Thu, 20 Mar 2008 15:11:44 -0700
6674600: (Escape Analysis) Optimize memory graph for instance's fields
Summary: EA gives opportunite to do more aggressive memory optimizations.
Reviewed-by: never, jrose
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 */
25 // Portions of code courtesy of Clifford Click
27 class MultiNode;
28 class PhaseCCP;
29 class PhaseTransform;
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 }
63 public:
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);
70 static Node *optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase);
71 static Node *optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase);
72 // This one should probably be a phase-specific function:
73 static bool detect_dominating_control(Node* dom, Node* sub);
75 // Is this Node a MemNode or some descendent? Default is YES.
76 virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp );
78 virtual const class TypePtr *adr_type() const; // returns bottom_type of address
80 // Shared code for Ideal methods:
81 Node *Ideal_common(PhaseGVN *phase, bool can_reshape); // Return -1 for short-circuit NULL.
83 // Helper function for adr_type() implementations.
84 static const TypePtr* calculate_adr_type(const Type* t, const TypePtr* cross_check = NULL);
86 // Raw access function, to allow copying of adr_type efficiently in
87 // product builds and retain the debug info for debug builds.
88 const TypePtr *raw_adr_type() const {
89 #ifdef ASSERT
90 return _adr_type;
91 #else
92 return 0;
93 #endif
94 }
96 // Map a load or store opcode to its corresponding store opcode.
97 // (Return -1 if unknown.)
98 virtual int store_Opcode() const { return -1; }
100 // What is the type of the value in memory? (T_VOID mean "unspecified".)
101 virtual BasicType memory_type() const = 0;
102 virtual int memory_size() const {
103 #ifdef ASSERT
104 return type2aelembytes(memory_type(), true);
105 #else
106 return type2aelembytes(memory_type());
107 #endif
108 }
110 // Search through memory states which precede this node (load or store).
111 // Look for an exact match for the address, with no intervening
112 // aliased stores.
113 Node* find_previous_store(PhaseTransform* phase);
115 // Can this node (load or store) accurately see a stored value in
116 // the given memory state? (The state may or may not be in(Memory).)
117 Node* can_see_stored_value(Node* st, PhaseTransform* phase) const;
119 #ifndef PRODUCT
120 static void dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st);
121 virtual void dump_spec(outputStream *st) const;
122 #endif
123 };
125 //------------------------------LoadNode---------------------------------------
126 // Load value; requires Memory and Address
127 class LoadNode : public MemNode {
128 protected:
129 virtual uint cmp( const Node &n ) const;
130 virtual uint size_of() const; // Size is bigger
131 const Type* const _type; // What kind of value is loaded?
132 public:
134 LoadNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt )
135 : MemNode(c,mem,adr,at), _type(rt) {
136 init_class_id(Class_Load);
137 }
139 // Polymorphic factory method:
140 static LoadNode* make( Compile *C, Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt, BasicType bt );
142 virtual uint hash() const; // Check the type
144 // Handle algebraic identities here. If we have an identity, return the Node
145 // we are equivalent to. We look for Load of a Store.
146 virtual Node *Identity( PhaseTransform *phase );
148 // If the load is from Field memory and the pointer is non-null, we can
149 // zero out the control input.
150 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
152 // Recover original value from boxed values
153 Node *eliminate_autobox(PhaseGVN *phase);
155 // Compute a new Type for this node. Basically we just do the pre-check,
156 // then call the virtual add() to set the type.
157 virtual const Type *Value( PhaseTransform *phase ) const;
159 virtual uint ideal_reg() const;
160 virtual const Type *bottom_type() const;
161 // Following method is copied from TypeNode:
162 void set_type(const Type* t) {
163 assert(t != NULL, "sanity");
164 debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH);
165 *(const Type**)&_type = t; // cast away const-ness
166 // If this node is in the hash table, make sure it doesn't need a rehash.
167 assert(check_hash == NO_HASH || check_hash == hash(), "type change must preserve hash code");
168 }
169 const Type* type() const { assert(_type != NULL, "sanity"); return _type; };
171 // Do not match memory edge
172 virtual uint match_edge(uint idx) const;
174 // Map a load opcode to its corresponding store opcode.
175 virtual int store_Opcode() const = 0;
177 // Check if the load's memory input is a Phi node with the same control.
