Fri, 29 Feb 2008 19:57:41 -0800
6667618: disable LoadL->ConvL2I ==> LoadI optimization
Summary: this optimization causes problems (sizes of Load and Store nodes do not match) for objects initialization code and Escape Analysis
Reviewed-by: jrose, never
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 // The code which unhooks non-raw memories from complete (macro-expanded)
112 // initializations was removed. After macro-expansion all stores catched
113 // by Initialize node became raw stores and there is no information
114 // which memory slices they modify. So it is unsafe to move any memory
115 // operation above these stores. Also in most cases hooked non-raw memories
116 // were already unhooked by using information from detect_ptr_independence()
117 // and find_previous_store().
119 if (mem->is_MergeMem()) {
120 MergeMemNode* mmem = mem->as_MergeMem();
121 const TypePtr *tp = t_adr->is_ptr();
122 uint alias_idx = phase->C->get_alias_index(tp);
123 #ifdef ASSERT
124 {
125 // Check that current type is consistent with the alias index used during graph construction
126 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
127 const TypePtr *adr_t = adr_type();
128 bool consistent = adr_t == NULL || adr_t->empty() || phase->C->must_alias(adr_t, alias_idx );
129 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
130 if( !consistent && adr_t != NULL && !adr_t->empty() &&
131 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
132 adr_t->isa_aryptr() && adr_t->offset() != Type::OffsetBot &&
133 ( adr_t->offset() == arrayOopDesc::length_offset_in_bytes() ||
134 adr_t->offset() == oopDesc::klass_offset_in_bytes() ||
135 adr_t->offset() == oopDesc::mark_offset_in_bytes() ) ) {
136 // don't assert if it is dead code.
137 consistent = true;
138 }
139 if( !consistent ) {
140 tty->print("alias_idx==%d, adr_type()==", alias_idx); if( adr_t == NULL ) { tty->print("NULL"); } else { adr_t->dump(); }
141 tty->cr();
142 print_alias_types();
143 assert(consistent, "adr_type must match alias idx");
144 }
145 }
146 #endif
147 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
148 // means an array I have not precisely typed yet. Do not do any
149 // alias stuff with it any time soon.
150 const TypeInstPtr *tinst = tp->isa_instptr();
151 if( tp->base() != Type::AnyPtr &&
152 !(tinst &&
153 tinst->klass()->is_java_lang_Object() &&
154 tinst->offset() == Type::OffsetBot) ) {
155 // compress paths and change unreachable cycles to TOP
156 // If not, we can update the input infinitely along a MergeMem cycle
157 // Equivalent code in PhiNode::Ideal
158 Node* m = phase->transform(mmem);
159 // If tranformed to a MergeMem, get the desired slice
160 // Otherwise the returned node represents memory for every slice
161 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
162 // Update input if it is progress over what we have now
163 }
164 }
166 if (mem != old_mem) {
167 set_req(MemNode::Memory, mem);
168 return this;
169 }
171 // let the subclass continue analyzing...
172 return NULL;
173 }
175 // Helper function for proving some simple control dominations.
176 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'.
177 // Already assumes that 'dom' is available at 'sub', and that 'sub'
178 // is not a constant (dominated by the method's StartNode).
179 // Used by MemNode::find_previous_store to prove that the
180 // control input of a memory operation predates (dominates)
181 // an allocation it wants to look past.
182 bool MemNode::detect_dominating_control(Node* dom, Node* sub) {
183 if (dom == NULL) return false;
184 if (dom->is_Proj()) dom = dom->in(0);
185 if (dom->is_Start()) return true; // anything inside the method
186 if (dom->is_Root()) return true; // dom 'controls' a constant
187 int cnt = 20; // detect cycle or too much effort
188 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
189 if (--cnt < 0) return false; // in a cycle or too complex
190 if (sub == dom) return true;
191 if (sub->is_Start()) return false;
192 if (sub->is_Root()) return false;
193 Node* up = sub->in(0);
194 if (sub == up && sub->is_Region()) {
195 for (uint i = 1; i < sub->req(); i++) {
196 Node* in = sub->in(i);
197 if (in != NULL && !in->is_top() && in != sub) {
198 up = in; break; // take any path on the way up to 'dom'
199 }
200 }
201 }
202 if (sub == up) return false; // some kind of tight cycle
203 sub = up;
204 }
205 return false;
206 }
208 //---------------------detect_ptr_independence---------------------------------
209 // Used by MemNode::find_previous_store to prove that two base
210 // pointers are never equal.
211 // The pointers are accompanied by their associated allocations,
212 // if any, which have been previously discovered by the caller.
213 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
214 Node* p2, AllocateNode* a2,
215 PhaseTransform* phase) {
216 // Attempt to prove that these two pointers cannot be aliased.
217 // They may both manifestly be allocations, and they should differ.
218 // Or, if they are not both allocations, they can be distinct constants.
219 // Otherwise, one is an allocation and the other a pre-existing value.
220 if (a1 == NULL && a2 == NULL) { // neither an allocation
221 return (p1 != p2) && p1->is_Con() && p2->is_Con();
222 } else if (a1 != NULL && a2 != NULL) { // both allocations
223 return (a1 != a2);
224 } else if (a1 != NULL) { // one allocation a1
225 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
226 return detect_dominating_control(p2->in(0), a1->in(0));
227 } else { //(a2 != NULL) // one allocation a2
228 return detect_dominating_control(p1->in(0), a2->in(0));
229 }
230 return false;
231 }
234 // The logic for reordering loads and stores uses four steps:
235 // (a) Walk carefully past stores and initializations which we
236 // can prove are independent of this load.
237 // (b) Observe that the next memory state makes an exact match
238 // with self (load or store), and locate the relevant store.
239 // (c) Ensure that, if we were to wire self directly to the store,
240 // the optimizer would fold it up somehow.
241 // (d) Do the rewiring, and return, depending on some other part of
242 // the optimizer to fold up the load.
243 // This routine handles steps (a) and (b). Steps (c) and (d) are
244 // specific to loads and stores, so they are handled by the callers.
245 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
246 //
247 Node* MemNode::find_previous_store(PhaseTransform* phase) {
248 Node* ctrl = in(MemNode::Control);
249 Node* adr = in(MemNode::Address);
250 intptr_t offset = 0;
251 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
252 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
254 if (offset == Type::OffsetBot)
255 return NULL; // cannot unalias unless there are precise offsets
257 intptr_t size_in_bytes = memory_size();
259 Node* mem = in(MemNode::Memory); // start searching here...
261 int cnt = 50; // Cycle limiter
262 for (;;) { // While we can dance past unrelated stores...
263 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
265 if (mem->is_Store()) {
266 Node* st_adr = mem->in(MemNode::Address);
267 intptr_t st_offset = 0;
268 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
269 if (st_base == NULL)
270 break; // inscrutable pointer
271 if (st_offset != offset && st_offset != Type::OffsetBot) {
272 const int MAX_STORE = BytesPerLong;
273 if (st_offset >= offset + size_in_bytes ||
274 st_offset <= offset - MAX_STORE ||
275 st_offset <= offset - mem->as_Store()->memory_size()) {
276 // Success: The offsets are provably independent.
277 // (You may ask, why not just test st_offset != offset and be done?
278 // The answer is that stores of different sizes can co-exist
279 // in the same sequence of RawMem effects. We sometimes initialize
280 // a whole 'tile' of array elements with a single jint or jlong.)
281 mem = mem->in(MemNode::Memory);
282 continue; // (a) advance through independent store memory
283 }
284 }
285 if (st_base != base &&
286 detect_ptr_independence(base, alloc,
287 st_base,
288 AllocateNode::Ideal_allocation(st_base, phase),
289 phase)) {
290 // Success: The bases are provably independent.
291 mem = mem->in(MemNode::Memory);
292 continue; // (a) advance through independent store memory
293 }
295 // (b) At this point, if the bases or offsets do not agree, we lose,
296 // since we have not managed to prove 'this' and 'mem' independent.
297 if (st_base == base && st_offset == offset) {
298 return mem; // let caller handle steps (c), (d)
299 }
301 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
302 InitializeNode* st_init = mem->in(0)->as_Initialize();
303 AllocateNode* st_alloc = st_init->allocation();
304 if (st_alloc == NULL)
305 break; // something degenerated
306 bool known_identical = false;
307 bool known_independent = false;
308 if (alloc == st_alloc)
309 known_identical = true;
310 else if (alloc != NULL)
311 known_independent = true;
312 else if (ctrl != NULL &&
313 detect_dominating_control(ctrl, st_alloc->in(0)))
314 known_independent = true;
316 if (known_independent) {
317 // The bases are provably independent: Either they are
318 // manifestly distinct allocations, or else the control
319 // of this load dominates the store's allocation.
320 int alias_idx = phase->C->get_alias_index(adr_type());
321 if (alias_idx == Compile::AliasIdxRaw) {
322 mem = st_alloc->in(TypeFunc::Memory);
323 } else {
324 mem = st_init->memory(alias_idx);
325 }
326 continue; // (a) advance through independent store memory
327 }
329 // (b) at this point, if we are not looking at a store initializing
330 // the same allocation we are loading from, we lose.
331 if (known_identical) {
332 // From caller, can_see_stored_value will consult find_captured_store.
333 return mem; // let caller handle steps (c), (d)
334 }
336 }
338 // Unless there is an explicit 'continue', we must bail out here,
339 // because 'mem' is an inscrutable memory state (e.g., a call).
340 break;
341 }
343 return NULL; // bail out
344 }
346 //----------------------calculate_adr_type-------------------------------------
347 // Helper function. Notices when the given type of address hits top or bottom.
348 // Also, asserts a cross-check of the type against the expected address type.
349 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
350 if (t == Type::TOP) return NULL; // does not touch memory any more?
351 #ifdef PRODUCT
352 cross_check = NULL;
353 #else
354 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
355 #endif
356 const TypePtr* tp = t->isa_ptr();
357 if (tp == NULL) {
358 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
359 return TypePtr::BOTTOM; // touches lots of memory
360 } else {
361 #ifdef ASSERT
362 // %%%% [phh] We don't check the alias index if cross_check is
363 // TypeRawPtr::BOTTOM. Needs to be investigated.
364 if (cross_check != NULL &&
365 cross_check != TypePtr::BOTTOM &&
366 cross_check != TypeRawPtr::BOTTOM) {
367 // Recheck the alias index, to see if it has changed (due to a bug).
368 Compile* C = Compile::current();
369 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
370 "must stay in the original alias category");
371 // The type of the address must be contained in the adr_type,
372 // disregarding "null"-ness.
373 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
374 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
375 assert(cross_check->meet(tp_notnull) == cross_check,
376 "real address must not escape from expected memory type");
377 }
378 #endif
379 return tp;
380 }
381 }
383 //------------------------adr_phi_is_loop_invariant----------------------------
384 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
385 // loop is loop invariant. Make a quick traversal of Phi and associated
386 // CastPP nodes, looking to see if they are a closed group within the loop.
387 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
388 // The idea is that the phi-nest must boil down to only CastPP nodes
389 // with the same data. This implies that any path into the loop already
390 // includes such a CastPP, and so the original cast, whatever its input,
391 // must be covered by an equivalent cast, with an earlier control input.
392 ResourceMark rm;
394 // The loop entry input of the phi should be the unique dominating
395 // node for every Phi/CastPP in the loop.
396 Unique_Node_List closure;
397 closure.push(adr_phi->in(LoopNode::EntryControl));
399 // Add the phi node and the cast to the worklist.
400 Unique_Node_List worklist;
401 worklist.push(adr_phi);
402 if( cast != NULL ){
403 if( !cast->is_ConstraintCast() ) return false;
404 worklist.push(cast);
405 }
407 // Begin recursive walk of phi nodes.
408 while( worklist.size() ){
409 // Take a node off the worklist
410 Node *n = worklist.pop();
411 if( !closure.member(n) ){
412 // Add it to the closure.
413 closure.push(n);
414 // Make a sanity check to ensure we don't waste too much time here.
415 if( closure.size() > 20) return false;
416 // This node is OK if:
417 // - it is a cast of an identical value
418 // - or it is a phi node (then we add its inputs to the worklist)
419 // Otherwise, the node is not OK, and we presume the cast is not invariant
420 if( n->is_ConstraintCast() ){
421 worklist.push(n->in(1));
422 } else if( n->is_Phi() ) {
423 for( uint i = 1; i < n->req(); i++ ) {
424 worklist.push(n->in(i));
425 }
426 } else {
427 return false;
428 }
429 }
430 }
432 // Quit when the worklist is empty, and we've found no offending nodes.
433 return true;
434 }
436 //------------------------------Ideal_DU_postCCP-------------------------------
437 // Find any cast-away of null-ness and keep its control. Null cast-aways are
438 // going away in this pass and we need to make this memory op depend on the
439 // gating null check.
441 // I tried to leave the CastPP's in. This makes the graph more accurate in
442 // some sense; we get to keep around the knowledge that an oop is not-null
443 // after some test. Alas, the CastPP's interfere with GVN (some values are
444 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
445 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
446 // some of the more trivial cases in the optimizer. Removing more useless
447 // Phi's started allowing Loads to illegally float above null checks. I gave
448 // up on this approach. CNC 10/20/2000
449 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
450 Node *ctr = in(MemNode::Control);
451 Node *mem = in(MemNode::Memory);
452 Node *adr = in(MemNode::Address);
453 Node *skipped_cast = NULL;
454 // Need a null check? Regular static accesses do not because they are
455 // from constant addresses. Array ops are gated by the range check (which
456 // always includes a NULL check). Just check field ops.
