Fri, 14 Mar 2008 15:26:33 -0700
6674588: (Escape Analysis) Improve Escape Analysis code
Summary: Current EA code has several problems which have to be fixed.
Reviewed-by: jrose, sgoldman
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 // This one should probably be a phase-specific function:
71 static bool detect_dominating_control(Node* dom, Node* sub);
73 // Is this Node a MemNode or some descendent? Default is YES.
74 virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp );
76 virtual const class TypePtr *adr_type() const; // returns bottom_type of address
78 // Shared code for Ideal methods:
79 Node *Ideal_common(PhaseGVN *phase, bool can_reshape); // Return -1 for short-circuit NULL.
81 // Helper function for adr_type() implementations.
82 static const TypePtr* calculate_adr_type(const Type* t, const TypePtr* cross_check = NULL);
84 // Raw access function, to allow copying of adr_type efficiently in
85 // product builds and retain the debug info for debug builds.
86 const TypePtr *raw_adr_type() const {
87 #ifdef ASSERT
88 return _adr_type;
89 #else
90 return 0;
91 #endif
92 }
94 // Map a load or store opcode to its corresponding store opcode.
95 // (Return -1 if unknown.)
96 virtual int store_Opcode() const { return -1; }
98 // What is the type of the value in memory? (T_VOID mean "unspecified".)
99 virtual BasicType memory_type() const = 0;
100 virtual int memory_size() const {
101 #ifdef ASSERT
102 return type2aelembytes(memory_type(), true);
103 #else
104 return type2aelembytes(memory_type());
105 #endif
106 }
108 // Search through memory states which precede this node (load or store).
109 // Look for an exact match for the address, with no intervening
110 // aliased stores.
111 Node* find_previous_store(PhaseTransform* phase);
113 // Can this node (load or store) accurately see a stored value in
114 // the given memory state? (The state may or may not be in(Memory).)
115 Node* can_see_stored_value(Node* st, PhaseTransform* phase) const;
117 #ifndef PRODUCT
118 static void dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st);
119 virtual void dump_spec(outputStream *st) const;
120 #endif
121 };
123 //------------------------------LoadNode---------------------------------------
124 // Load value; requires Memory and Address
125 class LoadNode : public MemNode {
126 protected:
127 virtual uint cmp( const Node &n ) const;
128 virtual uint size_of() const; // Size is bigger
129 const Type* const _type; // What kind of value is loaded?
130 public:
132 LoadNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt )
133 : MemNode(c,mem,adr,at), _type(rt) {
134 init_class_id(Class_Load);
135 }
137 // Polymorphic factory method:
138 static LoadNode* make( Compile *C, Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt, BasicType bt );
140 virtual uint hash() const; // Check the type
142 // Handle algebraic identities here. If we have an identity, return the Node
143 // we are equivalent to. We look for Load of a Store.
144 virtual Node *Identity( PhaseTransform *phase );
146 // If the load is from Field memory and the pointer is non-null, we can
147 // zero out the control input.
148 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
150 // Recover original value from boxed values
151 Node *eliminate_autobox(PhaseGVN *phase);
153 // Compute a new Type for this node. Basically we just do the pre-check,
154 // then call the virtual add() to set the type.
155 virtual const Type *Value( PhaseTransform *phase ) const;
157 virtual uint ideal_reg() const;
158 virtual const Type *bottom_type() const;
159 // Following method is copied from TypeNode:
160 void set_type(const Type* t) {
161 assert(t != NULL, "sanity");
162 debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH);
163 *(const Type**)&_type = t; // cast away const-ness
164 // If this node is in the hash table, make sure it doesn't need a rehash.
165 assert(check_hash == NO_HASH || check_hash == hash(), "type change must preserve hash code");
166 }
167 const Type* type() const { assert(_type != NULL, "sanity"); return _type; };
169 // Do not match memory edge
170 virtual uint match_edge(uint idx) const;
172 // Map a load opcode to its corresponding store opcode.
173 virtual int store_Opcode() const = 0;
175 // Check if the load's memory input is a Phi node with the same control.
