | |
33 class MemNode : public Node { | |
34 protected: | |
35 #ifdef ASSERT | |
36 const TypePtr* _adr_type; // What kind of memory is being addressed? | |
37 #endif | |
38 virtual uint size_of() const; // Size is bigger (ASSERT only) | |
39 public: | |
40 enum { Control, // When is it safe to do this load? | |
41 Memory, // Chunk of memory is being loaded from | |
42 Address, // Actually address, derived from base | |
43 ValueIn, // Value to store | |
44 OopStore // Preceeding oop store, only in StoreCM | |
45 }; | |
46 protected: | |
47 MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at ) | |
48 : Node(c0,c1,c2 ) { | |
49 init_class_id(Class_Mem); | |
50 debug_only(_adr_type=at; adr_type();) | |
51 } | |
52 MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at, Node *c3 ) | |
53 : Node(c0,c1,c2,c3) { | |
54 init_class_id(Class_Mem); | |
55 debug_only(_adr_type=at; adr_type();) | |
56 } | |
57 MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at, Node *c3, Node *c4) | |
58 : Node(c0,c1,c2,c3,c4) { | |
59 init_class_id(Class_Mem); | |
60 debug_only(_adr_type=at; adr_type();) | |
61 } | |
62 | |
33 | 63 public: |
0 | 64 // Helpers for the optimizer. Documented in memnode.cpp. |
65 static bool detect_ptr_independence(Node* p1, AllocateNode* a1, | |
66 Node* p2, AllocateNode* a2, | |
67 PhaseTransform* phase); | |
68 static bool adr_phi_is_loop_invariant(Node* adr_phi, Node* cast); | |
69 | |
74
2a9af0b9cb1c
6674600: (Escape Analysis) Optimize memory graph for instance's fields
kvn
parents:
64
diff
changeset
|
70 static Node *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:
64
diff
changeset
|
71 static Node *optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase); |
0 | 72 // This one should probably be a phase-specific function: |
85
f3b3fe64f59f
6692301: Side effect in NumberFormat tests with -server -Xcomp
kvn
parents:
74
diff
changeset
|
73 static bool all_controls_dominate(Node* dom, Node* sub); |
0 | 74 |
163 | 75 // Find any cast-away of null-ness and keep its control. |
76 static Node *Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ); | |
0 | 77 virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp ); |
78 | |
79 virtual const class TypePtr *adr_type() const; // returns bottom_type of address | |
80 | |
81 // Shared code for Ideal methods: | |
82 Node *Ideal_common(PhaseGVN *phase, bool can_reshape); // Return -1 for short-circuit NULL. | |
83 | |
84 // Helper function for adr_type() implementations. | |
85 static const TypePtr* calculate_adr_type(const Type* t, const TypePtr* cross_check = NULL); | |
86 | |
87 // Raw access function, to allow copying of adr_type efficiently in | |
88 // product builds and retain the debug info for debug builds. | |
89 const TypePtr *raw_adr_type() const { | |
90 #ifdef ASSERT | |
91 return _adr_type; | |
92 #else | |
93 return 0; | |
94 #endif | |
95 } | |
96 | |
97 // Map a load or store opcode to its corresponding store opcode. | |
98 // (Return -1 if unknown.) | |
99 virtual int store_Opcode() const { return -1; } | |
100 | |
101 // What is the type of the value in memory? (T_VOID mean "unspecified".) | |
102 virtual BasicType memory_type() const = 0; | |
29
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
103 virtual int memory_size() const { |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
104 #ifdef ASSERT |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
105 return type2aelembytes(memory_type(), true); |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
106 #else |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
107 return type2aelembytes(memory_type()); |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
108 #endif |
d5fc211aea19
6633953: type2aelembytes{T_ADDRESS} should be 8 bytes in 64 bit VM
kvn
parents:
17
diff
changeset
|
109 } |
0 | 110 |
111 // Search through memory states which precede this node (load or store). | |
112 // Look for an exact match for the address, with no intervening | |
113 // aliased stores. | |
114 Node* find_previous_store(PhaseTransform* phase); | |
115 | |
116 // Can this node (load or store) accurately see a stored value in | |
117 // the given memory state? (The state may or may not be in(Memory).) | |
118 Node* can_see_stored_value(Node* st, PhaseTransform* phase) const; | |
119 | |
120 #ifndef PRODUCT | |
121 static void dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st); | |
122 virtual void dump_spec(outputStream *st) const; | |
123 #endif | |
124 }; | |
125 | |
126 //------------------------------LoadNode--------------------------------------- | |
127 // Load value; requires Memory and Address | |
128 class LoadNode : public MemNode { | |
129 protected: | |
130 virtual uint cmp( const Node &n ) const; | |
131 virtual uint size_of() const; // Size is bigger | |
132 const Type* const _type; // What kind of value is loaded? | |
133 public: | |
134 | |
135 LoadNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt ) | |
136 : MemNode(c,mem,adr,at), _type(rt) { | |
137 init_class_id(Class_Load); | |
138 } | |
139 | |
140 // Polymorphic factory method: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
141 static Node* make( PhaseGVN& gvn, Node *c, Node *mem, Node *adr, |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
142 const TypePtr* at, const Type *rt, BasicType bt ); |
0 | 143 |
144 virtual uint hash() const; // Check the type | |
145 | |
146 // Handle algebraic identities here. If we have an identity, return the Node | |
147 // we are equivalent to. We look for Load of a Store. | |
148 virtual Node *Identity( PhaseTransform *phase ); | |
149 | |
150 // If the load is from Field memory and the pointer is non-null, we can | |
151 // zero out the control input. | |
152 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
153 | |
163 | 154 // Split instance field load through Phi. |
155 Node* split_through_phi(PhaseGVN *phase); | |
156 | |
17
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
157 // Recover original value from boxed values |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
158 Node *eliminate_autobox(PhaseGVN *phase); |
ff5961f4c095
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
never
parents:
0
diff
changeset
|
159 |
0 | 160 // Compute a new Type for this node. Basically we just do the pre-check, |
161 // then call the virtual add() to set the type. | |
162 virtual const Type *Value( PhaseTransform *phase ) const; | |
163 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
164 // Common methods for LoadKlass and LoadNKlass nodes. |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
165 const Type *klass_value_common( PhaseTransform *phase ) const; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
166 Node *klass_identity_common( PhaseTransform *phase ); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