457 if( !ctr ) {
458 // Scan upwards for the highest location we can place this memory op.
459 while( true ) {
460 switch( adr->Opcode() ) {
462 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
463 adr = adr->in(AddPNode::Base);
464 continue;
466 case Op_CastPP:
467 // If the CastPP is useless, just peek on through it.
468 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
469 // Remember the cast that we've peeked though. If we peek
470 // through more than one, then we end up remembering the highest
471 // one, that is, if in a loop, the one closest to the top.
472 skipped_cast = adr;
473 adr = adr->in(1);
474 continue;
475 }
476 // CastPP is going away in this pass! We need this memory op to be
477 // control-dependent on the test that is guarding the CastPP.
478 ccp->hash_delete(this);
479 set_req(MemNode::Control, adr->in(0));
480 ccp->hash_insert(this);
481 return this;
483 case Op_Phi:
484 // Attempt to float above a Phi to some dominating point.
485 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
486 // If we've already peeked through a Cast (which could have set the
487 // control), we can't float above a Phi, because the skipped Cast
488 // may not be loop invariant.
489 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
490 adr = adr->in(1);
491 continue;
492 }
493 }
495 // Intentional fallthrough!
497 // No obvious dominating point. The mem op is pinned below the Phi
498 // by the Phi itself. If the Phi goes away (no true value is merged)
499 // then the mem op can float, but not indefinitely. It must be pinned
500 // behind the controls leading to the Phi.
501 case Op_CheckCastPP:
502 // These usually stick around to change address type, however a
503 // useless one can be elided and we still need to pick up a control edge
504 if (adr->in(0) == NULL) {
505 // This CheckCastPP node has NO control and is likely useless. But we
506 // need check further up the ancestor chain for a control input to keep
507 // the node in place. 4959717.
508 skipped_cast = adr;
509 adr = adr->in(1);
510 continue;
511 }
512 ccp->hash_delete(this);
513 set_req(MemNode::Control, adr->in(0));
514 ccp->hash_insert(this);
515 return this;
517 // List of "safe" opcodes; those that implicitly block the memory
518 // op below any null check.
519 case Op_CastX2P: // no null checks on native pointers
520 case Op_Parm: // 'this' pointer is not null
521 case Op_LoadP: // Loading from within a klass
522 case Op_LoadKlass: // Loading from within a klass
523 case Op_ConP: // Loading from a klass
524 case Op_CreateEx: // Sucking up the guts of an exception oop
525 case Op_Con: // Reading from TLS
526 case Op_CMoveP: // CMoveP is pinned
527 break; // No progress
529 case Op_Proj: // Direct call to an allocation routine
530 case Op_SCMemProj: // Memory state from store conditional ops
531 #ifdef ASSERT
532 {
533 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
534 const Node* call = adr->in(0);
535 if (call->is_CallStaticJava()) {
536 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
537 assert(call_java && call_java->method() == NULL, "must be runtime call");
538 // We further presume that this is one of
539 // new_instance_Java, new_array_Java, or
540 // the like, but do not assert for this.
541 } else if (call->is_Allocate()) {
542 // similar case to new_instance_Java, etc.
543 } else if (!call->is_CallLeaf()) {
544 // Projections from fetch_oop (OSR) are allowed as well.
545 ShouldNotReachHere();
546 }
547 }
548 #endif
549 break;
550 default:
551 ShouldNotReachHere();
552 }
553 break;
554 }
555 }
557 return NULL; // No progress
558 }
561 //=============================================================================
562 uint LoadNode::size_of() const { return sizeof(*this); }
563 uint LoadNode::cmp( const Node &n ) const
564 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
565 const Type *LoadNode::bottom_type() const { return _type; }
566 uint LoadNode::ideal_reg() const {
567 return Matcher::base2reg[_type->base()];
568 }
570 #ifndef PRODUCT
571 void LoadNode::dump_spec(outputStream *st) const {
572 MemNode::dump_spec(st);
573 if( !Verbose && !WizardMode ) {
574 // standard dump does this in Verbose and WizardMode
575 st->print(" #"); _type->dump_on(st);
576 }
577 }
578 #endif
581 //----------------------------LoadNode::make-----------------------------------
582 // Polymorphic factory method:
583 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
584 // sanity check the alias category against the created node type
585 assert(!(adr_type->isa_oopptr() &&
586 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
587 "use LoadKlassNode instead");
588 assert(!(adr_type->isa_aryptr() &&
589 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
590 "use LoadRangeNode instead");
591 switch (bt) {
592 case T_BOOLEAN:
593 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
594 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
595 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
596 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
597 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
598 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
599 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
600 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
601 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
602 }
603 ShouldNotReachHere();
604 return (LoadNode*)NULL;
605 }
607 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
608 bool require_atomic = true;
609 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
610 }
615 //------------------------------hash-------------------------------------------
616 uint LoadNode::hash() const {
617 // unroll addition of interesting fields
618 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
619 }
621 //---------------------------can_see_stored_value------------------------------
622 // This routine exists to make sure this set of tests is done the same
623 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
624 // will change the graph shape in a way which makes memory alive twice at the
625 // same time (uses the Oracle model of aliasing), then some
626 // LoadXNode::Identity will fold things back to the equivalence-class model
627 // of aliasing.
628 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
629 Node* ld_adr = in(MemNode::Address);
631 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
632 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
633 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
634 atp->field() != NULL && !atp->field()->is_volatile()) {
635 uint alias_idx = atp->index();
636 bool final = atp->field()->is_final();
637 Node* result = NULL;
638 Node* current = st;
639 // Skip through chains of MemBarNodes checking the MergeMems for
640 // new states for the slice of this load. Stop once any other
641 // kind of node is encountered. Loads from final memory can skip
642 // through any kind of MemBar but normal loads shouldn't skip
643 // through MemBarAcquire since the could allow them to move out of
644 // a synchronized region.
645 while (current->is_Proj()) {
646 int opc = current->in(0)->Opcode();
647 if ((final && opc == Op_MemBarAcquire) ||
648 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
649 Node* mem = current->in(0)->in(TypeFunc::Memory);
650 if (mem->is_MergeMem()) {
651 MergeMemNode* merge = mem->as_MergeMem();
652 Node* new_st = merge->memory_at(alias_idx);
653 if (new_st == merge->base_memory()) {
654 // Keep searching
655 current = merge->base_memory();
656 continue;
657 }
658 // Save the new memory state for the slice and fall through
659 // to exit.
660 result = new_st;
661 }
662 }
663 break;
664 }
665 if (result != NULL) {
666 st = result;
667 }
668 }
671 // Loop around twice in the case Load -> Initialize -> Store.
672 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
673 for (int trip = 0; trip <= 1; trip++) {
675 if (st->is_Store()) {
676 Node* st_adr = st->in(MemNode::Address);
677 if (!phase->eqv(st_adr, ld_adr)) {
678 // Try harder before giving up... Match raw and non-raw pointers.
679 intptr_t st_off = 0;
680 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
681 if (alloc == NULL) return NULL;
682 intptr_t ld_off = 0;
683 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
684 if (alloc != allo2) return NULL;
685 if (ld_off != st_off) return NULL;
686 // At this point we have proven something like this setup:
687 // A = Allocate(...)
688 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
689 // S = StoreQ(, AddP(, A.Parm , #Off), V)
690 // (Actually, we haven't yet proven the Q's are the same.)
691 // In other words, we are loading from a casted version of
692 // the same pointer-and-offset that we stored to.
693 // Thus, we are able to replace L by V.
694 }
695 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
696 if (store_Opcode() != st->Opcode())
697 return NULL;
698 return st->in(MemNode::ValueIn);
699 }
701 intptr_t offset = 0; // scratch
703 // A load from a freshly-created object always returns zero.
704 // (This can happen after LoadNode::Ideal resets the load's memory input
705 // to find_captured_store, which returned InitializeNode::zero_memory.)
706 if (st->is_Proj() && st->in(0)->is_Allocate() &&
707 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
708 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
709 // return a zero value for the load's basic type
710 // (This is one of the few places where a generic PhaseTransform
711 // can create new nodes. Think of it as lazily manifesting
712 // virtually pre-existing constants.)
713 return phase->zerocon(memory_type());
714 }
716 // A load from an initialization barrier can match a captured store.
717 if (st->is_Proj() && st->in(0)->is_Initialize()) {
718 InitializeNode* init = st->in(0)->as_Initialize();
719 AllocateNode* alloc = init->allocation();
720 if (alloc != NULL &&
721 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
722 // examine a captured store value
723 st = init->find_captured_store(offset, memory_size(), phase);
724 if (st != NULL)
725 continue; // take one more trip around
726 }
727 }
729 break;
730 }
732 return NULL;
733 }
735 //------------------------------Identity---------------------------------------
736 // Loads are identity if previous store is to same address
737 Node *LoadNode::Identity( PhaseTransform *phase ) {
738 // If the previous store-maker is the right kind of Store, and the store is
739 // to the same address, then we are equal to the value stored.
740 Node* mem = in(MemNode::Memory);
741 Node* value = can_see_stored_value(mem, phase);
742 if( value ) {
743 // byte, short & char stores truncate naturally.
744 // A load has to load the truncated value which requires
745 // some sort of masking operation and that requires an
746 // Ideal call instead of an Identity call.
747 if (memory_size() < BytesPerInt) {
748 // If the input to the store does not fit with the load's result type,
749 // it must be truncated via an Ideal call.
750 if (!phase->type(value)->higher_equal(phase->type(this)))
751 return this;
752 }
753 // (This works even when value is a Con, but LoadNode::Value
754 // usually runs first, producing the singleton type of the Con.)
755 return value;
756 }
757 return this;
758 }
761 // Returns true if the AliasType refers to the field that holds the
762 // cached box array. Currently only handles the IntegerCache case.
763 static bool is_autobox_cache(Compile::AliasType* atp) {
764 if (atp != NULL && atp->field() != NULL) {
765 ciField* field = atp->field();
766 ciSymbol* klass = field->holder()->name();
767 if (field->name() == ciSymbol::cache_field_name() &&
768 field->holder()->uses_default_loader() &&
769 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
770 return true;
771 }
772 }
773 return false;
774 }
776 // Fetch the base value in the autobox array
777 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
778 if (atp != NULL && atp->field() != NULL) {
779 ciField* field = atp->field();
780 ciSymbol* klass = field->holder()->name();
781 if (field->name() == ciSymbol::cache_field_name() &&
782 field->holder()->uses_default_loader() &&
783 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
784 assert(field->is_constant(), "what?");
785 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
786 // Fetch the box object at the base of the array and get its value
787 ciInstance* box = array->obj_at(0)->as_instance();
788 ciInstanceKlass* ik = box->klass()->as_instance_klass();
789 if (ik->nof_nonstatic_fields() == 1) {
790 // This should be true nonstatic_field_at requires calling
791 // nof_nonstatic_fields so check it anyway
792 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
793 cache_offset = c.as_int();
794 }
795 return true;
796 }
797 }
798 return false;
799 }
801 // Returns true if the AliasType refers to the value field of an
802 // autobox object. Currently only handles Integer.
803 static bool is_autobox_object(Compile::AliasType* atp) {
804 if (atp != NULL && atp->field() != NULL) {
805 ciField* field = atp->field();
806 ciSymbol* klass = field->holder()->name();
807 if (field->name() == ciSymbol::value_name() &&
808 field->holder()->uses_default_loader() &&
809 klass == ciSymbol::java_lang_Integer()) {
810 return true;
811 }
812 }
813 return false;
814 }
817 // We're loading from an object which has autobox behaviour.
818 // If this object is result of a valueOf call we'll have a phi
819 // merging a newly allocated object and a load from the cache.
820 // We want to replace this load with the original incoming
821 // argument to the valueOf call.
822 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
823 Node* base = in(Address)->in(AddPNode::Base);
824 if (base->is_Phi() && base->req() == 3) {
825 AllocateNode* allocation = NULL;
826 int allocation_index = -1;
827 int load_index = -1;
828 for (uint i = 1; i < base->req(); i++) {
829 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
830 if (allocation != NULL) {
831 allocation_index = i;
832 load_index = 3 - allocation_index;
833 break;
834 }
835 }
836 LoadNode* load = NULL;
837 if (allocation != NULL && base->in(load_index)->is_Load()) {
838 load = base->in(load_index)->as_Load();
839 }
840 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
841 // Push the loads from the phi that comes from valueOf up
842 // through it to allow elimination of the loads and the recovery
843 // of the original value.
844 Node* mem_phi = in(Memory);
845 Node* offset = in(Address)->in(AddPNode::Offset);
847 Node* in1 = clone();
848 Node* in1_addr = in1->in(Address)->clone();
849 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
850 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
851 in1_addr->set_req(AddPNode::Offset, offset);
852 in1->set_req(0, base->in(allocation_index));
853 in1->set_req(Address, in1_addr);
854 in1->set_req(Memory, mem_phi->in(allocation_index));
856 Node* in2 = clone();
857 Node* in2_addr = in2->in(Address)->clone();
858 in2_addr->set_req(AddPNode::Base, base->in(load_index));
859 in2_addr->set_req(AddPNode::Address, base->in(load_index));
860 in2_addr->set_req(AddPNode::Offset, offset);
861 in2->set_req(0, base->in(load_index));
862 in2->set_req(Address, in2_addr);
863 in2->set_req(Memory, mem_phi->in(load_index));
865 in1_addr = phase->transform(in1_addr);
866 in1 = phase->transform(in1);
867 in2_addr = phase->transform(in2_addr);
868 in2 = phase->transform(in2);
870 PhiNode* result = PhiNode::make_blank(base->in(0), this);
871 result->set_req(allocation_index, in1);
872 result->set_req(load_index, in2);
873 return result;
874 }
875 } else if (base->is_Load()) {
876 // Eliminate the load of Integer.value for integers from the cache
877 // array by deriving the value from the index into the array.