176 bool is_instance_field_load_with_local_phi(Node* ctrl);
178 #ifndef PRODUCT
179 virtual void dump_spec(outputStream *st) const;
180 #endif
181 protected:
182 const Type* load_array_final_field(const TypeKlassPtr *tkls,
183 ciKlass* klass) const;
184 };
186 //------------------------------LoadBNode--------------------------------------
187 // Load a byte (8bits signed) from memory
188 class LoadBNode : public LoadNode {
189 public:
190 LoadBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::BYTE )
191 : LoadNode(c,mem,adr,at,ti) {}
192 virtual int Opcode() const;
193 virtual uint ideal_reg() const { return Op_RegI; }
194 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
195 virtual int store_Opcode() const { return Op_StoreB; }
196 virtual BasicType memory_type() const { return T_BYTE; }
197 };
199 //------------------------------LoadCNode--------------------------------------
200 // Load a char (16bits unsigned) from memory
201 class LoadCNode : public LoadNode {
202 public:
203 LoadCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::CHAR )
204 : LoadNode(c,mem,adr,at,ti) {}
205 virtual int Opcode() const;
206 virtual uint ideal_reg() const { return Op_RegI; }
207 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
208 virtual int store_Opcode() const { return Op_StoreC; }
209 virtual BasicType memory_type() const { return T_CHAR; }
210 };
212 //------------------------------LoadINode--------------------------------------
213 // Load an integer from memory
214 class LoadINode : public LoadNode {
215 public:
216 LoadINode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::INT )
217 : LoadNode(c,mem,adr,at,ti) {}
218 virtual int Opcode() const;
219 virtual uint ideal_reg() const { return Op_RegI; }
220 virtual int store_Opcode() const { return Op_StoreI; }
221 virtual BasicType memory_type() const { return T_INT; }
222 };
224 //------------------------------LoadRangeNode----------------------------------
225 // Load an array length from the array
226 class LoadRangeNode : public LoadINode {
227 public:
228 LoadRangeNode( Node *c, Node *mem, Node *adr, const TypeInt *ti = TypeInt::POS )
229 : LoadINode(c,mem,adr,TypeAryPtr::RANGE,ti) {}
230 virtual int Opcode() const;
231 virtual const Type *Value( PhaseTransform *phase ) const;
232 virtual Node *Identity( PhaseTransform *phase );
233 };
235 //------------------------------LoadLNode--------------------------------------
236 // Load a long from memory
237 class LoadLNode : public LoadNode {
238 virtual uint hash() const { return LoadNode::hash() + _require_atomic_access; }
239 virtual uint cmp( const Node &n ) const {
240 return _require_atomic_access == ((LoadLNode&)n)._require_atomic_access
241 && LoadNode::cmp(n);
242 }
243 virtual uint size_of() const { return sizeof(*this); }
244 const bool _require_atomic_access; // is piecewise load forbidden?
246 public:
247 LoadLNode( Node *c, Node *mem, Node *adr, const TypePtr* at,
248 const TypeLong *tl = TypeLong::LONG,
249 bool require_atomic_access = false )
250 : LoadNode(c,mem,adr,at,tl)
251 , _require_atomic_access(require_atomic_access)
252 {}
253 virtual int Opcode() const;
254 virtual uint ideal_reg() const { return Op_RegL; }
255 virtual int store_Opcode() const { return Op_StoreL; }
256 virtual BasicType memory_type() const { return T_LONG; }
257 bool require_atomic_access() { return _require_atomic_access; }
258 static LoadLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt);
259 #ifndef PRODUCT
260 virtual void dump_spec(outputStream *st) const {
261 LoadNode::dump_spec(st);
262 if (_require_atomic_access) st->print(" Atomic!");
263 }
264 #endif
265 };
267 //------------------------------LoadL_unalignedNode----------------------------
268 // Load a long from unaligned memory
269 class LoadL_unalignedNode : public LoadLNode {
270 public:
271 LoadL_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at )
272 : LoadLNode(c,mem,adr,at) {}
273 virtual int Opcode() const;
274 };
276 //------------------------------LoadFNode--------------------------------------
277 // Load a float (64 bits) from memory
278 class LoadFNode : public LoadNode {
279 public:
280 LoadFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::FLOAT )
281 : LoadNode(c,mem,adr,at,t) {}
282 virtual int Opcode() const;
283 virtual uint ideal_reg() const { return Op_RegF; }
284 virtual int store_Opcode() const { return Op_StoreF; }
285 virtual BasicType memory_type() const { return T_FLOAT; }
286 };
288 //------------------------------LoadDNode--------------------------------------
289 // Load a double (64 bits) from memory
290 class LoadDNode : public LoadNode {
291 public:
292 LoadDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t = Type::DOUBLE )
293 : LoadNode(c,mem,adr,at,t) {}
294 virtual int Opcode() const;
295 virtual uint ideal_reg() const { return Op_RegD; }
296 virtual int store_Opcode() const { return Op_StoreD; }
297 virtual BasicType memory_type() const { return T_DOUBLE; }
298 };
300 //------------------------------LoadD_unalignedNode----------------------------
301 // Load a double from unaligned memory
302 class LoadD_unalignedNode : public LoadDNode {
303 public:
304 LoadD_unalignedNode( Node *c, Node *mem, Node *adr, const TypePtr* at )
305 : LoadDNode(c,mem,adr,at) {}
306 virtual int Opcode() const;
307 };
309 //------------------------------LoadPNode--------------------------------------
310 // Load a pointer from memory (either object or array)
311 class LoadPNode : public LoadNode {
312 public:
313 LoadPNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypePtr* t )
314 : LoadNode(c,mem,adr,at,t) {}
315 virtual int Opcode() const;
316 virtual uint ideal_reg() const { return Op_RegP; }
317 virtual int store_Opcode() const { return Op_StoreP; }
318 virtual BasicType memory_type() const { return T_ADDRESS; }
319 // depends_only_on_test is almost always true, and needs to be almost always
320 // true to enable key hoisting & commoning optimizations. However, for the
321 // special case of RawPtr loads from TLS top & end, the control edge carries
322 // the dependence preventing hoisting past a Safepoint instead of the memory
323 // edge. (An unfortunate consequence of having Safepoints not set Raw
324 // Memory; itself an unfortunate consequence of having Nodes which produce
325 // results (new raw memory state) inside of loops preventing all manner of
326 // other optimizations). Basically, it's ugly but so is the alternative.