167 |
0 | 168 virtual uint ideal_reg() const; |
169 virtual const Type *bottom_type() const; | |
170 // Following method is copied from TypeNode: | |
171 void set_type(const Type* t) { | |
172 assert(t != NULL, "sanity"); | |
173 debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH); | |
174 *(const Type**)&_type = t; // cast away const-ness | |
175 // If this node is in the hash table, make sure it doesn't need a rehash. | |
176 assert(check_hash == NO_HASH || check_hash == hash(), "type change must preserve hash code"); | |
177 } | |
178 const Type* type() const { assert(_type != NULL, "sanity"); return _type; }; | |
179 | |
180 // Do not match memory edge | |
181 virtual uint match_edge(uint idx) const; | |
182 | |
183 // Map a load opcode to its corresponding store opcode. | |
184 virtual int store_Opcode() const = 0; | |
185 | |
64
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
33
diff
changeset
|
186 // Check if the load's memory input is a Phi node with the same control. |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
33
diff
changeset
|
187 bool is_instance_field_load_with_local_phi(Node* ctrl); |
b8f5ba577b02
6673473: (Escape Analysis) Add the instance's field information to PhiNode
kvn
parents:
33
diff
changeset
|
188 |
0 | 189 #ifndef PRODUCT |
190 virtual void dump_spec(outputStream *st) const; | |
191 #endif | |
192 protected: | |
193 const Type* load_array_final_field(const TypeKlassPtr *tkls, | |
194 ciKlass* klass) const; | |
195 }; | |
196 | |
197 //------------------------------LoadBNode-------------------------------------- | |
198 // Load a byte (8bits signed) from memory | |
199 class LoadBNode : public LoadNode { | |
200 public: | |
201 LoadBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::BYTE ) | |
202 : LoadNode(c,mem,adr,at,ti) {} | |
203 virtual int Opcode() const; | |
204 virtual uint ideal_reg() const { return Op_RegI; } | |
205 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
206 virtual int store_Opcode() const { return Op_StoreB; } | |
207 virtual BasicType memory_type() const { return T_BYTE; } | |
208 }; | |
209 | |
210 //------------------------------LoadCNode-------------------------------------- | |
211 // Load a char (16bits unsigned) from memory | |
212 class LoadCNode : public LoadNode { | |
213 public: | |
214 LoadCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::CHAR ) | |
215 : LoadNode(c,mem,adr,at,ti) {} | |
216 virtual int Opcode() const; | |
217 virtual uint ideal_reg() const { return Op_RegI; } | |
218 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
219 virtual int store_Opcode() const { return Op_StoreC; } | |
220 virtual BasicType memory_type() const { return T_CHAR; } | |
221 }; | |
222 | |
223 //------------------------------LoadINode-------------------------------------- | |
224 // Load an integer from memory | |
225 class LoadINode : public LoadNode { | |
226 public: | |
227 LoadINode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::INT ) | |
228 : LoadNode(c,mem,adr,at,ti) {} | |
229 virtual int Opcode() const; | |
230 virtual uint ideal_reg() const { return Op_RegI; } | |
231 virtual int store_Opcode() const { return Op_StoreI; } | |
232 virtual BasicType memory_type() const { return T_INT; } | |
233 }; | |
234 | |
235 //------------------------------LoadRangeNode---------------------------------- | |
236 // Load an array length from the array | |
237 class LoadRangeNode : public LoadINode { | |
238 public: | |
239 LoadRangeNode( Node *c, Node *mem, Node *adr, const TypeInt *ti = TypeInt::POS ) | |
240 : LoadINode(c,mem,adr,TypeAryPtr::RANGE,ti) {} | |
241 virtual int Opcode() const; | |
242 virtual const Type *Value( PhaseTransform *phase ) const; | |
243 virtual Node *Identity( PhaseTransform *phase ); | |
244 }; | |
245 | |
246 //------------------------------LoadLNode-------------------------------------- | |
247 // Load a long from memory | |
248 class LoadLNode : public LoadNode { | |
249 virtual uint hash() const { return LoadNode::hash() + _require_atomic_access; } | |
250 virtual uint cmp( const Node &n ) const { | |
251 return _require_atomic_access == ((LoadLNode&)n)._require_atomic_access | |
252 && LoadNode::cmp(n); | |
253 } | |
254 virtual uint size_of() const { return sizeof(*this); } | |
255 const bool _require_atomic_access; // is piecewise load forbidden? | |
256 | |
257 public: | |
258 LoadLNode( Node *c, Node *mem, Node *adr, const TypePtr* at, | |
259 const TypeLong *tl = TypeLong::LONG, | |
260 bool require_atomic_access = false ) | |
261 : LoadNode(c,mem,adr,at,tl) | |
262 , _require_atomic_access(require_atomic_access) | |
263 {} | |
264 virtual int Opcode() const; | |
265 virtual uint ideal_reg() const { return Op_RegL; } | |
266 virtual int store_Opcode() const { return Op_StoreL; } | |
267 virtual BasicType memory_type() const { return T_LONG; } | |
268 bool require_atomic_access() { return _require_atomic_access; } | |
269 static LoadLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt); | |
270 #ifndef PRODUCT | |
271 virtual void dump_spec(outputStream *st) const { | |
272 LoadNode::dump_spec(st); | |
273 if (_require_atomic_access) st->print(" Atomic!"); | |
274 } | |
275 #endif | |
276 }; | |
277 | |
278 //------------------------------LoadL_unalignedNode---------------------------- | |
279 // Load a long from unaligned memory | |
280 class LoadL_unalignedNode : public LoadLNode { | |
281 public: | |
282 LoadL_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at ) | |
283 : LoadLNode(c,mem,adr,at) {} | |
284 virtual int Opcode() const; | |
285 }; | |
286 | |
287 //------------------------------LoadFNode-------------------------------------- | |
288 // Load a float (64 bits) from memory | |
289 class LoadFNode : public LoadNode { | |
290 public: | |
291 LoadFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::FLOAT ) | |
292 : LoadNode(c,mem,adr,at,t) {} | |
293 virtual int Opcode() const; | |
294 virtual uint ideal_reg() const { return Op_RegF; } | |
295 virtual int store_Opcode() const { return Op_StoreF; } | |
296 virtual BasicType memory_type() const { return T_FLOAT; } | |
297 }; | |
298 | |
299 //------------------------------LoadDNode-------------------------------------- | |
300 // Load a double (64 bits) from memory | |
301 class LoadDNode : public LoadNode { | |
302 public: | |
303 LoadDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::DOUBLE ) | |
304 : LoadNode(c,mem,adr,at,t) {} | |
305 virtual int Opcode() const; | |
306 virtual uint ideal_reg() const { return Op_RegD; } | |
307 virtual int store_Opcode() const { return Op_StoreD; } | |
308 virtual BasicType memory_type() const { return T_DOUBLE; } | |
309 }; | |
310 | |
311 //------------------------------LoadD_unalignedNode---------------------------- | |
