Wed, 05 Dec 2007 09:01:00 -0800
6395208: Elide autoboxing for calls to HashMap.get(int) and HashMap.get(long)
Reviewed-by: kvn, rasbold
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 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
32 //=============================================================================
33 uint MemNode::size_of() const { return sizeof(*this); }
35 const TypePtr *MemNode::adr_type() const {
36 Node* adr = in(Address);
37 const TypePtr* cross_check = NULL;
38 DEBUG_ONLY(cross_check = _adr_type);
39 return calculate_adr_type(adr->bottom_type(), cross_check);
40 }
42 #ifndef PRODUCT
43 void MemNode::dump_spec(outputStream *st) const {
44 if (in(Address) == NULL) return; // node is dead
45 #ifndef ASSERT
46 // fake the missing field
47 const TypePtr* _adr_type = NULL;
48 if (in(Address) != NULL)
49 _adr_type = in(Address)->bottom_type()->isa_ptr();
50 #endif
51 dump_adr_type(this, _adr_type, st);
53 Compile* C = Compile::current();
54 if( C->alias_type(_adr_type)->is_volatile() )
55 st->print(" Volatile!");
56 }
58 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
59 st->print(" @");
60 if (adr_type == NULL) {
61 st->print("NULL");
62 } else {
63 adr_type->dump_on(st);
64 Compile* C = Compile::current();
65 Compile::AliasType* atp = NULL;
66 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
67 if (atp == NULL)
68 st->print(", idx=?\?;");
69 else if (atp->index() == Compile::AliasIdxBot)
70 st->print(", idx=Bot;");
71 else if (atp->index() == Compile::AliasIdxTop)
72 st->print(", idx=Top;");
73 else if (atp->index() == Compile::AliasIdxRaw)
74 st->print(", idx=Raw;");
75 else {
76 ciField* field = atp->field();
77 if (field) {
78 st->print(", name=");
79 field->print_name_on(st);
80 }
81 st->print(", idx=%d;", atp->index());
82 }
83 }
84 }
86 extern void print_alias_types();
88 #endif
90 //--------------------------Ideal_common---------------------------------------
91 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
92 // Unhook non-raw memories from complete (macro-expanded) initializations.
93 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
94 // If our control input is a dead region, kill all below the region
95 Node *ctl = in(MemNode::Control);
96 if (ctl && remove_dead_region(phase, can_reshape))
97 return this;
99 // Ignore if memory is dead, or self-loop
100 Node *mem = in(MemNode::Memory);
101 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
102 assert( mem != this, "dead loop in MemNode::Ideal" );
104 Node *address = in(MemNode::Address);
105 const Type *t_adr = phase->type( address );
106 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
108 // Avoid independent memory operations
109 Node* old_mem = mem;
111 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
112 InitializeNode* init = mem->in(0)->as_Initialize();
113 if (init->is_complete()) { // i.e., after macro expansion
114 const TypePtr* tp = t_adr->is_ptr();
115 uint alias_idx = phase->C->get_alias_index(tp);
116 // Free this slice from the init. It was hooked, temporarily,
117 // by GraphKit::set_output_for_allocation.
118 if (alias_idx > Compile::AliasIdxRaw) {
119 mem = init->memory(alias_idx);
120 // ...but not with the raw-pointer slice.
121 }
122 }
123 }
125 if (mem->is_MergeMem()) {
126 MergeMemNode* mmem = mem->as_MergeMem();
127 const TypePtr *tp = t_adr->is_ptr();
128 uint alias_idx = phase->C->get_alias_index(tp);
129 #ifdef ASSERT
130 {
131 // Check that current type is consistent with the alias index used during graph construction
132 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
133 const TypePtr *adr_t = adr_type();
134 bool consistent = adr_t == NULL || adr_t->empty() || phase->C->must_alias(adr_t, alias_idx );
135 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
136 if( !consistent && adr_t != NULL && !adr_t->empty() &&
137 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
138 adr_t->isa_aryptr() && adr_t->offset() != Type::OffsetBot &&
139 ( adr_t->offset() == arrayOopDesc::length_offset_in_bytes() ||
140 adr_t->offset() == oopDesc::klass_offset_in_bytes() ||
141 adr_t->offset() == oopDesc::mark_offset_in_bytes() ) ) {
142 // don't assert if it is dead code.
143 consistent = true;
144 }
145 if( !consistent ) {
146 tty->print("alias_idx==%d, adr_type()==", alias_idx); if( adr_t == NULL ) { tty->print("NULL"); } else { adr_t->dump(); }
147 tty->cr();
148 print_alias_types();
149 assert(consistent, "adr_type must match alias idx");
150 }
151 }
152 #endif
153 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
154 // means an array I have not precisely typed yet. Do not do any
155 // alias stuff with it any time soon.
156 const TypeInstPtr *tinst = tp->isa_instptr();
157 if( tp->base() != Type::AnyPtr &&
158 !(tinst &&
159 tinst->klass()->is_java_lang_Object() &&
160 tinst->offset() == Type::OffsetBot) ) {
161 // compress paths and change unreachable cycles to TOP
162 // If not, we can update the input infinitely along a MergeMem cycle
163 // Equivalent code in PhiNode::Ideal
164 Node* m = phase->transform(mmem);
165 // If tranformed to a MergeMem, get the desired slice
166 // Otherwise the returned node represents memory for every slice
167 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
168 // Update input if it is progress over what we have now
169 }
170 }
172 if (mem != old_mem) {
173 set_req(MemNode::Memory, mem);
174 return this;
175 }
177 // let the subclass continue analyzing...
178 return NULL;
179 }
181 // Helper function for proving some simple control dominations.
182 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'.
183 // Already assumes that 'dom' is available at 'sub', and that 'sub'
184 // is not a constant (dominated by the method's StartNode).
185 // Used by MemNode::find_previous_store to prove that the
186 // control input of a memory operation predates (dominates)
187 // an allocation it wants to look past.
188 bool MemNode::detect_dominating_control(Node* dom, Node* sub) {
189 if (dom == NULL) return false;
190 if (dom->is_Proj()) dom = dom->in(0);
191 if (dom->is_Start()) return true; // anything inside the method
192 if (dom->is_Root()) return true; // dom 'controls' a constant
193 int cnt = 20; // detect cycle or too much effort
194 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
195 if (--cnt < 0) return false; // in a cycle or too complex
196 if (sub == dom) return true;
197 if (sub->is_Start()) return false;
198 if (sub->is_Root()) return false;
199 Node* up = sub->in(0);
200 if (sub == up && sub->is_Region()) {
201 for (uint i = 1; i < sub->req(); i++) {
202 Node* in = sub->in(i);
203 if (in != NULL && !in->is_top() && in != sub) {
204 up = in; break; // take any path on the way up to 'dom'
205 }
206 }
207 }
208 if (sub == up) return false; // some kind of tight cycle
209 sub = up;
210 }
211 return false;
212 }
214 //---------------------detect_ptr_independence---------------------------------
215 // Used by MemNode::find_previous_store to prove that two base
216 // pointers are never equal.
217 // The pointers are accompanied by their associated allocations,
218 // if any, which have been previously discovered by the caller.
219 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
220 Node* p2, AllocateNode* a2,
221 PhaseTransform* phase) {
222 // Attempt to prove that these two pointers cannot be aliased.
223 // They may both manifestly be allocations, and they should differ.
224 // Or, if they are not both allocations, they can be distinct constants.
225 // Otherwise, one is an allocation and the other a pre-existing value.
226 if (a1 == NULL && a2 == NULL) { // neither an allocation
227 return (p1 != p2) && p1->is_Con() && p2->is_Con();
228 } else if (a1 != NULL && a2 != NULL) { // both allocations
229 return (a1 != a2);
230 } else if (a1 != NULL) { // one allocation a1
231 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
232 return detect_dominating_control(p2->in(0), a1->in(0));
233 } else { //(a2 != NULL) // one allocation a2
234 return detect_dominating_control(p1->in(0), a2->in(0));
235 }
236 return false;
237 }
240 // The logic for reordering loads and stores uses four steps:
241 // (a) Walk carefully past stores and initializations which we
242 // can prove are independent of this load.
243 // (b) Observe that the next memory state makes an exact match
244 // with self (load or store), and locate the relevant store.
245 // (c) Ensure that, if we were to wire self directly to the store,
246 // the optimizer would fold it up somehow.
247 // (d) Do the rewiring, and return, depending on some other part of
248 // the optimizer to fold up the load.
249 // This routine handles steps (a) and (b). Steps (c) and (d) are
250 // specific to loads and stores, so they are handled by the callers.
251 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
252 //
253 Node* MemNode::find_previous_store(PhaseTransform* phase) {
254 Node* ctrl = in(MemNode::Control);
255 Node* adr = in(MemNode::Address);
256 intptr_t offset = 0;
257 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
258 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
260 if (offset == Type::OffsetBot)
261 return NULL; // cannot unalias unless there are precise offsets
263 intptr_t size_in_bytes = memory_size();
265 Node* mem = in(MemNode::Memory); // start searching here...
267 int cnt = 50; // Cycle limiter
268 for (;;) { // While we can dance past unrelated stores...
269 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
271 if (mem->is_Store()) {
272 Node* st_adr = mem->in(MemNode::Address);
273 intptr_t st_offset = 0;
274 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
275 if (st_base == NULL)
276 break; // inscrutable pointer
277 if (st_offset != offset && st_offset != Type::OffsetBot) {
278 const int MAX_STORE = BytesPerLong;
279 if (st_offset >= offset + size_in_bytes ||
280 st_offset <= offset - MAX_STORE ||
281 st_offset <= offset - mem->as_Store()->memory_size()) {
282 // Success: The offsets are provably independent.
283 // (You may ask, why not just test st_offset != offset and be done?
284 // The answer is that stores of different sizes can co-exist
285 // in the same sequence of RawMem effects. We sometimes initialize
286 // a whole 'tile' of array elements with a single jint or jlong.)
287 mem = mem->in(MemNode::Memory);
288 continue; // (a) advance through independent store memory
289 }
290 }
291 if (st_base != base &&
292 detect_ptr_independence(base, alloc,
293 st_base,
294 AllocateNode::Ideal_allocation(st_base, phase),
295 phase)) {
296 // Success: The bases are provably independent.
297 mem = mem->in(MemNode::Memory);
298 continue; // (a) advance through independent store memory
299 }
301 // (b) At this point, if the bases or offsets do not agree, we lose,
302 // since we have not managed to prove 'this' and 'mem' independent.
303 if (st_base == base && st_offset == offset) {
304 return mem; // let caller handle steps (c), (d)
305 }
307 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
308 InitializeNode* st_init = mem->in(0)->as_Initialize();
309 AllocateNode* st_alloc = st_init->allocation();
310 if (st_alloc == NULL)
311 break; // something degenerated
312 bool known_identical = false;
313 bool known_independent = false;
314 if (alloc == st_alloc)
315 known_identical = true;
316 else if (alloc != NULL)
317 known_independent = true;
318 else if (ctrl != NULL &&
319 detect_dominating_control(ctrl, st_alloc->in(0)))
320 known_independent = true;
322 if (known_independent) {
323 // The bases are provably independent: Either they are
324 // manifestly distinct allocations, or else the control
325 // of this load dominates the store's allocation.
326 int alias_idx = phase->C->get_alias_index(adr_type());
327 if (alias_idx == Compile::AliasIdxRaw) {
328 mem = st_alloc->in(TypeFunc::Memory);
329 } else {
330 mem = st_init->memory(alias_idx);
331 }
332 continue; // (a) advance through independent store memory
333 }
335 // (b) at this point, if we are not looking at a store initializing
336 // the same allocation we are loading from, we lose.
337 if (known_identical) {
338 // From caller, can_see_stored_value will consult find_captured_store.
339 return mem; // let caller handle steps (c), (d)
340 }
342 }
344 // Unless there is an explicit 'continue', we must bail out here,
345 // because 'mem' is an inscrutable memory state (e.g., a call).
346 break;
347 }
349 return NULL; // bail out
350 }
352 //----------------------calculate_adr_type-------------------------------------
353 // Helper function. Notices when the given type of address hits top or bottom.
354 // Also, asserts a cross-check of the type against the expected address type.
355 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
356 if (t == Type::TOP) return NULL; // does not touch memory any more?
357 #ifdef PRODUCT
358 cross_check = NULL;
359 #else
360 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
361 #endif
362 const TypePtr* tp = t->isa_ptr();
363 if (tp == NULL) {
364 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
365 return TypePtr::BOTTOM; // touches lots of memory
366 } else {
367 #ifdef ASSERT
368 // %%%% [phh] We don't check the alias index if cross_check is
369 // TypeRawPtr::BOTTOM. Needs to be investigated.
370 if (cross_check != NULL &&
371 cross_check != TypePtr::BOTTOM &&
372 cross_check != TypeRawPtr::BOTTOM) {
373 // Recheck the alias index, to see if it has changed (due to a bug).
374 Compile* C = Compile::current();
375 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
376 "must stay in the original alias category");
377 // The type of the address must be contained in the adr_type,
378 // disregarding "null"-ness.
379 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
380 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
381 assert(cross_check->meet(tp_notnull) == cross_check,
382 "real address must not escape from expected memory type");
383 }
384 #endif
385 return tp;
386 }
387 }
389 //------------------------adr_phi_is_loop_invariant----------------------------
390 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
391 // loop is loop invariant. Make a quick traversal of Phi and associated
392 // CastPP nodes, looking to see if they are a closed group within the loop.
393 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
394 // The idea is that the phi-nest must boil down to only CastPP nodes
395 // with the same data. This implies that any path into the loop already
396 // includes such a CastPP, and so the original cast, whatever its input,
397 // must be covered by an equivalent cast, with an earlier control input.
398 ResourceMark rm;
400 // The loop entry input of the phi should be the unique dominating
401 // node for every Phi/CastPP in the loop.
402 Unique_Node_List closure;
403 closure.push(adr_phi->in(LoopNode::EntryControl));
405 // Add the phi node and the cast to the worklist.
406 Unique_Node_List worklist;
407 worklist.push(adr_phi);
408 if( cast != NULL ){
409 if( !cast->is_ConstraintCast() ) return false;
410 worklist.push(cast);
411 }
413 // Begin recursive walk of phi nodes.
414 while( worklist.size() ){
415 // Take a node off the worklist
416 Node *n = worklist.pop();
417 if( !closure.member(n) ){
418 // Add it to the closure.
419 closure.push(n);
420 // Make a sanity check to ensure we don't waste too much time here.