327 // See comment in macro.cpp, around line 125 expand_allocate_common().
328 virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
329 };
331 //------------------------------LoadKlassNode----------------------------------
332 // Load a Klass from an object
333 class LoadKlassNode : public LoadPNode {
334 public:
335 LoadKlassNode( Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeKlassPtr *tk = TypeKlassPtr::OBJECT )
336 : LoadPNode(c,mem,adr,at,tk) {}
337 virtual int Opcode() const;
338 virtual const Type *Value( PhaseTransform *phase ) const;
339 virtual Node *Identity( PhaseTransform *phase );
340 virtual bool depends_only_on_test() const { return true; }
341 };
343 //------------------------------LoadSNode--------------------------------------
344 // Load a short (16bits signed) from memory
345 class LoadSNode : public LoadNode {
346 public:
347 LoadSNode( Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti = TypeInt::SHORT )
348 : LoadNode(c,mem,adr,at,ti) {}
349 virtual int Opcode() const;
350 virtual uint ideal_reg() const { return Op_RegI; }
351 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
352 virtual int store_Opcode() const { return Op_StoreC; }
353 virtual BasicType memory_type() const { return T_SHORT; }
354 };
356 //------------------------------StoreNode--------------------------------------
357 // Store value; requires Store, Address and Value
358 class StoreNode : public MemNode {
359 protected:
360 virtual uint cmp( const Node &n ) const;
361 virtual bool depends_only_on_test() const { return false; }
363 Node *Ideal_masked_input (PhaseGVN *phase, uint mask);
364 Node *Ideal_sign_extended_input(PhaseGVN *phase, int num_bits);
366 public:
367 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val )
368 : MemNode(c,mem,adr,at,val) {
369 init_class_id(Class_Store);
370 }
371 StoreNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store )
372 : MemNode(c,mem,adr,at,val,oop_store) {
373 init_class_id(Class_Store);
374 }
376 // Polymorphic factory method:
377 static StoreNode* make( Compile *C, Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, BasicType bt );
379 virtual uint hash() const; // Check the type
381 // If the store is to Field memory and the pointer is non-null, we can
382 // zero out the control input.
383 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
385 // Compute a new Type for this node. Basically we just do the pre-check,
386 // then call the virtual add() to set the type.
387 virtual const Type *Value( PhaseTransform *phase ) const;
389 // Check for identity function on memory (Load then Store at same address)
390 virtual Node *Identity( PhaseTransform *phase );
392 // Do not match memory edge
393 virtual uint match_edge(uint idx) const;
395 virtual const Type *bottom_type() const; // returns Type::MEMORY
397 // Map a store opcode to its corresponding own opcode, trivially.
398 virtual int store_Opcode() const { return Opcode(); }
400 // have all possible loads of the value stored been optimized away?
401 bool value_never_loaded(PhaseTransform *phase) const;
402 };
404 //------------------------------StoreBNode-------------------------------------
405 // Store byte to memory
406 class StoreBNode : public StoreNode {
407 public:
408 StoreBNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
409 virtual int Opcode() const;
410 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
411 virtual BasicType memory_type() const { return T_BYTE; }
412 };
414 //------------------------------StoreCNode-------------------------------------
415 // Store char/short to memory
416 class StoreCNode : public StoreNode {
417 public:
418 StoreCNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
419 virtual int Opcode() const;
420 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
421 virtual BasicType memory_type() const { return T_CHAR; }
422 };
424 //------------------------------StoreINode-------------------------------------
425 // Store int to memory
426 class StoreINode : public StoreNode {
427 public:
428 StoreINode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
429 virtual int Opcode() const;
430 virtual BasicType memory_type() const { return T_INT; }
431 };
433 //------------------------------StoreLNode-------------------------------------
434 // Store long to memory
435 class StoreLNode : public StoreNode {
436 virtual uint hash() const { return StoreNode::hash() + _require_atomic_access; }
437 virtual uint cmp( const Node &n ) const {
438 return _require_atomic_access == ((StoreLNode&)n)._require_atomic_access
439 && StoreNode::cmp(n);
440 }
441 virtual uint size_of() const { return sizeof(*this); }
442 const bool _require_atomic_access; // is piecewise store forbidden?