312 // Load a double from unaligned memory | |
313 class LoadD_unalignedNode : public LoadDNode { | |
314 public: | |
315 LoadD_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at ) | |
316 : LoadDNode(c,mem,adr,at) {} | |
317 virtual int Opcode() const; | |
318 }; | |
319 | |
320 //------------------------------LoadPNode-------------------------------------- | |
321 // Load a pointer from memory (either object or array) | |
322 class LoadPNode : public LoadNode { | |
323 public: | |
324 LoadPNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypePtr* t ) | |
325 : LoadNode(c,mem,adr,at,t) {} | |
326 virtual int Opcode() const; | |
327 virtual uint ideal_reg() const { return Op_RegP; } | |
328 virtual int store_Opcode() const { return Op_StoreP; } | |
329 virtual BasicType memory_type() const { return T_ADDRESS; } | |
330 // depends_only_on_test is almost always true, and needs to be almost always | |
331 // true to enable key hoisting & commoning optimizations. However, for the | |
332 // special case of RawPtr loads from TLS top & end, the control edge carries | |
333 // the dependence preventing hoisting past a Safepoint instead of the memory | |
334 // edge. (An unfortunate consequence of having Safepoints not set Raw | |
335 // Memory; itself an unfortunate consequence of having Nodes which produce | |
336 // results (new raw memory state) inside of loops preventing all manner of | |
337 // other optimizations). Basically, it's ugly but so is the alternative. | |
338 // See comment in macro.cpp, around line 125 expand_allocate_common(). | |
339 virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; } | |
340 }; | |
341 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
342 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
343 //------------------------------LoadNNode-------------------------------------- |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
344 // Load a narrow oop from memory (either object or array) |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
345 class LoadNNode : public LoadNode { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
346 public: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
347 LoadNNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const Type* t ) |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
348 : LoadNode(c,mem,adr,at,t) {} |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
349 virtual int Opcode() const; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
350 virtual uint ideal_reg() const { return Op_RegN; } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
351 virtual int store_Opcode() const { return Op_StoreN; } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
352 virtual BasicType memory_type() const { return T_NARROWOOP; } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
353 // depends_only_on_test is almost always true, and needs to be almost always |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
354 // true to enable key hoisting & commoning optimizations. However, for the |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
355 // special case of RawPtr loads from TLS top & end, the control edge carries |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
356 // the dependence preventing hoisting past a Safepoint instead of the memory |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
357 // edge. (An unfortunate consequence of having Safepoints not set Raw |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
358 // Memory; itself an unfortunate consequence of having Nodes which produce |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
359 // results (new raw memory state) inside of loops preventing all manner of |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
360 // other optimizations). Basically, it's ugly but so is the alternative. |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
361 // See comment in macro.cpp, around line 125 expand_allocate_common(). |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
362 virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
363 }; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
364 |
0 | 365 //------------------------------LoadKlassNode---------------------------------- |
366 // Load a Klass from an object | |
367 class LoadKlassNode : public LoadPNode { | |
368 public: | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
369 LoadKlassNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeKlassPtr *tk ) |
0 | 370 : LoadPNode(c,mem,adr,at,tk) {} |
371 virtual int Opcode() const; | |
372 virtual const Type *Value( PhaseTransform *phase ) const; | |
373 virtual Node *Identity( PhaseTransform *phase ); | |
374 virtual bool depends_only_on_test() const { return true; } | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
375 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
376 // Polymorphic factory method: |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
377 static Node* make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
378 const TypeKlassPtr *tk = TypeKlassPtr::OBJECT ); |
0 | 379 }; |
380 | |
164
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
381 //------------------------------LoadNKlassNode--------------------------------- |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
382 // Load a narrow Klass from an object. |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
383 class LoadNKlassNode : public LoadNNode { |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
384 public: |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
385 LoadNKlassNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeNarrowOop *tk ) |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
386 : LoadNNode(c,mem,adr,at,tk) {} |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
387 virtual int Opcode() const; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
388 virtual uint ideal_reg() const { return Op_RegN; } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
389 virtual int store_Opcode() const { return Op_StoreN; } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
390 virtual BasicType memory_type() const { return T_NARROWOOP; } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
391 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
392 virtual const Type *Value( PhaseTransform *phase ) const; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
393 virtual Node *Identity( PhaseTransform *phase ); |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
394 virtual bool depends_only_on_test() const { return true; } |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
395 }; |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
396 |
c436414a719e
6703890: Compressed Oops: add LoadNKlass node to generate narrow oops (32-bits) compare instructions