421 if( closure.size() > 20) return false;
422 // This node is OK if:
423 // - it is a cast of an identical value
424 // - or it is a phi node (then we add its inputs to the worklist)
425 // Otherwise, the node is not OK, and we presume the cast is not invariant
426 if( n->is_ConstraintCast() ){
427 worklist.push(n->in(1));
428 } else if( n->is_Phi() ) {
429 for( uint i = 1; i < n->req(); i++ ) {
430 worklist.push(n->in(i));
431 }
432 } else {
433 return false;
434 }
435 }
436 }
438 // Quit when the worklist is empty, and we've found no offending nodes.
439 return true;
440 }
442 //------------------------------Ideal_DU_postCCP-------------------------------
443 // Find any cast-away of null-ness and keep its control. Null cast-aways are
444 // going away in this pass and we need to make this memory op depend on the
445 // gating null check.
447 // I tried to leave the CastPP's in. This makes the graph more accurate in
448 // some sense; we get to keep around the knowledge that an oop is not-null
449 // after some test. Alas, the CastPP's interfere with GVN (some values are
450 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
451 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
452 // some of the more trivial cases in the optimizer. Removing more useless
453 // Phi's started allowing Loads to illegally float above null checks. I gave
454 // up on this approach. CNC 10/20/2000
455 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
456 Node *ctr = in(MemNode::Control);
457 Node *mem = in(MemNode::Memory);
458 Node *adr = in(MemNode::Address);
459 Node *skipped_cast = NULL;
460 // Need a null check? Regular static accesses do not because they are
461 // from constant addresses. Array ops are gated by the range check (which
462 // always includes a NULL check). Just check field ops.
463 if( !ctr ) {
464 // Scan upwards for the highest location we can place this memory op.
465 while( true ) {
466 switch( adr->Opcode() ) {
468 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
469 adr = adr->in(AddPNode::Base);
470 continue;
472 case Op_CastPP:
473 // If the CastPP is useless, just peek on through it.
474 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
475 // Remember the cast that we've peeked though. If we peek
476 // through more than one, then we end up remembering the highest
477 // one, that is, if in a loop, the one closest to the top.
478 skipped_cast = adr;
479 adr = adr->in(1);
480 continue;
481 }
482 // CastPP is going away in this pass! We need this memory op to be
483 // control-dependent on the test that is guarding the CastPP.
484 ccp->hash_delete(this);
485 set_req(MemNode::Control, adr->in(0));
486 ccp->hash_insert(this);
487 return this;
489 case Op_Phi:
490 // Attempt to float above a Phi to some dominating point.
491 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
492 // If we've already peeked through a Cast (which could have set the
493 // control), we can't float above a Phi, because the skipped Cast
494 // may not be loop invariant.
495 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
496 adr = adr->in(1);
497 continue;
498 }
499 }
501 // Intentional fallthrough!
503 // No obvious dominating point. The mem op is pinned below the Phi
504 // by the Phi itself. If the Phi goes away (no true value is merged)
505 // then the mem op can float, but not indefinitely. It must be pinned
506 // behind the controls leading to the Phi.
507 case Op_CheckCastPP:
508 // These usually stick around to change address type, however a
509 // useless one can be elided and we still need to pick up a control edge
510 if (adr->in(0) == NULL) {
511 // This CheckCastPP node has NO control and is likely useless. But we
512 // need check further up the ancestor chain for a control input to keep
513 // the node in place. 4959717.
514 skipped_cast = adr;
515 adr = adr->in(1);
516 continue;
517 }
518 ccp->hash_delete(this);
519 set_req(MemNode::Control, adr->in(0));
520 ccp->hash_insert(this);
521 return this;
523 // List of "safe" opcodes; those that implicitly block the memory
524 // op below any null check.
525 case Op_CastX2P: // no null checks on native pointers
526 case Op_Parm: // 'this' pointer is not null
527 case Op_LoadP: // Loading from within a klass
528 case Op_LoadKlass: // Loading from within a klass
529 case Op_ConP: // Loading from a klass
530 case Op_CreateEx: // Sucking up the guts of an exception oop
531 case Op_Con: // Reading from TLS
532 case Op_CMoveP: // CMoveP is pinned
533 break; // No progress
535 case Op_Proj: // Direct call to an allocation routine
536 case Op_SCMemProj: // Memory state from store conditional ops
537 #ifdef ASSERT
538 {
539 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
540 const Node* call = adr->in(0);
541 if (call->is_CallStaticJava()) {
542 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
543 assert(call_java && call_java->method() == NULL, "must be runtime call");
544 // We further presume that this is one of
545 // new_instance_Java, new_array_Java, or
546 // the like, but do not assert for this.
547 } else if (call->is_Allocate()) {
548 // similar case to new_instance_Java, etc.
549 } else if (!call->is_CallLeaf()) {
550 // Projections from fetch_oop (OSR) are allowed as well.
551 ShouldNotReachHere();
552 }
553 }
554 #endif
555 break;
556 default:
557 ShouldNotReachHere();
558 }
559 break;
560 }
561 }
563 return NULL; // No progress
564 }
567 //=============================================================================
568 uint LoadNode::size_of() const { return sizeof(*this); }
569 uint LoadNode::cmp( const Node &n ) const
570 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
571 const Type *LoadNode::bottom_type() const { return _type; }
572 uint LoadNode::ideal_reg() const {
573 return Matcher::base2reg[_type->base()];
574 }
576 #ifndef PRODUCT
577 void LoadNode::dump_spec(outputStream *st) const {
578 MemNode::dump_spec(st);
579 if( !Verbose && !WizardMode ) {
580 // standard dump does this in Verbose and WizardMode
581 st->print(" #"); _type->dump_on(st);
582 }
583 }
584 #endif
587 //----------------------------LoadNode::make-----------------------------------
588 // Polymorphic factory method:
589 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
590 // sanity check the alias category against the created node type
591 assert(!(adr_type->isa_oopptr() &&
592 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
593 "use LoadKlassNode instead");
594 assert(!(adr_type->isa_aryptr() &&
595 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
596 "use LoadRangeNode instead");
597 switch (bt) {
598 case T_BOOLEAN:
599 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
600 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
601 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
602 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
603 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
604 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
605 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
606 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
607 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
608 }
609 ShouldNotReachHere();
610 return (LoadNode*)NULL;
611 }
613 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
614 bool require_atomic = true;
615 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
616 }
621 //------------------------------hash-------------------------------------------
622 uint LoadNode::hash() const {
623 // unroll addition of interesting fields
624 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
625 }
627 //---------------------------can_see_stored_value------------------------------
628 // This routine exists to make sure this set of tests is done the same
629 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
630 // will change the graph shape in a way which makes memory alive twice at the
631 // same time (uses the Oracle model of aliasing), then some
632 // LoadXNode::Identity will fold things back to the equivalence-class model
633 // of aliasing.
634 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
635 Node* ld_adr = in(MemNode::Address);
637 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
638 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
639 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
640 atp->field() != NULL && !atp->field()->is_volatile()) {
641 uint alias_idx = atp->index();
642 bool final = atp->field()->is_final();
643 Node* result = NULL;
644 Node* current = st;
645 // Skip through chains of MemBarNodes checking the MergeMems for
646 // new states for the slice of this load. Stop once any other
647 // kind of node is encountered. Loads from final memory can skip
648 // through any kind of MemBar but normal loads shouldn't skip
649 // through MemBarAcquire since the could allow them to move out of
650 // a synchronized region.
651 while (current->is_Proj()) {
652 int opc = current->in(0)->Opcode();
653 if ((final && opc == Op_MemBarAcquire) ||
654 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
655 Node* mem = current->in(0)->in(TypeFunc::Memory);
656 if (mem->is_MergeMem()) {
657 MergeMemNode* merge = mem->as_MergeMem();
658 Node* new_st = merge->memory_at(alias_idx);
659 if (new_st == merge->base_memory()) {
660 // Keep searching
661 current = merge->base_memory();
662 continue;
663 }
664 // Save the new memory state for the slice and fall through
665 // to exit.
666 result = new_st;
667 }
668 }
669 break;
670 }
671 if (result != NULL) {
672 st = result;
673 }
674 }
677 // Loop around twice in the case Load -> Initialize -> Store.
678 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
679 for (int trip = 0; trip <= 1; trip++) {
681 if (st->is_Store()) {
682 Node* st_adr = st->in(MemNode::Address);
683 if (!phase->eqv(st_adr, ld_adr)) {
684 // Try harder before giving up... Match raw and non-raw pointers.
685 intptr_t st_off = 0;
686 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
687 if (alloc == NULL) return NULL;
688 intptr_t ld_off = 0;
689 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
690 if (alloc != allo2) return NULL;
691 if (ld_off != st_off) return NULL;
692 // At this point we have proven something like this setup:
693 // A = Allocate(...)
694 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
695 // S = StoreQ(, AddP(, A.Parm , #Off), V)
696 // (Actually, we haven't yet proven the Q's are the same.)
697 // In other words, we are loading from a casted version of
698 // the same pointer-and-offset that we stored to.
699 // Thus, we are able to replace L by V.
700 }
701 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
702 if (store_Opcode() != st->Opcode())
703 return NULL;
704 return st->in(MemNode::ValueIn);
705 }
707 intptr_t offset = 0; // scratch
709 // A load from a freshly-created object always returns zero.
710 // (This can happen after LoadNode::Ideal resets the load's memory input
711 // to find_captured_store, which returned InitializeNode::zero_memory.)
712 if (st->is_Proj() && st->in(0)->is_Allocate() &&
713 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
714 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
715 // return a zero value for the load's basic type
716 // (This is one of the few places where a generic PhaseTransform
717 // can create new nodes. Think of it as lazily manifesting
718 // virtually pre-existing constants.)
719 return phase->zerocon(memory_type());
720 }
722 // A load from an initialization barrier can match a captured store.
723 if (st->is_Proj() && st->in(0)->is_Initialize()) {
724 InitializeNode* init = st->in(0)->as_Initialize();
725 AllocateNode* alloc = init->allocation();
726 if (alloc != NULL &&
727 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
728 // examine a captured store value
729 st = init->find_captured_store(offset, memory_size(), phase);
730 if (st != NULL)
731 continue; // take one more trip around
732 }
733 }
735 break;
736 }
738 return NULL;
739 }
741 //------------------------------Identity---------------------------------------
742 // Loads are identity if previous store is to same address
743 Node *LoadNode::Identity( PhaseTransform *phase ) {
744 // If the previous store-maker is the right kind of Store, and the store is
745 // to the same address, then we are equal to the value stored.
746 Node* mem = in(MemNode::Memory);
747 Node* value = can_see_stored_value(mem, phase);
748 if( value ) {
749 // byte, short & char stores truncate naturally.
750 // A load has to load the truncated value which requires
751 // some sort of masking operation and that requires an
752 // Ideal call instead of an Identity call.
753 if (memory_size() < BytesPerInt) {
754 // If the input to the store does not fit with the load's result type,
755 // it must be truncated via an Ideal call.
756 if (!phase->type(value)->higher_equal(phase->type(this)))
757 return this;
758 }
759 // (This works even when value is a Con, but LoadNode::Value
760 // usually runs first, producing the singleton type of the Con.)
761 return value;
762 }
763 return this;
764 }
767 // Returns true if the AliasType refers to the field that holds the
768 // cached box array. Currently only handles the IntegerCache case.
769 static bool is_autobox_cache(Compile::AliasType* atp) {
770 if (atp != NULL && atp->field() != NULL) {
771 ciField* field = atp->field();
772 ciSymbol* klass = field->holder()->name();
773 if (field->name() == ciSymbol::cache_field_name() &&
774 field->holder()->uses_default_loader() &&
775 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
776 return true;
777 }
778 }
779 return false;
780 }
782 // Fetch the base value in the autobox array
783 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
784 if (atp != NULL && atp->field() != NULL) {
785 ciField* field = atp->field();
786 ciSymbol* klass = field->holder()->name();
787 if (field->name() == ciSymbol::cache_field_name() &&
788 field->holder()->uses_default_loader() &&
789 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
790 assert(field->is_constant(), "what?");
791 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
792 // Fetch the box object at the base of the array and get its value
793 ciInstance* box = array->obj_at(0)->as_instance();
794 ciInstanceKlass* ik = box->klass()->as_instance_klass();
795 if (ik->nof_nonstatic_fields() == 1) {
796 // This should be true nonstatic_field_at requires calling
797 // nof_nonstatic_fields so check it anyway
798 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
799 cache_offset = c.as_int();
800 }
801 return true;
802 }
803 }
804 return false;
805 }
807 // Returns true if the AliasType refers to the value field of an
808 // autobox object. Currently only handles Integer.
809 static bool is_autobox_object(Compile::AliasType* atp) {
810 if (atp != NULL && atp->field() != NULL) {
811 ciField* field = atp->field();
812 ciSymbol* klass = field->holder()->name();
813 if (field->name() == ciSymbol::value_name() &&
814 field->holder()->uses_default_loader() &&
815 klass == ciSymbol::java_lang_Integer()) {
816 return true;
817 }
818 }
819 return false;
820 }
823 // We're loading from an object which has autobox behaviour.
824 // If this object is result of a valueOf call we'll have a phi
825 // merging a newly allocated object and a load from the cache.
826 // We want to replace this load with the original incoming
827 // argument to the valueOf call.
828 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
829 Node* base = in(Address)->in(AddPNode::Base);
830 if (base->is_Phi() && base->req() == 3) {
831 AllocateNode* allocation = NULL;
832 int allocation_index = -1;
833 int load_index = -1;
834 for (uint i = 1; i < base->req(); i++) {
835 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
836 if (allocation != NULL) {
837 allocation_index = i;
838 load_index = 3 - allocation_index;
839 break;
840 }
841 }
842 LoadNode* load = NULL;
843 if (allocation != NULL && base->in(load_index)->is_Load()) {
844 load = base->in(load_index)->as_Load();
845 }
846 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
847 // Push the loads from the phi that comes from valueOf up
848 // through it to allow elimination of the loads and the recovery
849 // of the original value.
850 Node* mem_phi = in(Memory);
851 Node* offset = in(Address)->in(AddPNode::Offset);
853 Node* in1 = clone();
854 Node* in1_addr = in1->in(Address)->clone();
855 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
856 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
857 in1_addr->set_req(AddPNode::Offset, offset);
858 in1->set_req(0, base->in(allocation_index));
859 in1->set_req(Address, in1_addr);
860 in1->set_req(Memory, mem_phi->in(allocation_index));
862 Node* in2 = clone();
863 Node* in2_addr = in2->in(Address)->clone();
864 in2_addr->set_req(AddPNode::Base, base->in(load_index));
865 in2_addr->set_req(AddPNode::Address, base->in(load_index));
866 in2_addr->set_req(AddPNode::Offset, offset);
867 in2->set_req(0, base->in(load_index));
868 in2->set_req(Address, in2_addr);
869 in2->set_req(Memory, mem_phi->in(load_index));
871 in1_addr = phase->transform(in1_addr);
872 in1 = phase->transform(in1);
873 in2_addr = phase->transform(in2_addr);
874 in2 = phase->transform(in2);
876 PhiNode* result = PhiNode::make_blank(base->in(0), this);
877 result->set_req(allocation_index, in1);
878 result->set_req(load_index, in2);
879 return result;
880 }
881 } else if (base->is_Load()) {
882 // Eliminate the load of Integer.value for integers from the cache
883 // array by deriving the value from the index into the array.