444 public:
445 StoreLNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val,
446 bool require_atomic_access = false )
447 : StoreNode(c,mem,adr,at,val)
448 , _require_atomic_access(require_atomic_access)
449 {}
450 virtual int Opcode() const;
451 virtual BasicType memory_type() const { return T_LONG; }
452 bool require_atomic_access() { return _require_atomic_access; }
453 static StoreLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val);
454 #ifndef PRODUCT
455 virtual void dump_spec(outputStream *st) const {
456 StoreNode::dump_spec(st);
457 if (_require_atomic_access) st->print(" Atomic!");
458 }
459 #endif
460 };
462 //------------------------------StoreFNode-------------------------------------
463 // Store float to memory
464 class StoreFNode : public StoreNode {
465 public:
466 StoreFNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
467 virtual int Opcode() const;
468 virtual BasicType memory_type() const { return T_FLOAT; }
469 };
471 //------------------------------StoreDNode-------------------------------------
472 // Store double to memory
473 class StoreDNode : public StoreNode {
474 public:
475 StoreDNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val ) : StoreNode(c,mem,adr,at,val) {}
476 virtual int Opcode() const;
477 virtual BasicType memory_type() const { return T_DOUBLE; }
478 };
480 //------------------------------StorePNode-------------------------------------
481 // Store pointer to memory
482 class StorePNode : public StoreNode {
483 public:
484 StorePNode( 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_ADDRESS; }
487 };
489 //------------------------------StoreCMNode-----------------------------------
490 // Store card-mark byte to memory for CM
491 // The last StoreCM before a SafePoint must be preserved and occur after its "oop" store
492 // Preceeding equivalent StoreCMs may be eliminated.
493 class StoreCMNode : public StoreNode {
494 public:
495 StoreCMNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store ) : StoreNode(c,mem,adr,at,val,oop_store) {}
496 virtual int Opcode() const;
497 virtual Node *Identity( PhaseTransform *phase );
498 virtual const Type *Value( PhaseTransform *phase ) const;
499 virtual BasicType memory_type() const { return T_VOID; } // unspecific
500 };
502 //------------------------------LoadPLockedNode---------------------------------
503 // Load-locked a pointer from memory (either object or array).
504 // On Sparc & Intel this is implemented as a normal pointer load.
505 // On PowerPC and friends it's a real load-locked.
506 class LoadPLockedNode : public LoadPNode {
507 public:
508 LoadPLockedNode( Node *c, Node *mem, Node *adr )
509 : LoadPNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) {}
510 virtual int Opcode() const;
511 virtual int store_Opcode() const { return Op_StorePConditional; }
512 virtual bool depends_only_on_test() const { return true; }
513 };
515 //------------------------------LoadLLockedNode---------------------------------
516 // Load-locked a pointer from memory (either object or array).
517 // On Sparc & Intel this is implemented as a normal long load.
518 class LoadLLockedNode : public LoadLNode {
519 public:
520 LoadLLockedNode( Node *c, Node *mem, Node *adr )
521 : LoadLNode(c,mem,adr,TypeRawPtr::BOTTOM, TypeLong::LONG) {}
522 virtual int Opcode() const;
523 virtual int store_Opcode() const { return Op_StoreLConditional; }
524 };
526 //------------------------------SCMemProjNode---------------------------------------
527 // This class defines a projection of the memory state of a store conditional node.
528 // These nodes return a value, but also update memory.
529 class SCMemProjNode : public ProjNode {
530 public:
531 enum {SCMEMPROJCON = (uint)-2};
532 SCMemProjNode( Node *src) : ProjNode( src, SCMEMPROJCON) { }
533 virtual int Opcode() const;
534 virtual bool is_CFG() const { return false; }
535 virtual const Type *bottom_type() const {return Type::MEMORY;}
536 virtual const TypePtr *adr_type() const { return in(0)->in(MemNode::Memory)->adr_type();}
537 virtual uint ideal_reg() const { return 0;} // memory projections don't have a register
538 virtual const Type *Value( PhaseTransform *phase ) const;
539 #ifndef PRODUCT
540 virtual void dump_spec(outputStream *st) const {};
541 #endif
542 };
544 //------------------------------LoadStoreNode---------------------------
545 class LoadStoreNode : public Node {
546 public:
547 enum {
548 ExpectedIn = MemNode::ValueIn+1 // One more input than MemNode
549 };
550 LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex);
551 virtual bool depends_only_on_test() const { return false; }
552 virtual const Type *bottom_type() const { return TypeInt::BOOL; }
553 virtual uint ideal_reg() const { return Op_RegI; }
554 virtual uint match_edge(uint idx) const { return idx == MemNode::Address || idx == MemNode::ValueIn; }
555 };
557 //------------------------------StorePConditionalNode---------------------------
558 // Conditionally store pointer to memory, if no change since prior
559 // load-locked. Sets flags for success or failure of the store.
560 class StorePConditionalNode : public LoadStoreNode {
561 public:
562 StorePConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { }
563 virtual int Opcode() const;
564 // Produces flags
565 virtual uint ideal_reg() const { return Op_RegFlags; }
566 };
568 //------------------------------StoreLConditionalNode---------------------------
569 // Conditionally store long to memory, if no change since prior
570 // load-locked. Sets flags for success or failure of the store.