kvn
parents:
163
diff
changeset
|
397 |
0 | 398 //------------------------------LoadSNode-------------------------------------- |
399 // Load a short (16bits signed) from memory | |
400 class LoadSNode : public LoadNode { | |
401 public: | |
402 LoadSNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::SHORT ) | |
403 : LoadNode(c,mem,adr,at,ti) {} | |
404 virtual int Opcode() const; | |
405 virtual uint ideal_reg() const { return Op_RegI; } | |
406 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
407 virtual int store_Opcode() const { return Op_StoreC; } | |
408 virtual BasicType memory_type() const { return T_SHORT; } | |
409 }; | |
410 | |
411 //------------------------------StoreNode-------------------------------------- | |
412 // Store value; requires Store, Address and Value | |
413 class StoreNode : public MemNode { | |
414 protected: | |
415 virtual uint cmp( const Node &n ) const; | |
416 virtual bool depends_only_on_test() const { return false; } | |
417 | |
418 Node *Ideal_masked_input (PhaseGVN *phase, uint mask); | |
419 Node *Ideal_sign_extended_input(PhaseGVN *phase, int num_bits); | |
420 | |
421 public: | |
422 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) | |
423 : MemNode(c,mem,adr,at,val) { | |
424 init_class_id(Class_Store); | |
425 } | |
426 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store ) | |
427 : MemNode(c,mem,adr,at,val,oop_store) { | |
428 init_class_id(Class_Store); | |
429 } | |
430 | |
431 // Polymorphic factory method: | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
432 static StoreNode* make( PhaseGVN& gvn, Node *c, Node *mem, Node *adr, |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
433 const TypePtr* at, Node *val, BasicType bt ); |
0 | 434 |
435 virtual uint hash() const; // Check the type | |
436 | |
437 // If the store is to Field memory and the pointer is non-null, we can | |
438 // zero out the control input. | |
439 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
440 | |
441 // Compute a new Type for this node. Basically we just do the pre-check, | |
442 // then call the virtual add() to set the type. | |
443 virtual const Type *Value( PhaseTransform *phase ) const; | |
444 | |
445 // Check for identity function on memory (Load then Store at same address) | |
446 virtual Node *Identity( PhaseTransform *phase ); | |
447 | |
448 // Do not match memory edge | |
449 virtual uint match_edge(uint idx) const; | |
450 | |
451 virtual const Type *bottom_type() const; // returns Type::MEMORY | |
452 | |
453 // Map a store opcode to its corresponding own opcode, trivially. | |
454 virtual int store_Opcode() const { return Opcode(); } | |
455 | |
456 // have all possible loads of the value stored been optimized away? | |
457 bool value_never_loaded(PhaseTransform *phase) const; | |
458 }; | |
459 | |
460 //------------------------------StoreBNode------------------------------------- | |
461 // Store byte to memory | |
462 class StoreBNode : public StoreNode { | |
463 public: | |
464 StoreBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
465 virtual int Opcode() const; | |
466 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
467 virtual BasicType memory_type() const { return T_BYTE; } | |
468 }; | |
469 | |
470 //------------------------------StoreCNode------------------------------------- | |
471 // Store char/short to memory | |
472 class StoreCNode : public StoreNode { | |
473 public: | |
474 StoreCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
475 virtual int Opcode() const; | |
476 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
477 virtual BasicType memory_type() const { return T_CHAR; } | |
478 }; | |
479 | |
480 //------------------------------StoreINode------------------------------------- | |
481 // Store int to memory | |
482 class StoreINode : public StoreNode { | |
483 public: | |
484 StoreINode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
485 virtual int Opcode() const; | |
486 virtual BasicType memory_type() const { return T_INT; } | |
487 }; | |
488 | |
489 //------------------------------StoreLNode------------------------------------- | |
490 // Store long to memory | |
491 class StoreLNode : public StoreNode { | |
492 virtual uint hash() const { return StoreNode::hash() + _require_atomic_access; } | |
493 virtual uint cmp( const Node &n ) const { | |
494 return _require_atomic_access == ((StoreLNode&)n)._require_atomic_access | |
495 && StoreNode::cmp(n); | |
496 } | |
497 virtual uint size_of() const { return sizeof(*this); } | |
498 const bool _require_atomic_access; // is piecewise store forbidden? | |
499 | |
500 public: | |
501 StoreLNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, | |
502 bool require_atomic_access = false ) | |
503 : StoreNode(c,mem,adr,at,val) | |
504 , _require_atomic_access(require_atomic_access) | |
505 {} | |
506 virtual int Opcode() const; | |
507 virtual BasicType memory_type() const { return T_LONG; } | |
508 bool require_atomic_access() { return _require_atomic_access; } | |
509 static StoreLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val); | |
510 #ifndef PRODUCT | |
511 virtual void dump_spec(outputStream *st) const { | |
512 StoreNode::dump_spec(st); | |
513 if (_require_atomic_access) st->print(" Atomic!"); | |
514 } | |
515 #endif | |
516 }; | |
517 | |
518 //------------------------------StoreFNode------------------------------------- | |
519 // Store float to memory | |
520 class StoreFNode : public StoreNode { | |
521 public: | |
522 StoreFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
523 virtual int Opcode() const; | |
524 virtual BasicType memory_type() const { return T_FLOAT; } | |
525 }; | |
526 | |
527 //------------------------------StoreDNode------------------------------------- | |
528 // Store double to memory | |
529 class StoreDNode : public StoreNode { | |
530 public: | |
531 StoreDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
532 virtual int Opcode() const; | |
533 virtual BasicType memory_type() const { return T_DOUBLE; } | |
534 }; | |
535 | |
536 //------------------------------StorePNode------------------------------------- | |
537 // Store pointer to memory | |
538 class StorePNode : public StoreNode { | |
539 public: | |
540 StorePNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} | |
541 virtual int Opcode() const; | |
542 virtual BasicType memory_type() const { return T_ADDRESS; } | |
543 }; | |
544 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
545 //------------------------------StoreNNode------------------------------------- |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
546 // Store narrow oop to memory |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
547 class StoreNNode : public StoreNode { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