178 bool is_instance_field_load_with_local_phi(Node* ctrl);
180 #ifndef PRODUCT
181 virtual void dump_spec(outputStream *st) const;
182 #endif
183 protected:
184 const Type* load_array_final_field(const TypeKlassPtr *tkls,
185 ciKlass* klass) const;
186 };
188 //------------------------------LoadBNode--------------------------------------
189 // Load a byte (8bits signed) from memory
190 class LoadBNode : public LoadNode {
191 public:
192 LoadBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::BYTE )
193 : LoadNode(c,mem,adr,at,ti) {}
194 virtual int Opcode() const;
195 virtual uint ideal_reg() const { return Op_RegI; }
196 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
197 virtual int store_Opcode() const { return Op_StoreB; }
198 virtual BasicType memory_type() const { return T_BYTE; }
199 };
201 //------------------------------LoadCNode--------------------------------------
202 // Load a char (16bits unsigned) from memory
203 class LoadCNode : public LoadNode {
204 public:
205 LoadCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::CHAR )
206 : LoadNode(c,mem,adr,at,ti) {}
207 virtual int Opcode() const;
208 virtual uint ideal_reg() const { return Op_RegI; }
209 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
210 virtual int store_Opcode() const { return Op_StoreC; }
211 virtual BasicType memory_type() const { return T_CHAR; }
212 };
214 //------------------------------LoadINode--------------------------------------
215 // Load an integer from memory
216 class LoadINode : public LoadNode {
217 public:
218 LoadINode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::INT )
219 : LoadNode(c,mem,adr,at,ti) {}
220 virtual int Opcode() const;
221 virtual uint ideal_reg() const { return Op_RegI; }
222 virtual int store_Opcode() const { return Op_StoreI; }
223 virtual BasicType memory_type() const { return T_INT; }
224 };
226 //------------------------------LoadRangeNode----------------------------------
227 // Load an array length from the array
228 class LoadRangeNode : public LoadINode {
229 public:
230 LoadRangeNode( Node *c, Node *mem, Node *adr, const TypeInt *ti = TypeInt::POS )
231 : LoadINode(c,mem,adr,TypeAryPtr::RANGE,ti) {}
232 virtual int Opcode() const;
233 virtual const Type *Value( PhaseTransform *phase ) const;
234 virtual Node *Identity( PhaseTransform *phase );
235 };
237 //------------------------------LoadLNode--------------------------------------
238 // Load a long from memory
239 class LoadLNode : public LoadNode {
240 virtual uint hash() const { return LoadNode::hash() + _require_atomic_access; }
241 virtual uint cmp( const Node &n ) const {
242 return _require_atomic_access == ((LoadLNode&)n)._require_atomic_access
243 && LoadNode::cmp(n);
244 }
245 virtual uint size_of() const { return sizeof(*this); }
246 const bool _require_atomic_access; // is piecewise load forbidden?
248 public:
249 LoadLNode( Node *c, Node *mem, Node *adr, const TypePtr* at,
250 const TypeLong *tl = TypeLong::LONG,
251 bool require_atomic_access = false )
252 : LoadNode(c,mem,adr,at,tl)
253 , _require_atomic_access(require_atomic_access)
254 {}
255 virtual int Opcode() const;
256 virtual uint ideal_reg() const { return Op_RegL; }
257 virtual int store_Opcode() const { return Op_StoreL; }
258 virtual BasicType memory_type() const { return T_LONG; }
259 bool require_atomic_access() { return _require_atomic_access; }
260 static LoadLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt);
261 #ifndef PRODUCT
262 virtual void dump_spec(outputStream *st) const {
263 LoadNode::dump_spec(st);
264 if (_require_atomic_access) st->print(" Atomic!");
265 }
266 #endif
267 };
269 //------------------------------LoadL_unalignedNode----------------------------
270 // Load a long from unaligned memory
271 class LoadL_unalignedNode : public LoadLNode {
272 public:
273 LoadL_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at )
274 : LoadLNode(c,mem,adr,at) {}
275 virtual int Opcode() const;
276 };
278 //------------------------------LoadFNode--------------------------------------
279 // Load a float (64 bits) from memory
280 class LoadFNode : public LoadNode {
281 public:
282 LoadFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::FLOAT )
283 : LoadNode(c,mem,adr,at,t) {}
284 virtual int Opcode() const;
285 virtual uint ideal_reg() const { return Op_RegF; }
286 virtual int store_Opcode() const { return Op_StoreF; }
287 virtual BasicType memory_type() const { return T_FLOAT; }
288 };
290 //------------------------------LoadDNode--------------------------------------
291 // Load a double (64 bits) from memory
292 class LoadDNode : public LoadNode {
293 public:
294 LoadDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::DOUBLE )
295 : LoadNode(c,mem,adr,at,t) {}
296 virtual int Opcode() const;
297 virtual uint ideal_reg() const { return Op_RegD; }
298 virtual int store_Opcode() const { return Op_StoreD; }
299 virtual BasicType memory_type() const { return T_DOUBLE; }
300 };
302 //------------------------------LoadD_unalignedNode----------------------------
303 // Load a double from unaligned memory
304 class LoadD_unalignedNode : public LoadDNode {
305 public:
306 LoadD_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at )
307 : LoadDNode(c,mem,adr,at) {}
308 virtual int Opcode() const;
309 };
311 //------------------------------LoadPNode--------------------------------------
312 // Load a pointer from memory (either object or array)
313 class LoadPNode : public LoadNode {
314 public:
315 LoadPNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypePtr* t )
316 : LoadNode(c,mem,adr,at,t) {}
317 virtual int Opcode() const;
318 virtual uint ideal_reg() const { return Op_RegP; }
319 virtual int store_Opcode() const { return Op_StoreP; }
320 virtual BasicType memory_type() const { return T_ADDRESS; }
321 // depends_only_on_test is almost always true, and needs to be almost always
322 // true to enable key hoisting & commoning optimizations. However, for the
323 // special case of RawPtr loads from TLS top & end, the control edge carries
324 // the dependence preventing hoisting past a Safepoint instead of the memory
325 // edge. (An unfortunate consequence of having Safepoints not set Raw
326 // Memory; itself an unfortunate consequence of having Nodes which produce
327 // results (new raw memory state) inside of loops preventing all manner of
328 // other optimizations). Basically, it's ugly but so is the alternative.