878 // Capture the offset of the load and then reverse the computation.
879 Node* load_base = base->in(Address)->in(AddPNode::Base);
880 if (load_base != NULL) {
881 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
882 intptr_t cache_offset;
883 int shift = -1;
884 Node* cache = NULL;
885 if (is_autobox_cache(atp)) {
886 shift = exact_log2(type2aelembytes(T_OBJECT));
887 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
888 }
889 if (cache != NULL && base->in(Address)->is_AddP()) {
890 Node* elements[4];
891 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
892 int cache_low;
893 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
894 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
895 // Add up all the offsets making of the address of the load
896 Node* result = elements[0];
897 for (int i = 1; i < count; i++) {
898 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
899 }
900 // Remove the constant offset from the address and then
901 // remove the scaling of the offset to recover the original index.
902 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
903 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
904 // Peel the shift off directly but wrap it in a dummy node
905 // since Ideal can't return existing nodes
906 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
907 } else {
908 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
909 }
910 #ifdef _LP64
911 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
912 #endif
913 return result;
914 }
915 }
916 }
917 }
918 return NULL;
919 }
922 //------------------------------Ideal------------------------------------------
923 // If the load is from Field memory and the pointer is non-null, we can
924 // zero out the control input.
925 // If the offset is constant and the base is an object allocation,
926 // try to hook me up to the exact initializing store.
927 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
928 Node* p = MemNode::Ideal_common(phase, can_reshape);
929 if (p) return (p == NodeSentinel) ? NULL : p;
931 Node* ctrl = in(MemNode::Control);
932 Node* address = in(MemNode::Address);
934 // Skip up past a SafePoint control. Cannot do this for Stores because
935 // pointer stores & cardmarks must stay on the same side of a SafePoint.
936 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
937 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
938 ctrl = ctrl->in(0);
939 set_req(MemNode::Control,ctrl);
940 }
942 // Check for useless control edge in some common special cases
943 if (in(MemNode::Control) != NULL) {
944 intptr_t ignore = 0;
945 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
946 if (base != NULL
947 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
948 && detect_dominating_control(base->in(0), phase->C->start())) {
949 // A method-invariant, non-null address (constant or 'this' argument).
950 set_req(MemNode::Control, NULL);
951 }
952 }
954 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
955 Node* base = in(Address)->in(AddPNode::Base);
956 if (base != NULL) {
957 Compile::AliasType* atp = phase->C->alias_type(adr_type());
958 if (is_autobox_object(atp)) {
959 Node* result = eliminate_autobox(phase);
960 if (result != NULL) return result;
961 }
962 }
963 }
965 // Check for prior store with a different base or offset; make Load
966 // independent. Skip through any number of them. Bail out if the stores
967 // are in an endless dead cycle and report no progress. This is a key
968 // transform for Reflection. However, if after skipping through the Stores
969 // we can't then fold up against a prior store do NOT do the transform as
970 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
971 // array memory alive twice: once for the hoisted Load and again after the
972 // bypassed Store. This situation only works if EVERYBODY who does
973 // anti-dependence work knows how to bypass. I.e. we need all
974 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
975 // the alias index stuff. So instead, peek through Stores and IFF we can
976 // fold up, do so.
977 Node* prev_mem = find_previous_store(phase);
978 // Steps (a), (b): Walk past independent stores to find an exact match.
979 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
980 // (c) See if we can fold up on the spot, but don't fold up here.
981 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
982 // just return a prior value, which is done by Identity calls.
983 if (can_see_stored_value(prev_mem, phase)) {
984 // Make ready for step (d):
985 set_req(MemNode::Memory, prev_mem);
986 return this;
987 }
988 }
990 return NULL; // No further progress
991 }
993 // Helper to recognize certain Klass fields which are invariant across
994 // some group of array types (e.g., int[] or all T[] where T < Object).
995 const Type*
996 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
997 ciKlass* klass) const {
998 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
999 // The field is Klass::_modifier_flags. Return its (constant) value.
1000 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1001 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1002 return TypeInt::make(klass->modifier_flags());
1003 }
1004 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1005 // The field is Klass::_access_flags. Return its (constant) value.
1006 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1007 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1008 return TypeInt::make(klass->access_flags());
1009 }
1010 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1011 // The field is Klass::_layout_helper. Return its constant value if known.
1012 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1013 return TypeInt::make(klass->layout_helper());
1014 }
1016 // No match.
1017 return NULL;
1018 }
1020 //------------------------------Value-----------------------------------------
1021 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1022 // Either input is TOP ==> the result is TOP
1023 Node* mem = in(MemNode::Memory);
1024 const Type *t1 = phase->type(mem);
1025 if (t1 == Type::TOP) return Type::TOP;
1026 Node* adr = in(MemNode::Address);
1027 const TypePtr* tp = phase->type(adr)->isa_ptr();
1028 if (tp == NULL || tp->empty()) return Type::TOP;
1029 int off = tp->offset();
1030 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1032 // Try to guess loaded type from pointer type
1033 if (tp->base() == Type::AryPtr) {
1034 const Type *t = tp->is_aryptr()->elem();
1035 // Don't do this for integer types. There is only potential profit if
1036 // the element type t is lower than _type; that is, for int types, if _type is
1037 // more restrictive than t. This only happens here if one is short and the other
1038 // char (both 16 bits), and in those cases we've made an intentional decision
1039 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1040 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1041 //
1042 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1043 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1044 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1045 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1046 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1047 // In fact, that could have been the original type of p1, and p1 could have
1048 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1049 // expression (LShiftL quux 3) independently optimized to the constant 8.
1050 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1051 && Opcode() != Op_LoadKlass) {
1052 // t might actually be lower than _type, if _type is a unique
1053 // concrete subclass of abstract class t.
1054 // Make sure the reference is not into the header, by comparing
1055 // the offset against the offset of the start of the array's data.
1056 // Different array types begin at slightly different offsets (12 vs. 16).
1057 // We choose T_BYTE as an example base type that is least restrictive
1058 // as to alignment, which will therefore produce the smallest
1059 // possible base offset.
1060 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1061 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1062 const Type* jt = t->join(_type);
1063 // In any case, do not allow the join, per se, to empty out the type.
1064 if (jt->empty() && !t->empty()) {
1065 // This can happen if a interface-typed array narrows to a class type.
1066 jt = _type;
1067 }
1069 if (EliminateAutoBox) {
1070 // The pointers in the autobox arrays are always non-null
1071 Node* base = in(Address)->in(AddPNode::Base);
1072 if (base != NULL) {
1073 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1074 if (is_autobox_cache(atp)) {
1075 return jt->join(TypePtr::NOTNULL)->is_ptr();
1076 }
1077 }
1078 }
1079 return jt;
1080 }
1081 }
1082 } else if (tp->base() == Type::InstPtr) {
1083 assert( off != Type::OffsetBot ||
1084 // arrays can be cast to Objects
1085 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1086 // unsafe field access may not have a constant offset
1087 phase->C->has_unsafe_access(),
1088 "Field accesses must be precise" );
1089 // For oop loads, we expect the _type to be precise
1090 } else if (tp->base() == Type::KlassPtr) {
1091 assert( off != Type::OffsetBot ||
1092 // arrays can be cast to Objects
1093 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1094 // also allow array-loading from the primary supertype
1095 // array during subtype checks
1096 Opcode() == Op_LoadKlass,
1097 "Field accesses must be precise" );
1098 // For klass/static loads, we expect the _type to be precise
1099 }
1101 const TypeKlassPtr *tkls = tp->isa_klassptr();
1102 if (tkls != NULL && !StressReflectiveCode) {
1103 ciKlass* klass = tkls->klass();
1104 if (klass->is_loaded() && tkls->klass_is_exact()) {
1105 // We are loading a field from a Klass metaobject whose identity
1106 // is known at compile time (the type is "exact" or "precise").
1107 // Check for fields we know are maintained as constants by the VM.
1108 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1109 // The field is Klass::_super_check_offset. Return its (constant) value.
1110 // (Folds up type checking code.)
1111 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1112 return TypeInt::make(klass->super_check_offset());
1113 }
1114 // Compute index into primary_supers array
1115 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1116 // Check for overflowing; use unsigned compare to handle the negative case.
1117 if( depth < ciKlass::primary_super_limit() ) {
1118 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1119 // (Folds up type checking code.)
1120 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1121 ciKlass *ss = klass->super_of_depth(depth);
1122 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1123 }
1124 const Type* aift = load_array_final_field(tkls, klass);
1125 if (aift != NULL) return aift;
1126 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1127 && klass->is_array_klass()) {
1128 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1129 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1130 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1131 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1132 }
1133 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1134 // The field is Klass::_java_mirror. Return its (constant) value.
1135 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1136 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1137 return TypeInstPtr::make(klass->java_mirror());
1138 }
1139 }
1141 // We can still check if we are loading from the primary_supers array at a
1142 // shallow enough depth. Even though the klass is not exact, entries less
1143 // than or equal to its super depth are correct.
1144 if (klass->is_loaded() ) {
1145 ciType *inner = klass->klass();
1146 while( inner->is_obj_array_klass() )
1147 inner = inner->as_obj_array_klass()->base_element_type();
1148 if( inner->is_instance_klass() &&
1149 !inner->as_instance_klass()->flags().is_interface() ) {
1150 // Compute index into primary_supers array
1151 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1152 // Check for overflowing; use unsigned compare to handle the negative case.
1153 if( depth < ciKlass::primary_super_limit() &&
1154 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1155 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1156 // (Folds up type checking code.)
1157 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1158 ciKlass *ss = klass->super_of_depth(depth);
1159 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1160 }
1161 }
1162 }
1164 // If the type is enough to determine that the thing is not an array,
1165 // we can give the layout_helper a positive interval type.
1166 // This will help short-circuit some reflective code.
1167 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1168 && !klass->is_array_klass() // not directly typed as an array
1169 && !klass->is_interface() // specifically not Serializable & Cloneable
1170 && !klass->is_java_lang_Object() // not the supertype of all T[]
1171 ) {
1172 // Note: When interfaces are reliable, we can narrow the interface
1173 // test to (klass != Serializable && klass != Cloneable).
1174 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1175 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1176 // The key property of this type is that it folds up tests
1177 // for array-ness, since it proves that the layout_helper is positive.
1178 // Thus, a generic value like the basic object layout helper works fine.
1179 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1180 }
1181 }
1183 // If we are loading from a freshly-allocated object, produce a zero,
1184 // if the load is provably beyond the header of the object.
1185 // (Also allow a variable load from a fresh array to produce zero.)
1186 if (ReduceFieldZeroing) {
1187 Node* value = can_see_stored_value(mem,phase);
1188 if (value != NULL && value->is_Con())
1189 return value->bottom_type();
1190 }
1192 return _type;
1193 }
1195 //------------------------------match_edge-------------------------------------
1196 // Do we Match on this edge index or not? Match only the address.
1197 uint LoadNode::match_edge(uint idx) const {
1198 return idx == MemNode::Address;
1199 }
1201 //--------------------------LoadBNode::Ideal--------------------------------------
1202 //
1203 // If the previous store is to the same address as this load,
1204 // and the value stored was larger than a byte, replace this load
1205 // with the value stored truncated to a byte. If no truncation is
1206 // needed, the replacement is done in LoadNode::Identity().
1207 //
1208 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1209 Node* mem = in(MemNode::Memory);
1210 Node* value = can_see_stored_value(mem,phase);
1211 if( value && !phase->type(value)->higher_equal( _type ) ) {
1212 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1213 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1214 }
1215 // Identity call will handle the case where truncation is not needed.
1216 return LoadNode::Ideal(phase, can_reshape);
1217 }
1219 //--------------------------LoadCNode::Ideal--------------------------------------
1220 //
1221 // If the previous store is to the same address as this load,
1222 // and the value stored was larger than a char, replace this load
1223 // with the value stored truncated to a char. If no truncation is
1224 // needed, the replacement is done in LoadNode::Identity().
1225 //
1226 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1227 Node* mem = in(MemNode::Memory);
1228 Node* value = can_see_stored_value(mem,phase);
1229 if( value && !phase->type(value)->higher_equal( _type ) )
1230 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1231 // Identity call will handle the case where truncation is not needed.
1232 return LoadNode::Ideal(phase, can_reshape);
1233 }
1235 //--------------------------LoadSNode::Ideal--------------------------------------
1236 //
1237 // If the previous store is to the same address as this load,
1238 // and the value stored was larger than a short, replace this load
1239 // with the value stored truncated to a short. If no truncation is
1240 // needed, the replacement is done in LoadNode::Identity().
1241 //
1242 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1243 Node* mem = in(MemNode::Memory);
1244 Node* value = can_see_stored_value(mem,phase);
1245 if( value && !phase->type(value)->higher_equal( _type ) ) {
1246 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1247 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1248 }
1249 // Identity call will handle the case where truncation is not needed.