548 public: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
549 StoreNNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {} |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
550 virtual int Opcode() const; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
551 virtual BasicType memory_type() const { return T_NARROWOOP; } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
552 }; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
553 |
0 | 554 //------------------------------StoreCMNode----------------------------------- |
555 // Store card-mark byte to memory for CM | |
556 // The last StoreCM before a SafePoint must be preserved and occur after its "oop" store | |
557 // Preceeding equivalent StoreCMs may be eliminated. | |
558 class StoreCMNode : public StoreNode { | |
559 public: | |
560 StoreCMNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store ) : StoreNode(c,mem,adr,at,val,oop_store) {} | |
561 virtual int Opcode() const; | |
562 virtual Node *Identity( PhaseTransform *phase ); | |
563 virtual const Type *Value( PhaseTransform *phase ) const; | |
564 virtual BasicType memory_type() const { return T_VOID; } // unspecific | |
565 }; | |
566 | |
567 //------------------------------LoadPLockedNode--------------------------------- | |
568 // Load-locked a pointer from memory (either object or array). | |
569 // On Sparc & Intel this is implemented as a normal pointer load. | |
570 // On PowerPC and friends it's a real load-locked. | |
571 class LoadPLockedNode : public LoadPNode { | |
572 public: | |
573 LoadPLockedNode( Node *c, Node *mem, Node *adr ) | |
574 : LoadPNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) {} | |
575 virtual int Opcode() const; | |
576 virtual int store_Opcode() const { return Op_StorePConditional; } | |
577 virtual bool depends_only_on_test() const { return true; } | |
578 }; | |
579 | |
580 //------------------------------LoadLLockedNode--------------------------------- | |
581 // Load-locked a pointer from memory (either object or array). | |
582 // On Sparc & Intel this is implemented as a normal long load. | |
583 class LoadLLockedNode : public LoadLNode { | |
584 public: | |
585 LoadLLockedNode( Node *c, Node *mem, Node *adr ) | |
586 : LoadLNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeLong::LONG) {} | |
587 virtual int Opcode() const; | |
588 virtual int store_Opcode() const { return Op_StoreLConditional; } | |
589 }; | |
590 | |
591 //------------------------------SCMemProjNode--------------------------------------- | |
592 // This class defines a projection of the memory state of a store conditional node. | |
593 // These nodes return a value, but also update memory. | |
594 class SCMemProjNode : public ProjNode { | |
595 public: | |
596 enum {SCMEMPROJCON = (uint)-2}; | |
597 SCMemProjNode( Node *src) : ProjNode( src, SCMEMPROJCON) { } | |
598 virtual int Opcode() const; | |
599 virtual bool is_CFG() const { return false; } | |
600 virtual const Type *bottom_type() const {return Type::MEMORY;} | |
601 virtual const TypePtr *adr_type() const { return in(0)->in(MemNode::Memory)->adr_type();} | |
602 virtual uint ideal_reg() const { return 0;} // memory projections don't have a register | |
603 virtual const Type *Value( PhaseTransform *phase ) const; | |
604 #ifndef PRODUCT | |
605 virtual void dump_spec(outputStream *st) const {}; | |
606 #endif | |
607 }; | |
608 | |
609 //------------------------------LoadStoreNode--------------------------- | |
610 class LoadStoreNode : public Node { | |
611 public: | |
612 enum { | |
613 ExpectedIn = MemNode::ValueIn+1 // One more input than MemNode | |
614 }; | |
615 LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex); | |
616 virtual bool depends_only_on_test() const { return false; } | |
617 virtual const Type *bottom_type() const { return TypeInt::BOOL; } | |
618 virtual uint ideal_reg() const { return Op_RegI; } | |
619 virtual uint match_edge(uint idx) const { return idx == MemNode::Address || idx == MemNode::ValueIn; } | |
620 }; | |
621 | |
622 //------------------------------StorePConditionalNode--------------------------- | |
623 // Conditionally store pointer to memory, if no change since prior | |
624 // load-locked. Sets flags for success or failure of the store. | |
625 class StorePConditionalNode : public LoadStoreNode { | |
626 public: | |
627 StorePConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { } | |
628 virtual int Opcode() const; | |
629 // Produces flags | |
630 virtual uint ideal_reg() const { return Op_RegFlags; } | |
631 }; | |
632 | |
633 //------------------------------StoreLConditionalNode--------------------------- | |
634 // Conditionally store long to memory, if no change since prior | |
635 // load-locked. Sets flags for success or failure of the store. | |
636 class StoreLConditionalNode : public LoadStoreNode { | |
637 public: | |
638 StoreLConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { } | |
639 virtual int Opcode() const; | |
640 }; | |
641 | |
642 | |
643 //------------------------------CompareAndSwapLNode--------------------------- | |
644 class CompareAndSwapLNode : public LoadStoreNode { | |
645 public: | |
646 CompareAndSwapLNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { } | |
647 virtual int Opcode() const; | |
648 }; | |
649 | |
650 | |
651 //------------------------------CompareAndSwapINode--------------------------- | |
652 class CompareAndSwapINode : public LoadStoreNode { | |
653 public: | |
654 CompareAndSwapINode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { } | |
655 virtual int Opcode() const; | |
656 }; | |
657 | |
658 | |
659 //------------------------------CompareAndSwapPNode--------------------------- | |
660 class CompareAndSwapPNode : public LoadStoreNode { | |
661 public: | |
662 CompareAndSwapPNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { } | |
663 virtual int Opcode() const; | |
664 }; | |
665 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
666 //------------------------------CompareAndSwapNNode--------------------------- |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
667 class CompareAndSwapNNode : public LoadStoreNode { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
668 public: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
669 CompareAndSwapNNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
670 virtual int Opcode() const; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
671 }; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
74
diff
changeset
|
672 |
0 | 673 //------------------------------ClearArray------------------------------------- |
674 class ClearArrayNode: public Node { | |
675 public: | |
676 ClearArrayNode( Node *ctrl, Node *arymem, Node *word_cnt, Node *base ) : Node(ctrl,arymem,word_cnt,base) {} | |