571 class StoreLConditionalNode : public LoadStoreNode {
572 public:
573 StoreLConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreNode(c, mem, adr, val, ll) { }
574 virtual int Opcode() const;
575 };
578 //------------------------------CompareAndSwapLNode---------------------------
579 class CompareAndSwapLNode : public LoadStoreNode {
580 public:
581 CompareAndSwapLNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
582 virtual int Opcode() const;
583 };
586 //------------------------------CompareAndSwapINode---------------------------
587 class CompareAndSwapINode : public LoadStoreNode {
588 public:
589 CompareAndSwapINode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
590 virtual int Opcode() const;
591 };
594 //------------------------------CompareAndSwapPNode---------------------------
595 class CompareAndSwapPNode : public LoadStoreNode {
596 public:
597 CompareAndSwapPNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreNode(c, mem, adr, val, ex) { }
598 virtual int Opcode() const;
599 };
601 //------------------------------ClearArray-------------------------------------
602 class ClearArrayNode: public Node {
603 public:
604 ClearArrayNode( Node *ctrl, Node *arymem, Node *word_cnt, Node *base ) : Node(ctrl,arymem,word_cnt,base) {}
605 virtual int Opcode() const;
606 virtual const Type *bottom_type() const { return Type::MEMORY; }
607 // ClearArray modifies array elements, and so affects only the
608 // array memory addressed by the bottom_type of its base address.
609 virtual const class TypePtr *adr_type() const;
610 virtual Node *Identity( PhaseTransform *phase );
611 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
612 virtual uint match_edge(uint idx) const;
614 // Clear the given area of an object or array.
615 // The start offset must always be aligned mod BytesPerInt.
616 // The end offset must always be aligned mod BytesPerLong.
617 // Return the new memory.
618 static Node* clear_memory(Node* control, Node* mem, Node* dest,
619 intptr_t start_offset,
620 intptr_t end_offset,
621 PhaseGVN* phase);
622 static Node* clear_memory(Node* control, Node* mem, Node* dest,
623 intptr_t start_offset,
624 Node* end_offset,
625 PhaseGVN* phase);
626 static Node* clear_memory(Node* control, Node* mem, Node* dest,
627 Node* start_offset,
628 Node* end_offset,
629 PhaseGVN* phase);
630 };
632 //------------------------------StrComp-------------------------------------
633 class StrCompNode: public Node {
634 public:
635 StrCompNode(Node *control,
636 Node* char_array_mem,
637 Node* value_mem,
638 Node* count_mem,
639 Node* offset_mem,
640 Node* s1, Node* s2): Node(control,
641 char_array_mem,
642 value_mem,
643 count_mem,
644 offset_mem,
645 s1, s2) {};
646 virtual int Opcode() const;
647 virtual bool depends_only_on_test() const { return false; }
648 virtual const Type* bottom_type() const { return TypeInt::INT; }
649 // a StrCompNode (conservatively) aliases with everything:
650 virtual const TypePtr* adr_type() const { return TypePtr::BOTTOM; }
651 virtual uint match_edge(uint idx) const;
652 virtual uint ideal_reg() const { return Op_RegI; }
653 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
654 };
656 //------------------------------MemBar-----------------------------------------
657 // There are different flavors of Memory Barriers to match the Java Memory
658 // Model. Monitor-enter and volatile-load act as Aquires: no following ref
659 // can be moved to before them. We insert a MemBar-Acquire after a FastLock or
660 // volatile-load. Monitor-exit and volatile-store act as Release: no
661 // preceeding ref can be moved to after them. We insert a MemBar-Release
662 // before a FastUnlock or volatile-store. All volatiles need to be
663 // serialized, so we follow all volatile-stores with a MemBar-Volatile to
664 // seperate it from any following volatile-load.
665 class MemBarNode: public MultiNode {
666 virtual uint hash() const ; // { return NO_HASH; }
667 virtual uint cmp( const Node &n ) const ; // Always fail, except on self
669 virtual uint size_of() const { return sizeof(*this); }
670 // Memory type this node is serializing. Usually either rawptr or bottom.
671 const TypePtr* _adr_type;
673 public:
674 enum {
675 Precedent = TypeFunc::Parms // optional edge to force precedence
676 };
677 MemBarNode(Compile* C, int alias_idx, Node* precedent);
678 virtual int Opcode() const = 0;
679 virtual const class TypePtr *adr_type() const { return _adr_type; }
680 virtual const Type *Value( PhaseTransform *phase ) const;
681 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
682 virtual uint match_edge(uint idx) const { return 0; }
683 virtual const Type *bottom_type() const { return TypeTuple::MEMBAR; }
684 virtual Node *match( const ProjNode *proj, const Matcher *m );
685 // Factory method. Builds a wide or narrow membar.
686 // Optional 'precedent' becomes an extra edge if not null.
687 static MemBarNode* make(Compile* C, int opcode,
688 int alias_idx = Compile::AliasIdxBot,
689 Node* precedent = NULL);
690 };
692 // "Acquire" - no following ref can move before (but earlier refs can
693 // follow, like an early Load stalled in cache). Requires multi-cpu
694 // visibility. Inserted after a volatile load or FastLock.
695 class MemBarAcquireNode: public MemBarNode {
696 public:
697 MemBarAcquireNode(Compile* C, int alias_idx, Node* precedent)
698 : MemBarNode(C, alias_idx, precedent) {}
699 virtual int Opcode() const;
700 };
702 // "Release" - no earlier ref can move after (but later refs can move
703 // up, like a speculative pipelined cache-hitting Load). Requires
704 // multi-cpu visibility. Inserted before a volatile store or FastUnLock.