329 // See comment in macro.cpp, around line 125 expand_allocate_common().
330 virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
331 };
333 //------------------------------LoadKlassNode----------------------------------
334 // Load a Klass from an object
335 class LoadKlassNode : public LoadPNode {
336 public:
337 LoadKlassNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeKlassPtr *tk = TypeKlassPtr::OBJECT )
338 : LoadPNode(c,mem,adr,at,tk) {}
339 virtual int Opcode() const;
340 virtual const Type *Value( PhaseTransform *phase ) const;
341 virtual Node *Identity( PhaseTransform *phase );
342 virtual bool depends_only_on_test() const { return true; }
343 };
345 //------------------------------LoadSNode--------------------------------------
346 // Load a short (16bits signed) from memory
347 class LoadSNode : public LoadNode {
348 public:
349 LoadSNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::SHORT )
350 : LoadNode(c,mem,adr,at,ti) {}
351 virtual int Opcode() const;
352 virtual uint ideal_reg() const { return Op_RegI; }
353 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
354 virtual int store_Opcode() const { return Op_StoreC; }
355 virtual BasicType memory_type() const { return T_SHORT; }
356 };
358 //------------------------------StoreNode--------------------------------------
359 // Store value; requires Store, Address and Value
360 class StoreNode : public MemNode {
361 protected:
362 virtual uint cmp( const Node &n ) const;
363 virtual bool depends_only_on_test() const { return false; }
365 Node *Ideal_masked_input (PhaseGVN *phase, uint mask);
366 Node *Ideal_sign_extended_input(PhaseGVN *phase, int num_bits);
368 public:
369 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val )
370 : MemNode(c,mem,adr,at,val) {
371 init_class_id(Class_Store);
372 }
373 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store )
374 : MemNode(c,mem,adr,at,val,oop_store) {
375 init_class_id(Class_Store);
376 }
378 // Polymorphic factory method:
379 static StoreNode* make( Compile *C, Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, BasicType bt );
381 virtual uint hash() const; // Check the type
383 // If the store is to Field memory and the pointer is non-null, we can
384 // zero out the control input.
385 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
387 // Compute a new Type for this node. Basically we just do the pre-check,
388 // then call the virtual add() to set the type.
389 virtual const Type *Value( PhaseTransform *phase ) const;
391 // Check for identity function on memory (Load then Store at same address)
392 virtual Node *Identity( PhaseTransform *phase );
394 // Do not match memory edge
395 virtual uint match_edge(uint idx) const;
397 virtual const Type *bottom_type() const; // returns Type::MEMORY
399 // Map a store opcode to its corresponding own opcode, trivially.
400 virtual int store_Opcode() const { return Opcode(); }
402 // have all possible loads of the value stored been optimized away?
403 bool value_never_loaded(PhaseTransform *phase) const;
404 };
406 //------------------------------StoreBNode-------------------------------------
407 // Store byte to memory
408 class StoreBNode : public StoreNode {
409 public:
410 StoreBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
411 virtual int Opcode() const;
412 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
413 virtual BasicType memory_type() const { return T_BYTE; }
414 };
416 //------------------------------StoreCNode-------------------------------------
417 // Store char/short to memory
418 class StoreCNode : public StoreNode {
419 public:
420 StoreCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
421 virtual int Opcode() const;
422 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
423 virtual BasicType memory_type() const { return T_CHAR; }
424 };
426 //------------------------------StoreINode-------------------------------------
427 // Store int to memory
428 class StoreINode : public StoreNode {
429 public:
430 StoreINode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
431 virtual int Opcode() const;
432 virtual BasicType memory_type() const { return T_INT; }
433 };
435 //------------------------------StoreLNode-------------------------------------
436 // Store long to memory
437 class StoreLNode : public StoreNode {
438 virtual uint hash() const { return StoreNode::hash() + _require_atomic_access; }
439 virtual uint cmp( const Node &n ) const {
440 return _require_atomic_access == ((StoreLNode&)n)._require_atomic_access
441 && StoreNode::cmp(n);
442 }
443 virtual uint size_of() const { return sizeof(*this); }
444 const bool _require_atomic_access; // is piecewise store forbidden?