677 virtual int Opcode() const; | |
678 virtual const Type *bottom_type() const { return Type::MEMORY; } | |
679 // ClearArray modifies array elements, and so affects only the | |
680 // array memory addressed by the bottom_type of its base address. | |
681 virtual const class TypePtr *adr_type() const; | |
682 virtual Node *Identity( PhaseTransform *phase ); | |
683 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
684 virtual uint match_edge(uint idx) const; | |
685 | |
686 // Clear the given area of an object or array. | |
687 // The start offset must always be aligned mod BytesPerInt. | |
688 // The end offset must always be aligned mod BytesPerLong. | |
689 // Return the new memory. | |
690 static Node* clear_memory(Node* control, Node* mem, Node* dest, | |
691 intptr_t start_offset, | |
692 intptr_t end_offset, | |
693 PhaseGVN* phase); | |
694 static Node* clear_memory(Node* control, Node* mem, Node* dest, | |
695 intptr_t start_offset, | |
696 Node* end_offset, | |
697 PhaseGVN* phase); | |
698 static Node* clear_memory(Node* control, Node* mem, Node* dest, | |
699 Node* start_offset, | |
700 Node* end_offset, | |
701 PhaseGVN* phase); | |
702 }; | |
703 | |
704 //------------------------------StrComp------------------------------------- | |
705 class StrCompNode: public Node { | |
706 public: | |
707 StrCompNode(Node *control, | |
708 Node* char_array_mem, | |
709 Node* value_mem, | |
710 Node* count_mem, | |
711 Node* offset_mem, | |
712 Node* s1, Node* s2): Node(control, | |
713 char_array_mem, | |
714 value_mem, | |
715 count_mem, | |
716 offset_mem, | |
717 s1, s2) {}; | |
718 virtual int Opcode() const; | |
719 virtual bool depends_only_on_test() const { return false; } | |
720 virtual const Type* bottom_type() const { return TypeInt::INT; } | |
721 // a StrCompNode (conservatively) aliases with everything: | |
722 virtual const TypePtr* adr_type() const { return TypePtr::BOTTOM; } | |
723 virtual uint match_edge(uint idx) const; | |
724 virtual uint ideal_reg() const { return Op_RegI; } | |
725 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
726 }; | |
727 | |
728 //------------------------------MemBar----------------------------------------- | |
729 // There are different flavors of Memory Barriers to match the Java Memory | |
730 // Model. Monitor-enter and volatile-load act as Aquires: no following ref | |
731 // can be moved to before them. We insert a MemBar-Acquire after a FastLock or | |
732 // volatile-load. Monitor-exit and volatile-store act as Release: no | |
733 // preceeding ref can be moved to after them. We insert a MemBar-Release | |
734 // before a FastUnlock or volatile-store. All volatiles need to be | |
735 // serialized, so we follow all volatile-stores with a MemBar-Volatile to | |
736 // seperate it from any following volatile-load. | |
737 class MemBarNode: public MultiNode { | |
738 virtual uint hash() const ; // { return NO_HASH; } | |
739 virtual uint cmp( const Node &n ) const ; // Always fail, except on self | |
740 | |
741 virtual uint size_of() const { return sizeof(*this); } | |
742 // Memory type this node is serializing. Usually either rawptr or bottom. | |
743 const TypePtr* _adr_type; | |
744 | |
745 public: | |
746 enum { | |
747 Precedent = TypeFunc::Parms // optional edge to force precedence | |
748 }; | |
749 MemBarNode(Compile* C, int alias_idx, Node* precedent); | |
750 virtual int Opcode() const = 0; | |
751 virtual const class TypePtr *adr_type() const { return _adr_type; } | |
752 virtual const Type *Value( PhaseTransform *phase ) const; | |
753 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
754 virtual uint match_edge(uint idx) const { return 0; } | |
755 virtual const Type *bottom_type() const { return TypeTuple::MEMBAR; } | |
756 virtual Node *match( const ProjNode *proj, const Matcher *m ); | |
757 // Factory method. Builds a wide or narrow membar. | |
758 // Optional 'precedent' becomes an extra edge if not null. | |
759 static MemBarNode* make(Compile* C, int opcode, | |
760 int alias_idx = Compile::AliasIdxBot, | |
761 Node* precedent = NULL); | |
762 }; | |
763 | |
764 // "Acquire" - no following ref can move before (but earlier refs can | |
765 // follow, like an early Load stalled in cache). Requires multi-cpu | |
766 // visibility. Inserted after a volatile load or FastLock. | |
767 class MemBarAcquireNode: public MemBarNode { | |
768 public: | |
769 MemBarAcquireNode(Compile* C, int alias_idx, Node* precedent) | |
770 : MemBarNode(C, alias_idx, precedent) {} | |
771 virtual int Opcode() const; | |
772 }; | |
773 | |
774 // "Release" - no earlier ref can move after (but later refs can move | |
775 // up, like a speculative pipelined cache-hitting Load). Requires | |
776 // multi-cpu visibility. Inserted before a volatile store or FastUnLock. | |
777 class MemBarReleaseNode: public MemBarNode { | |
778 public: | |
779 MemBarReleaseNode(Compile* C, int alias_idx, Node* precedent) | |
780 : MemBarNode(C, alias_idx, precedent) {} | |
781 virtual int Opcode() const; | |
782 }; | |
783 | |
784 // Ordering between a volatile store and a following volatile load. | |
785 // Requires multi-CPU visibility? | |
786 class MemBarVolatileNode: public MemBarNode { | |
787 public: | |
788 MemBarVolatileNode(Compile* C, int alias_idx, Node* precedent) | |
789 : MemBarNode(C, alias_idx, precedent) {} | |
790 virtual int Opcode() const; | |
791 }; | |
792 | |
793 // Ordering within the same CPU. Used to order unsafe memory references | |
794 // inside the compiler when we lack alias info. Not needed "outside" the | |
795 // compiler because the CPU does all the ordering for us. | |
796 class MemBarCPUOrderNode: public MemBarNode { | |
797 public: | |
798 MemBarCPUOrderNode(Compile* C, int alias_idx, Node* precedent) | |
799 : MemBarNode(C, alias_idx, precedent) {} | |
800 virtual int Opcode() const; | |
801 virtual uint ideal_reg() const { return 0; } // not matched in the AD file | |
802 }; | |
803 | |
804 // Isolation of object setup after an AllocateNode and before next safepoint. | |
805 // (See comment in memnode.cpp near InitializeNode::InitializeNode for semantics.) | |
806 class InitializeNode: public MemBarNode { | |
807 friend class AllocateNode; | |
808 | |
809 bool _is_complete; | |
810 | |
811 public: | |
812 enum { | |
813 Control = TypeFunc::Control, | |
814 Memory = TypeFunc::Memory, // MergeMem for states affected by this op | |
815 RawAddress = TypeFunc::Parms+0, // the newly-allocated raw address | |
816 RawStores = TypeFunc::Parms+1 // zero or more stores (or TOP) | |
817 }; | |
818 | |
819 InitializeNode(Compile* C, int adr_type, Node* rawoop); | |
820 virtual int Opcode() const; | |
821 virtual uint size_of() const { return sizeof(*this); } | |