884 // Capture the offset of the load and then reverse the computation.
885 Node* load_base = base->in(Address)->in(AddPNode::Base);
886 if (load_base != NULL) {
887 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
888 intptr_t cache_offset;
889 int shift = -1;
890 Node* cache = NULL;
891 if (is_autobox_cache(atp)) {
892 shift = exact_log2(type2aelembytes[T_OBJECT]);
893 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
894 }
895 if (cache != NULL && base->in(Address)->is_AddP()) {
896 Node* elements[4];
897 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
898 int cache_low;
899 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
900 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
901 // Add up all the offsets making of the address of the load
902 Node* result = elements[0];
903 for (int i = 1; i < count; i++) {
904 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
905 }
906 // Remove the constant offset from the address and then
907 // remove the scaling of the offset to recover the original index.
908 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
909 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
910 // Peel the shift off directly but wrap it in a dummy node
911 // since Ideal can't return existing nodes
912 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
913 } else {
914 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
915 }
916 #ifdef _LP64
917 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
918 #endif
919 return result;
920 }
921 }
922 }
923 }
924 return NULL;
925 }
928 //------------------------------Ideal------------------------------------------
929 // If the load is from Field memory and the pointer is non-null, we can
930 // zero out the control input.
931 // If the offset is constant and the base is an object allocation,
932 // try to hook me up to the exact initializing store.
933 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
934 Node* p = MemNode::Ideal_common(phase, can_reshape);
935 if (p) return (p == NodeSentinel) ? NULL : p;
937 Node* ctrl = in(MemNode::Control);
938 Node* address = in(MemNode::Address);
940 // Skip up past a SafePoint control. Cannot do this for Stores because
941 // pointer stores & cardmarks must stay on the same side of a SafePoint.
942 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
943 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
944 ctrl = ctrl->in(0);
945 set_req(MemNode::Control,ctrl);
946 }
948 // Check for useless control edge in some common special cases
949 if (in(MemNode::Control) != NULL) {
950 intptr_t ignore = 0;
951 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
952 if (base != NULL
953 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
954 && detect_dominating_control(base->in(0), phase->C->start())) {
955 // A method-invariant, non-null address (constant or 'this' argument).
956 set_req(MemNode::Control, NULL);
957 }
958 }
960 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
961 Node* base = in(Address)->in(AddPNode::Base);
962 if (base != NULL) {
963 Compile::AliasType* atp = phase->C->alias_type(adr_type());
964 if (is_autobox_object(atp)) {
965 Node* result = eliminate_autobox(phase);
966 if (result != NULL) return result;
967 }
968 }
969 }
971 // Check for prior store with a different base or offset; make Load
972 // independent. Skip through any number of them. Bail out if the stores
973 // are in an endless dead cycle and report no progress. This is a key
974 // transform for Reflection. However, if after skipping through the Stores
975 // we can't then fold up against a prior store do NOT do the transform as
976 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
977 // array memory alive twice: once for the hoisted Load and again after the
978 // bypassed Store. This situation only works if EVERYBODY who does
979 // anti-dependence work knows how to bypass. I.e. we need all
980 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
981 // the alias index stuff. So instead, peek through Stores and IFF we can
982 // fold up, do so.
983 Node* prev_mem = find_previous_store(phase);
984 // Steps (a), (b): Walk past independent stores to find an exact match.
985 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
986 // (c) See if we can fold up on the spot, but don't fold up here.
987 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
988 // just return a prior value, which is done by Identity calls.
989 if (can_see_stored_value(prev_mem, phase)) {
990 // Make ready for step (d):
991 set_req(MemNode::Memory, prev_mem);
992 return this;
993 }
994 }
996 return NULL; // No further progress
997 }
999 // Helper to recognize certain Klass fields which are invariant across
1000 // some group of array types (e.g., int[] or all T[] where T < Object).
1001 const Type*
1002 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1003 ciKlass* klass) const {
1004 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1005 // The field is Klass::_modifier_flags. Return its (constant) value.
1006 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1007 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1008 return TypeInt::make(klass->modifier_flags());
1009 }
1010 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1011 // The field is Klass::_access_flags. Return its (constant) value.
1012 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1013 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1014 return TypeInt::make(klass->access_flags());
1015 }
1016 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1017 // The field is Klass::_layout_helper. Return its constant value if known.
1018 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1019 return TypeInt::make(klass->layout_helper());
1020 }
1022 // No match.
1023 return NULL;
1024 }
1026 //------------------------------Value-----------------------------------------
1027 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1028 // Either input is TOP ==> the result is TOP
1029 Node* mem = in(MemNode::Memory);
1030 const Type *t1 = phase->type(mem);
1031 if (t1 == Type::TOP) return Type::TOP;
1032 Node* adr = in(MemNode::Address);
1033 const TypePtr* tp = phase->type(adr)->isa_ptr();
1034 if (tp == NULL || tp->empty()) return Type::TOP;
1035 int off = tp->offset();
1036 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1038 // Try to guess loaded type from pointer type
1039 if (tp->base() == Type::AryPtr) {
1040 const Type *t = tp->is_aryptr()->elem();
1041 // Don't do this for integer types. There is only potential profit if
1042 // the element type t is lower than _type; that is, for int types, if _type is
1043 // more restrictive than t. This only happens here if one is short and the other
1044 // char (both 16 bits), and in those cases we've made an intentional decision
1045 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1046 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1047 //
1048 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1049 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1050 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1051 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1052 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1053 // In fact, that could have been the original type of p1, and p1 could have
1054 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1055 // expression (LShiftL quux 3) independently optimized to the constant 8.
1056 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1057 && Opcode() != Op_LoadKlass) {
1058 // t might actually be lower than _type, if _type is a unique
1059 // concrete subclass of abstract class t.
1060 // Make sure the reference is not into the header, by comparing
1061 // the offset against the offset of the start of the array's data.
1062 // Different array types begin at slightly different offsets (12 vs. 16).
1063 // We choose T_BYTE as an example base type that is least restrictive
1064 // as to alignment, which will therefore produce the smallest
1065 // possible base offset.
1066 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1067 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1068 const Type* jt = t->join(_type);
1069 // In any case, do not allow the join, per se, to empty out the type.
1070 if (jt->empty() && !t->empty()) {
1071 // This can happen if a interface-typed array narrows to a class type.
1072 jt = _type;
1073 }
1075 if (EliminateAutoBox) {
1076 // The pointers in the autobox arrays are always non-null
1077 Node* base = in(Address)->in(AddPNode::Base);
1078 if (base != NULL) {
1079 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1080 if (is_autobox_cache(atp)) {
1081 return jt->join(TypePtr::NOTNULL)->is_ptr();
1082 }
1083 }
1084 }
1085 return jt;
1086 }
1087 }
1088 } else if (tp->base() == Type::InstPtr) {
1089 assert( off != Type::OffsetBot ||
1090 // arrays can be cast to Objects
1091 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1092 // unsafe field access may not have a constant offset
1093 phase->C->has_unsafe_access(),
1094 "Field accesses must be precise" );
1095 // For oop loads, we expect the _type to be precise
1096 } else if (tp->base() == Type::KlassPtr) {
1097 assert( off != Type::OffsetBot ||
1098 // arrays can be cast to Objects
1099 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1100 // also allow array-loading from the primary supertype
1101 // array during subtype checks
1102 Opcode() == Op_LoadKlass,
1103 "Field accesses must be precise" );
1104 // For klass/static loads, we expect the _type to be precise
1105 }
1107 const TypeKlassPtr *tkls = tp->isa_klassptr();
1108 if (tkls != NULL && !StressReflectiveCode) {
1109 ciKlass* klass = tkls->klass();
1110 if (klass->is_loaded() && tkls->klass_is_exact()) {
1111 // We are loading a field from a Klass metaobject whose identity
1112 // is known at compile time (the type is "exact" or "precise").
1113 // Check for fields we know are maintained as constants by the VM.
1114 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1115 // The field is Klass::_super_check_offset. Return its (constant) value.
1116 // (Folds up type checking code.)
1117 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1118 return TypeInt::make(klass->super_check_offset());
1119 }
1120 // Compute index into primary_supers array
1121 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1122 // Check for overflowing; use unsigned compare to handle the negative case.
1123 if( depth < ciKlass::primary_super_limit() ) {
1124 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1125 // (Folds up type checking code.)
1126 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1127 ciKlass *ss = klass->super_of_depth(depth);
1128 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1129 }
1130 const Type* aift = load_array_final_field(tkls, klass);
1131 if (aift != NULL) return aift;
1132 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1133 && klass->is_array_klass()) {
1134 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1135 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1136 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1137 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1138 }
1139 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1140 // The field is Klass::_java_mirror. Return its (constant) value.
1141 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1142 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1143 return TypeInstPtr::make(klass->java_mirror());
1144 }
1145 }
1147 // We can still check if we are loading from the primary_supers array at a
1148 // shallow enough depth. Even though the klass is not exact, entries less
1149 // than or equal to its super depth are correct.
1150 if (klass->is_loaded() ) {
1151 ciType *inner = klass->klass();
1152 while( inner->is_obj_array_klass() )
1153 inner = inner->as_obj_array_klass()->base_element_type();
1154 if( inner->is_instance_klass() &&
1155 !inner->as_instance_klass()->flags().is_interface() ) {
1156 // Compute index into primary_supers array
1157 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1158 // Check for overflowing; use unsigned compare to handle the negative case.
1159 if( depth < ciKlass::primary_super_limit() &&
1160 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1161 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1162 // (Folds up type checking code.)
1163 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1164 ciKlass *ss = klass->super_of_depth(depth);
1165 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1166 }
1167 }
1168 }
1170 // If the type is enough to determine that the thing is not an array,
1171 // we can give the layout_helper a positive interval type.
1172 // This will help short-circuit some reflective code.
1173 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1174 && !klass->is_array_klass() // not directly typed as an array
1175 && !klass->is_interface() // specifically not Serializable & Cloneable
1176 && !klass->is_java_lang_Object() // not the supertype of all T[]
1177 ) {
1178 // Note: When interfaces are reliable, we can narrow the interface
1179 // test to (klass != Serializable && klass != Cloneable).
1180 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1181 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1182 // The key property of this type is that it folds up tests
1183 // for array-ness, since it proves that the layout_helper is positive.
1184 // Thus, a generic value like the basic object layout helper works fine.
1185 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1186 }
1187 }
1189 // If we are loading from a freshly-allocated object, produce a zero,
1190 // if the load is provably beyond the header of the object.
1191 // (Also allow a variable load from a fresh array to produce zero.)
1192 if (ReduceFieldZeroing) {
1193 Node* value = can_see_stored_value(mem,phase);
1194 if (value != NULL && value->is_Con())
1195 return value->bottom_type();
1196 }
1198 return _type;
1199 }
1201 //------------------------------match_edge-------------------------------------
1202 // Do we Match on this edge index or not? Match only the address.
1203 uint LoadNode::match_edge(uint idx) const {
1204 return idx == MemNode::Address;
1205 }
1207 //--------------------------LoadBNode::Ideal--------------------------------------
1208 //
1209 // If the previous store is to the same address as this load,
1210 // and the value stored was larger than a byte, replace this load
1211 // with the value stored truncated to a byte. If no truncation is
1212 // needed, the replacement is done in LoadNode::Identity().
1213 //
1214 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1215 Node* mem = in(MemNode::Memory);
1216 Node* value = can_see_stored_value(mem,phase);
1217 if( value && !phase->type(value)->higher_equal( _type ) ) {
1218 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1219 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1220 }
1221 // Identity call will handle the case where truncation is not needed.
1222 return LoadNode::Ideal(phase, can_reshape);
1223 }
1225 //--------------------------LoadCNode::Ideal--------------------------------------
1226 //
1227 // If the previous store is to the same address as this load,
1228 // and the value stored was larger than a char, replace this load
1229 // with the value stored truncated to a char. If no truncation is
1230 // needed, the replacement is done in LoadNode::Identity().
1231 //
1232 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1233 Node* mem = in(MemNode::Memory);
1234 Node* value = can_see_stored_value(mem,phase);
1235 if( value && !phase->type(value)->higher_equal( _type ) )
1236 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1237 // Identity call will handle the case where truncation is not needed.
1238 return LoadNode::Ideal(phase, can_reshape);
1239 }
1241 //--------------------------LoadSNode::Ideal--------------------------------------
1242 //
1243 // If the previous store is to the same address as this load,
1244 // and the value stored was larger than a short, replace this load
1245 // with the value stored truncated to a short. If no truncation is
1246 // needed, the replacement is done in LoadNode::Identity().
1247 //
1248 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1249 Node* mem = in(MemNode::Memory);
1250 Node* value = can_see_stored_value(mem,phase);
1251 if( value && !phase->type(value)->higher_equal( _type ) ) {
1252 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1253 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1254 }
1255 // Identity call will handle the case where truncation is not needed.
1256 return LoadNode::Ideal(phase, can_reshape);
1257 }
1259 //=============================================================================
1260 //------------------------------Value------------------------------------------
1261 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1262 // Either input is TOP ==> the result is TOP
1263 const Type *t1 = phase->type( in(MemNode::Memory) );
1264 if (t1 == Type::TOP) return Type::TOP;
1265 Node *adr = in(MemNode::Address);
1266 const Type *t2 = phase->type( adr );
1267 if (t2 == Type::TOP) return Type::TOP;
1268 const TypePtr *tp = t2->is_ptr();
1269 if (TypePtr::above_centerline(tp->ptr()) ||
1270 tp->ptr() == TypePtr::Null) return Type::TOP;
1272 // Return a more precise klass, if possible
1273 const TypeInstPtr *tinst = tp->isa_instptr();
1274 if (tinst != NULL) {
1275 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1276 int offset = tinst->offset();
1277 if (ik == phase->C->env()->Class_klass()
1278 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1279 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1280 // We are loading a special hidden field from a Class mirror object,
1281 // the field which points to the VM's Klass metaobject.
1282 ciType* t = tinst->java_mirror_type();
1283 // java_mirror_type returns non-null for compile-time Class constants.