446 public:
447 StoreLNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val,
448 bool require_atomic_access = false )
449 : StoreNode(c,mem,adr,at,val)
450 , _require_atomic_access(require_atomic_access)
451 {}
452 virtual int Opcode() const;
453 virtual BasicType memory_type() const { return T_LONG; }
454 bool require_atomic_access() { return _require_atomic_access; }
455 static StoreLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val);
456 #ifndef PRODUCT
457 virtual void dump_spec(outputStream *st) const {
458 StoreNode::dump_spec(st);
459 if (_require_atomic_access) st->print(" Atomic!");
460 }
461 #endif
462 };
464 //------------------------------StoreFNode-------------------------------------
465 // Store float to memory
466 class StoreFNode : public StoreNode {
467 public:
468 StoreFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
469 virtual int Opcode() const;
470 virtual BasicType memory_type() const { return T_FLOAT; }
471 };
473 //------------------------------StoreDNode-------------------------------------
474 // Store double to memory
475 class StoreDNode : public StoreNode {
476 public:
477 StoreDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
478 virtual int Opcode() const;
479 virtual BasicType memory_type() const { return T_DOUBLE; }
480 };
482 //------------------------------StorePNode-------------------------------------
483 // Store pointer to memory
484 class StorePNode : public StoreNode {
485 public:
486 StorePNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
487 virtual int Opcode() const;
488 virtual BasicType memory_type() const { return T_ADDRESS; }
489 };
491 //------------------------------StoreCMNode-----------------------------------
492 // Store card-mark byte to memory for CM
493 // The last StoreCM before a SafePoint must be preserved and occur after its "oop" store
494 // Preceeding equivalent StoreCMs may be eliminated.
495 class StoreCMNode : public StoreNode {
496 public:
497 StoreCMNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store ) : StoreNode(c,mem,adr,at,val,oop_store) {}
498 virtual int Opcode() const;
499 virtual Node *Identity( PhaseTransform *phase );
500 virtual const Type *Value( PhaseTransform *phase ) const;
501 virtual BasicType memory_type() const { return T_VOID; } // unspecific
502 };
504 //------------------------------LoadPLockedNode---------------------------------
505 // Load-locked a pointer from memory (either object or array).
506 // On Sparc & Intel this is implemented as a normal pointer load.
507 // On PowerPC and friends it's a real load-locked.
508 class LoadPLockedNode : public LoadPNode {
509 public:
510 LoadPLockedNode( Node *c, Node *mem, Node *adr )
511 : LoadPNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) {}
512 virtual int Opcode() const;
513 virtual int store_Opcode() const { return Op_StorePConditional; }
514 virtual bool depends_only_on_test() const { return true; }
515 };
517 //------------------------------LoadLLockedNode---------------------------------
518 // Load-locked a pointer from memory (either object or array).
519 // On Sparc & Intel this is implemented as a normal long load.
520 class LoadLLockedNode : public LoadLNode {
521 public:
522 LoadLLockedNode( Node *c, Node *mem, Node *adr )
523 : LoadLNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeLong::LONG) {}
524 virtual int Opcode() const;
525 virtual int store_Opcode() const { return Op_StoreLConditional; }
526 };
528 //------------------------------SCMemProjNode---------------------------------------
529 // This class defines a projection of the memory state of a store conditional node.
530 // These nodes return a value, but also update memory.
531 class SCMemProjNode : public ProjNode {
532 public:
533 enum {SCMEMPROJCON = (uint)-2};
534 SCMemProjNode( Node *src) : ProjNode( src, SCMEMPROJCON) { }
535 virtual int Opcode() const;
536 virtual bool is_CFG() const { return false; }
537 virtual const Type *bottom_type() const {return Type::MEMORY;}
538 virtual const TypePtr *adr_type() const { return in(0)->in(MemNode::Memory)->adr_type();}
539 virtual uint ideal_reg() const { return 0;} // memory projections don't have a register
540 virtual const Type *Value( PhaseTransform *phase ) const;
541 #ifndef PRODUCT
542 virtual void dump_spec(outputStream *st) const {};
543 #endif
544 };
546 //------------------------------LoadStoreNode---------------------------
547 class LoadStoreNode : public Node {
548 public:
549 enum {
550 ExpectedIn = MemNode::ValueIn+1 // One more input than MemNode
551 };
552 LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex);
553 virtual bool depends_only_on_test() const { return false; }
554 virtual const Type *bottom_type() const { return TypeInt::BOOL; }
555 virtual uint ideal_reg() const { return Op_RegI; }
556 virtual uint match_edge(uint idx) const { return idx == MemNode::Address || idx == MemNode::ValueIn; }
557 };
559 //------------------------------StorePConditionalNode---------------------------
560 // Conditionally store pointer to memory, if no change since prior
561 // load-locked. Sets flags for success or failure of the store.
562 class StorePConditionalNode : public LoadStoreNode {
563 public:
564 StorePConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { }
565 virtual int Opcode() const;
566 // Produces flags
567 virtual uint ideal_reg() const { return Op_RegFlags; }
568 };
570 //------------------------------StoreLConditionalNode---------------------------
571 // Conditionally store long to memory, if no change since prior
572 // load-locked. Sets flags for success or failure of the store.