822 virtual uint ideal_reg() const { return 0; } // not matched in the AD file | |
823 virtual const RegMask &in_RegMask(uint) const; // mask for RawAddress | |
824 | |
825 // Manage incoming memory edges via a MergeMem on in(Memory): | |
826 Node* memory(uint alias_idx); | |
827 | |
828 // The raw memory edge coming directly from the Allocation. | |
829 // The contents of this memory are *always* all-zero-bits. | |
830 Node* zero_memory() { return memory(Compile::AliasIdxRaw); } | |
831 | |
832 // Return the corresponding allocation for this initialization (or null if none). | |
833 // (Note: Both InitializeNode::allocation and AllocateNode::initialization | |
834 // are defined in graphKit.cpp, which sets up the bidirectional relation.) | |
835 AllocateNode* allocation(); | |
836 | |
837 // Anything other than zeroing in this init? | |
838 bool is_non_zero(); | |
839 | |
840 // An InitializeNode must completed before macro expansion is done. | |
841 // Completion requires that the AllocateNode must be followed by | |
842 // initialization of the new memory to zero, then to any initializers. | |
843 bool is_complete() { return _is_complete; } | |
844 | |
845 // Mark complete. (Must not yet be complete.) | |
846 void set_complete(PhaseGVN* phase); | |
847 | |
848 #ifdef ASSERT | |
849 // ensure all non-degenerate stores are ordered and non-overlapping | |
850 bool stores_are_sane(PhaseTransform* phase); | |
851 #endif //ASSERT | |
852 | |
853 // See if this store can be captured; return offset where it initializes. | |
854 // Return 0 if the store cannot be moved (any sort of problem). | |
855 intptr_t can_capture_store(StoreNode* st, PhaseTransform* phase); | |
856 | |
857 // Capture another store; reformat it to write my internal raw memory. | |
858 // Return the captured copy, else NULL if there is some sort of problem. | |
859 Node* capture_store(StoreNode* st, intptr_t start, PhaseTransform* phase); | |
860 | |
861 // Find captured store which corresponds to the range [start..start+size). | |
862 // Return my own memory projection (meaning the initial zero bits) | |
863 // if there is no such store. Return NULL if there is a problem. | |
864 Node* find_captured_store(intptr_t start, int size_in_bytes, PhaseTransform* phase); | |
865 | |
866 // Called when the associated AllocateNode is expanded into CFG. | |
867 Node* complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, | |
868 intptr_t header_size, Node* size_in_bytes, | |
869 PhaseGVN* phase); | |
870 | |
871 private: | |
872 void remove_extra_zeroes(); | |
873 | |
874 // Find out where a captured store should be placed (or already is placed). | |
875 int captured_store_insertion_point(intptr_t start, int size_in_bytes, | |
876 PhaseTransform* phase); | |
877 | |
878 static intptr_t get_store_offset(Node* st, PhaseTransform* phase); | |
879 | |
880 Node* make_raw_address(intptr_t offset, PhaseTransform* phase); | |
881 | |
882 bool detect_init_independence(Node* n, bool st_is_pinned, int& count); | |
883 | |
884 void coalesce_subword_stores(intptr_t header_size, Node* size_in_bytes, | |
885 PhaseGVN* phase); | |
886 | |
887 intptr_t find_next_fullword_store(uint i, PhaseGVN* phase); | |
888 }; | |
889 | |
890 //------------------------------MergeMem--------------------------------------- | |
891 // (See comment in memnode.cpp near MergeMemNode::MergeMemNode for semantics.) | |
892 class MergeMemNode: public Node { | |
893 virtual uint hash() const ; // { return NO_HASH; } | |
894 virtual uint cmp( const Node &n ) const ; // Always fail, except on self | |
895 friend class MergeMemStream; | |
896 MergeMemNode(Node* def); // clients use MergeMemNode::make | |
897 | |
898 public: | |
899 // If the input is a whole memory state, clone it with all its slices intact. | |
900 // Otherwise, make a new memory state with just that base memory input. | |
901 // In either case, the result is a newly created MergeMem. | |
902 static MergeMemNode* make(Compile* C, Node* base_memory); | |
903 | |
904 virtual int Opcode() const; | |
905 virtual Node *Identity( PhaseTransform *phase ); | |
906 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
907 virtual uint ideal_reg() const { return NotAMachineReg; } | |
908 virtual uint match_edge(uint idx) const { return 0; } | |
909 virtual const RegMask &out_RegMask() const; | |
910 virtual const Type *bottom_type() const { return Type::MEMORY; } | |
911 virtual const TypePtr *adr_type() const { return TypePtr::BOTTOM; } | |
912 // sparse accessors | |
913 // Fetch the previously stored "set_memory_at", or else the base memory. | |
914 // (Caller should clone it if it is a phi-nest.) | |
915 Node* memory_at(uint alias_idx) const; | |
916 // set the memory, regardless of its previous value | |
917 void set_memory_at(uint alias_idx, Node* n); | |
918 // the "base" is the memory that provides the non-finite support | |
919 Node* base_memory() const { return in(Compile::AliasIdxBot); } | |
920 // warning: setting the base can implicitly set any of the other slices too | |
921 void set_base_memory(Node* def); | |
922 // sentinel value which denotes a copy of the base memory: | |
923 Node* empty_memory() const { return in(Compile::AliasIdxTop); } | |
924 static Node* make_empty_memory(); // where the sentinel comes from | |
925 bool is_empty_memory(Node* n) const { assert((n == empty_memory()) == n->is_top(), "sanity"); return n->is_top(); } | |
926 // hook for the iterator, to perform any necessary setup | |
927 void iteration_setup(const MergeMemNode* other = NULL); | |
928 // push sentinels until I am at least as long as the other (semantic no-op) | |
929 void grow_to_match(const MergeMemNode* other); | |
930 bool verify_sparse() const PRODUCT_RETURN0; | |
931 #ifndef PRODUCT | |
932 virtual void dump_spec(outputStream *st) const; | |
933 #endif | |
934 }; | |
935 | |
936 class MergeMemStream : public StackObj { | |
937 private: | |
938 MergeMemNode* _mm; | |
939 const MergeMemNode* _mm2; // optional second guy, contributes non-empty iterations | |
940 Node* _mm_base; // loop-invariant base memory of _mm | |
941 int _idx; | |
942 int _cnt; | |
943 Node* _mem; | |
944 Node* _mem2; | |
945 int _cnt2; | |
946 | |
947 void init(MergeMemNode* mm, const MergeMemNode* mm2 = NULL) { | |
948 // subsume_node will break sparseness at times, whenever a memory slice | |
949 // folds down to a copy of the base ("fat") memory. In such a case, | |
950 // the raw edge will update to base, although it should be top. | |
951 // This iterator will recognize either top or base_memory as an | |
952 // "empty" slice. See is_empty, is_empty2, and next below. | |
953 // | |
954 // The sparseness property is repaired in MergeMemNode::Ideal. | |
955 // As long as access to a MergeMem goes through this iterator | |