1284 if (t != NULL) {
1285 // constant oop => constant klass
1286 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1287 return TypeKlassPtr::make(ciArrayKlass::make(t));
1288 }
1289 if (!t->is_klass()) {
1290 // a primitive Class (e.g., int.class) has NULL for a klass field
1291 return TypePtr::NULL_PTR;
1292 }
1293 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1294 return TypeKlassPtr::make(t->as_klass());
1295 }
1296 // non-constant mirror, so we can't tell what's going on
1297 }
1298 if( !ik->is_loaded() )
1299 return _type; // Bail out if not loaded
1300 if (offset == oopDesc::klass_offset_in_bytes()) {
1301 if (tinst->klass_is_exact()) {
1302 return TypeKlassPtr::make(ik);
1303 }
1304 // See if we can become precise: no subklasses and no interface
1305 // (Note: We need to support verified interfaces.)
1306 if (!ik->is_interface() && !ik->has_subklass()) {
1307 //assert(!UseExactTypes, "this code should be useless with exact types");
1308 // Add a dependence; if any subclass added we need to recompile
1309 if (!ik->is_final()) {
1310 // %%% should use stronger assert_unique_concrete_subtype instead
1311 phase->C->dependencies()->assert_leaf_type(ik);
1312 }
1313 // Return precise klass
1314 return TypeKlassPtr::make(ik);
1315 }
1317 // Return root of possible klass
1318 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1319 }
1320 }
1322 // Check for loading klass from an array
1323 const TypeAryPtr *tary = tp->isa_aryptr();
1324 if( tary != NULL ) {
1325 ciKlass *tary_klass = tary->klass();
1326 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1327 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1328 if (tary->klass_is_exact()) {
1329 return TypeKlassPtr::make(tary_klass);
1330 }
1331 ciArrayKlass *ak = tary->klass()->as_array_klass();
1332 // If the klass is an object array, we defer the question to the
1333 // array component klass.
1334 if( ak->is_obj_array_klass() ) {
1335 assert( ak->is_loaded(), "" );
1336 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1337 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1338 ciInstanceKlass* ik = base_k->as_instance_klass();
1339 // See if we can become precise: no subklasses and no interface
1340 if (!ik->is_interface() && !ik->has_subklass()) {
1341 //assert(!UseExactTypes, "this code should be useless with exact types");
1342 // Add a dependence; if any subclass added we need to recompile
1343 if (!ik->is_final()) {
1344 phase->C->dependencies()->assert_leaf_type(ik);
1345 }
1346 // Return precise array klass
1347 return TypeKlassPtr::make(ak);
1348 }
1349 }
1350 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1351 } else { // Found a type-array?
1352 //assert(!UseExactTypes, "this code should be useless with exact types");
1353 assert( ak->is_type_array_klass(), "" );
1354 return TypeKlassPtr::make(ak); // These are always precise
1355 }
1356 }
1357 }
1359 // Check for loading klass from an array klass
1360 const TypeKlassPtr *tkls = tp->isa_klassptr();
1361 if (tkls != NULL && !StressReflectiveCode) {
1362 ciKlass* klass = tkls->klass();
1363 if( !klass->is_loaded() )
1364 return _type; // Bail out if not loaded
1365 if( klass->is_obj_array_klass() &&
1366 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1367 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1368 // // Always returning precise element type is incorrect,
1369 // // e.g., element type could be object and array may contain strings
1370 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1372 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1373 // according to the element type's subclassing.
1374 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1375 }
1376 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1377 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1378 ciKlass* sup = klass->as_instance_klass()->super();
1379 // The field is Klass::_super. Return its (constant) value.
1380 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1381 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1382 }
1383 }
1385 // Bailout case
1386 return LoadNode::Value(phase);
1387 }
1389 //------------------------------Identity---------------------------------------
1390 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1391 // Also feed through the klass in Allocate(...klass...)._klass.
1392 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1393 Node* x = LoadNode::Identity(phase);
1394 if (x != this) return x;
1396 // Take apart the address into an oop and and offset.
1397 // Return 'this' if we cannot.
1398 Node* adr = in(MemNode::Address);
1399 intptr_t offset = 0;
1400 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1401 if (base == NULL) return this;
1402 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1403 if (toop == NULL) return this;
1405 // We can fetch the klass directly through an AllocateNode.
1406 // This works even if the klass is not constant (clone or newArray).
1407 if (offset == oopDesc::klass_offset_in_bytes()) {
1408 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1409 if (allocated_klass != NULL) {
1410 return allocated_klass;
1411 }
1412 }
1414 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1415 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1416 // See inline_native_Class_query for occurrences of these patterns.
1417 // Java Example: x.getClass().isAssignableFrom(y)
1418 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1419 //
1420 // This improves reflective code, often making the Class
1421 // mirror go completely dead. (Current exception: Class
1422 // mirrors may appear in debug info, but we could clean them out by
1423 // introducing a new debug info operator for klassOop.java_mirror).
1424 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1425 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1426 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1427 // We are loading a special hidden field from a Class mirror,
1428 // the field which points to its Klass or arrayKlass metaobject.
1429 if (base->is_Load()) {
1430 Node* adr2 = base->in(MemNode::Address);
1431 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1432 if (tkls != NULL && !tkls->empty()
1433 && (tkls->klass()->is_instance_klass() ||
1434 tkls->klass()->is_array_klass())
1435 && adr2->is_AddP()
1436 ) {
1437 int mirror_field = Klass::java_mirror_offset_in_bytes();
1438 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1439 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1440 }
1441 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1442 return adr2->in(AddPNode::Base);
1443 }
1444 }
1445 }
1446 }
1448 return this;
1449 }
1451 //------------------------------Value-----------------------------------------
1452 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1453 // Either input is TOP ==> the result is TOP
1454 const Type *t1 = phase->type( in(MemNode::Memory) );
1455 if( t1 == Type::TOP ) return Type::TOP;
1456 Node *adr = in(MemNode::Address);
1457 const Type *t2 = phase->type( adr );
1458 if( t2 == Type::TOP ) return Type::TOP;
1459 const TypePtr *tp = t2->is_ptr();
1460 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1461 const TypeAryPtr *tap = tp->isa_aryptr();
1462 if( !tap ) return _type;
1463 return tap->size();
1464 }
1466 //------------------------------Identity---------------------------------------
1467 // Feed through the length in AllocateArray(...length...)._length.
1468 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1469 Node* x = LoadINode::Identity(phase);
1470 if (x != this) return x;
1472 // Take apart the address into an oop and and offset.
1473 // Return 'this' if we cannot.
1474 Node* adr = in(MemNode::Address);
1475 intptr_t offset = 0;
1476 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1477 if (base == NULL) return this;
1478 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1479 if (tary == NULL) return this;
1481 // We can fetch the length directly through an AllocateArrayNode.
1482 // This works even if the length is not constant (clone or newArray).
1483 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1484 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1485 if (allocated_length != NULL) {
1486 return allocated_length;
1487 }
1488 }
1490 return this;
1492 }
1493 //=============================================================================
1494 //---------------------------StoreNode::make-----------------------------------
1495 // Polymorphic factory method:
1496 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1497 switch (bt) {
1498 case T_BOOLEAN:
1499 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1500 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1501 case T_CHAR:
1502 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1503 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1504 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1505 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1506 case T_ADDRESS:
1507 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1508 }
1509 ShouldNotReachHere();
1510 return (StoreNode*)NULL;
1511 }
1513 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1514 bool require_atomic = true;
1515 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1516 }
1519 //--------------------------bottom_type----------------------------------------
1520 const Type *StoreNode::bottom_type() const {
1521 return Type::MEMORY;
1522 }
1524 //------------------------------hash-------------------------------------------
1525 uint StoreNode::hash() const {
1526 // unroll addition of interesting fields
1527 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1529 // Since they are not commoned, do not hash them:
1530 return NO_HASH;
1531 }
1533 //------------------------------Ideal------------------------------------------
1534 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1535 // When a store immediately follows a relevant allocation/initialization,
1536 // try to capture it into the initialization, or hoist it above.
1537 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1538 Node* p = MemNode::Ideal_common(phase, can_reshape);
1539 if (p) return (p == NodeSentinel) ? NULL : p;
1541 Node* mem = in(MemNode::Memory);
1542 Node* address = in(MemNode::Address);
1544 // Back-to-back stores to same address? Fold em up.
1545 // Generally unsafe if I have intervening uses...
1546 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1547 // Looking at a dead closed cycle of memory?
1548 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1550 assert(Opcode() == mem->Opcode() ||
1551 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1552 "no mismatched stores, except on raw memory");
1554 if (mem->outcnt() == 1 && // check for intervening uses
1555 mem->as_Store()->memory_size() <= this->memory_size()) {
1556 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1557 // For example, 'mem' might be the final state at a conditional return.
1558 // Or, 'mem' might be used by some node which is live at the same time
1559 // 'this' is live, which might be unschedulable. So, require exactly
1560 // ONE user, the 'this' store, until such time as we clone 'mem' for
1561 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1562 if (can_reshape) { // (%%% is this an anachronism?)
1563 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1564 phase->is_IterGVN());
1565 } else {
1566 // It's OK to do this in the parser, since DU info is always accurate,
1567 // and the parser always refers to nodes via SafePointNode maps.
1568 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1569 }
1570 return this;
1571 }
1572 }
1574 // Capture an unaliased, unconditional, simple store into an initializer.
1575 // Or, if it is independent of the allocation, hoist it above the allocation.
1576 if (ReduceFieldZeroing && /*can_reshape &&*/
1577 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1578 InitializeNode* init = mem->in(0)->as_Initialize();
1579 intptr_t offset = init->can_capture_store(this, phase);
1580 if (offset > 0) {
1581 Node* moved = init->capture_store(this, offset, phase);
1582 // If the InitializeNode captured me, it made a raw copy of me,
1583 // and I need to disappear.
1584 if (moved != NULL) {
1585 // %%% hack to ensure that Ideal returns a new node:
1586 mem = MergeMemNode::make(phase->C, mem);
1587 return mem; // fold me away
1588 }
1589 }
1590 }
1592 return NULL; // No further progress
1593 }
1595 //------------------------------Value-----------------------------------------
1596 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1597 // Either input is TOP ==> the result is TOP
1598 const Type *t1 = phase->type( in(MemNode::Memory) );
1599 if( t1 == Type::TOP ) return Type::TOP;
1600 const Type *t2 = phase->type( in(MemNode::Address) );
1601 if( t2 == Type::TOP ) return Type::TOP;
1602 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1603 if( t3 == Type::TOP ) return Type::TOP;
1604 return Type::MEMORY;
1605 }
1607 //------------------------------Identity---------------------------------------
1608 // Remove redundant stores:
1609 // Store(m, p, Load(m, p)) changes to m.
1610 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1611 Node *StoreNode::Identity( PhaseTransform *phase ) {
1612 Node* mem = in(MemNode::Memory);
1613 Node* adr = in(MemNode::Address);
1614 Node* val = in(MemNode::ValueIn);
1616 // Load then Store? Then the Store is useless
1617 if (val->is_Load() &&
1618 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1619 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1620 val->as_Load()->store_Opcode() == Opcode()) {
1621 return mem;
1622 }
1624 // Two stores in a row of the same value?
1625 if (mem->is_Store() &&
1626 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1627 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1628 mem->Opcode() == Opcode()) {
1629 return mem;
1630 }
1632 // Store of zero anywhere into a freshly-allocated object?
1633 // Then the store is useless.
1634 // (It must already have been captured by the InitializeNode.)
1635 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1636 // a newly allocated object is already all-zeroes everywhere
1637 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1638 return mem;
1639 }
1641 // the store may also apply to zero-bits in an earlier object
1642 Node* prev_mem = find_previous_store(phase);
1643 // Steps (a), (b): Walk past independent stores to find an exact match.
1644 if (prev_mem != NULL) {
1645 Node* prev_val = can_see_stored_value(prev_mem, phase);
1646 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1647 // prev_val and val might differ by a cast; it would be good
1648 // to keep the more informative of the two.
1649 return mem;
1650 }
1651 }
1652 }
1654 return this;
1655 }
1657 //------------------------------match_edge-------------------------------------
1658 // Do we Match on this edge index or not? Match only memory & value
1659 uint StoreNode::match_edge(uint idx) const {
1660 return idx == MemNode::Address || idx == MemNode::ValueIn;
1661 }
1663 //------------------------------cmp--------------------------------------------
1664 // Do not common stores up together. They generally have to be split
1665 // back up anyways, so do not bother.
1666 uint StoreNode::cmp( const Node &n ) const {
1667 return (&n == this); // Always fail except on self
1668 }
1670 //------------------------------Ideal_masked_input-----------------------------
1671 // Check for a useless mask before a partial-word store
1672 // (StoreB ... (AndI valIn conIa) )
1673 // If (conIa & mask == mask) this simplifies to
1674 // (StoreB ... (valIn) )
1675 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1676 Node *val = in(MemNode::ValueIn);
1677 if( val->Opcode() == Op_AndI ) {
1678 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1679 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1680 set_req(MemNode::ValueIn, val->in(1));
1681 return this;
1682 }
1683 }
1684 return NULL;
1685 }
1688 //------------------------------Ideal_sign_extended_input----------------------
1689 // Check for useless sign-extension before a partial-word store
1690 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1691 // If (conIL == conIR && conIR <= num_bits) this simplifies to
1692 // (StoreB ... (valIn) )
1693 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1694 Node *val = in(MemNode::ValueIn);
1695 if( val->Opcode() == Op_RShiftI ) {
1696 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1697 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
1698 Node *shl = val->in(1);
1699 if( shl->Opcode() == Op_LShiftI ) {
1700 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
1701 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
1702 set_req(MemNode::ValueIn, shl->in(1));
1703 return this;
1704 }
1705 }
1706 }
1707 }
1708 return NULL;
1709 }
1711 //------------------------------value_never_loaded-----------------------------------
1712 // Determine whether there are any possible loads of the value stored.
1713 // For simplicity, we actually check if there are any loads from the
1714 // address stored to, not just for loads of the value stored by this node.
1715 //
1716 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
1717 Node *adr = in(Address);
1718 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
1719 if (adr_oop == NULL)
1720 return false;
1721 if (!adr_oop->is_instance())
1722 return false; // if not a distinct instance, there may be aliases of the address
1723 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
1724 Node *use = adr->fast_out(i);
1725 int opc = use->Opcode();
1726 if (use->is_Load() || use->is_LoadStore()) {
1727 return false;
1728 }
1729 }
1730 return true;
1731 }
1733 //=============================================================================
1734 //------------------------------Ideal------------------------------------------
1735 // If the store is from an AND mask that leaves the low bits untouched, then
1736 // we can skip the AND operation. If the store is from a sign-extension
1737 // (a left shift, then right shift) we can skip both.