573 class StoreLConditionalNode : public LoadStoreNode {
574 public:
575 StoreLConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { }
576 virtual int Opcode() const;
577 };
580 //------------------------------CompareAndSwapLNode---------------------------
581 class CompareAndSwapLNode : public LoadStoreNode {
582 public:
583 CompareAndSwapLNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
584 virtual int Opcode() const;
585 };
588 //------------------------------CompareAndSwapINode---------------------------
589 class CompareAndSwapINode : public LoadStoreNode {
590 public:
591 CompareAndSwapINode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
592 virtual int Opcode() const;
593 };
596 //------------------------------CompareAndSwapPNode---------------------------
597 class CompareAndSwapPNode : public LoadStoreNode {
598 public:
599 CompareAndSwapPNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
600 virtual int Opcode() const;
601 };
603 //------------------------------ClearArray-------------------------------------
604 class ClearArrayNode: public Node {
605 public:
606 ClearArrayNode( Node *ctrl, Node *arymem, Node *word_cnt, Node *base ) : Node(ctrl,arymem,word_cnt,base) {}
607 virtual int Opcode() const;
608 virtual const Type *bottom_type() const { return Type::MEMORY; }
609 // ClearArray modifies array elements, and so affects only the
610 // array memory addressed by the bottom_type of its base address.
611 virtual const class TypePtr *adr_type() const;
612 virtual Node *Identity( PhaseTransform *phase );
613 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
614 virtual uint match_edge(uint idx) const;
616 // Clear the given area of an object or array.
617 // The start offset must always be aligned mod BytesPerInt.
618 // The end offset must always be aligned mod BytesPerLong.
619 // Return the new memory.
620 static Node* clear_memory(Node* control, Node* mem, Node* dest,
621 intptr_t start_offset,
622 intptr_t end_offset,
623 PhaseGVN* phase);
624 static Node* clear_memory(Node* control, Node* mem, Node* dest,
625 intptr_t start_offset,
626 Node* end_offset,
627 PhaseGVN* phase);
628 static Node* clear_memory(Node* control, Node* mem, Node* dest,
629 Node* start_offset,
630 Node* end_offset,
631 PhaseGVN* phase);
632 };
634 //------------------------------StrComp-------------------------------------
635 class StrCompNode: public Node {
636 public:
637 StrCompNode(Node *control,
638 Node* char_array_mem,
639 Node* value_mem,
640 Node* count_mem,
641 Node* offset_mem,
642 Node* s1, Node* s2): Node(control,
643 char_array_mem,
644 value_mem,
645 count_mem,
646 offset_mem,
647 s1, s2) {};
648 virtual int Opcode() const;
649 virtual bool depends_only_on_test() const { return false; }
650 virtual const Type* bottom_type() const { return TypeInt::INT; }
651 // a StrCompNode (conservatively) aliases with everything:
652 virtual const TypePtr* adr_type() const { return TypePtr::BOTTOM; }
653 virtual uint match_edge(uint idx) const;
654 virtual uint ideal_reg() const { return Op_RegI; }
655 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
656 };
658 //------------------------------MemBar-----------------------------------------
659 // There are different flavors of Memory Barriers to match the Java Memory
660 // Model. Monitor-enter and volatile-load act as Aquires: no following ref
661 // can be moved to before them. We insert a MemBar-Acquire after a FastLock or
662 // volatile-load. Monitor-exit and volatile-store act as Release: no
663 // preceeding ref can be moved to after them. We insert a MemBar-Release
664 // before a FastUnlock or volatile-store. All volatiles need to be
665 // serialized, so we follow all volatile-stores with a MemBar-Volatile to
666 // seperate it from any following volatile-load.
667 class MemBarNode: public MultiNode {
668 virtual uint hash() const ; // { return NO_HASH; }
669 virtual uint cmp( const Node &n ) const ; // Always fail, except on self
671 virtual uint size_of() const { return sizeof(*this); }
672 // Memory type this node is serializing. Usually either rawptr or bottom.
673 const TypePtr* _adr_type;
675 public:
676 enum {
677 Precedent = TypeFunc::Parms // optional edge to force precedence
678 };
679 MemBarNode(Compile* C, int alias_idx, Node* precedent);
680 virtual int Opcode() const = 0;
681 virtual const class TypePtr *adr_type() const { return _adr_type; }
682 virtual const Type *Value( PhaseTransform *phase ) const;
683 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
684 virtual uint match_edge(uint idx) const { return 0; }
685 virtual const Type *bottom_type() const { return TypeTuple::MEMBAR; }
686 virtual Node *match( const ProjNode *proj, const Matcher *m );
687 // Factory method. Builds a wide or narrow membar.
688 // Optional 'precedent' becomes an extra edge if not null.
689 static MemBarNode* make(Compile* C, int opcode,
690 int alias_idx = Compile::AliasIdxBot,
691 Node* precedent = NULL);
692 };
694 // "Acquire" - no following ref can move before (but earlier refs can
695 // follow, like an early Load stalled in cache). Requires multi-cpu
696 // visibility. Inserted after a volatile load or FastLock.
697 class MemBarAcquireNode: public MemBarNode {
698 public:
699 MemBarAcquireNode(Compile* C, int alias_idx, Node* precedent)
700 : MemBarNode(C, alias_idx, precedent) {}
701 virtual int Opcode() const;
702 };
704 // "Release" - no earlier ref can move after (but later refs can move
705 // up, like a speculative pipelined cache-hitting Load). Requires
706 // multi-cpu visibility. Inserted before a volatile store or FastUnLock.