1250 return LoadNode::Ideal(phase, can_reshape);
1251 }
1253 //=============================================================================
1254 //------------------------------Value------------------------------------------
1255 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1256 // Either input is TOP ==> the result is TOP
1257 const Type *t1 = phase->type( in(MemNode::Memory) );
1258 if (t1 == Type::TOP) return Type::TOP;
1259 Node *adr = in(MemNode::Address);
1260 const Type *t2 = phase->type( adr );
1261 if (t2 == Type::TOP) return Type::TOP;
1262 const TypePtr *tp = t2->is_ptr();
1263 if (TypePtr::above_centerline(tp->ptr()) ||
1264 tp->ptr() == TypePtr::Null) return Type::TOP;
1266 // Return a more precise klass, if possible
1267 const TypeInstPtr *tinst = tp->isa_instptr();
1268 if (tinst != NULL) {
1269 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1270 int offset = tinst->offset();
1271 if (ik == phase->C->env()->Class_klass()
1272 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1273 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1274 // We are loading a special hidden field from a Class mirror object,
1275 // the field which points to the VM's Klass metaobject.
1276 ciType* t = tinst->java_mirror_type();
1277 // java_mirror_type returns non-null for compile-time Class constants.
1278 if (t != NULL) {
1279 // constant oop => constant klass
1280 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1281 return TypeKlassPtr::make(ciArrayKlass::make(t));
1282 }
1283 if (!t->is_klass()) {
1284 // a primitive Class (e.g., int.class) has NULL for a klass field
1285 return TypePtr::NULL_PTR;
1286 }
1287 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1288 return TypeKlassPtr::make(t->as_klass());
1289 }
1290 // non-constant mirror, so we can't tell what's going on
1291 }
1292 if( !ik->is_loaded() )
1293 return _type; // Bail out if not loaded
1294 if (offset == oopDesc::klass_offset_in_bytes()) {
1295 if (tinst->klass_is_exact()) {
1296 return TypeKlassPtr::make(ik);
1297 }
1298 // See if we can become precise: no subklasses and no interface
1299 // (Note: We need to support verified interfaces.)
1300 if (!ik->is_interface() && !ik->has_subklass()) {
1301 //assert(!UseExactTypes, "this code should be useless with exact types");
1302 // Add a dependence; if any subclass added we need to recompile
1303 if (!ik->is_final()) {
1304 // %%% should use stronger assert_unique_concrete_subtype instead
1305 phase->C->dependencies()->assert_leaf_type(ik);
1306 }
1307 // Return precise klass
1308 return TypeKlassPtr::make(ik);
1309 }
1311 // Return root of possible klass
1312 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1313 }
1314 }
1316 // Check for loading klass from an array
1317 const TypeAryPtr *tary = tp->isa_aryptr();
1318 if( tary != NULL ) {
1319 ciKlass *tary_klass = tary->klass();
1320 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1321 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1322 if (tary->klass_is_exact()) {
1323 return TypeKlassPtr::make(tary_klass);
1324 }
1325 ciArrayKlass *ak = tary->klass()->as_array_klass();
1326 // If the klass is an object array, we defer the question to the
1327 // array component klass.
1328 if( ak->is_obj_array_klass() ) {
1329 assert( ak->is_loaded(), "" );
1330 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1331 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1332 ciInstanceKlass* ik = base_k->as_instance_klass();
1333 // See if we can become precise: no subklasses and no interface
1334 if (!ik->is_interface() && !ik->has_subklass()) {
1335 //assert(!UseExactTypes, "this code should be useless with exact types");
1336 // Add a dependence; if any subclass added we need to recompile
1337 if (!ik->is_final()) {
1338 phase->C->dependencies()->assert_leaf_type(ik);
1339 }
1340 // Return precise array klass
1341 return TypeKlassPtr::make(ak);
1342 }
1343 }
1344 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1345 } else { // Found a type-array?
1346 //assert(!UseExactTypes, "this code should be useless with exact types");
1347 assert( ak->is_type_array_klass(), "" );
1348 return TypeKlassPtr::make(ak); // These are always precise
1349 }
1350 }
1351 }
1353 // Check for loading klass from an array klass
1354 const TypeKlassPtr *tkls = tp->isa_klassptr();
1355 if (tkls != NULL && !StressReflectiveCode) {
1356 ciKlass* klass = tkls->klass();
1357 if( !klass->is_loaded() )
1358 return _type; // Bail out if not loaded
1359 if( klass->is_obj_array_klass() &&
1360 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1361 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1362 // // Always returning precise element type is incorrect,
1363 // // e.g., element type could be object and array may contain strings
1364 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1366 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1367 // according to the element type's subclassing.
1368 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1369 }
1370 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1371 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1372 ciKlass* sup = klass->as_instance_klass()->super();
1373 // The field is Klass::_super. Return its (constant) value.
1374 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1375 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1376 }
1377 }
1379 // Bailout case
1380 return LoadNode::Value(phase);
1381 }
1383 //------------------------------Identity---------------------------------------
1384 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1385 // Also feed through the klass in Allocate(...klass...)._klass.
1386 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1387 Node* x = LoadNode::Identity(phase);
1388 if (x != this) return x;
1390 // Take apart the address into an oop and and offset.
1391 // Return 'this' if we cannot.
1392 Node* adr = in(MemNode::Address);
1393 intptr_t offset = 0;
1394 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1395 if (base == NULL) return this;
1396 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1397 if (toop == NULL) return this;
1399 // We can fetch the klass directly through an AllocateNode.
1400 // This works even if the klass is not constant (clone or newArray).
1401 if (offset == oopDesc::klass_offset_in_bytes()) {
1402 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1403 if (allocated_klass != NULL) {
1404 return allocated_klass;
1405 }
1406 }
1408 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1409 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1410 // See inline_native_Class_query for occurrences of these patterns.
1411 // Java Example: x.getClass().isAssignableFrom(y)
1412 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1413 //
1414 // This improves reflective code, often making the Class
1415 // mirror go completely dead. (Current exception: Class
1416 // mirrors may appear in debug info, but we could clean them out by
1417 // introducing a new debug info operator for klassOop.java_mirror).
1418 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1419 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1420 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1421 // We are loading a special hidden field from a Class mirror,
1422 // the field which points to its Klass or arrayKlass metaobject.
1423 if (base->is_Load()) {
1424 Node* adr2 = base->in(MemNode::Address);
1425 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1426 if (tkls != NULL && !tkls->empty()
1427 && (tkls->klass()->is_instance_klass() ||
1428 tkls->klass()->is_array_klass())
1429 && adr2->is_AddP()
1430 ) {
1431 int mirror_field = Klass::java_mirror_offset_in_bytes();
1432 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1433 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1434 }
1435 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1436 return adr2->in(AddPNode::Base);
1437 }
1438 }
1439 }
1440 }
1442 return this;
1443 }
1445 //------------------------------Value-----------------------------------------
1446 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1447 // Either input is TOP ==> the result is TOP
1448 const Type *t1 = phase->type( in(MemNode::Memory) );
1449 if( t1 == Type::TOP ) return Type::TOP;
1450 Node *adr = in(MemNode::Address);
1451 const Type *t2 = phase->type( adr );
1452 if( t2 == Type::TOP ) return Type::TOP;
1453 const TypePtr *tp = t2->is_ptr();
1454 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1455 const TypeAryPtr *tap = tp->isa_aryptr();
1456 if( !tap ) return _type;
1457 return tap->size();
1458 }
1460 //------------------------------Identity---------------------------------------
1461 // Feed through the length in AllocateArray(...length...)._length.
1462 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1463 Node* x = LoadINode::Identity(phase);
1464 if (x != this) return x;
1466 // Take apart the address into an oop and and offset.
1467 // Return 'this' if we cannot.
1468 Node* adr = in(MemNode::Address);
1469 intptr_t offset = 0;
1470 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1471 if (base == NULL) return this;
1472 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1473 if (tary == NULL) return this;
1475 // We can fetch the length directly through an AllocateArrayNode.
1476 // This works even if the length is not constant (clone or newArray).
1477 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1478 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1479 if (allocated_length != NULL) {
1480 return allocated_length;
1481 }
1482 }
1484 return this;
1486 }
1487 //=============================================================================
1488 //---------------------------StoreNode::make-----------------------------------
1489 // Polymorphic factory method:
1490 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1491 switch (bt) {
1492 case T_BOOLEAN:
1493 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1494 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1495 case T_CHAR:
1496 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1497 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1498 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1499 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1500 case T_ADDRESS:
1501 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1502 }
1503 ShouldNotReachHere();
1504 return (StoreNode*)NULL;
1505 }
1507 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1508 bool require_atomic = true;
1509 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1510 }
1513 //--------------------------bottom_type----------------------------------------
1514 const Type *StoreNode::bottom_type() const {
1515 return Type::MEMORY;
1516 }
1518 //------------------------------hash-------------------------------------------
1519 uint StoreNode::hash() const {
1520 // unroll addition of interesting fields
1521 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1523 // Since they are not commoned, do not hash them:
1524 return NO_HASH;
1525 }
1527 //------------------------------Ideal------------------------------------------
1528 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1529 // When a store immediately follows a relevant allocation/initialization,
1530 // try to capture it into the initialization, or hoist it above.
1531 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1532 Node* p = MemNode::Ideal_common(phase, can_reshape);
1533 if (p) return (p == NodeSentinel) ? NULL : p;
1535 Node* mem = in(MemNode::Memory);
1536 Node* address = in(MemNode::Address);
1538 // Back-to-back stores to same address? Fold em up.
1539 // Generally unsafe if I have intervening uses...
1540 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1541 // Looking at a dead closed cycle of memory?
1542 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1544 assert(Opcode() == mem->Opcode() ||
1545 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1546 "no mismatched stores, except on raw memory");
1548 if (mem->outcnt() == 1 && // check for intervening uses
1549 mem->as_Store()->memory_size() <= this->memory_size()) {
1550 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1551 // For example, 'mem' might be the final state at a conditional return.
1552 // Or, 'mem' might be used by some node which is live at the same time
1553 // 'this' is live, which might be unschedulable. So, require exactly
1554 // ONE user, the 'this' store, until such time as we clone 'mem' for
1555 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1556 if (can_reshape) { // (%%% is this an anachronism?)
1557 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1558 phase->is_IterGVN());
1559 } else {
1560 // It's OK to do this in the parser, since DU info is always accurate,
1561 // and the parser always refers to nodes via SafePointNode maps.
1562 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1563 }
1564 return this;
1565 }
1566 }
1568 // Capture an unaliased, unconditional, simple store into an initializer.
1569 // Or, if it is independent of the allocation, hoist it above the allocation.
1570 if (ReduceFieldZeroing && /*can_reshape &&*/
1571 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1572 InitializeNode* init = mem->in(0)->as_Initialize();
1573 intptr_t offset = init->can_capture_store(this, phase);
1574 if (offset > 0) {
1575 Node* moved = init->capture_store(this, offset, phase);
1576 // If the InitializeNode captured me, it made a raw copy of me,
1577 // and I need to disappear.
1578 if (moved != NULL) {
1579 // %%% hack to ensure that Ideal returns a new node:
1580 mem = MergeMemNode::make(phase->C, mem);
1581 return mem; // fold me away
1582 }
1583 }
1584 }
1586 return NULL; // No further progress
1587 }
1589 //------------------------------Value-----------------------------------------
1590 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1591 // Either input is TOP ==> the result is TOP
1592 const Type *t1 = phase->type( in(MemNode::Memory) );
1593 if( t1 == Type::TOP ) return Type::TOP;
1594 const Type *t2 = phase->type( in(MemNode::Address) );
1595 if( t2 == Type::TOP ) return Type::TOP;
1596 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1597 if( t3 == Type::TOP ) return Type::TOP;
1598 return Type::MEMORY;
1599 }
1601 //------------------------------Identity---------------------------------------
1602 // Remove redundant stores:
1603 // Store(m, p, Load(m, p)) changes to m.
1604 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1605 Node *StoreNode::Identity( PhaseTransform *phase ) {
1606 Node* mem = in(MemNode::Memory);
1607 Node* adr = in(MemNode::Address);
1608 Node* val = in(MemNode::ValueIn);
1610 // Load then Store? Then the Store is useless
1611 if (val->is_Load() &&
1612 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1613 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1614 val->as_Load()->store_Opcode() == Opcode()) {
1615 return mem;
1616 }
1618 // Two stores in a row of the same value?
1619 if (mem->is_Store() &&
1620 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1621 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1622 mem->Opcode() == Opcode()) {
1623 return mem;
1624 }
1626 // Store of zero anywhere into a freshly-allocated object?
1627 // Then the store is useless.
1628 // (It must already have been captured by the InitializeNode.)
1629 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1630 // a newly allocated object is already all-zeroes everywhere
1631 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1632 return mem;
1633 }
1635 // the store may also apply to zero-bits in an earlier object
1636 Node* prev_mem = find_previous_store(phase);
1637 // Steps (a), (b): Walk past independent stores to find an exact match.
1638 if (prev_mem != NULL) {
1639 Node* prev_val = can_see_stored_value(prev_mem, phase);
1640 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1641 // prev_val and val might differ by a cast; it would be good
1642 // to keep the more informative of the two.
1643 return mem;
1644 }
1645 }
1646 }
1648 return this;
1649 }
1651 //------------------------------match_edge-------------------------------------
1652 // Do we Match on this edge index or not? Match only memory & value
1653 uint StoreNode::match_edge(uint idx) const {
1654 return idx == MemNode::Address || idx == MemNode::ValueIn;
1655 }
1657 //------------------------------cmp--------------------------------------------
1658 // Do not common stores up together. They generally have to be split
1659 // back up anyways, so do not bother.