1738 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
1739 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
1740 if( progress != NULL ) return progress;
1742 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
1743 if( progress != NULL ) return progress;
1745 // Finally check the default case
1746 return StoreNode::Ideal(phase, can_reshape);
1747 }
1749 //=============================================================================
1750 //------------------------------Ideal------------------------------------------
1751 // If the store is from an AND mask that leaves the low bits untouched, then
1752 // we can skip the AND operation
1753 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
1754 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
1755 if( progress != NULL ) return progress;
1757 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
1758 if( progress != NULL ) return progress;
1760 // Finally check the default case
1761 return StoreNode::Ideal(phase, can_reshape);
1762 }
1764 //=============================================================================
1765 //------------------------------Identity---------------------------------------
1766 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
1767 // No need to card mark when storing a null ptr
1768 Node* my_store = in(MemNode::OopStore);
1769 if (my_store->is_Store()) {
1770 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
1771 if( t1 == TypePtr::NULL_PTR ) {
1772 return in(MemNode::Memory);
1773 }
1774 }
1775 return this;
1776 }
1778 //------------------------------Value-----------------------------------------
1779 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
1780 // If extra input is TOP ==> the result is TOP
1781 const Type *t1 = phase->type( in(MemNode::OopStore) );
1782 if( t1 == Type::TOP ) return Type::TOP;
1784 return StoreNode::Value( phase );
1785 }
1788 //=============================================================================
1789 //----------------------------------SCMemProjNode------------------------------
1790 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
1791 {
1792 return bottom_type();
1793 }
1795 //=============================================================================
1796 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
1797 init_req(MemNode::Control, c );
1798 init_req(MemNode::Memory , mem);
1799 init_req(MemNode::Address, adr);
1800 init_req(MemNode::ValueIn, val);
1801 init_req( ExpectedIn, ex );
1802 init_class_id(Class_LoadStore);
1804 }
1806 //=============================================================================
1807 //-------------------------------adr_type--------------------------------------
1808 // Do we Match on this edge index or not? Do not match memory
1809 const TypePtr* ClearArrayNode::adr_type() const {
1810 Node *adr = in(3);
1811 return MemNode::calculate_adr_type(adr->bottom_type());
1812 }
1814 //------------------------------match_edge-------------------------------------
1815 // Do we Match on this edge index or not? Do not match memory
1816 uint ClearArrayNode::match_edge(uint idx) const {
1817 return idx > 1;
1818 }
1820 //------------------------------Identity---------------------------------------
1821 // Clearing a zero length array does nothing
1822 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
1823 return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this;
1824 }
1826 //------------------------------Idealize---------------------------------------
1827 // Clearing a short array is faster with stores
1828 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
1829 const int unit = BytesPerLong;
1830 const TypeX* t = phase->type(in(2))->isa_intptr_t();
1831 if (!t) return NULL;
1832 if (!t->is_con()) return NULL;
1833 intptr_t raw_count = t->get_con();
1834 intptr_t size = raw_count;
1835 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
1836 // Clearing nothing uses the Identity call.
1837 // Negative clears are possible on dead ClearArrays
1838 // (see jck test stmt114.stmt11402.val).
1839 if (size <= 0 || size % unit != 0) return NULL;
1840 intptr_t count = size / unit;
1841 // Length too long; use fast hardware clear
1842 if (size > Matcher::init_array_short_size) return NULL;
1843 Node *mem = in(1);
1844 if( phase->type(mem)==Type::TOP ) return NULL;
1845 Node *adr = in(3);
1846 const Type* at = phase->type(adr);
1847 if( at==Type::TOP ) return NULL;
1848 const TypePtr* atp = at->isa_ptr();
1849 // adjust atp to be the correct array element address type
1850 if (atp == NULL) atp = TypePtr::BOTTOM;
1851 else atp = atp->add_offset(Type::OffsetBot);
1852 // Get base for derived pointer purposes
1853 if( adr->Opcode() != Op_AddP ) Unimplemented();
1854 Node *base = adr->in(1);
1856 Node *zero = phase->makecon(TypeLong::ZERO);
1857 Node *off = phase->MakeConX(BytesPerLong);
1858 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1859 count--;
1860 while( count-- ) {
1861 mem = phase->transform(mem);
1862 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
1863 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1864 }
1865 return mem;
1866 }
1868 //----------------------------clear_memory-------------------------------------
1869 // Generate code to initialize object storage to zero.
1870 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1871 intptr_t start_offset,
1872 Node* end_offset,
1873 PhaseGVN* phase) {
1874 Compile* C = phase->C;
1875 intptr_t offset = start_offset;
1877 int unit = BytesPerLong;
1878 if ((offset % unit) != 0) {
1879 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
1880 adr = phase->transform(adr);
1881 const TypePtr* atp = TypeRawPtr::BOTTOM;
1882 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1883 mem = phase->transform(mem);
1884 offset += BytesPerInt;
1885 }
1886 assert((offset % unit) == 0, "");
1888 // Initialize the remaining stuff, if any, with a ClearArray.
1889 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
1890 }
1892 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1893 Node* start_offset,
1894 Node* end_offset,
1895 PhaseGVN* phase) {
1896 Compile* C = phase->C;
1897 int unit = BytesPerLong;
1898 Node* zbase = start_offset;
1899 Node* zend = end_offset;
1901 // Scale to the unit required by the CPU:
1902 if (!Matcher::init_array_count_is_in_bytes) {
1903 Node* shift = phase->intcon(exact_log2(unit));
1904 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
1905 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
1906 }
1908 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
1909 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
1911 // Bulk clear double-words
1912 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
1913 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
1914 return phase->transform(mem);
1915 }
1917 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1918 intptr_t start_offset,
1919 intptr_t end_offset,
1920 PhaseGVN* phase) {
1921 Compile* C = phase->C;
1922 assert((end_offset % BytesPerInt) == 0, "odd end offset");
1923 intptr_t done_offset = end_offset;
1924 if ((done_offset % BytesPerLong) != 0) {
1925 done_offset -= BytesPerInt;
1926 }
1927 if (done_offset > start_offset) {
1928 mem = clear_memory(ctl, mem, dest,
1929 start_offset, phase->MakeConX(done_offset), phase);
1930 }
1931 if (done_offset < end_offset) { // emit the final 32-bit store
1932 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
1933 adr = phase->transform(adr);
1934 const TypePtr* atp = TypeRawPtr::BOTTOM;
1935 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1936 mem = phase->transform(mem);
1937 done_offset += BytesPerInt;
1938 }
1939 assert(done_offset == end_offset, "");
1940 return mem;
1941 }
1943 //=============================================================================
1944 // Do we match on this edge? No memory edges
1945 uint StrCompNode::match_edge(uint idx) const {
1946 return idx == 5 || idx == 6;
1947 }
1949 //------------------------------Ideal------------------------------------------
1950 // Return a node which is more "ideal" than the current node. Strip out
1951 // control copies
1952 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
1953 return remove_dead_region(phase, can_reshape) ? this : NULL;
1954 }
1957 //=============================================================================
1958 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
1959 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
1960 _adr_type(C->get_adr_type(alias_idx))
1961 {
1962 init_class_id(Class_MemBar);
1963 Node* top = C->top();
1964 init_req(TypeFunc::I_O,top);
1965 init_req(TypeFunc::FramePtr,top);
1966 init_req(TypeFunc::ReturnAdr,top);
1967 if (precedent != NULL)
1968 init_req(TypeFunc::Parms, precedent);
1969 }
1971 //------------------------------cmp--------------------------------------------
1972 uint MemBarNode::hash() const { return NO_HASH; }
1973 uint MemBarNode::cmp( const Node &n ) const {
1974 return (&n == this); // Always fail except on self
1975 }
1977 //------------------------------make-------------------------------------------
1978 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
1979 int len = Precedent + (pn == NULL? 0: 1);
1980 switch (opcode) {
1981 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
1982 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
1983 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
1984 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
1985 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
1986 default: ShouldNotReachHere(); return NULL;
1987 }
1988 }
1990 //------------------------------Ideal------------------------------------------
1991 // Return a node which is more "ideal" than the current node. Strip out
1992 // control copies
1993 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1994 if (remove_dead_region(phase, can_reshape)) return this;
1995 return NULL;
1996 }
1998 //------------------------------Value------------------------------------------
1999 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2000 if( !in(0) ) return Type::TOP;
2001 if( phase->type(in(0)) == Type::TOP )
2002 return Type::TOP;
2003 return TypeTuple::MEMBAR;
2004 }
2006 //------------------------------match------------------------------------------
2007 // Construct projections for memory.
2008 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2009 switch (proj->_con) {
2010 case TypeFunc::Control:
2011 case TypeFunc::Memory:
2012 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2013 }
2014 ShouldNotReachHere();
2015 return NULL;
2016 }
2018 //===========================InitializeNode====================================
2019 // SUMMARY:
2020 // This node acts as a memory barrier on raw memory, after some raw stores.
2021 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2022 // The Initialize can 'capture' suitably constrained stores as raw inits.
2023 // It can coalesce related raw stores into larger units (called 'tiles').
2024 // It can avoid zeroing new storage for memory units which have raw inits.
2025 // At macro-expansion, it is marked 'complete', and does not optimize further.
2026 //
2027 // EXAMPLE:
2028 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2029 // ctl = incoming control; mem* = incoming memory
2030 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2031 // First allocate uninitialized memory and fill in the header:
2032 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2033 // ctl := alloc.Control; mem* := alloc.Memory*
2034 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2035 // Then initialize to zero the non-header parts of the raw memory block:
2036 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2037 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2038 // After the initialize node executes, the object is ready for service:
2039 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2040 // Suppose its body is immediately initialized as {1,2}:
2041 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2042 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2043 // mem.SLICE(#short[*]) := store2
2044 //
2045 // DETAILS:
2046 // An InitializeNode collects and isolates object initialization after
2047 // an AllocateNode and before the next possible safepoint. As a
2048 // memory barrier (MemBarNode), it keeps critical stores from drifting
2049 // down past any safepoint or any publication of the allocation.
2050 // Before this barrier, a newly-allocated object may have uninitialized bits.
2051 // After this barrier, it may be treated as a real oop, and GC is allowed.
2052 //
2053 // The semantics of the InitializeNode include an implicit zeroing of
2054 // the new object from object header to the end of the object.
2055 // (The object header and end are determined by the AllocateNode.)
2056 //
2057 // Certain stores may be added as direct inputs to the InitializeNode.
2058 // These stores must update raw memory, and they must be to addresses
2059 // derived from the raw address produced by AllocateNode, and with
2060 // a constant offset. They must be ordered by increasing offset.
2061 // The first one is at in(RawStores), the last at in(req()-1).
2062 // Unlike most memory operations, they are not linked in a chain,
2063 // but are displayed in parallel as users of the rawmem output of
2064 // the allocation.
2065 //
2066 // (See comments in InitializeNode::capture_store, which continue
2067 // the example given above.)
2068 //
2069 // When the associated Allocate is macro-expanded, the InitializeNode
2070 // may be rewritten to optimize collected stores. A ClearArrayNode
2071 // may also be created at that point to represent any required zeroing.
2072 // The InitializeNode is then marked 'complete', prohibiting further
2073 // capturing of nearby memory operations.
2074 //
2075 // During macro-expansion, all captured initializations which store
2076 // constant values of 32 bits or smaller are coalesced (if advantagous)
2077 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2078 // initialized in fewer memory operations. Memory words which are
2079 // covered by neither tiles nor non-constant stores are pre-zeroed
2080 // by explicit stores of zero. (The code shape happens to do all
2081 // zeroing first, then all other stores, with both sequences occurring
2082 // in order of ascending offsets.)
2083 //
2084 // Alternatively, code may be inserted between an AllocateNode and its
2085 // InitializeNode, to perform arbitrary initialization of the new object.
2086 // E.g., the object copying intrinsics insert complex data transfers here.
2087 // The initialization must then be marked as 'complete' disable the
2088 // built-in zeroing semantics and the collection of initializing stores.
2089 //
2090 // While an InitializeNode is incomplete, reads from the memory state
2091 // produced by it are optimizable if they match the control edge and
2092 // new oop address associated with the allocation/initialization.
2093 // They return a stored value (if the offset matches) or else zero.
2094 // A write to the memory state, if it matches control and address,
2095 // and if it is to a constant offset, may be 'captured' by the
2096 // InitializeNode. It is cloned as a raw memory operation and rewired
2097 // inside the initialization, to the raw oop produced by the allocation.
2098 // Operations on addresses which are provably distinct (e.g., to
2099 // other AllocateNodes) are allowed to bypass the initialization.
2100 //
2101 // The effect of all this is to consolidate object initialization
2102 // (both arrays and non-arrays, both piecewise and bulk) into a
2103 // single location, where it can be optimized as a unit.
2104 //
2105 // Only stores with an offset less than TrackedInitializationLimit words
2106 // will be considered for capture by an InitializeNode. This puts a
2107 // reasonable limit on the complexity of optimized initializations.
2109 //---------------------------InitializeNode------------------------------------
2110 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2111 : _is_complete(false),
2112 MemBarNode(C, adr_type, rawoop)
2113 {
2114 init_class_id(Class_Initialize);
2116 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2117 assert(in(RawAddress) == rawoop, "proper init");
2118 // Note: allocation() can be NULL, for secondary initialization barriers
2119 }
2121 // Since this node is not matched, it will be processed by the
2122 // register allocator. Declare that there are no constraints
2123 // on the allocation of the RawAddress edge.
2124 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2125 // This edge should be set to top, by the set_complete. But be conservative.
2126 if (idx == InitializeNode::RawAddress)
2127 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2128 return RegMask::Empty;
2129 }
2131 Node* InitializeNode::memory(uint alias_idx) {
2132 Node* mem = in(Memory);
2133 if (mem->is_MergeMem()) {
2134 return mem->as_MergeMem()->memory_at(alias_idx);
2135 } else {
2136 // incoming raw memory is not split
2137 return mem;
2138 }
2139 }
2141 bool InitializeNode::is_non_zero() {
2142 if (is_complete()) return false;
2143 remove_extra_zeroes();
2144 return (req() > RawStores);
2145 }
2147 void InitializeNode::set_complete(PhaseGVN* phase) {
2148 assert(!is_complete(), "caller responsibility");
2149 _is_complete = true;
2151 // After this node is complete, it contains a bunch of
2152 // raw-memory initializations. There is no need for
2153 // it to have anything to do with non-raw memory effects.
2154 // Therefore, tell all non-raw users to re-optimize themselves,
2155 // after skipping the memory effects of this initialization.