705 class MemBarReleaseNode: public MemBarNode {
706 public:
707 MemBarReleaseNode(Compile* C, int alias_idx, Node* precedent)
708 : MemBarNode(C, alias_idx, precedent) {}
709 virtual int Opcode() const;
710 };
712 // Ordering between a volatile store and a following volatile load.
713 // Requires multi-CPU visibility?
714 class MemBarVolatileNode: public MemBarNode {
715 public:
716 MemBarVolatileNode(Compile* C, int alias_idx, Node* precedent)
717 : MemBarNode(C, alias_idx, precedent) {}
718 virtual int Opcode() const;
719 };
721 // Ordering within the same CPU. Used to order unsafe memory references
722 // inside the compiler when we lack alias info. Not needed "outside" the
723 // compiler because the CPU does all the ordering for us.
724 class MemBarCPUOrderNode: public MemBarNode {
725 public:
726 MemBarCPUOrderNode(Compile* C, int alias_idx, Node* precedent)
727 : MemBarNode(C, alias_idx, precedent) {}
728 virtual int Opcode() const;
729 virtual uint ideal_reg() const { return 0; } // not matched in the AD file
730 };
732 // Isolation of object setup after an AllocateNode and before next safepoint.
733 // (See comment in memnode.cpp near InitializeNode::InitializeNode for semantics.)
734 class InitializeNode: public MemBarNode {
735 friend class AllocateNode;
737 bool _is_complete;
739 public:
740 enum {
741 Control = TypeFunc::Control,
742 Memory = TypeFunc::Memory, // MergeMem for states affected by this op
743 RawAddress = TypeFunc::Parms+0, // the newly-allocated raw address
744 RawStores = TypeFunc::Parms+1 // zero or more stores (or TOP)
745 };
747 InitializeNode(Compile* C, int adr_type, Node* rawoop);
748 virtual int Opcode() const;
749 virtual uint size_of() const { return sizeof(*this); }
750 virtual uint ideal_reg() const { return 0; } // not matched in the AD file
751 virtual const RegMask &in_RegMask(uint) const; // mask for RawAddress
753 // Manage incoming memory edges via a MergeMem on in(Memory):
754 Node* memory(uint alias_idx);
756 // The raw memory edge coming directly from the Allocation.
757 // The contents of this memory are *always* all-zero-bits.
758 Node* zero_memory() { return memory(Compile::AliasIdxRaw); }
760 // Return the corresponding allocation for this initialization (or null if none).
761 // (Note: Both InitializeNode::allocation and AllocateNode::initialization
762 // are defined in graphKit.cpp, which sets up the bidirectional relation.)
763 AllocateNode* allocation();
765 // Anything other than zeroing in this init?
766 bool is_non_zero();
768 // An InitializeNode must completed before macro expansion is done.
769 // Completion requires that the AllocateNode must be followed by
770 // initialization of the new memory to zero, then to any initializers.
771 bool is_complete() { return _is_complete; }
773 // Mark complete. (Must not yet be complete.)
774 void set_complete(PhaseGVN* phase);
776 #ifdef ASSERT
777 // ensure all non-degenerate stores are ordered and non-overlapping
778 bool stores_are_sane(PhaseTransform* phase);
779 #endif //ASSERT
781 // See if this store can be captured; return offset where it initializes.
782 // Return 0 if the store cannot be moved (any sort of problem).
783 intptr_t can_capture_store(StoreNode* st, PhaseTransform* phase);
785 // Capture another store; reformat it to write my internal raw memory.
786 // Return the captured copy, else NULL if there is some sort of problem.
787 Node* capture_store(StoreNode* st, intptr_t start, PhaseTransform* phase);
789 // Find captured store which corresponds to the range [start..start+size).
790 // Return my own memory projection (meaning the initial zero bits)
791 // if there is no such store. Return NULL if there is a problem.
792 Node* find_captured_store(intptr_t start, int size_in_bytes, PhaseTransform* phase);
794 // Called when the associated AllocateNode is expanded into CFG.
795 Node* complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
796 intptr_t header_size, Node* size_in_bytes,
797 PhaseGVN* phase);
799 private:
800 void remove_extra_zeroes();
802 // Find out where a captured store should be placed (or already is placed).
803 int captured_store_insertion_point(intptr_t start, int size_in_bytes,
804 PhaseTransform* phase);
806 static intptr_t get_store_offset(Node* st, PhaseTransform* phase);
808 Node* make_raw_address(intptr_t offset, PhaseTransform* phase);
810 bool detect_init_independence(Node* n, bool st_is_pinned, int& count);
812 void coalesce_subword_stores(intptr_t header_size, Node* size_in_bytes,
813 PhaseGVN* phase);
815 intptr_t find_next_fullword_store(uint i, PhaseGVN* phase);
816 };
818 //------------------------------MergeMem---------------------------------------
819 // (See comment in memnode.cpp near MergeMemNode::MergeMemNode for semantics.)