1660 uint StoreNode::cmp( const Node &n ) const {
1661 return (&n == this); // Always fail except on self
1662 }
1664 //------------------------------Ideal_masked_input-----------------------------
1665 // Check for a useless mask before a partial-word store
1666 // (StoreB ... (AndI valIn conIa) )
1667 // If (conIa & mask == mask) this simplifies to
1668 // (StoreB ... (valIn) )
1669 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1670 Node *val = in(MemNode::ValueIn);
1671 if( val->Opcode() == Op_AndI ) {
1672 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1673 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1674 set_req(MemNode::ValueIn, val->in(1));
1675 return this;
1676 }
1677 }
1678 return NULL;
1679 }
1682 //------------------------------Ideal_sign_extended_input----------------------
1683 // Check for useless sign-extension before a partial-word store
1684 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1685 // If (conIL == conIR && conIR <= num_bits) this simplifies to
1686 // (StoreB ... (valIn) )
1687 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1688 Node *val = in(MemNode::ValueIn);
1689 if( val->Opcode() == Op_RShiftI ) {
1690 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1691 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
1692 Node *shl = val->in(1);
1693 if( shl->Opcode() == Op_LShiftI ) {
1694 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
1695 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
1696 set_req(MemNode::ValueIn, shl->in(1));
1697 return this;
1698 }
1699 }
1700 }
1701 }
1702 return NULL;
1703 }
1705 //------------------------------value_never_loaded-----------------------------------
1706 // Determine whether there are any possible loads of the value stored.
1707 // For simplicity, we actually check if there are any loads from the
1708 // address stored to, not just for loads of the value stored by this node.
1709 //
1710 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
1711 Node *adr = in(Address);
1712 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
1713 if (adr_oop == NULL)
1714 return false;
1715 if (!adr_oop->is_instance())
1716 return false; // if not a distinct instance, there may be aliases of the address
1717 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
1718 Node *use = adr->fast_out(i);
1719 int opc = use->Opcode();
1720 if (use->is_Load() || use->is_LoadStore()) {
1721 return false;
1722 }
1723 }
1724 return true;
1725 }
1727 //=============================================================================
1728 //------------------------------Ideal------------------------------------------
1729 // If the store is from an AND mask that leaves the low bits untouched, then
1730 // we can skip the AND operation. If the store is from a sign-extension
1731 // (a left shift, then right shift) we can skip both.
1732 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
1733 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
1734 if( progress != NULL ) return progress;
1736 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
1737 if( progress != NULL ) return progress;
1739 // Finally check the default case
1740 return StoreNode::Ideal(phase, can_reshape);
1741 }
1743 //=============================================================================
1744 //------------------------------Ideal------------------------------------------
1745 // If the store is from an AND mask that leaves the low bits untouched, then
1746 // we can skip the AND operation
1747 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
1748 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
1749 if( progress != NULL ) return progress;
1751 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
1752 if( progress != NULL ) return progress;
1754 // Finally check the default case
1755 return StoreNode::Ideal(phase, can_reshape);
1756 }
1758 //=============================================================================
1759 //------------------------------Identity---------------------------------------
1760 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
1761 // No need to card mark when storing a null ptr
1762 Node* my_store = in(MemNode::OopStore);
1763 if (my_store->is_Store()) {
1764 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
1765 if( t1 == TypePtr::NULL_PTR ) {
1766 return in(MemNode::Memory);
1767 }
1768 }
1769 return this;
1770 }
1772 //------------------------------Value-----------------------------------------
1773 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
1774 // If extra input is TOP ==> the result is TOP
1775 const Type *t1 = phase->type( in(MemNode::OopStore) );
1776 if( t1 == Type::TOP ) return Type::TOP;
1778 return StoreNode::Value( phase );
1779 }
1782 //=============================================================================
1783 //----------------------------------SCMemProjNode------------------------------
1784 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
1785 {
1786 return bottom_type();
1787 }
1789 //=============================================================================
1790 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
1791 init_req(MemNode::Control, c );
1792 init_req(MemNode::Memory , mem);
1793 init_req(MemNode::Address, adr);
1794 init_req(MemNode::ValueIn, val);
1795 init_req( ExpectedIn, ex );
1796 init_class_id(Class_LoadStore);
1798 }
1800 //=============================================================================
1801 //-------------------------------adr_type--------------------------------------
1802 // Do we Match on this edge index or not? Do not match memory
1803 const TypePtr* ClearArrayNode::adr_type() const {
1804 Node *adr = in(3);
1805 return MemNode::calculate_adr_type(adr->bottom_type());
1806 }
1808 //------------------------------match_edge-------------------------------------
1809 // Do we Match on this edge index or not? Do not match memory
1810 uint ClearArrayNode::match_edge(uint idx) const {
1811 return idx > 1;
1812 }
1814 //------------------------------Identity---------------------------------------
1815 // Clearing a zero length array does nothing
1816 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
1817 return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this;
1818 }
1820 //------------------------------Idealize---------------------------------------
1821 // Clearing a short array is faster with stores
1822 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
1823 const int unit = BytesPerLong;
1824 const TypeX* t = phase->type(in(2))->isa_intptr_t();
1825 if (!t) return NULL;
1826 if (!t->is_con()) return NULL;
1827 intptr_t raw_count = t->get_con();
1828 intptr_t size = raw_count;
1829 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
1830 // Clearing nothing uses the Identity call.
1831 // Negative clears are possible on dead ClearArrays
1832 // (see jck test stmt114.stmt11402.val).
1833 if (size <= 0 || size % unit != 0) return NULL;
1834 intptr_t count = size / unit;
1835 // Length too long; use fast hardware clear
1836 if (size > Matcher::init_array_short_size) return NULL;
1837 Node *mem = in(1);
1838 if( phase->type(mem)==Type::TOP ) return NULL;
1839 Node *adr = in(3);
1840 const Type* at = phase->type(adr);
1841 if( at==Type::TOP ) return NULL;
1842 const TypePtr* atp = at->isa_ptr();
1843 // adjust atp to be the correct array element address type
1844 if (atp == NULL) atp = TypePtr::BOTTOM;
1845 else atp = atp->add_offset(Type::OffsetBot);
1846 // Get base for derived pointer purposes
1847 if( adr->Opcode() != Op_AddP ) Unimplemented();
1848 Node *base = adr->in(1);
1850 Node *zero = phase->makecon(TypeLong::ZERO);
1851 Node *off = phase->MakeConX(BytesPerLong);
1852 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1853 count--;
1854 while( count-- ) {
1855 mem = phase->transform(mem);
1856 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
1857 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1858 }
1859 return mem;
1860 }
1862 //----------------------------clear_memory-------------------------------------
1863 // Generate code to initialize object storage to zero.
1864 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1865 intptr_t start_offset,
1866 Node* end_offset,
1867 PhaseGVN* phase) {
1868 Compile* C = phase->C;
1869 intptr_t offset = start_offset;
1871 int unit = BytesPerLong;
1872 if ((offset % unit) != 0) {
1873 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
1874 adr = phase->transform(adr);
1875 const TypePtr* atp = TypeRawPtr::BOTTOM;
1876 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1877 mem = phase->transform(mem);
1878 offset += BytesPerInt;
1879 }
1880 assert((offset % unit) == 0, "");
1882 // Initialize the remaining stuff, if any, with a ClearArray.
1883 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
1884 }
1886 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1887 Node* start_offset,
1888 Node* end_offset,
1889 PhaseGVN* phase) {
1890 Compile* C = phase->C;
1891 int unit = BytesPerLong;
1892 Node* zbase = start_offset;
1893 Node* zend = end_offset;
1895 // Scale to the unit required by the CPU:
1896 if (!Matcher::init_array_count_is_in_bytes) {
1897 Node* shift = phase->intcon(exact_log2(unit));
1898 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
1899 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
1900 }
1902 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
1903 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
1905 // Bulk clear double-words
1906 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
1907 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
1908 return phase->transform(mem);
1909 }
1911 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1912 intptr_t start_offset,
1913 intptr_t end_offset,
1914 PhaseGVN* phase) {
1915 Compile* C = phase->C;
1916 assert((end_offset % BytesPerInt) == 0, "odd end offset");
1917 intptr_t done_offset = end_offset;
1918 if ((done_offset % BytesPerLong) != 0) {
1919 done_offset -= BytesPerInt;
1920 }
1921 if (done_offset > start_offset) {
1922 mem = clear_memory(ctl, mem, dest,
1923 start_offset, phase->MakeConX(done_offset), phase);
1924 }
1925 if (done_offset < end_offset) { // emit the final 32-bit store
1926 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
1927 adr = phase->transform(adr);
1928 const TypePtr* atp = TypeRawPtr::BOTTOM;
1929 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1930 mem = phase->transform(mem);
1931 done_offset += BytesPerInt;
1932 }
1933 assert(done_offset == end_offset, "");
1934 return mem;
1935 }
1937 //=============================================================================
1938 // Do we match on this edge? No memory edges
1939 uint StrCompNode::match_edge(uint idx) const {
1940 return idx == 5 || idx == 6;
1941 }
1943 //------------------------------Ideal------------------------------------------
1944 // Return a node which is more "ideal" than the current node. Strip out
1945 // control copies
1946 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
1947 return remove_dead_region(phase, can_reshape) ? this : NULL;
1948 }
1951 //=============================================================================
1952 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
1953 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
1954 _adr_type(C->get_adr_type(alias_idx))
1955 {
1956 init_class_id(Class_MemBar);
1957 Node* top = C->top();
1958 init_req(TypeFunc::I_O,top);
1959 init_req(TypeFunc::FramePtr,top);
1960 init_req(TypeFunc::ReturnAdr,top);
1961 if (precedent != NULL)
1962 init_req(TypeFunc::Parms, precedent);
1963 }
1965 //------------------------------cmp--------------------------------------------
1966 uint MemBarNode::hash() const { return NO_HASH; }
1967 uint MemBarNode::cmp( const Node &n ) const {
1968 return (&n == this); // Always fail except on self
1969 }
1971 //------------------------------make-------------------------------------------
1972 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
1973 int len = Precedent + (pn == NULL? 0: 1);
1974 switch (opcode) {
1975 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
1976 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
1977 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
1978 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
1979 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
1980 default: ShouldNotReachHere(); return NULL;
1981 }
1982 }
1984 //------------------------------Ideal------------------------------------------
1985 // Return a node which is more "ideal" than the current node. Strip out
1986 // control copies
1987 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1988 if (remove_dead_region(phase, can_reshape)) return this;
1989 return NULL;
1990 }
1992 //------------------------------Value------------------------------------------
1993 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
1994 if( !in(0) ) return Type::TOP;
1995 if( phase->type(in(0)) == Type::TOP )
1996 return Type::TOP;
1997 return TypeTuple::MEMBAR;
1998 }
2000 //------------------------------match------------------------------------------
2001 // Construct projections for memory.
2002 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2003 switch (proj->_con) {
2004 case TypeFunc::Control:
2005 case TypeFunc::Memory:
2006 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2007 }
2008 ShouldNotReachHere();
2009 return NULL;
2010 }
2012 //===========================InitializeNode====================================
2013 // SUMMARY:
2014 // This node acts as a memory barrier on raw memory, after some raw stores.
2015 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2016 // The Initialize can 'capture' suitably constrained stores as raw inits.
2017 // It can coalesce related raw stores into larger units (called 'tiles').
2018 // It can avoid zeroing new storage for memory units which have raw inits.
2019 // At macro-expansion, it is marked 'complete', and does not optimize further.
2020 //
2021 // EXAMPLE:
2022 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2023 // ctl = incoming control; mem* = incoming memory
2024 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2025 // First allocate uninitialized memory and fill in the header:
2026 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2027 // ctl := alloc.Control; mem* := alloc.Memory*
2028 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2029 // Then initialize to zero the non-header parts of the raw memory block:
2030 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2031 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2032 // After the initialize node executes, the object is ready for service:
2033 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2034 // Suppose its body is immediately initialized as {1,2}:
2035 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2036 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2037 // mem.SLICE(#short[*]) := store2
2038 //
2039 // DETAILS:
2040 // An InitializeNode collects and isolates object initialization after
2041 // an AllocateNode and before the next possible safepoint. As a
2042 // memory barrier (MemBarNode), it keeps critical stores from drifting
2043 // down past any safepoint or any publication of the allocation.
2044 // Before this barrier, a newly-allocated object may have uninitialized bits.
2045 // After this barrier, it may be treated as a real oop, and GC is allowed.
2046 //
2047 // The semantics of the InitializeNode include an implicit zeroing of
2048 // the new object from object header to the end of the object.
2049 // (The object header and end are determined by the AllocateNode.)
2050 //
2051 // Certain stores may be added as direct inputs to the InitializeNode.
2052 // These stores must update raw memory, and they must be to addresses
2053 // derived from the raw address produced by AllocateNode, and with
2054 // a constant offset. They must be ordered by increasing offset.
2055 // The first one is at in(RawStores), the last at in(req()-1).
2056 // Unlike most memory operations, they are not linked in a chain,
2057 // but are displayed in parallel as users of the rawmem output of
2058 // the allocation.
2059 //
2060 // (See comments in InitializeNode::capture_store, which continue
2061 // the example given above.)
2062 //
2063 // When the associated Allocate is macro-expanded, the InitializeNode
2064 // may be rewritten to optimize collected stores. A ClearArrayNode
2065 // may also be created at that point to represent any required zeroing.
2066 // The InitializeNode is then marked 'complete', prohibiting further
2067 // capturing of nearby memory operations.
2068 //
2069 // During macro-expansion, all captured initializations which store
2070 // constant values of 32 bits or smaller are coalesced (if advantagous)
2071 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2072 // initialized in fewer memory operations. Memory words which are
2073 // covered by neither tiles nor non-constant stores are pre-zeroed
2074 // by explicit stores of zero. (The code shape happens to do all
2075 // zeroing first, then all other stores, with both sequences occurring
2076 // in order of ascending offsets.)
2077 //
2078 // Alternatively, code may be inserted between an AllocateNode and its
2079 // InitializeNode, to perform arbitrary initialization of the new object.