2156 PhaseIterGVN* igvn = phase->is_IterGVN();
2157 if (igvn) igvn->add_users_to_worklist(this);
2158 }
2160 // convenience function
2161 // return false if the init contains any stores already
2162 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2163 InitializeNode* init = initialization();
2164 if (init == NULL || init->is_complete()) return false;
2165 init->remove_extra_zeroes();
2166 // for now, if this allocation has already collected any inits, bail:
2167 if (init->is_non_zero()) return false;
2168 init->set_complete(phase);
2169 return true;
2170 }
2172 void InitializeNode::remove_extra_zeroes() {
2173 if (req() == RawStores) return;
2174 Node* zmem = zero_memory();
2175 uint fill = RawStores;
2176 for (uint i = fill; i < req(); i++) {
2177 Node* n = in(i);
2178 if (n->is_top() || n == zmem) continue; // skip
2179 if (fill < i) set_req(fill, n); // compact
2180 ++fill;
2181 }
2182 // delete any empty spaces created:
2183 while (fill < req()) {
2184 del_req(fill);
2185 }
2186 }
2188 // Helper for remembering which stores go with which offsets.
2189 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2190 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2191 intptr_t offset = -1;
2192 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2193 phase, offset);
2194 if (base == NULL) return -1; // something is dead,
2195 if (offset < 0) return -1; // dead, dead
2196 return offset;
2197 }
2199 // Helper for proving that an initialization expression is
2200 // "simple enough" to be folded into an object initialization.
2201 // Attempts to prove that a store's initial value 'n' can be captured
2202 // within the initialization without creating a vicious cycle, such as:
2203 // { Foo p = new Foo(); p.next = p; }
2204 // True for constants and parameters and small combinations thereof.
2205 bool InitializeNode::detect_init_independence(Node* n,
2206 bool st_is_pinned,
2207 int& count) {
2208 if (n == NULL) return true; // (can this really happen?)
2209 if (n->is_Proj()) n = n->in(0);
2210 if (n == this) return false; // found a cycle
2211 if (n->is_Con()) return true;
2212 if (n->is_Start()) return true; // params, etc., are OK
2213 if (n->is_Root()) return true; // even better
2215 Node* ctl = n->in(0);
2216 if (ctl != NULL && !ctl->is_top()) {
2217 if (ctl->is_Proj()) ctl = ctl->in(0);
2218 if (ctl == this) return false;
2220 // If we already know that the enclosing memory op is pinned right after
2221 // the init, then any control flow that the store has picked up
2222 // must have preceded the init, or else be equal to the init.
2223 // Even after loop optimizations (which might change control edges)
2224 // a store is never pinned *before* the availability of its inputs.
2225 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
2226 return false; // failed to prove a good control
2228 }
2230 // Check data edges for possible dependencies on 'this'.
2231 if ((count += 1) > 20) return false; // complexity limit
2232 for (uint i = 1; i < n->req(); i++) {
2233 Node* m = n->in(i);
2234 if (m == NULL || m == n || m->is_top()) continue;
2235 uint first_i = n->find_edge(m);
2236 if (i != first_i) continue; // process duplicate edge just once
2237 if (!detect_init_independence(m, st_is_pinned, count)) {
2238 return false;
2239 }
2240 }
2242 return true;
2243 }
2245 // Here are all the checks a Store must pass before it can be moved into
2246 // an initialization. Returns zero if a check fails.
2247 // On success, returns the (constant) offset to which the store applies,
2248 // within the initialized memory.
2249 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2250 const int FAIL = 0;
2251 if (st->req() != MemNode::ValueIn + 1)
2252 return FAIL; // an inscrutable StoreNode (card mark?)
2253 Node* ctl = st->in(MemNode::Control);
2254 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2255 return FAIL; // must be unconditional after the initialization
2256 Node* mem = st->in(MemNode::Memory);
2257 if (!(mem->is_Proj() && mem->in(0) == this))
2258 return FAIL; // must not be preceded by other stores
2259 Node* adr = st->in(MemNode::Address);
2260 intptr_t offset;
2261 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2262 if (alloc == NULL)
2263 return FAIL; // inscrutable address
2264 if (alloc != allocation())
2265 return FAIL; // wrong allocation! (store needs to float up)
2266 Node* val = st->in(MemNode::ValueIn);
2267 int complexity_count = 0;
2268 if (!detect_init_independence(val, true, complexity_count))
2269 return FAIL; // stored value must be 'simple enough'
2271 return offset; // success
2272 }
2274 // Find the captured store in(i) which corresponds to the range
2275 // [start..start+size) in the initialized object.
2276 // If there is one, return its index i. If there isn't, return the
2277 // negative of the index where it should be inserted.
2278 // Return 0 if the queried range overlaps an initialization boundary
2279 // or if dead code is encountered.
2280 // If size_in_bytes is zero, do not bother with overlap checks.
2281 int InitializeNode::captured_store_insertion_point(intptr_t start,
2282 int size_in_bytes,
2283 PhaseTransform* phase) {
2284 const int FAIL = 0, MAX_STORE = BytesPerLong;
2286 if (is_complete())
2287 return FAIL; // arraycopy got here first; punt
2289 assert(allocation() != NULL, "must be present");
2291 // no negatives, no header fields:
2292 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
2293 if (start < (intptr_t) sizeof(arrayOopDesc) &&
2294 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2296 // after a certain size, we bail out on tracking all the stores:
2297 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2298 if (start >= ti_limit) return FAIL;
2300 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2301 if (i >= limit) return -(int)i; // not found; here is where to put it
2303 Node* st = in(i);
2304 intptr_t st_off = get_store_offset(st, phase);
2305 if (st_off < 0) {
2306 if (st != zero_memory()) {
2307 return FAIL; // bail out if there is dead garbage
2308 }
2309 } else if (st_off > start) {
2310 // ...we are done, since stores are ordered
2311 if (st_off < start + size_in_bytes) {
2312 return FAIL; // the next store overlaps
2313 }
2314 return -(int)i; // not found; here is where to put it
2315 } else if (st_off < start) {
2316 if (size_in_bytes != 0 &&
2317 start < st_off + MAX_STORE &&
2318 start < st_off + st->as_Store()->memory_size()) {
2319 return FAIL; // the previous store overlaps
2320 }
2321 } else {
2322 if (size_in_bytes != 0 &&
2323 st->as_Store()->memory_size() != size_in_bytes) {
2324 return FAIL; // mismatched store size
2325 }
2326 return i;
2327 }
2329 ++i;
2330 }
2331 }
2333 // Look for a captured store which initializes at the offset 'start'
2334 // with the given size. If there is no such store, and no other
2335 // initialization interferes, then return zero_memory (the memory
2336 // projection of the AllocateNode).
2337 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2338 PhaseTransform* phase) {
2339 assert(stores_are_sane(phase), "");
2340 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2341 if (i == 0) {
2342 return NULL; // something is dead
2343 } else if (i < 0) {
2344 return zero_memory(); // just primordial zero bits here
2345 } else {
2346 Node* st = in(i); // here is the store at this position
2347 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2348 return st;
2349 }
2350 }
2352 // Create, as a raw pointer, an address within my new object at 'offset'.
2353 Node* InitializeNode::make_raw_address(intptr_t offset,
2354 PhaseTransform* phase) {
2355 Node* addr = in(RawAddress);
2356 if (offset != 0) {
2357 Compile* C = phase->C;
2358 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2359 phase->MakeConX(offset)) );
2360 }
2361 return addr;
2362 }
2364 // Clone the given store, converting it into a raw store
2365 // initializing a field or element of my new object.
2366 // Caller is responsible for retiring the original store,
2367 // with subsume_node or the like.
2368 //
2369 // From the example above InitializeNode::InitializeNode,
2370 // here are the old stores to be captured:
2371 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2372 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2373 //
2374 // Here is the changed code; note the extra edges on init:
2375 // alloc = (Allocate ...)
2376 // rawoop = alloc.RawAddress
2377 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2378 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2379 // init = (Initialize alloc.Control alloc.Memory rawoop
2380 // rawstore1 rawstore2)
2381 //
2382 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2383 PhaseTransform* phase) {
2384 assert(stores_are_sane(phase), "");
2386 if (start < 0) return NULL;
2387 assert(can_capture_store(st, phase) == start, "sanity");
2389 Compile* C = phase->C;
2390 int size_in_bytes = st->memory_size();
2391 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2392 if (i == 0) return NULL; // bail out
2393 Node* prev_mem = NULL; // raw memory for the captured store
2394 if (i > 0) {
2395 prev_mem = in(i); // there is a pre-existing store under this one
2396 set_req(i, C->top()); // temporarily disconnect it
2397 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2398 } else {
2399 i = -i; // no pre-existing store
2400 prev_mem = zero_memory(); // a slice of the newly allocated object
2401 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2402 set_req(--i, C->top()); // reuse this edge; it has been folded away
2403 else
2404 ins_req(i, C->top()); // build a new edge
2405 }
2406 Node* new_st = st->clone();
2407 new_st->set_req(MemNode::Control, in(Control));
2408 new_st->set_req(MemNode::Memory, prev_mem);
2409 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2410 new_st = phase->transform(new_st);
2412 // At this point, new_st might have swallowed a pre-existing store
2413 // at the same offset, or perhaps new_st might have disappeared,
2414 // if it redundantly stored the same value (or zero to fresh memory).
2416 // In any case, wire it in:
2417 set_req(i, new_st);
2419 // The caller may now kill the old guy.
2420 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2421 assert(check_st == new_st || check_st == NULL, "must be findable");
2422 assert(!is_complete(), "");
2423 return new_st;
2424 }
2426 static bool store_constant(jlong* tiles, int num_tiles,
2427 intptr_t st_off, int st_size,
2428 jlong con) {
2429 if ((st_off & (st_size-1)) != 0)
2430 return false; // strange store offset (assume size==2**N)
2431 address addr = (address)tiles + st_off;
2432 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2433 switch (st_size) {
2434 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2435 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2436 case sizeof(jint): *(jint*) addr = (jint) con; break;
2437 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2438 default: return false; // strange store size (detect size!=2**N here)
2439 }
2440 return true; // return success to caller
2441 }
2443 // Coalesce subword constants into int constants and possibly
2444 // into long constants. The goal, if the CPU permits,
2445 // is to initialize the object with a small number of 64-bit tiles.
2446 // Also, convert floating-point constants to bit patterns.
2447 // Non-constants are not relevant to this pass.
2448 //
2449 // In terms of the running example on InitializeNode::InitializeNode
2450 // and InitializeNode::capture_store, here is the transformation
2451 // of rawstore1 and rawstore2 into rawstore12:
2452 // alloc = (Allocate ...)
2453 // rawoop = alloc.RawAddress
2454 // tile12 = 0x00010002
2455 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2456 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2457 //
2458 void
2459 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2460 Node* size_in_bytes,
2461 PhaseGVN* phase) {
2462 Compile* C = phase->C;
2464 assert(stores_are_sane(phase), "");
2465 // Note: After this pass, they are not completely sane,
2466 // since there may be some overlaps.
2468 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2470 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2471 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2472 size_limit = MIN2(size_limit, ti_limit);
2473 size_limit = align_size_up(size_limit, BytesPerLong);
2474 int num_tiles = size_limit / BytesPerLong;
2476 // allocate space for the tile map:
2477 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2478 jlong tiles_buf[small_len];
2479 Node* nodes_buf[small_len];
2480 jlong inits_buf[small_len];
2481 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2482 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2483 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2484 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2485 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2486 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2487 // tiles: exact bitwise model of all primitive constants
2488 // nodes: last constant-storing node subsumed into the tiles model
2489 // inits: which bytes (in each tile) are touched by any initializations
2491 //// Pass A: Fill in the tile model with any relevant stores.
2493 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2494 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2495 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2496 Node* zmem = zero_memory(); // initially zero memory state
2497 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2498 Node* st = in(i);
2499 intptr_t st_off = get_store_offset(st, phase);
2501 // Figure out the store's offset and constant value:
2502 if (st_off < header_size) continue; //skip (ignore header)
2503 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2504 int st_size = st->as_Store()->memory_size();
2505 if (st_off + st_size > size_limit) break;
2507 // Record which bytes are touched, whether by constant or not.
2508 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2509 continue; // skip (strange store size)
2511 const Type* val = phase->type(st->in(MemNode::ValueIn));
2512 if (!val->singleton()) continue; //skip (non-con store)
2513 BasicType type = val->basic_type();
2515 jlong con = 0;
2516 switch (type) {
2517 case T_INT: con = val->is_int()->get_con(); break;
2518 case T_LONG: con = val->is_long()->get_con(); break;
2519 case T_FLOAT: con = jint_cast(val->getf()); break;
2520 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2521 default: continue; //skip (odd store type)
2522 }
2524 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2525 st->Opcode() == Op_StoreL) {
2526 continue; // This StoreL is already optimal.
2527 }
2529 // Store down the constant.
2530 store_constant(tiles, num_tiles, st_off, st_size, con);
2532 intptr_t j = st_off >> LogBytesPerLong;
2534 if (type == T_INT && st_size == BytesPerInt
2535 && (st_off & BytesPerInt) == BytesPerInt) {
2536 jlong lcon = tiles[j];
2537 if (!Matcher::isSimpleConstant64(lcon) &&
2538 st->Opcode() == Op_StoreI) {
2539 // This StoreI is already optimal by itself.
2540 jint* intcon = (jint*) &tiles[j];
2541 intcon[1] = 0; // undo the store_constant()
2543 // If the previous store is also optimal by itself, back up and
2544 // undo the action of the previous loop iteration... if we can.
2545 // But if we can't, just let the previous half take care of itself.
2546 st = nodes[j];
2547 st_off -= BytesPerInt;
2548 con = intcon[0];
2549 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2550 assert(st_off >= header_size, "still ignoring header");
2551 assert(get_store_offset(st, phase) == st_off, "must be");
2552 assert(in(i-1) == zmem, "must be");
2553 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2554 assert(con == tcon->is_int()->get_con(), "must be");
2555 // Undo the effects of the previous loop trip, which swallowed st:
2556 intcon[0] = 0; // undo store_constant()
2557 set_req(i-1, st); // undo set_req(i, zmem)
2558 nodes[j] = NULL; // undo nodes[j] = st
2559 --old_subword; // undo ++old_subword
2560 }
2561 continue; // This StoreI is already optimal.
2562 }
2563 }
2565 // This store is not needed.
2566 set_req(i, zmem);
2567 nodes[j] = st; // record for the moment
2568 if (st_size < BytesPerLong) // something has changed
2569 ++old_subword; // includes int/float, but who's counting...