707 class MemBarReleaseNode: public MemBarNode {
708 public:
709 MemBarReleaseNode(Compile* C, int alias_idx, Node* precedent)
710 : MemBarNode(C, alias_idx, precedent) {}
711 virtual int Opcode() const;
712 };
714 // Ordering between a volatile store and a following volatile load.
715 // Requires multi-CPU visibility?
716 class MemBarVolatileNode: public MemBarNode {
717 public:
718 MemBarVolatileNode(Compile* C, int alias_idx, Node* precedent)
719 : MemBarNode(C, alias_idx, precedent) {}
720 virtual int Opcode() const;
721 };
723 // Ordering within the same CPU. Used to order unsafe memory references
724 // inside the compiler when we lack alias info. Not needed "outside" the
725 // compiler because the CPU does all the ordering for us.
726 class MemBarCPUOrderNode: public MemBarNode {
727 public:
728 MemBarCPUOrderNode(Compile* C, int alias_idx, Node* precedent)
729 : MemBarNode(C, alias_idx, precedent) {}
730 virtual int Opcode() const;
731 virtual uint ideal_reg() const { return 0; } // not matched in the AD file
732 };
734 // Isolation of object setup after an AllocateNode and before next safepoint.
735 // (See comment in memnode.cpp near InitializeNode::InitializeNode for semantics.)
736 class InitializeNode: public MemBarNode {
737 friend class AllocateNode;
739 bool _is_complete;
741 public:
742 enum {
743 Control = TypeFunc::Control,
744 Memory = TypeFunc::Memory, // MergeMem for states affected by this op
745 RawAddress = TypeFunc::Parms+0, // the newly-allocated raw address
746 RawStores = TypeFunc::Parms+1 // zero or more stores (or TOP)
747 };
749 InitializeNode(Compile* C, int adr_type, Node* rawoop);
750 virtual int Opcode() const;
751 virtual uint size_of() const { return sizeof(*this); }
752 virtual uint ideal_reg() const { return 0; } // not matched in the AD file
753 virtual const RegMask &in_RegMask(uint) const; // mask for RawAddress
755 // Manage incoming memory edges via a MergeMem on in(Memory):
756 Node* memory(uint alias_idx);
758 // The raw memory edge coming directly from the Allocation.
759 // The contents of this memory are *always* all-zero-bits.
760 Node* zero_memory() { return memory(Compile::AliasIdxRaw); }
762 // Return the corresponding allocation for this initialization (or null if none).
763 // (Note: Both InitializeNode::allocation and AllocateNode::initialization
764 // are defined in graphKit.cpp, which sets up the bidirectional relation.)
765 AllocateNode* allocation();
767 // Anything other than zeroing in this init?
768 bool is_non_zero();
770 // An InitializeNode must completed before macro expansion is done.
771 // Completion requires that the AllocateNode must be followed by
772 // initialization of the new memory to zero, then to any initializers.
773 bool is_complete() { return _is_complete; }
775 // Mark complete. (Must not yet be complete.)
776 void set_complete(PhaseGVN* phase);
778 #ifdef ASSERT
779 // ensure all non-degenerate stores are ordered and non-overlapping
780 bool stores_are_sane(PhaseTransform* phase);
781 #endif //ASSERT
783 // See if this store can be captured; return offset where it initializes.
784 // Return 0 if the store cannot be moved (any sort of problem).
785 intptr_t can_capture_store(StoreNode* st, PhaseTransform* phase);
787 // Capture another store; reformat it to write my internal raw memory.
788 // Return the captured copy, else NULL if there is some sort of problem.
789 Node* capture_store(StoreNode* st, intptr_t start, PhaseTransform* phase);
791 // Find captured store which corresponds to the range [start..start+size).
792 // Return my own memory projection (meaning the initial zero bits)
793 // if there is no such store. Return NULL if there is a problem.
794 Node* find_captured_store(intptr_t start, int size_in_bytes, PhaseTransform* phase);
796 // Called when the associated AllocateNode is expanded into CFG.
797 Node* complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
798 intptr_t header_size, Node* size_in_bytes,
799 PhaseGVN* phase);
801 private:
802 void remove_extra_zeroes();
804 // Find out where a captured store should be placed (or already is placed).
805 int captured_store_insertion_point(intptr_t start, int size_in_bytes,
806 PhaseTransform* phase);
808 static intptr_t get_store_offset(Node* st, PhaseTransform* phase);
810 Node* make_raw_address(intptr_t offset, PhaseTransform* phase);
812 bool detect_init_independence(Node* n, bool st_is_pinned, int& count);
814 void coalesce_subword_stores(intptr_t header_size, Node* size_in_bytes,
815 PhaseGVN* phase);
817 intptr_t find_next_fullword_store(uint i, PhaseGVN* phase);
818 };
820 //------------------------------MergeMem---------------------------------------
821 // (See comment in memnode.cpp near MergeMemNode::MergeMemNode for semantics.)