956 // or the memory_at accessor, flaws in the sparseness will | |
957 // never be observed. | |
958 // | |
959 // Also, iteration_setup repairs sparseness. | |
960 assert(mm->verify_sparse(), "please, no dups of base"); | |
961 assert(mm2==NULL || mm2->verify_sparse(), "please, no dups of base"); | |
962 | |
963 _mm = mm; | |
964 _mm_base = mm->base_memory(); | |
965 _mm2 = mm2; | |
966 _cnt = mm->req(); | |
967 _idx = Compile::AliasIdxBot-1; // start at the base memory | |
968 _mem = NULL; | |
969 _mem2 = NULL; | |
970 } | |
971 | |
972 #ifdef ASSERT | |
973 Node* check_memory() const { | |
974 if (at_base_memory()) | |
975 return _mm->base_memory(); | |
976 else if ((uint)_idx < _mm->req() && !_mm->in(_idx)->is_top()) | |
977 return _mm->memory_at(_idx); | |
978 else | |
979 return _mm_base; | |
980 } | |
981 Node* check_memory2() const { | |
982 return at_base_memory()? _mm2->base_memory(): _mm2->memory_at(_idx); | |
983 } | |
984 #endif | |
985 | |
986 static bool match_memory(Node* mem, const MergeMemNode* mm, int idx) PRODUCT_RETURN0; | |
987 void assert_synch() const { | |
988 assert(!_mem || _idx >= _cnt || match_memory(_mem, _mm, _idx), | |
989 "no side-effects except through the stream"); | |
990 } | |
991 | |
992 public: | |
993 | |
994 // expected usages: | |
995 // for (MergeMemStream mms(mem->is_MergeMem()); next_non_empty(); ) { ... } | |
996 // for (MergeMemStream mms(mem1, mem2); next_non_empty2(); ) { ... } | |
997 | |
998 // iterate over one merge | |
999 MergeMemStream(MergeMemNode* mm) { | |
1000 mm->iteration_setup(); | |
1001 init(mm); | |
1002 debug_only(_cnt2 = 999); | |
1003 } | |
1004 // iterate in parallel over two merges | |
1005 // only iterates through non-empty elements of mm2 | |
1006 MergeMemStream(MergeMemNode* mm, const MergeMemNode* mm2) { | |
1007 assert(mm2, "second argument must be a MergeMem also"); | |
1008 ((MergeMemNode*)mm2)->iteration_setup(); // update hidden state | |
1009 mm->iteration_setup(mm2); | |
1010 init(mm, mm2); | |
1011 _cnt2 = mm2->req(); | |
1012 } | |
1013 #ifdef ASSERT | |
1014 ~MergeMemStream() { | |
1015 assert_synch(); | |
1016 } | |
1017 #endif | |
1018 | |
1019 MergeMemNode* all_memory() const { | |
1020 return _mm; | |
1021 } | |
1022 Node* base_memory() const { | |
1023 assert(_mm_base == _mm->base_memory(), "no update to base memory, please"); | |
1024 return _mm_base; | |
1025 } | |
1026 const MergeMemNode* all_memory2() const { | |
1027 assert(_mm2 != NULL, ""); | |
1028 return _mm2; | |
1029 } | |
1030 bool at_base_memory() const { | |
1031 return _idx == Compile::AliasIdxBot; | |
1032 } | |
1033 int alias_idx() const { | |
1034 assert(_mem, "must call next 1st"); | |
1035 return _idx; | |
1036 } | |
1037 | |
1038 const TypePtr* adr_type() const { | |
1039 return Compile::current()->get_adr_type(alias_idx()); | |
1040 } | |
1041 | |
1042 const TypePtr* adr_type(Compile* C) const { | |
1043 return C->get_adr_type(alias_idx()); | |
1044 } | |
1045 bool is_empty() const { | |
1046 assert(_mem, "must call next 1st"); | |
1047 assert(_mem->is_top() == (_mem==_mm->empty_memory()), "correct sentinel"); | |
1048 return _mem->is_top(); | |
1049 } | |
1050 bool is_empty2() const { | |
1051 assert(_mem2, "must call next 1st"); | |
1052 assert(_mem2->is_top() == (_mem2==_mm2->empty_memory()), "correct sentinel"); | |
1053 return _mem2->is_top(); | |
1054 } | |
1055 Node* memory() const { | |
1056 assert(!is_empty(), "must not be empty"); | |
1057 assert_synch(); | |
1058 return _mem; | |
1059 } | |
1060 // get the current memory, regardless of empty or non-empty status | |
1061 Node* force_memory() const { | |
1062 assert(!is_empty() || !at_base_memory(), ""); | |
1063 // Use _mm_base to defend against updates to _mem->base_memory(). | |
1064 Node *mem = _mem->is_top() ? _mm_base : _mem; | |
1065 assert(mem == check_memory(), ""); | |
1066 return mem; | |
1067 } | |
1068 Node* memory2() const { | |
1069 assert(_mem2 == check_memory2(), ""); | |
1070 return _mem2; | |
1071 } | |
1072 void set_memory(Node* mem) { | |
1073 if (at_base_memory()) { | |
1074 // Note that this does not change the invariant _mm_base. | |
1075 _mm->set_base_memory(mem); | |
1076 } else { | |
1077 _mm->set_memory_at(_idx, mem); | |
1078 } | |
1079 _mem = mem; | |
1080 assert_synch(); | |
1081 } | |
1082 | |
1083 // Recover from a side effect to the MergeMemNode. | |
1084 void set_memory() { | |
1085 _mem = _mm->in(_idx); | |
1086 } | |
1087 | |
1088 bool next() { return next(false); } | |
1089 bool next2() { return next(true); } | |
1090 | |
1091 bool next_non_empty() { return next_non_empty(false); } | |
1092 bool next_non_empty2() { return next_non_empty(true); } | |
1093 // next_non_empty2 can yield states where is_empty() is true | |
1094 | |
1095 private: | |
1096 // find the next item, which might be empty | |
1097 bool next(bool have_mm2) { | |
1098 assert((_mm2 != NULL) == have_mm2, "use other next"); | |
1099 assert_synch(); | |
1100 if (++_idx < _cnt) { | |
1101 // Note: This iterator allows _mm to be non-sparse. | |
1102 // It behaves the same whether _mem is top or base_memory. | |
1103 _mem = _mm->in(_idx); | |
1104 if (have_mm2) | |
1105 _mem2 = _mm2->in((_idx < _cnt2) ? _idx : Compile::AliasIdxTop); | |
1106 return true; | |
1107 } | |
1108 return false; | |
1109 } | |
1110 | |
1111 // find the next non-empty item | |
1112 bool next_non_empty(bool have_mm2) { | |
1113 while (next(have_mm2)) { | |
1114 if (!is_empty()) { | |
1115 // make sure _mem2 is filled in sensibly | |
1116 if (have_mm2 && _mem2->is_top()) _mem2 = _mm2->base_memory(); | |
1117 return true; | |
1118 } else if (have_mm2 && !is_empty2()) { | |
1119 return true; // is_empty() == true | |
1120 } | |
1121 } | |
1122 return false; | |
1123 } | |
1124 }; | |
1125 | |
1126 //------------------------------Prefetch--------------------------------------- | |
1127 | |
1128 // Non-faulting prefetch load. Prefetch for many reads. | |
1129 class PrefetchReadNode : public Node { | |
1130 public: | |
1131 PrefetchReadNode(Node *abio, Node *adr) : Node(0,abio,adr) {} | |
1132 virtual int Opcode() const; | |
1133 virtual uint ideal_reg() const { return NotAMachineReg; } | |
1134 virtual uint match_edge(uint idx) const { return idx==2; } | |
1135 virtual const Type *bottom_type() const { return Type::ABIO; } | |
1136 }; | |
1137 | |
1138 // Non-faulting prefetch load. Prefetch for many reads & many writes. | |
1139 class PrefetchWriteNode : public Node { | |
1140 public: | |
1141 PrefetchWriteNode(Node *abio, Node *adr) : Node(0,abio,adr) {} | |
1142 virtual int Opcode() const; | |
1143 virtual uint ideal_reg() const { return NotAMachineReg; } | |
1144 virtual uint match_edge(uint idx) const { return idx==2; } | |
1145 virtual const Type *bottom_type() const { return Type::ABIO; } | |
1146 }; |