2570 else ++old_long;
2571 }
2573 if ((old_subword + old_long) == 0)
2574 return; // nothing more to do
2576 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2577 // Be sure to insert them before overlapping non-constant stores.
2578 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2579 for (int j = 0; j < num_tiles; j++) {
2580 jlong con = tiles[j];
2581 jlong init = inits[j];
2582 if (con == 0) continue;
2583 jint con0, con1; // split the constant, address-wise
2584 jint init0, init1; // split the init map, address-wise
2585 { union { jlong con; jint intcon[2]; } u;
2586 u.con = con;
2587 con0 = u.intcon[0];
2588 con1 = u.intcon[1];
2589 u.con = init;
2590 init0 = u.intcon[0];
2591 init1 = u.intcon[1];
2592 }
2594 Node* old = nodes[j];
2595 assert(old != NULL, "need the prior store");
2596 intptr_t offset = (j * BytesPerLong);
2598 bool split = !Matcher::isSimpleConstant64(con);
2600 if (offset < header_size) {
2601 assert(offset + BytesPerInt >= header_size, "second int counts");
2602 assert(*(jint*)&tiles[j] == 0, "junk in header");
2603 split = true; // only the second word counts
2604 // Example: int a[] = { 42 ... }
2605 } else if (con0 == 0 && init0 == -1) {
2606 split = true; // first word is covered by full inits
2607 // Example: int a[] = { ... foo(), 42 ... }
2608 } else if (con1 == 0 && init1 == -1) {
2609 split = true; // second word is covered by full inits
2610 // Example: int a[] = { ... 42, foo() ... }
2611 }
2613 // Here's a case where init0 is neither 0 nor -1:
2614 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2615 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2616 // In this case the tile is not split; it is (jlong)42.
2617 // The big tile is stored down, and then the foo() value is inserted.
2618 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2620 Node* ctl = old->in(MemNode::Control);
2621 Node* adr = make_raw_address(offset, phase);
2622 const TypePtr* atp = TypeRawPtr::BOTTOM;
2624 // One or two coalesced stores to plop down.
2625 Node* st[2];
2626 intptr_t off[2];
2627 int nst = 0;
2628 if (!split) {
2629 ++new_long;
2630 off[nst] = offset;
2631 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2632 phase->longcon(con), T_LONG);
2633 } else {
2634 // Omit either if it is a zero.
2635 if (con0 != 0) {
2636 ++new_int;
2637 off[nst] = offset;
2638 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2639 phase->intcon(con0), T_INT);
2640 }
2641 if (con1 != 0) {
2642 ++new_int;
2643 offset += BytesPerInt;
2644 adr = make_raw_address(offset, phase);
2645 off[nst] = offset;
2646 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2647 phase->intcon(con1), T_INT);
2648 }
2649 }
2651 // Insert second store first, then the first before the second.
2652 // Insert each one just before any overlapping non-constant stores.
2653 while (nst > 0) {
2654 Node* st1 = st[--nst];
2655 C->copy_node_notes_to(st1, old);
2656 st1 = phase->transform(st1);
2657 offset = off[nst];
2658 assert(offset >= header_size, "do not smash header");
2659 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
2660 guarantee(ins_idx != 0, "must re-insert constant store");
2661 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
2662 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
2663 set_req(--ins_idx, st1);
2664 else
2665 ins_req(ins_idx, st1);
2666 }
2667 }
2669 if (PrintCompilation && WizardMode)
2670 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
2671 old_subword, old_long, new_int, new_long);
2672 if (C->log() != NULL)
2673 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
2674 old_subword, old_long, new_int, new_long);
2676 // Clean up any remaining occurrences of zmem:
2677 remove_extra_zeroes();
2678 }
2680 // Explore forward from in(start) to find the first fully initialized
2681 // word, and return its offset. Skip groups of subword stores which
2682 // together initialize full words. If in(start) is itself part of a
2683 // fully initialized word, return the offset of in(start). If there
2684 // are no following full-word stores, or if something is fishy, return
2685 // a negative value.
2686 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
2687 int int_map = 0;
2688 intptr_t int_map_off = 0;
2689 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
2691 for (uint i = start, limit = req(); i < limit; i++) {
2692 Node* st = in(i);
2694 intptr_t st_off = get_store_offset(st, phase);
2695 if (st_off < 0) break; // return conservative answer
2697 int st_size = st->as_Store()->memory_size();
2698 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
2699 return st_off; // we found a complete word init
2700 }
2702 // update the map:
2704 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
2705 if (this_int_off != int_map_off) {
2706 // reset the map:
2707 int_map = 0;
2708 int_map_off = this_int_off;
2709 }
2711 int subword_off = st_off - this_int_off;
2712 int_map |= right_n_bits(st_size) << subword_off;
2713 if ((int_map & FULL_MAP) == FULL_MAP) {
2714 return this_int_off; // we found a complete word init
2715 }
2717 // Did this store hit or cross the word boundary?
2718 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
2719 if (next_int_off == this_int_off + BytesPerInt) {
2720 // We passed the current int, without fully initializing it.
2721 int_map_off = next_int_off;
2722 int_map >>= BytesPerInt;
2723 } else if (next_int_off > this_int_off + BytesPerInt) {
2724 // We passed the current and next int.
2725 return this_int_off + BytesPerInt;
2726 }
2727 }
2729 return -1;
2730 }
2733 // Called when the associated AllocateNode is expanded into CFG.
2734 // At this point, we may perform additional optimizations.
2735 // Linearize the stores by ascending offset, to make memory
2736 // activity as coherent as possible.
2737 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
2738 intptr_t header_size,
2739 Node* size_in_bytes,
2740 PhaseGVN* phase) {
2741 assert(!is_complete(), "not already complete");
2742 assert(stores_are_sane(phase), "");
2743 assert(allocation() != NULL, "must be present");
2745 remove_extra_zeroes();
2747 if (ReduceFieldZeroing || ReduceBulkZeroing)
2748 // reduce instruction count for common initialization patterns
2749 coalesce_subword_stores(header_size, size_in_bytes, phase);
2751 Node* zmem = zero_memory(); // initially zero memory state
2752 Node* inits = zmem; // accumulating a linearized chain of inits
2753 #ifdef ASSERT
2754 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
2755 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
2756 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
2757 #endif
2758 intptr_t zeroes_done = header_size;
2760 bool do_zeroing = true; // we might give up if inits are very sparse
2761 int big_init_gaps = 0; // how many large gaps have we seen?
2763 if (ZeroTLAB) do_zeroing = false;
2764 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
2766 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2767 Node* st = in(i);
2768 intptr_t st_off = get_store_offset(st, phase);
2769 if (st_off < 0)
2770 break; // unknown junk in the inits
2771 if (st->in(MemNode::Memory) != zmem)
2772 break; // complicated store chains somehow in list
2774 int st_size = st->as_Store()->memory_size();
2775 intptr_t next_init_off = st_off + st_size;
2777 if (do_zeroing && zeroes_done < next_init_off) {
2778 // See if this store needs a zero before it or under it.
2779 intptr_t zeroes_needed = st_off;
2781 if (st_size < BytesPerInt) {
2782 // Look for subword stores which only partially initialize words.
2783 // If we find some, we must lay down some word-level zeroes first,
2784 // underneath the subword stores.
2785 //
2786 // Examples:
2787 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
2788 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
2789 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
2790 //
2791 // Note: coalesce_subword_stores may have already done this,
2792 // if it was prompted by constant non-zero subword initializers.
2793 // But this case can still arise with non-constant stores.
2795 intptr_t next_full_store = find_next_fullword_store(i, phase);
2797 // In the examples above:
2798 // in(i) p q r s x y z
2799 // st_off 12 13 14 15 12 13 14
2800 // st_size 1 1 1 1 1 1 1
2801 // next_full_s. 12 16 16 16 16 16 16
2802 // z's_done 12 16 16 16 12 16 12
2803 // z's_needed 12 16 16 16 16 16 16
2804 // zsize 0 0 0 0 4 0 4
2805 if (next_full_store < 0) {
2806 // Conservative tack: Zero to end of current word.
2807 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
2808 } else {
2809 // Zero to beginning of next fully initialized word.
2810 // Or, don't zero at all, if we are already in that word.
2811 assert(next_full_store >= zeroes_needed, "must go forward");
2812 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
2813 zeroes_needed = next_full_store;
2814 }
2815 }
2817 if (zeroes_needed > zeroes_done) {
2818 intptr_t zsize = zeroes_needed - zeroes_done;
2819 // Do some incremental zeroing on rawmem, in parallel with inits.
2820 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2821 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2822 zeroes_done, zeroes_needed,
2823 phase);
2824 zeroes_done = zeroes_needed;
2825 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
2826 do_zeroing = false; // leave the hole, next time
2827 }
2828 }
2830 // Collect the store and move on:
2831 st->set_req(MemNode::Memory, inits);
2832 inits = st; // put it on the linearized chain
2833 set_req(i, zmem); // unhook from previous position
2835 if (zeroes_done == st_off)
2836 zeroes_done = next_init_off;
2838 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
2840 #ifdef ASSERT
2841 // Various order invariants. Weaker than stores_are_sane because
2842 // a large constant tile can be filled in by smaller non-constant stores.
2843 assert(st_off >= last_init_off, "inits do not reverse");
2844 last_init_off = st_off;
2845 const Type* val = NULL;
2846 if (st_size >= BytesPerInt &&
2847 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
2848 (int)val->basic_type() < (int)T_OBJECT) {
2849 assert(st_off >= last_tile_end, "tiles do not overlap");
2850 assert(st_off >= last_init_end, "tiles do not overwrite inits");
2851 last_tile_end = MAX2(last_tile_end, next_init_off);
2852 } else {
2853 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
2854 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
2855 assert(st_off >= last_init_end, "inits do not overlap");
2856 last_init_end = next_init_off; // it's a non-tile
2857 }
2858 #endif //ASSERT
2859 }
2861 remove_extra_zeroes(); // clear out all the zmems left over
2862 add_req(inits);
2864 if (!ZeroTLAB) {
2865 // If anything remains to be zeroed, zero it all now.
2866 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2867 // if it is the last unused 4 bytes of an instance, forget about it
2868 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
2869 if (zeroes_done + BytesPerLong >= size_limit) {
2870 assert(allocation() != NULL, "");
2871 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
2872 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
2873 if (zeroes_done == k->layout_helper())
2874 zeroes_done = size_limit;
2875 }
2876 if (zeroes_done < size_limit) {
2877 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2878 zeroes_done, size_in_bytes, phase);
2879 }
2880 }
2882 set_complete(phase);
2883 return rawmem;
2884 }
2887 #ifdef ASSERT
2888 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
2889 if (is_complete())
2890 return true; // stores could be anything at this point
2891 intptr_t last_off = sizeof(oopDesc);
2892 for (uint i = InitializeNode::RawStores; i < req(); i++) {
2893 Node* st = in(i);
2894 intptr_t st_off = get_store_offset(st, phase);
2895 if (st_off < 0) continue; // ignore dead garbage
2896 if (last_off > st_off) {
2897 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
2898 this->dump(2);
2899 assert(false, "ascending store offsets");
2900 return false;
2901 }
2902 last_off = st_off + st->as_Store()->memory_size();
2903 }
2904 return true;
2905 }
2906 #endif //ASSERT
2911 //============================MergeMemNode=====================================
2912 //
2913 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
2914 // contributing store or call operations. Each contributor provides the memory
2915 // state for a particular "alias type" (see Compile::alias_type). For example,
2916 // if a MergeMem has an input X for alias category #6, then any memory reference
2917 // to alias category #6 may use X as its memory state input, as an exact equivalent
2918 // to using the MergeMem as a whole.
2919 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
2920 //
2921 // (Here, the <N> notation gives the index of the relevant adr_type.)
2922 //
2923 // In one special case (and more cases in the future), alias categories overlap.
2924 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
2925 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
2926 // it is exactly equivalent to that state W:
2927 // MergeMem(<Bot>: W) <==> W
2928 //
2929 // Usually, the merge has more than one input. In that case, where inputs
2930 // overlap (i.e., one is Bot), the narrower alias type determines the memory
2931 // state for that type, and the wider alias type (Bot) fills in everywhere else:
2932 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
2933 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
2934 //
2935 // A merge can take a "wide" memory state as one of its narrow inputs.
2936 // This simply means that the merge observes out only the relevant parts of
2937 // the wide input. That is, wide memory states arriving at narrow merge inputs
2938 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
2939 //
2940 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
2941 // and that memory slices "leak through":
2942 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
2943 //
2944 // But, in such a cascade, repeated memory slices can "block the leak":
2945 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
2946 //
2947 // In the last example, Y is not part of the combined memory state of the
2948 // outermost MergeMem. The system must, of course, prevent unschedulable
2949 // memory states from arising, so you can be sure that the state Y is somehow
2950 // a precursor to state Y'.
2951 //
2952 //
2953 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
2954 // of each MergeMemNode array are exactly the numerical alias indexes, including
2955 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
2956 // Compile::alias_type (and kin) produce and manage these indexes.
2957 //
2958 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
2959 // (Note that this provides quick access to the top node inside MergeMem methods,
2960 // without the need to reach out via TLS to Compile::current.)
2961 //
2962 // As a consequence of what was just described, a MergeMem that represents a full
2963 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
2964 // containing all alias categories.
2965 //
2966 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
2967 //
2968 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
2969 // a memory state for the alias type <N>, or else the top node, meaning that
2970 // there is no particular input for that alias type. Note that the length of
2971 // a MergeMem is variable, and may be extended at any time to accommodate new
2972 // memory states at larger alias indexes. When merges grow, they are of course
2973 // filled with "top" in the unused in() positions.
2974 //
2975 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
2976 // (Top was chosen because it works smoothly with passes like GCM.)
2977 //
2978 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
2979 // the type of random VM bits like TLS references.) Since it is always the
2980 // first non-Bot memory slice, some low-level loops use it to initialize an
2981 // index variable: for (i = AliasIdxRaw; i < req(); i++).
2982 //
2983 //
2984 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
2985 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
2986 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
2987 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
2988 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
2989 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
2990 //
2991 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
2992 // really that different from the other memory inputs. An abbreviation called
2993 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
2994 //
2995 //
2996 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
2997 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
2998 // that "emerges though" the base memory will be marked as excluding the alias types
2999 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3000 //
3001 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3002 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3003 //
3004 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3005 // (It is currently unimplemented.) As you can see, the resulting merge is
3006 // actually a disjoint union of memory states, rather than an overlay.