2080 // E.g., the object copying intrinsics insert complex data transfers here.
2081 // The initialization must then be marked as 'complete' disable the
2082 // built-in zeroing semantics and the collection of initializing stores.
2083 //
2084 // While an InitializeNode is incomplete, reads from the memory state
2085 // produced by it are optimizable if they match the control edge and
2086 // new oop address associated with the allocation/initialization.
2087 // They return a stored value (if the offset matches) or else zero.
2088 // A write to the memory state, if it matches control and address,
2089 // and if it is to a constant offset, may be 'captured' by the
2090 // InitializeNode. It is cloned as a raw memory operation and rewired
2091 // inside the initialization, to the raw oop produced by the allocation.
2092 // Operations on addresses which are provably distinct (e.g., to
2093 // other AllocateNodes) are allowed to bypass the initialization.
2094 //
2095 // The effect of all this is to consolidate object initialization
2096 // (both arrays and non-arrays, both piecewise and bulk) into a
2097 // single location, where it can be optimized as a unit.
2098 //
2099 // Only stores with an offset less than TrackedInitializationLimit words
2100 // will be considered for capture by an InitializeNode. This puts a
2101 // reasonable limit on the complexity of optimized initializations.
2103 //---------------------------InitializeNode------------------------------------
2104 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2105 : _is_complete(false),
2106 MemBarNode(C, adr_type, rawoop)
2107 {
2108 init_class_id(Class_Initialize);
2110 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2111 assert(in(RawAddress) == rawoop, "proper init");
2112 // Note: allocation() can be NULL, for secondary initialization barriers
2113 }
2115 // Since this node is not matched, it will be processed by the
2116 // register allocator. Declare that there are no constraints
2117 // on the allocation of the RawAddress edge.
2118 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2119 // This edge should be set to top, by the set_complete. But be conservative.
2120 if (idx == InitializeNode::RawAddress)
2121 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2122 return RegMask::Empty;
2123 }
2125 Node* InitializeNode::memory(uint alias_idx) {
2126 Node* mem = in(Memory);
2127 if (mem->is_MergeMem()) {
2128 return mem->as_MergeMem()->memory_at(alias_idx);
2129 } else {
2130 // incoming raw memory is not split
2131 return mem;
2132 }
2133 }
2135 bool InitializeNode::is_non_zero() {
2136 if (is_complete()) return false;
2137 remove_extra_zeroes();
2138 return (req() > RawStores);
2139 }
2141 void InitializeNode::set_complete(PhaseGVN* phase) {
2142 assert(!is_complete(), "caller responsibility");
2143 _is_complete = true;
2145 // After this node is complete, it contains a bunch of
2146 // raw-memory initializations. There is no need for
2147 // it to have anything to do with non-raw memory effects.
2148 // Therefore, tell all non-raw users to re-optimize themselves,
2149 // after skipping the memory effects of this initialization.
2150 PhaseIterGVN* igvn = phase->is_IterGVN();
2151 if (igvn) igvn->add_users_to_worklist(this);
2152 }
2154 // convenience function
2155 // return false if the init contains any stores already
2156 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2157 InitializeNode* init = initialization();
2158 if (init == NULL || init->is_complete()) return false;
2159 init->remove_extra_zeroes();
2160 // for now, if this allocation has already collected any inits, bail:
2161 if (init->is_non_zero()) return false;
2162 init->set_complete(phase);
2163 return true;
2164 }
2166 void InitializeNode::remove_extra_zeroes() {
2167 if (req() == RawStores) return;
2168 Node* zmem = zero_memory();
2169 uint fill = RawStores;
2170 for (uint i = fill; i < req(); i++) {
2171 Node* n = in(i);
2172 if (n->is_top() || n == zmem) continue; // skip
2173 if (fill < i) set_req(fill, n); // compact
2174 ++fill;
2175 }
2176 // delete any empty spaces created:
2177 while (fill < req()) {
2178 del_req(fill);
2179 }
2180 }
2182 // Helper for remembering which stores go with which offsets.
2183 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2184 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2185 intptr_t offset = -1;
2186 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2187 phase, offset);
2188 if (base == NULL) return -1; // something is dead,
2189 if (offset < 0) return -1; // dead, dead
2190 return offset;
2191 }
2193 // Helper for proving that an initialization expression is
2194 // "simple enough" to be folded into an object initialization.
2195 // Attempts to prove that a store's initial value 'n' can be captured
2196 // within the initialization without creating a vicious cycle, such as:
2197 // { Foo p = new Foo(); p.next = p; }
2198 // True for constants and parameters and small combinations thereof.
2199 bool InitializeNode::detect_init_independence(Node* n,
2200 bool st_is_pinned,
2201 int& count) {
2202 if (n == NULL) return true; // (can this really happen?)
2203 if (n->is_Proj()) n = n->in(0);
2204 if (n == this) return false; // found a cycle
2205 if (n->is_Con()) return true;
2206 if (n->is_Start()) return true; // params, etc., are OK
2207 if (n->is_Root()) return true; // even better
2209 Node* ctl = n->in(0);
2210 if (ctl != NULL && !ctl->is_top()) {
2211 if (ctl->is_Proj()) ctl = ctl->in(0);
2212 if (ctl == this) return false;
2214 // If we already know that the enclosing memory op is pinned right after
2215 // the init, then any control flow that the store has picked up
2216 // must have preceded the init, or else be equal to the init.
2217 // Even after loop optimizations (which might change control edges)
2218 // a store is never pinned *before* the availability of its inputs.
2219 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
2220 return false; // failed to prove a good control
2222 }
2224 // Check data edges for possible dependencies on 'this'.
2225 if ((count += 1) > 20) return false; // complexity limit
2226 for (uint i = 1; i < n->req(); i++) {
2227 Node* m = n->in(i);
2228 if (m == NULL || m == n || m->is_top()) continue;
2229 uint first_i = n->find_edge(m);
2230 if (i != first_i) continue; // process duplicate edge just once
2231 if (!detect_init_independence(m, st_is_pinned, count)) {
2232 return false;
2233 }
2234 }
2236 return true;
2237 }
2239 // Here are all the checks a Store must pass before it can be moved into
2240 // an initialization. Returns zero if a check fails.
2241 // On success, returns the (constant) offset to which the store applies,
2242 // within the initialized memory.
2243 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2244 const int FAIL = 0;
2245 if (st->req() != MemNode::ValueIn + 1)
2246 return FAIL; // an inscrutable StoreNode (card mark?)
2247 Node* ctl = st->in(MemNode::Control);
2248 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2249 return FAIL; // must be unconditional after the initialization
2250 Node* mem = st->in(MemNode::Memory);
2251 if (!(mem->is_Proj() && mem->in(0) == this))
2252 return FAIL; // must not be preceded by other stores
2253 Node* adr = st->in(MemNode::Address);
2254 intptr_t offset;
2255 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2256 if (alloc == NULL)
2257 return FAIL; // inscrutable address
2258 if (alloc != allocation())
2259 return FAIL; // wrong allocation! (store needs to float up)
2260 Node* val = st->in(MemNode::ValueIn);
2261 int complexity_count = 0;
2262 if (!detect_init_independence(val, true, complexity_count))
2263 return FAIL; // stored value must be 'simple enough'
2265 return offset; // success
2266 }
2268 // Find the captured store in(i) which corresponds to the range
2269 // [start..start+size) in the initialized object.
2270 // If there is one, return its index i. If there isn't, return the
2271 // negative of the index where it should be inserted.
2272 // Return 0 if the queried range overlaps an initialization boundary
2273 // or if dead code is encountered.
2274 // If size_in_bytes is zero, do not bother with overlap checks.
2275 int InitializeNode::captured_store_insertion_point(intptr_t start,
2276 int size_in_bytes,
2277 PhaseTransform* phase) {
2278 const int FAIL = 0, MAX_STORE = BytesPerLong;
2280 if (is_complete())
2281 return FAIL; // arraycopy got here first; punt
2283 assert(allocation() != NULL, "must be present");
2285 // no negatives, no header fields:
2286 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
2287 if (start < (intptr_t) sizeof(arrayOopDesc) &&
2288 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2290 // after a certain size, we bail out on tracking all the stores:
2291 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2292 if (start >= ti_limit) return FAIL;
2294 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2295 if (i >= limit) return -(int)i; // not found; here is where to put it
2297 Node* st = in(i);
2298 intptr_t st_off = get_store_offset(st, phase);
2299 if (st_off < 0) {
2300 if (st != zero_memory()) {
2301 return FAIL; // bail out if there is dead garbage
2302 }
2303 } else if (st_off > start) {
2304 // ...we are done, since stores are ordered
2305 if (st_off < start + size_in_bytes) {
2306 return FAIL; // the next store overlaps
2307 }
2308 return -(int)i; // not found; here is where to put it
2309 } else if (st_off < start) {
2310 if (size_in_bytes != 0 &&
2311 start < st_off + MAX_STORE &&
2312 start < st_off + st->as_Store()->memory_size()) {
2313 return FAIL; // the previous store overlaps
2314 }
2315 } else {
2316 if (size_in_bytes != 0 &&
2317 st->as_Store()->memory_size() != size_in_bytes) {
2318 return FAIL; // mismatched store size
2319 }
2320 return i;
2321 }
2323 ++i;
2324 }
2325 }
2327 // Look for a captured store which initializes at the offset 'start'
2328 // with the given size. If there is no such store, and no other
2329 // initialization interferes, then return zero_memory (the memory
2330 // projection of the AllocateNode).
2331 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2332 PhaseTransform* phase) {
2333 assert(stores_are_sane(phase), "");
2334 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2335 if (i == 0) {
2336 return NULL; // something is dead
2337 } else if (i < 0) {
2338 return zero_memory(); // just primordial zero bits here
2339 } else {
2340 Node* st = in(i); // here is the store at this position
2341 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2342 return st;
2343 }
2344 }
2346 // Create, as a raw pointer, an address within my new object at 'offset'.
2347 Node* InitializeNode::make_raw_address(intptr_t offset,
2348 PhaseTransform* phase) {
2349 Node* addr = in(RawAddress);
2350 if (offset != 0) {
2351 Compile* C = phase->C;
2352 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2353 phase->MakeConX(offset)) );
2354 }
2355 return addr;
2356 }
2358 // Clone the given store, converting it into a raw store
2359 // initializing a field or element of my new object.
2360 // Caller is responsible for retiring the original store,
2361 // with subsume_node or the like.
2362 //
2363 // From the example above InitializeNode::InitializeNode,
2364 // here are the old stores to be captured:
2365 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2366 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2367 //
2368 // Here is the changed code; note the extra edges on init:
2369 // alloc = (Allocate ...)
2370 // rawoop = alloc.RawAddress
2371 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2372 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2373 // init = (Initialize alloc.Control alloc.Memory rawoop
2374 // rawstore1 rawstore2)
2375 //
2376 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2377 PhaseTransform* phase) {
2378 assert(stores_are_sane(phase), "");
2380 if (start < 0) return NULL;
2381 assert(can_capture_store(st, phase) == start, "sanity");
2383 Compile* C = phase->C;
2384 int size_in_bytes = st->memory_size();
2385 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2386 if (i == 0) return NULL; // bail out
2387 Node* prev_mem = NULL; // raw memory for the captured store
2388 if (i > 0) {
2389 prev_mem = in(i); // there is a pre-existing store under this one
2390 set_req(i, C->top()); // temporarily disconnect it
2391 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2392 } else {
2393 i = -i; // no pre-existing store
2394 prev_mem = zero_memory(); // a slice of the newly allocated object
2395 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2396 set_req(--i, C->top()); // reuse this edge; it has been folded away
2397 else
2398 ins_req(i, C->top()); // build a new edge
2399 }
2400 Node* new_st = st->clone();
2401 new_st->set_req(MemNode::Control, in(Control));
2402 new_st->set_req(MemNode::Memory, prev_mem);
2403 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2404 new_st = phase->transform(new_st);
2406 // At this point, new_st might have swallowed a pre-existing store
2407 // at the same offset, or perhaps new_st might have disappeared,
2408 // if it redundantly stored the same value (or zero to fresh memory).
2410 // In any case, wire it in:
2411 set_req(i, new_st);
2413 // The caller may now kill the old guy.
2414 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2415 assert(check_st == new_st || check_st == NULL, "must be findable");
2416 assert(!is_complete(), "");
2417 return new_st;
2418 }
2420 static bool store_constant(jlong* tiles, int num_tiles,
2421 intptr_t st_off, int st_size,
2422 jlong con) {
2423 if ((st_off & (st_size-1)) != 0)
2424 return false; // strange store offset (assume size==2**N)
2425 address addr = (address)tiles + st_off;
2426 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2427 switch (st_size) {
2428 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2429 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2430 case sizeof(jint): *(jint*) addr = (jint) con; break;
2431 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2432 default: return false; // strange store size (detect size!=2**N here)
2433 }
2434 return true; // return success to caller
2435 }
2437 // Coalesce subword constants into int constants and possibly
2438 // into long constants. The goal, if the CPU permits,
2439 // is to initialize the object with a small number of 64-bit tiles.
2440 // Also, convert floating-point constants to bit patterns.
2441 // Non-constants are not relevant to this pass.
2442 //
2443 // In terms of the running example on InitializeNode::InitializeNode
2444 // and InitializeNode::capture_store, here is the transformation
2445 // of rawstore1 and rawstore2 into rawstore12:
2446 // alloc = (Allocate ...)