822 class MergeMemNode: public Node {
823 virtual uint hash() const ; // { return NO_HASH; }
824 virtual uint cmp( const Node &n ) const ; // Always fail, except on self
825 friend class MergeMemStream;
826 MergeMemNode(Node* def); // clients use MergeMemNode::make
828 public:
829 // If the input is a whole memory state, clone it with all its slices intact.
830 // Otherwise, make a new memory state with just that base memory input.
831 // In either case, the result is a newly created MergeMem.
832 static MergeMemNode* make(Compile* C, Node* base_memory);
834 virtual int Opcode() const;
835 virtual Node *Identity( PhaseTransform *phase );
836 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
837 virtual uint ideal_reg() const { return NotAMachineReg; }
838 virtual uint match_edge(uint idx) const { return 0; }
839 virtual const RegMask &out_RegMask() const;
840 virtual const Type *bottom_type() const { return Type::MEMORY; }
841 virtual const TypePtr *adr_type() const { return TypePtr::BOTTOM; }
842 // sparse accessors
843 // Fetch the previously stored "set_memory_at", or else the base memory.
844 // (Caller should clone it if it is a phi-nest.)
845 Node* memory_at(uint alias_idx) const;
846 // set the memory, regardless of its previous value
847 void set_memory_at(uint alias_idx, Node* n);
848 // the "base" is the memory that provides the non-finite support
849 Node* base_memory() const { return in(Compile::AliasIdxBot); }
850 // warning: setting the base can implicitly set any of the other slices too
851 void set_base_memory(Node* def);
852 // sentinel value which denotes a copy of the base memory:
853 Node* empty_memory() const { return in(Compile::AliasIdxTop); }
854 static Node* make_empty_memory(); // where the sentinel comes from
855 bool is_empty_memory(Node* n) const { assert((n == empty_memory()) == n->is_top(), "sanity"); return n->is_top(); }
856 // hook for the iterator, to perform any necessary setup
857 void iteration_setup(const MergeMemNode* other = NULL);
858 // push sentinels until I am at least as long as the other (semantic no-op)
859 void grow_to_match(const MergeMemNode* other);
860 bool verify_sparse() const PRODUCT_RETURN0;
861 #ifndef PRODUCT
862 virtual void dump_spec(outputStream *st) const;
863 #endif
864 };
866 class MergeMemStream : public StackObj {
867 private:
868 MergeMemNode* _mm;
869 const MergeMemNode* _mm2; // optional second guy, contributes non-empty iterations
870 Node* _mm_base; // loop-invariant base memory of _mm
871 int _idx;
872 int _cnt;
873 Node* _mem;
874 Node* _mem2;
875 int _cnt2;
877 void init(MergeMemNode* mm, const MergeMemNode* mm2 = NULL) {
878 // subsume_node will break sparseness at times, whenever a memory slice
879 // folds down to a copy of the base ("fat") memory. In such a case,
880 // the raw edge will update to base, although it should be top.
881 // This iterator will recognize either top or base_memory as an
882 // "empty" slice. See is_empty, is_empty2, and next below.
883 //
884 // The sparseness property is repaired in MergeMemNode::Ideal.
885 // As long as access to a MergeMem goes through this iterator
886 // or the memory_at accessor, flaws in the sparseness will
887 // never be observed.
888 //
889 // Also, iteration_setup repairs sparseness.
890 assert(mm->verify_sparse(), "please, no dups of base");
891 assert(mm2==NULL || mm2->verify_sparse(), "please, no dups of base");
893 _mm = mm;
894 _mm_base = mm->base_memory();
895 _mm2 = mm2;
896 _cnt = mm->req();
897 _idx = Compile::AliasIdxBot-1; // start at the base memory
898 _mem = NULL;
899 _mem2 = NULL;
900 }
902 #ifdef ASSERT
903 Node* check_memory() const {
904 if (at_base_memory())
905 return _mm->base_memory();
906 else if ((uint)_idx < _mm->req() && !_mm->in(_idx)->is_top())
907 return _mm->memory_at(_idx);
908 else
909 return _mm_base;
910 }
911 Node* check_memory2() const {
912 return at_base_memory()? _mm2->base_memory(): _mm2->memory_at(_idx);
913 }
914 #endif
916 static bool match_memory(Node* mem, const MergeMemNode* mm, int idx) PRODUCT_RETURN0;
917 void assert_synch() const {
918 assert(!_mem || _idx >= _cnt || match_memory(_mem, _mm, _idx),
919 "no side-effects except through the stream");
920 }
922 public:
924 // expected usages:
925 // for (MergeMemStream mms(mem->is_MergeMem()); next_non_empty(); ) { ... }
926 // for (MergeMemStream mms(mem1, mem2); next_non_empty2(); ) { ... }
928 // iterate over one merge
929 MergeMemStream(MergeMemNode* mm) {
930 mm->iteration_setup();
931 init(mm);
932 debug_only(_cnt2 = 999);