3007 //
3009 //------------------------------MergeMemNode-----------------------------------
3010 Node* MergeMemNode::make_empty_memory() {
3011 Node* empty_memory = (Node*) Compile::current()->top();
3012 assert(empty_memory->is_top(), "correct sentinel identity");
3013 return empty_memory;
3014 }
3016 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3017 init_class_id(Class_MergeMem);
3018 // all inputs are nullified in Node::Node(int)
3019 // set_input(0, NULL); // no control input
3021 // Initialize the edges uniformly to top, for starters.
3022 Node* empty_mem = make_empty_memory();
3023 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3024 init_req(i,empty_mem);
3025 }
3026 assert(empty_memory() == empty_mem, "");
3028 if( new_base != NULL && new_base->is_MergeMem() ) {
3029 MergeMemNode* mdef = new_base->as_MergeMem();
3030 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3031 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3032 mms.set_memory(mms.memory2());
3033 }
3034 assert(base_memory() == mdef->base_memory(), "");
3035 } else {
3036 set_base_memory(new_base);
3037 }
3038 }
3040 // Make a new, untransformed MergeMem with the same base as 'mem'.
3041 // If mem is itself a MergeMem, populate the result with the same edges.
3042 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3043 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3044 }
3046 //------------------------------cmp--------------------------------------------
3047 uint MergeMemNode::hash() const { return NO_HASH; }
3048 uint MergeMemNode::cmp( const Node &n ) const {
3049 return (&n == this); // Always fail except on self
3050 }
3052 //------------------------------Identity---------------------------------------
3053 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3054 // Identity if this merge point does not record any interesting memory
3055 // disambiguations.
3056 Node* base_mem = base_memory();
3057 Node* empty_mem = empty_memory();
3058 if (base_mem != empty_mem) { // Memory path is not dead?
3059 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3060 Node* mem = in(i);
3061 if (mem != empty_mem && mem != base_mem) {
3062 return this; // Many memory splits; no change
3063 }
3064 }
3065 }
3066 return base_mem; // No memory splits; ID on the one true input
3067 }
3069 //------------------------------Ideal------------------------------------------
3070 // This method is invoked recursively on chains of MergeMem nodes
3071 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3072 // Remove chain'd MergeMems
3073 //
3074 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3075 // relative to the "in(Bot)". Since we are patching both at the same time,
3076 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3077 // but rewrite each "in(i)" relative to the new "in(Bot)".
3078 Node *progress = NULL;
3081 Node* old_base = base_memory();
3082 Node* empty_mem = empty_memory();
3083 if (old_base == empty_mem)
3084 return NULL; // Dead memory path.
3086 MergeMemNode* old_mbase;
3087 if (old_base != NULL && old_base->is_MergeMem())
3088 old_mbase = old_base->as_MergeMem();
3089 else
3090 old_mbase = NULL;
3091 Node* new_base = old_base;
3093 // simplify stacked MergeMems in base memory
3094 if (old_mbase) new_base = old_mbase->base_memory();
3096 // the base memory might contribute new slices beyond my req()
3097 if (old_mbase) grow_to_match(old_mbase);
3099 // Look carefully at the base node if it is a phi.
3100 PhiNode* phi_base;
3101 if (new_base != NULL && new_base->is_Phi())
3102 phi_base = new_base->as_Phi();
3103 else
3104 phi_base = NULL;
3106 Node* phi_reg = NULL;
3107 uint phi_len = (uint)-1;
3108 if (phi_base != NULL && !phi_base->is_copy()) {
3109 // do not examine phi if degraded to a copy
3110 phi_reg = phi_base->region();
3111 phi_len = phi_base->req();
3112 // see if the phi is unfinished
3113 for (uint i = 1; i < phi_len; i++) {
3114 if (phi_base->in(i) == NULL) {
3115 // incomplete phi; do not look at it yet!
3116 phi_reg = NULL;
3117 phi_len = (uint)-1;
3118 break;
3119 }
3120 }
3121 }
3123 // Note: We do not call verify_sparse on entry, because inputs
3124 // can normalize to the base_memory via subsume_node or similar
3125 // mechanisms. This method repairs that damage.
3127 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3129 // Look at each slice.
3130 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3131 Node* old_in = in(i);
3132 // calculate the old memory value
3133 Node* old_mem = old_in;
3134 if (old_mem == empty_mem) old_mem = old_base;
3135 assert(old_mem == memory_at(i), "");
3137 // maybe update (reslice) the old memory value
3139 // simplify stacked MergeMems
3140 Node* new_mem = old_mem;
3141 MergeMemNode* old_mmem;
3142 if (old_mem != NULL && old_mem->is_MergeMem())
3143 old_mmem = old_mem->as_MergeMem();
3144 else
3145 old_mmem = NULL;
3146 if (old_mmem == this) {
3147 // This can happen if loops break up and safepoints disappear.
3148 // A merge of BotPtr (default) with a RawPtr memory derived from a
3149 // safepoint can be rewritten to a merge of the same BotPtr with
3150 // the BotPtr phi coming into the loop. If that phi disappears
3151 // also, we can end up with a self-loop of the mergemem.
3152 // In general, if loops degenerate and memory effects disappear,
3153 // a mergemem can be left looking at itself. This simply means
3154 // that the mergemem's default should be used, since there is
3155 // no longer any apparent effect on this slice.
3156 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3157 // from start. Update the input to TOP.
3158 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3159 }
3160 else if (old_mmem != NULL) {
3161 new_mem = old_mmem->memory_at(i);
3162 }
3163 // else preceeding memory was not a MergeMem
3165 // replace equivalent phis (unfortunately, they do not GVN together)
3166 if (new_mem != NULL && new_mem != new_base &&
3167 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3168 if (new_mem->is_Phi()) {
3169 PhiNode* phi_mem = new_mem->as_Phi();
3170 for (uint i = 1; i < phi_len; i++) {
3171 if (phi_base->in(i) != phi_mem->in(i)) {
3172 phi_mem = NULL;
3173 break;
3174 }
3175 }
3176 if (phi_mem != NULL) {
3177 // equivalent phi nodes; revert to the def
3178 new_mem = new_base;
3179 }
3180 }
3181 }
3183 // maybe store down a new value
3184 Node* new_in = new_mem;
3185 if (new_in == new_base) new_in = empty_mem;
3187 if (new_in != old_in) {
3188 // Warning: Do not combine this "if" with the previous "if"
3189 // A memory slice might have be be rewritten even if it is semantically
3190 // unchanged, if the base_memory value has changed.
3191 set_req(i, new_in);
3192 progress = this; // Report progress
3193 }
3194 }
3196 if (new_base != old_base) {
3197 set_req(Compile::AliasIdxBot, new_base);
3198 // Don't use set_base_memory(new_base), because we need to update du.
3199 assert(base_memory() == new_base, "");
3200 progress = this;
3201 }
3203 if( base_memory() == this ) {
3204 // a self cycle indicates this memory path is dead
3205 set_req(Compile::AliasIdxBot, empty_mem);
3206 }
3208 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3209 // Recursion must occur after the self cycle check above
3210 if( base_memory()->is_MergeMem() ) {
3211 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3212 Node *m = phase->transform(new_mbase); // Rollup any cycles
3213 if( m != NULL && (m->is_top() ||
3214 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3215 // propagate rollup of dead cycle to self
3216 set_req(Compile::AliasIdxBot, empty_mem);
3217 }
3218 }
3220 if( base_memory() == empty_mem ) {
3221 progress = this;
3222 // Cut inputs during Parse phase only.
3223 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3224 if( !can_reshape ) {
3225 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3226 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3227 }
3228 }
3229 }
3231 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3232 // Check if PhiNode::Ideal's "Split phis through memory merges"
3233 // transform should be attempted. Look for this->phi->this cycle.
3234 uint merge_width = req();
3235 if (merge_width > Compile::AliasIdxRaw) {
3236 PhiNode* phi = base_memory()->as_Phi();
3237 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3238 if (phi->in(i) == this) {
3239 phase->is_IterGVN()->_worklist.push(phi);
3240 break;
3241 }
3242 }
3243 }
3244 }
3246 assert(verify_sparse(), "please, no dups of base");
3247 return progress;
3248 }
3250 //-------------------------set_base_memory-------------------------------------
3251 void MergeMemNode::set_base_memory(Node *new_base) {
3252 Node* empty_mem = empty_memory();
3253 set_req(Compile::AliasIdxBot, new_base);
3254 assert(memory_at(req()) == new_base, "must set default memory");
3255 // Clear out other occurrences of new_base:
3256 if (new_base != empty_mem) {
3257 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3258 if (in(i) == new_base) set_req(i, empty_mem);
3259 }
3260 }
3261 }
3263 //------------------------------out_RegMask------------------------------------
3264 const RegMask &MergeMemNode::out_RegMask() const {
3265 return RegMask::Empty;
3266 }
3268 //------------------------------dump_spec--------------------------------------
3269 #ifndef PRODUCT
3270 void MergeMemNode::dump_spec(outputStream *st) const {
3271 st->print(" {");
3272 Node* base_mem = base_memory();
3273 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3274 Node* mem = memory_at(i);
3275 if (mem == base_mem) { st->print(" -"); continue; }
3276 st->print( " N%d:", mem->_idx );
3277 Compile::current()->get_adr_type(i)->dump_on(st);
3278 }
3279 st->print(" }");
3280 }
3281 #endif // !PRODUCT
3284 #ifdef ASSERT
3285 static bool might_be_same(Node* a, Node* b) {
3286 if (a == b) return true;
3287 if (!(a->is_Phi() || b->is_Phi())) return false;
3288 // phis shift around during optimization
3289 return true; // pretty stupid...
3290 }
3292 // verify a narrow slice (either incoming or outgoing)
3293 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3294 if (!VerifyAliases) return; // don't bother to verify unless requested
3295 if (is_error_reported()) return; // muzzle asserts when debugging an error
3296 if (Node::in_dump()) return; // muzzle asserts when printing
3297 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3298 assert(n != NULL, "");
3299 // Elide intervening MergeMem's
3300 while (n->is_MergeMem()) {
3301 n = n->as_MergeMem()->memory_at(alias_idx);
3302 }
3303 Compile* C = Compile::current();
3304 const TypePtr* n_adr_type = n->adr_type();
3305 if (n == m->empty_memory()) {
3306 // Implicit copy of base_memory()
3307 } else if (n_adr_type != TypePtr::BOTTOM) {
3308 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3309 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3310 } else {
3311 // A few places like make_runtime_call "know" that VM calls are narrow,
3312 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3313 bool expected_wide_mem = false;
3314 if (n == m->base_memory()) {
3315 expected_wide_mem = true;
3316 } else if (alias_idx == Compile::AliasIdxRaw ||
3317 n == m->memory_at(Compile::AliasIdxRaw)) {
3318 expected_wide_mem = true;
3319 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3320 // memory can "leak through" calls on channels that
3321 // are write-once. Allow this also.
3322 expected_wide_mem = true;
3323 }
3324 assert(expected_wide_mem, "expected narrow slice replacement");
3325 }
3326 }
3327 #else // !ASSERT
3328 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3329 #endif
3332 //-----------------------------memory_at---------------------------------------
3333 Node* MergeMemNode::memory_at(uint alias_idx) const {
3334 assert(alias_idx >= Compile::AliasIdxRaw ||
3335 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3336 "must avoid base_memory and AliasIdxTop");
3338 // Otherwise, it is a narrow slice.
3339 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3340 Compile *C = Compile::current();
3341 if (is_empty_memory(n)) {
3342 // the array is sparse; empty slots are the "top" node
3343 n = base_memory();
3344 assert(Node::in_dump()
3345 || n == NULL || n->bottom_type() == Type::TOP
3346 || n->adr_type() == TypePtr::BOTTOM
3347 || n->adr_type() == TypeRawPtr::BOTTOM
3348 || Compile::current()->AliasLevel() == 0,
3349 "must be a wide memory");
3350 // AliasLevel == 0 if we are organizing the memory states manually.
3351 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3352 } else {
3353 // make sure the stored slice is sane
3354 #ifdef ASSERT
3355 if (is_error_reported() || Node::in_dump()) {
3356 } else if (might_be_same(n, base_memory())) {
3357 // Give it a pass: It is a mostly harmless repetition of the base.
3358 // This can arise normally from node subsumption during optimization.
3359 } else {
3360 verify_memory_slice(this, alias_idx, n);
3361 }
3362 #endif
3363 }
3364 return n;
3365 }
3367 //---------------------------set_memory_at-------------------------------------
3368 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3369 verify_memory_slice(this, alias_idx, n);
3370 Node* empty_mem = empty_memory();
3371 if (n == base_memory()) n = empty_mem; // collapse default
3372 uint need_req = alias_idx+1;
3373 if (req() < need_req) {
3374 if (n == empty_mem) return; // already the default, so do not grow me
3375 // grow the sparse array
3376 do {
3377 add_req(empty_mem);
3378 } while (req() < need_req);
3379 }
3380 set_req( alias_idx, n );
3381 }
3385 //--------------------------iteration_setup------------------------------------
3386 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3387 if (other != NULL) {
3388 grow_to_match(other);
3389 // invariant: the finite support of mm2 is within mm->req()
3390 #ifdef ASSERT
3391 for (uint i = req(); i < other->req(); i++) {
3392 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3393 }
3394 #endif
3395 }
3396 // Replace spurious copies of base_memory by top.
3397 Node* base_mem = base_memory();
3398 if (base_mem != NULL && !base_mem->is_top()) {
3399 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3400 if (in(i) == base_mem)
3401 set_req(i, empty_memory());
3402 }
3403 }
3404 }
3406 //---------------------------grow_to_match-------------------------------------
3407 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3408 Node* empty_mem = empty_memory();
3409 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3410 // look for the finite support of the other memory
3411 for (uint i = other->req(); --i >= req(); ) {
3412 if (other->in(i) != empty_mem) {
3413 uint new_len = i+1;
3414 while (req() < new_len) add_req(empty_mem);
3415 break;
3416 }
3417 }
3418 }
3420 //---------------------------verify_sparse-------------------------------------
3421 #ifndef PRODUCT
3422 bool MergeMemNode::verify_sparse() const {
3423 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3424 Node* base_mem = base_memory();
3425 // The following can happen in degenerate cases, since empty==top.
3426 if (is_empty_memory(base_mem)) return true;
3427 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3428 assert(in(i) != NULL, "sane slice");
3429 if (in(i) == base_mem) return false; // should have been the sentinel value!
3430 }
3431 return true;
3432 }
3434 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3435 Node* n;
3436 n = mm->in(idx);
3437 if (mem == n) return true; // might be empty_memory()
3438 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3439 if (mem == n) return true;
3440 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3441 if (mem == n) return true;
3442 if (n == NULL) break;
3443 }
3444 return false;
3445 }
3446 #endif // !PRODUCT