2447 // rawoop = alloc.RawAddress
2448 // tile12 = 0x00010002
2449 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2450 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2451 //
2452 void
2453 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2454 Node* size_in_bytes,
2455 PhaseGVN* phase) {
2456 Compile* C = phase->C;
2458 assert(stores_are_sane(phase), "");
2459 // Note: After this pass, they are not completely sane,
2460 // since there may be some overlaps.
2462 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2464 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2465 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2466 size_limit = MIN2(size_limit, ti_limit);
2467 size_limit = align_size_up(size_limit, BytesPerLong);
2468 int num_tiles = size_limit / BytesPerLong;
2470 // allocate space for the tile map:
2471 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2472 jlong tiles_buf[small_len];
2473 Node* nodes_buf[small_len];
2474 jlong inits_buf[small_len];
2475 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2476 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2477 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2478 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2479 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2480 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2481 // tiles: exact bitwise model of all primitive constants
2482 // nodes: last constant-storing node subsumed into the tiles model
2483 // inits: which bytes (in each tile) are touched by any initializations
2485 //// Pass A: Fill in the tile model with any relevant stores.
2487 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2488 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2489 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2490 Node* zmem = zero_memory(); // initially zero memory state
2491 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2492 Node* st = in(i);
2493 intptr_t st_off = get_store_offset(st, phase);
2495 // Figure out the store's offset and constant value:
2496 if (st_off < header_size) continue; //skip (ignore header)
2497 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2498 int st_size = st->as_Store()->memory_size();
2499 if (st_off + st_size > size_limit) break;
2501 // Record which bytes are touched, whether by constant or not.
2502 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2503 continue; // skip (strange store size)
2505 const Type* val = phase->type(st->in(MemNode::ValueIn));
2506 if (!val->singleton()) continue; //skip (non-con store)
2507 BasicType type = val->basic_type();
2509 jlong con = 0;
2510 switch (type) {
2511 case T_INT: con = val->is_int()->get_con(); break;
2512 case T_LONG: con = val->is_long()->get_con(); break;
2513 case T_FLOAT: con = jint_cast(val->getf()); break;
2514 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2515 default: continue; //skip (odd store type)
2516 }
2518 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2519 st->Opcode() == Op_StoreL) {
2520 continue; // This StoreL is already optimal.
2521 }
2523 // Store down the constant.
2524 store_constant(tiles, num_tiles, st_off, st_size, con);
2526 intptr_t j = st_off >> LogBytesPerLong;
2528 if (type == T_INT && st_size == BytesPerInt
2529 && (st_off & BytesPerInt) == BytesPerInt) {
2530 jlong lcon = tiles[j];
2531 if (!Matcher::isSimpleConstant64(lcon) &&
2532 st->Opcode() == Op_StoreI) {
2533 // This StoreI is already optimal by itself.
2534 jint* intcon = (jint*) &tiles[j];
2535 intcon[1] = 0; // undo the store_constant()
2537 // If the previous store is also optimal by itself, back up and
2538 // undo the action of the previous loop iteration... if we can.
2539 // But if we can't, just let the previous half take care of itself.
2540 st = nodes[j];
2541 st_off -= BytesPerInt;
2542 con = intcon[0];
2543 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2544 assert(st_off >= header_size, "still ignoring header");
2545 assert(get_store_offset(st, phase) == st_off, "must be");
2546 assert(in(i-1) == zmem, "must be");
2547 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2548 assert(con == tcon->is_int()->get_con(), "must be");
2549 // Undo the effects of the previous loop trip, which swallowed st:
2550 intcon[0] = 0; // undo store_constant()
2551 set_req(i-1, st); // undo set_req(i, zmem)
2552 nodes[j] = NULL; // undo nodes[j] = st
2553 --old_subword; // undo ++old_subword
2554 }
2555 continue; // This StoreI is already optimal.
2556 }
2557 }
2559 // This store is not needed.
2560 set_req(i, zmem);
2561 nodes[j] = st; // record for the moment
2562 if (st_size < BytesPerLong) // something has changed
2563 ++old_subword; // includes int/float, but who's counting...
2564 else ++old_long;
2565 }
2567 if ((old_subword + old_long) == 0)
2568 return; // nothing more to do
2570 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2571 // Be sure to insert them before overlapping non-constant stores.
2572 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2573 for (int j = 0; j < num_tiles; j++) {
2574 jlong con = tiles[j];
2575 jlong init = inits[j];
2576 if (con == 0) continue;
2577 jint con0, con1; // split the constant, address-wise
2578 jint init0, init1; // split the init map, address-wise
2579 { union { jlong con; jint intcon[2]; } u;
2580 u.con = con;
2581 con0 = u.intcon[0];
2582 con1 = u.intcon[1];
2583 u.con = init;
2584 init0 = u.intcon[0];
2585 init1 = u.intcon[1];
2586 }
2588 Node* old = nodes[j];
2589 assert(old != NULL, "need the prior store");
2590 intptr_t offset = (j * BytesPerLong);
2592 bool split = !Matcher::isSimpleConstant64(con);
2594 if (offset < header_size) {
2595 assert(offset + BytesPerInt >= header_size, "second int counts");
2596 assert(*(jint*)&tiles[j] == 0, "junk in header");
2597 split = true; // only the second word counts
2598 // Example: int a[] = { 42 ... }
2599 } else if (con0 == 0 && init0 == -1) {
2600 split = true; // first word is covered by full inits
2601 // Example: int a[] = { ... foo(), 42 ... }
2602 } else if (con1 == 0 && init1 == -1) {
2603 split = true; // second word is covered by full inits
2604 // Example: int a[] = { ... 42, foo() ... }
2605 }
2607 // Here's a case where init0 is neither 0 nor -1:
2608 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2609 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2610 // In this case the tile is not split; it is (jlong)42.
2611 // The big tile is stored down, and then the foo() value is inserted.
2612 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2614 Node* ctl = old->in(MemNode::Control);
2615 Node* adr = make_raw_address(offset, phase);
2616 const TypePtr* atp = TypeRawPtr::BOTTOM;
2618 // One or two coalesced stores to plop down.
2619 Node* st[2];
2620 intptr_t off[2];
2621 int nst = 0;
2622 if (!split) {
2623 ++new_long;
2624 off[nst] = offset;
2625 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2626 phase->longcon(con), T_LONG);
2627 } else {
2628 // Omit either if it is a zero.
2629 if (con0 != 0) {
2630 ++new_int;
2631 off[nst] = offset;
2632 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2633 phase->intcon(con0), T_INT);
2634 }
2635 if (con1 != 0) {
2636 ++new_int;
2637 offset += BytesPerInt;
2638 adr = make_raw_address(offset, phase);
2639 off[nst] = offset;
2640 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2641 phase->intcon(con1), T_INT);
2642 }
2643 }
2645 // Insert second store first, then the first before the second.
2646 // Insert each one just before any overlapping non-constant stores.
2647 while (nst > 0) {
2648 Node* st1 = st[--nst];
2649 C->copy_node_notes_to(st1, old);
2650 st1 = phase->transform(st1);
2651 offset = off[nst];
2652 assert(offset >= header_size, "do not smash header");
2653 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
2654 guarantee(ins_idx != 0, "must re-insert constant store");
2655 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
2656 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
2657 set_req(--ins_idx, st1);
2658 else
2659 ins_req(ins_idx, st1);
2660 }
2661 }
2663 if (PrintCompilation && WizardMode)
2664 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
2665 old_subword, old_long, new_int, new_long);
2666 if (C->log() != NULL)
2667 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
2668 old_subword, old_long, new_int, new_long);
2670 // Clean up any remaining occurrences of zmem:
2671 remove_extra_zeroes();
2672 }
2674 // Explore forward from in(start) to find the first fully initialized
2675 // word, and return its offset. Skip groups of subword stores which
2676 // together initialize full words. If in(start) is itself part of a
2677 // fully initialized word, return the offset of in(start). If there
2678 // are no following full-word stores, or if something is fishy, return
2679 // a negative value.
2680 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
2681 int int_map = 0;
2682 intptr_t int_map_off = 0;
2683 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
2685 for (uint i = start, limit = req(); i < limit; i++) {
2686 Node* st = in(i);
2688 intptr_t st_off = get_store_offset(st, phase);
2689 if (st_off < 0) break; // return conservative answer
2691 int st_size = st->as_Store()->memory_size();
2692 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
2693 return st_off; // we found a complete word init
2694 }
2696 // update the map:
2698 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
2699 if (this_int_off != int_map_off) {
2700 // reset the map:
2701 int_map = 0;
2702 int_map_off = this_int_off;
2703 }
2705 int subword_off = st_off - this_int_off;
2706 int_map |= right_n_bits(st_size) << subword_off;
2707 if ((int_map & FULL_MAP) == FULL_MAP) {
2708 return this_int_off; // we found a complete word init
2709 }
2711 // Did this store hit or cross the word boundary?
2712 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
2713 if (next_int_off == this_int_off + BytesPerInt) {
2714 // We passed the current int, without fully initializing it.
2715 int_map_off = next_int_off;
2716 int_map >>= BytesPerInt;
2717 } else if (next_int_off > this_int_off + BytesPerInt) {
2718 // We passed the current and next int.
2719 return this_int_off + BytesPerInt;
2720 }
2721 }
2723 return -1;
2724 }
2727 // Called when the associated AllocateNode is expanded into CFG.
2728 // At this point, we may perform additional optimizations.
2729 // Linearize the stores by ascending offset, to make memory
2730 // activity as coherent as possible.
2731 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
2732 intptr_t header_size,
2733 Node* size_in_bytes,
2734 PhaseGVN* phase) {
2735 assert(!is_complete(), "not already complete");
2736 assert(stores_are_sane(phase), "");
2737 assert(allocation() != NULL, "must be present");
2739 remove_extra_zeroes();
2741 if (ReduceFieldZeroing || ReduceBulkZeroing)
2742 // reduce instruction count for common initialization patterns
2743 coalesce_subword_stores(header_size, size_in_bytes, phase);
2745 Node* zmem = zero_memory(); // initially zero memory state
2746 Node* inits = zmem; // accumulating a linearized chain of inits
2747 #ifdef ASSERT
2748 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
2749 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
2750 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
2751 #endif
2752 intptr_t zeroes_done = header_size;
2754 bool do_zeroing = true; // we might give up if inits are very sparse
2755 int big_init_gaps = 0; // how many large gaps have we seen?
2757 if (ZeroTLAB) do_zeroing = false;
2758 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
2760 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2761 Node* st = in(i);
2762 intptr_t st_off = get_store_offset(st, phase);
2763 if (st_off < 0)
2764 break; // unknown junk in the inits
2765 if (st->in(MemNode::Memory) != zmem)
2766 break; // complicated store chains somehow in list
2768 int st_size = st->as_Store()->memory_size();
2769 intptr_t next_init_off = st_off + st_size;
2771 if (do_zeroing && zeroes_done < next_init_off) {
2772 // See if this store needs a zero before it or under it.
2773 intptr_t zeroes_needed = st_off;
2775 if (st_size < BytesPerInt) {
2776 // Look for subword stores which only partially initialize words.
2777 // If we find some, we must lay down some word-level zeroes first,
2778 // underneath the subword stores.
2779 //
2780 // Examples:
2781 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
2782 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
2783 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
2784 //
2785 // Note: coalesce_subword_stores may have already done this,
2786 // if it was prompted by constant non-zero subword initializers.
2787 // But this case can still arise with non-constant stores.
2789 intptr_t next_full_store = find_next_fullword_store(i, phase);
2791 // In the examples above:
2792 // in(i) p q r s x y z
2793 // st_off 12 13 14 15 12 13 14
2794 // st_size 1 1 1 1 1 1 1
2795 // next_full_s. 12 16 16 16 16 16 16
2796 // z's_done 12 16 16 16 12 16 12
2797 // z's_needed 12 16 16 16 16 16 16
2798 // zsize 0 0 0 0 4 0 4
2799 if (next_full_store < 0) {
2800 // Conservative tack: Zero to end of current word.
2801 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
2802 } else {
2803 // Zero to beginning of next fully initialized word.
2804 // Or, don't zero at all, if we are already in that word.
2805 assert(next_full_store >= zeroes_needed, "must go forward");
2806 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
2807 zeroes_needed = next_full_store;
2808 }
2809 }
2811 if (zeroes_needed > zeroes_done) {
2812 intptr_t zsize = zeroes_needed - zeroes_done;
2813 // Do some incremental zeroing on rawmem, in parallel with inits.
2814 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2815 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2816 zeroes_done, zeroes_needed,
2817 phase);
2818 zeroes_done = zeroes_needed;
2819 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
2820 do_zeroing = false; // leave the hole, next time
2821 }
2822 }
2824 // Collect the store and move on:
2825 st->set_req(MemNode::Memory, inits);
2826 inits = st; // put it on the linearized chain
2827 set_req(i, zmem); // unhook from previous position
2829 if (zeroes_done == st_off)
2830 zeroes_done = next_init_off;
2832 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
2834 #ifdef ASSERT
2835 // Various order invariants. Weaker than stores_are_sane because
2836 // a large constant tile can be filled in by smaller non-constant stores.
2837 assert(st_off >= last_init_off, "inits do not reverse");
2838 last_init_off = st_off;
2839 const Type* val = NULL;
2840 if (st_size >= BytesPerInt &&
2841 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
2842 (int)val->basic_type() < (int)T_OBJECT) {
2843 assert(st_off >= last_tile_end, "tiles do not overlap");
2844 assert(st_off >= last_init_end, "tiles do not overwrite inits");
2845 last_tile_end = MAX2(last_tile_end, next_init_off);
2846 } else {
2847 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
2848 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
2849 assert(st_off >= last_init_end, "inits do not overlap");
2850 last_init_end = next_init_off; // it's a non-tile
2851 }
2852 #endif //ASSERT
2853 }
2855 remove_extra_zeroes(); // clear out all the zmems left over
2856 add_req(inits);
2858 if (!ZeroTLAB) {
2859 // If anything remains to be zeroed, zero it all now.