820 class MergeMemNode: public Node {
821 virtual uint hash() const ; // { return NO_HASH; }
822 virtual uint cmp( const Node &n ) const ; // Always fail, except on self
823 friend class MergeMemStream;
824 MergeMemNode(Node* def); // clients use MergeMemNode::make
826 public:
827 // If the input is a whole memory state, clone it with all its slices intact.
828 // Otherwise, make a new memory state with just that base memory input.
829 // In either case, the result is a newly created MergeMem.
830 static MergeMemNode* make(Compile* C, Node* base_memory);
832 virtual int Opcode() const;
833 virtual Node *Identity( PhaseTransform *phase );
834 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
835 virtual uint ideal_reg() const { return NotAMachineReg; }
836 virtual uint match_edge(uint idx) const { return 0; }
837 virtual const RegMask &out_RegMask() const;
838 virtual const Type *bottom_type() const { return Type::MEMORY; }
839 virtual const TypePtr *adr_type() const { return TypePtr::BOTTOM; }
840 // sparse accessors
841 // Fetch the previously stored "set_memory_at", or else the base memory.
842 // (Caller should clone it if it is a phi-nest.)
843 Node* memory_at(uint alias_idx) const;
844 // set the memory, regardless of its previous value
845 void set_memory_at(uint alias_idx, Node* n);
846 // the "base" is the memory that provides the non-finite support
847 Node* base_memory() const { return in(Compile::AliasIdxBot); }
848 // warning: setting the base can implicitly set any of the other slices too
849 void set_base_memory(Node* def);
850 // sentinel value which denotes a copy of the base memory:
851 Node* empty_memory() const { return in(Compile::AliasIdxTop); }
852 static Node* make_empty_memory(); // where the sentinel comes from
853 bool is_empty_memory(Node* n) const { assert((n == empty_memory()) == n->is_top(), "sanity"); return n->is_top(); }
854 // hook for the iterator, to perform any necessary setup
855 void iteration_setup(const MergeMemNode* other = NULL);
856 // push sentinels until I am at least as long as the other (semantic no-op)
857 void grow_to_match(const MergeMemNode* other);
858 bool verify_sparse() const PRODUCT_RETURN0;
859 #ifndef PRODUCT
860 virtual void dump_spec(outputStream *st) const;
861 #endif
862 };
864 class MergeMemStream : public StackObj {
865 private:
866 MergeMemNode* _mm;
867 const MergeMemNode* _mm2; // optional second guy, contributes non-empty iterations
868 Node* _mm_base; // loop-invariant base memory of _mm
869 int _idx;
870 int _cnt;
871 Node* _mem;
872 Node* _mem2;
873 int _cnt2;
875 void init(MergeMemNode* mm, const MergeMemNode* mm2 = NULL) {
876 // subsume_node will break sparseness at times, whenever a memory slice
877 // folds down to a copy of the base ("fat") memory. In such a case,
878 // the raw edge will update to base, although it should be top.
879 // This iterator will recognize either top or base_memory as an
880 // "empty" slice. See is_empty, is_empty2, and next below.
881 //
882 // The sparseness property is repaired in MergeMemNode::Ideal.
883 // As long as access to a MergeMem goes through this iterator
884 // or the memory_at accessor, flaws in the sparseness will
885 // never be observed.
886 //
887 // Also, iteration_setup repairs sparseness.
888 assert(mm->verify_sparse(), "please, no dups of base");
889 assert(mm2==NULL || mm2->verify_sparse(), "please, no dups of base");
891 _mm = mm;
892 _mm_base = mm->base_memory();
893 _mm2 = mm2;
894 _cnt = mm->req();
895 _idx = Compile::AliasIdxBot-1; // start at the base memory
896 _mem = NULL;
897 _mem2 = NULL;
898 }
900 #ifdef ASSERT
901 Node* check_memory() const {
902 if (at_base_memory())
903 return _mm->base_memory();
904 else if ((uint)_idx < _mm->req() && !_mm->in(_idx)->is_top())
905 return _mm->memory_at(_idx);
906 else
907 return _mm_base;
908 }
909 Node* check_memory2() const {
910 return at_base_memory()? _mm2->base_memory(): _mm2->memory_at(_idx);
911 }
912 #endif
914 static bool match_memory(Node* mem, const MergeMemNode* mm, int idx) PRODUCT_RETURN0;
915 void assert_synch() const {
916 assert(!_mem || _idx >= _cnt || match_memory(_mem, _mm, _idx),
917 "no side-effects except through the stream");
918 }
920 public:
922 // expected usages:
923 // for (MergeMemStream mms(mem->is_MergeMem()); next_non_empty(); ) { ... }
924 // for (MergeMemStream mms(mem1, mem2); next_non_empty2(); ) { ... }
926 // iterate over one merge
927 MergeMemStream(MergeMemNode* mm) {
928 mm->iteration_setup();
929 init(mm);