933 }
934 // iterate in parallel over two merges
935 // only iterates through non-empty elements of mm2
936 MergeMemStream(MergeMemNode* mm, const MergeMemNode* mm2) {
937 assert(mm2, "second argument must be a MergeMem also");
938 ((MergeMemNode*)mm2)->iteration_setup(); // update hidden state
939 mm->iteration_setup(mm2);
940 init(mm, mm2);
941 _cnt2 = mm2->req();
942 }
943 #ifdef ASSERT
944 ~MergeMemStream() {
945 assert_synch();
946 }
947 #endif
949 MergeMemNode* all_memory() const {
950 return _mm;
951 }
952 Node* base_memory() const {
953 assert(_mm_base == _mm->base_memory(), "no update to base memory, please");
954 return _mm_base;
955 }
956 const MergeMemNode* all_memory2() const {
957 assert(_mm2 != NULL, "");
958 return _mm2;
959 }
960 bool at_base_memory() const {
961 return _idx == Compile::AliasIdxBot;
962 }
963 int alias_idx() const {
964 assert(_mem, "must call next 1st");
965 return _idx;
966 }
968 const TypePtr* adr_type() const {
969 return Compile::current()->get_adr_type(alias_idx());
970 }
972 const TypePtr* adr_type(Compile* C) const {
973 return C->get_adr_type(alias_idx());
974 }
975 bool is_empty() const {
976 assert(_mem, "must call next 1st");
977 assert(_mem->is_top() == (_mem==_mm->empty_memory()), "correct sentinel");
978 return _mem->is_top();
979 }
980 bool is_empty2() const {
981 assert(_mem2, "must call next 1st");
982 assert(_mem2->is_top() == (_mem2==_mm2->empty_memory()), "correct sentinel");
983 return _mem2->is_top();
984 }
985 Node* memory() const {
986 assert(!is_empty(), "must not be empty");
987 assert_synch();
988 return _mem;
989 }
990 // get the current memory, regardless of empty or non-empty status
991 Node* force_memory() const {
992 assert(!is_empty() || !at_base_memory(), "");
993 // Use _mm_base to defend against updates to _mem->base_memory().
994 Node *mem = _mem->is_top() ? _mm_base : _mem;
995 assert(mem == check_memory(), "");
996 return mem;
997 }
998 Node* memory2() const {
999 assert(_mem2 == check_memory2(), "");
1000 return _mem2;
1001 }
1002 void set_memory(Node* mem) {
1003 if (at_base_memory()) {
1004 // Note that this does not change the invariant _mm_base.
1005 _mm->set_base_memory(mem);
1006 } else {
1007 _mm->set_memory_at(_idx, mem);
1008 }
1009 _mem = mem;
1010 assert_synch();
1011 }
1013 // Recover from a side effect to the MergeMemNode.
1014 void set_memory() {
1015 _mem = _mm->in(_idx);
1016 }
1018 bool next() { return next(false); }
1019 bool next2() { return next(true); }
1021 bool next_non_empty() { return next_non_empty(false); }
1022 bool next_non_empty2() { return next_non_empty(true); }
1023 // next_non_empty2 can yield states where is_empty() is true
1025 private:
1026 // find the next item, which might be empty
1027 bool next(bool have_mm2) {
1028 assert((_mm2 != NULL) == have_mm2, "use other next");
1029 assert_synch();
1030 if (++_idx < _cnt) {
1031 // Note: This iterator allows _mm to be non-sparse.
1032 // It behaves the same whether _mem is top or base_memory.
1033 _mem = _mm->in(_idx);
1034 if (have_mm2)
1035 _mem2 = _mm2->in((_idx < _cnt2) ? _idx : Compile::AliasIdxTop);
1036 return true;
1037 }
1038 return false;
1039 }
1041 // find the next non-empty item
1042 bool next_non_empty(bool have_mm2) {
1043 while (next(have_mm2)) {
1044 if (!is_empty()) {
1045 // make sure _mem2 is filled in sensibly
1046 if (have_mm2 && _mem2->is_top()) _mem2 = _mm2->base_memory();
1047 return true;
1048 } else if (have_mm2 && !is_empty2()) {
1049 return true; // is_empty() == true
1050 }
1051 }
1052 return false;
1053 }
1054 };
1056 //------------------------------Prefetch---------------------------------------
1058 // Non-faulting prefetch load. Prefetch for many reads.
1059 class PrefetchReadNode : public Node {
1060 public:
1061 PrefetchReadNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
1062 virtual int Opcode() const;
1063 virtual uint ideal_reg() const { return NotAMachineReg; }
1064 virtual uint match_edge(uint idx) const { return idx==2; }
1065 virtual const Type *bottom_type() const { return Type::ABIO; }
1066 };
1068 // Non-faulting prefetch load. Prefetch for many reads & many writes.
1069 class PrefetchWriteNode : public Node {
1070 public:
1071 PrefetchWriteNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
1072 virtual int Opcode() const;
1073 virtual uint ideal_reg() const { return NotAMachineReg; }
1074 virtual uint match_edge(uint idx) const { return idx==2; }
1075 virtual const Type *bottom_type() const { return Type::ABIO; }
1076 };