930 debug_only(_cnt2 = 999);
931 }
932 // iterate in parallel over two merges
933 // only iterates through non-empty elements of mm2
934 MergeMemStream(MergeMemNode* mm, const MergeMemNode* mm2) {
935 assert(mm2, "second argument must be a MergeMem also");
936 ((MergeMemNode*)mm2)->iteration_setup(); // update hidden state
937 mm->iteration_setup(mm2);
938 init(mm, mm2);
939 _cnt2 = mm2->req();
940 }
941 #ifdef ASSERT
942 ~MergeMemStream() {
943 assert_synch();
944 }
945 #endif
947 MergeMemNode* all_memory() const {
948 return _mm;
949 }
950 Node* base_memory() const {
951 assert(_mm_base == _mm->base_memory(), "no update to base memory, please");
952 return _mm_base;
953 }
954 const MergeMemNode* all_memory2() const {
955 assert(_mm2 != NULL, "");
956 return _mm2;
957 }
958 bool at_base_memory() const {
959 return _idx == Compile::AliasIdxBot;
960 }
961 int alias_idx() const {
962 assert(_mem, "must call next 1st");
963 return _idx;
964 }
966 const TypePtr* adr_type() const {
967 return Compile::current()->get_adr_type(alias_idx());
968 }
970 const TypePtr* adr_type(Compile* C) const {
971 return C->get_adr_type(alias_idx());
972 }
973 bool is_empty() const {
974 assert(_mem, "must call next 1st");
975 assert(_mem->is_top() == (_mem==_mm->empty_memory()), "correct sentinel");
976 return _mem->is_top();
977 }
978 bool is_empty2() const {
979 assert(_mem2, "must call next 1st");
980 assert(_mem2->is_top() == (_mem2==_mm2->empty_memory()), "correct sentinel");
981 return _mem2->is_top();
982 }
983 Node* memory() const {
984 assert(!is_empty(), "must not be empty");
985 assert_synch();
986 return _mem;
987 }
988 // get the current memory, regardless of empty or non-empty status
989 Node* force_memory() const {
990 assert(!is_empty() || !at_base_memory(), "");
991 // Use _mm_base to defend against updates to _mem->base_memory().
992 Node *mem = _mem->is_top() ? _mm_base : _mem;
993 assert(mem == check_memory(), "");
994 return mem;
995 }
996 Node* memory2() const {
997 assert(_mem2 == check_memory2(), "");
998 return _mem2;
999 }
1000 void set_memory(Node* mem) {
1001 if (at_base_memory()) {
1002 // Note that this does not change the invariant _mm_base.
1003 _mm->set_base_memory(mem);
1004 } else {
1005 _mm->set_memory_at(_idx, mem);
1006 }
1007 _mem = mem;
1008 assert_synch();
1009 }
1011 // Recover from a side effect to the MergeMemNode.
1012 void set_memory() {
1013 _mem = _mm->in(_idx);
1014 }
1016 bool next() { return next(false); }
1017 bool next2() { return next(true); }
1019 bool next_non_empty() { return next_non_empty(false); }
1020 bool next_non_empty2() { return next_non_empty(true); }
1021 // next_non_empty2 can yield states where is_empty() is true
1023 private:
1024 // find the next item, which might be empty
1025 bool next(bool have_mm2) {
1026 assert((_mm2 != NULL) == have_mm2, "use other next");
1027 assert_synch();
1028 if (++_idx < _cnt) {
1029 // Note: This iterator allows _mm to be non-sparse.
1030 // It behaves the same whether _mem is top or base_memory.
1031 _mem = _mm->in(_idx);
1032 if (have_mm2)
1033 _mem2 = _mm2->in((_idx < _cnt2) ? _idx : Compile::AliasIdxTop);
1034 return true;
1035 }
1036 return false;
1037 }
1039 // find the next non-empty item
1040 bool next_non_empty(bool have_mm2) {
1041 while (next(have_mm2)) {
1042 if (!is_empty()) {
1043 // make sure _mem2 is filled in sensibly
1044 if (have_mm2 && _mem2->is_top()) _mem2 = _mm2->base_memory();
1045 return true;
1046 } else if (have_mm2 && !is_empty2()) {
1047 return true; // is_empty() == true
1048 }
1049 }
1050 return false;
1051 }
1052 };
1054 //------------------------------Prefetch---------------------------------------
1056 // Non-faulting prefetch load. Prefetch for many reads.
1057 class PrefetchReadNode : public Node {
1058 public:
1059 PrefetchReadNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
1060 virtual int Opcode() const;
1061 virtual uint ideal_reg() const { return NotAMachineReg; }
1062 virtual uint match_edge(uint idx) const { return idx==2; }
1063 virtual const Type *bottom_type() const { return Type::ABIO; }
1064 };
1066 // Non-faulting prefetch load. Prefetch for many reads & many writes.
1067 class PrefetchWriteNode : public Node {
1068 public:
1069 PrefetchWriteNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
1070 virtual int Opcode() const;
1071 virtual uint ideal_reg() const { return NotAMachineReg; }
1072 virtual uint match_edge(uint idx) const { return idx==2; }
1073 virtual const Type *bottom_type() const { return Type::ABIO; }
1074 };