2860 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2861 // if it is the last unused 4 bytes of an instance, forget about it
2862 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
2863 if (zeroes_done + BytesPerLong >= size_limit) {
2864 assert(allocation() != NULL, "");
2865 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
2866 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
2867 if (zeroes_done == k->layout_helper())
2868 zeroes_done = size_limit;
2869 }
2870 if (zeroes_done < size_limit) {
2871 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2872 zeroes_done, size_in_bytes, phase);
2873 }
2874 }
2876 set_complete(phase);
2877 return rawmem;
2878 }
2881 #ifdef ASSERT
2882 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
2883 if (is_complete())
2884 return true; // stores could be anything at this point
2885 intptr_t last_off = sizeof(oopDesc);
2886 for (uint i = InitializeNode::RawStores; i < req(); i++) {
2887 Node* st = in(i);
2888 intptr_t st_off = get_store_offset(st, phase);
2889 if (st_off < 0) continue; // ignore dead garbage
2890 if (last_off > st_off) {
2891 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
2892 this->dump(2);
2893 assert(false, "ascending store offsets");
2894 return false;
2895 }
2896 last_off = st_off + st->as_Store()->memory_size();
2897 }
2898 return true;
2899 }
2900 #endif //ASSERT
2905 //============================MergeMemNode=====================================
2906 //
2907 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
2908 // contributing store or call operations. Each contributor provides the memory
2909 // state for a particular "alias type" (see Compile::alias_type). For example,
2910 // if a MergeMem has an input X for alias category #6, then any memory reference
2911 // to alias category #6 may use X as its memory state input, as an exact equivalent
2912 // to using the MergeMem as a whole.
2913 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
2914 //
2915 // (Here, the <N> notation gives the index of the relevant adr_type.)
2916 //
2917 // In one special case (and more cases in the future), alias categories overlap.
2918 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
2919 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
2920 // it is exactly equivalent to that state W:
2921 // MergeMem(<Bot>: W) <==> W
2922 //
2923 // Usually, the merge has more than one input. In that case, where inputs
2924 // overlap (i.e., one is Bot), the narrower alias type determines the memory
2925 // state for that type, and the wider alias type (Bot) fills in everywhere else:
2926 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
2927 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
2928 //
2929 // A merge can take a "wide" memory state as one of its narrow inputs.
2930 // This simply means that the merge observes out only the relevant parts of
2931 // the wide input. That is, wide memory states arriving at narrow merge inputs
2932 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
2933 //
2934 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
2935 // and that memory slices "leak through":
2936 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
2937 //
2938 // But, in such a cascade, repeated memory slices can "block the leak":
2939 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
2940 //
2941 // In the last example, Y is not part of the combined memory state of the
2942 // outermost MergeMem. The system must, of course, prevent unschedulable
2943 // memory states from arising, so you can be sure that the state Y is somehow
2944 // a precursor to state Y'.
2945 //
2946 //
2947 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
2948 // of each MergeMemNode array are exactly the numerical alias indexes, including
2949 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
2950 // Compile::alias_type (and kin) produce and manage these indexes.
2951 //
2952 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
2953 // (Note that this provides quick access to the top node inside MergeMem methods,
2954 // without the need to reach out via TLS to Compile::current.)
2955 //
2956 // As a consequence of what was just described, a MergeMem that represents a full
2957 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
2958 // containing all alias categories.
2959 //
2960 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
2961 //
2962 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
2963 // a memory state for the alias type <N>, or else the top node, meaning that
2964 // there is no particular input for that alias type. Note that the length of
2965 // a MergeMem is variable, and may be extended at any time to accommodate new
2966 // memory states at larger alias indexes. When merges grow, they are of course
2967 // filled with "top" in the unused in() positions.
2968 //
2969 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
2970 // (Top was chosen because it works smoothly with passes like GCM.)
2971 //
2972 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
2973 // the type of random VM bits like TLS references.) Since it is always the
2974 // first non-Bot memory slice, some low-level loops use it to initialize an
2975 // index variable: for (i = AliasIdxRaw; i < req(); i++).
2976 //
2977 //
2978 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
2979 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
2980 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
2981 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
2982 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
2983 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
2984 //
2985 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
2986 // really that different from the other memory inputs. An abbreviation called
2987 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
2988 //
2989 //
2990 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
2991 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
2992 // that "emerges though" the base memory will be marked as excluding the alias types
2993 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
2994 //
2995 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
2996 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
2997 //
2998 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
2999 // (It is currently unimplemented.) As you can see, the resulting merge is
3000 // actually a disjoint union of memory states, rather than an overlay.
3001 //
3003 //------------------------------MergeMemNode-----------------------------------
3004 Node* MergeMemNode::make_empty_memory() {
3005 Node* empty_memory = (Node*) Compile::current()->top();
3006 assert(empty_memory->is_top(), "correct sentinel identity");
3007 return empty_memory;
3008 }
3010 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3011 init_class_id(Class_MergeMem);
3012 // all inputs are nullified in Node::Node(int)
3013 // set_input(0, NULL); // no control input
3015 // Initialize the edges uniformly to top, for starters.
3016 Node* empty_mem = make_empty_memory();
3017 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3018 init_req(i,empty_mem);
3019 }
3020 assert(empty_memory() == empty_mem, "");
3022 if( new_base != NULL && new_base->is_MergeMem() ) {
3023 MergeMemNode* mdef = new_base->as_MergeMem();
3024 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3025 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3026 mms.set_memory(mms.memory2());
3027 }
3028 assert(base_memory() == mdef->base_memory(), "");
3029 } else {
3030 set_base_memory(new_base);
3031 }
3032 }
3034 // Make a new, untransformed MergeMem with the same base as 'mem'.
3035 // If mem is itself a MergeMem, populate the result with the same edges.
3036 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3037 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3038 }
3040 //------------------------------cmp--------------------------------------------
3041 uint MergeMemNode::hash() const { return NO_HASH; }
3042 uint MergeMemNode::cmp( const Node &n ) const {
3043 return (&n == this); // Always fail except on self
3044 }
3046 //------------------------------Identity---------------------------------------
3047 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3048 // Identity if this merge point does not record any interesting memory
3049 // disambiguations.
3050 Node* base_mem = base_memory();
3051 Node* empty_mem = empty_memory();
3052 if (base_mem != empty_mem) { // Memory path is not dead?
3053 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3054 Node* mem = in(i);
3055 if (mem != empty_mem && mem != base_mem) {
3056 return this; // Many memory splits; no change
3057 }
3058 }
3059 }
3060 return base_mem; // No memory splits; ID on the one true input
3061 }
3063 //------------------------------Ideal------------------------------------------
3064 // This method is invoked recursively on chains of MergeMem nodes
3065 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3066 // Remove chain'd MergeMems
3067 //
3068 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3069 // relative to the "in(Bot)". Since we are patching both at the same time,
3070 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3071 // but rewrite each "in(i)" relative to the new "in(Bot)".
3072 Node *progress = NULL;
3075 Node* old_base = base_memory();
3076 Node* empty_mem = empty_memory();
3077 if (old_base == empty_mem)
3078 return NULL; // Dead memory path.
3080 MergeMemNode* old_mbase;
3081 if (old_base != NULL && old_base->is_MergeMem())
3082 old_mbase = old_base->as_MergeMem();
3083 else
3084 old_mbase = NULL;
3085 Node* new_base = old_base;
3087 // simplify stacked MergeMems in base memory
3088 if (old_mbase) new_base = old_mbase->base_memory();
3090 // the base memory might contribute new slices beyond my req()
3091 if (old_mbase) grow_to_match(old_mbase);
3093 // Look carefully at the base node if it is a phi.
3094 PhiNode* phi_base;
3095 if (new_base != NULL && new_base->is_Phi())
3096 phi_base = new_base->as_Phi();
3097 else
3098 phi_base = NULL;
3100 Node* phi_reg = NULL;
3101 uint phi_len = (uint)-1;
3102 if (phi_base != NULL && !phi_base->is_copy()) {
3103 // do not examine phi if degraded to a copy
3104 phi_reg = phi_base->region();
3105 phi_len = phi_base->req();
3106 // see if the phi is unfinished
3107 for (uint i = 1; i < phi_len; i++) {
3108 if (phi_base->in(i) == NULL) {
3109 // incomplete phi; do not look at it yet!
3110 phi_reg = NULL;
3111 phi_len = (uint)-1;
3112 break;
3113 }
3114 }
3115 }
3117 // Note: We do not call verify_sparse on entry, because inputs
3118 // can normalize to the base_memory via subsume_node or similar
3119 // mechanisms. This method repairs that damage.
3121 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3123 // Look at each slice.
3124 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3125 Node* old_in = in(i);
3126 // calculate the old memory value
3127 Node* old_mem = old_in;
3128 if (old_mem == empty_mem) old_mem = old_base;
3129 assert(old_mem == memory_at(i), "");
3131 // maybe update (reslice) the old memory value
3133 // simplify stacked MergeMems
3134 Node* new_mem = old_mem;
3135 MergeMemNode* old_mmem;
3136 if (old_mem != NULL && old_mem->is_MergeMem())
3137 old_mmem = old_mem->as_MergeMem();
3138 else
3139 old_mmem = NULL;
3140 if (old_mmem == this) {
3141 // This can happen if loops break up and safepoints disappear.
3142 // A merge of BotPtr (default) with a RawPtr memory derived from a
3143 // safepoint can be rewritten to a merge of the same BotPtr with
3144 // the BotPtr phi coming into the loop. If that phi disappears
3145 // also, we can end up with a self-loop of the mergemem.
3146 // In general, if loops degenerate and memory effects disappear,
3147 // a mergemem can be left looking at itself. This simply means
3148 // that the mergemem's default should be used, since there is
3149 // no longer any apparent effect on this slice.
3150 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3151 // from start. Update the input to TOP.
3152 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3153 }
3154 else if (old_mmem != NULL) {
3155 new_mem = old_mmem->memory_at(i);
3156 }
3157 // else preceeding memory was not a MergeMem
3159 // replace equivalent phis (unfortunately, they do not GVN together)
3160 if (new_mem != NULL && new_mem != new_base &&
3161 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3162 if (new_mem->is_Phi()) {
3163 PhiNode* phi_mem = new_mem->as_Phi();
3164 for (uint i = 1; i < phi_len; i++) {
3165 if (phi_base->in(i) != phi_mem->in(i)) {
3166 phi_mem = NULL;
3167 break;
3168 }
3169 }
3170 if (phi_mem != NULL) {
3171 // equivalent phi nodes; revert to the def
3172 new_mem = new_base;
3173 }
3174 }
3175 }
3177 // maybe store down a new value
3178 Node* new_in = new_mem;
3179 if (new_in == new_base) new_in = empty_mem;
3181 if (new_in != old_in) {
3182 // Warning: Do not combine this "if" with the previous "if"
3183 // A memory slice might have be be rewritten even if it is semantically
3184 // unchanged, if the base_memory value has changed.
3185 set_req(i, new_in);
3186 progress = this; // Report progress
3187 }
3188 }
3190 if (new_base != old_base) {
3191 set_req(Compile::AliasIdxBot, new_base);
3192 // Don't use set_base_memory(new_base), because we need to update du.
3193 assert(base_memory() == new_base, "");
3194 progress = this;
3195 }
3197 if( base_memory() == this ) {
3198 // a self cycle indicates this memory path is dead
3199 set_req(Compile::AliasIdxBot, empty_mem);
3200 }
3202 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3203 // Recursion must occur after the self cycle check above
3204 if( base_memory()->is_MergeMem() ) {
3205 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3206 Node *m = phase->transform(new_mbase); // Rollup any cycles
3207 if( m != NULL && (m->is_top() ||
3208 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3209 // propagate rollup of dead cycle to self
3210 set_req(Compile::AliasIdxBot, empty_mem);
3211 }
3212 }
3214 if( base_memory() == empty_mem ) {
3215 progress = this;
3216 // Cut inputs during Parse phase only.
3217 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3218 if( !can_reshape ) {
3219 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3220 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3221 }
3222 }
3223 }
3225 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3226 // Check if PhiNode::Ideal's "Split phis through memory merges"
3227 // transform should be attempted. Look for this->phi->this cycle.
3228 uint merge_width = req();
3229 if (merge_width > Compile::AliasIdxRaw) {
3230 PhiNode* phi = base_memory()->as_Phi();
3231 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3232 if (phi->in(i) == this) {
3233 phase->is_IterGVN()->_worklist.push(phi);
3234 break;
3235 }
3236 }
3237 }
3238 }
3240 assert(verify_sparse(), "please, no dups of base");
3241 return progress;
3242 }
3244 //-------------------------set_base_memory-------------------------------------
3245 void MergeMemNode::set_base_memory(Node *new_base) {
3246 Node* empty_mem = empty_memory();
3247 set_req(Compile::AliasIdxBot, new_base);
3248 assert(memory_at(req()) == new_base, "must set default memory");
3249 // Clear out other occurrences of new_base:
3250 if (new_base != empty_mem) {
3251 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3252 if (in(i) == new_base) set_req(i, empty_mem);
3253 }
3254 }
3255 }
3257 //------------------------------out_RegMask------------------------------------
3258 const RegMask &MergeMemNode::out_RegMask() const {
3259 return RegMask::Empty;
3260 }
3262 //------------------------------dump_spec--------------------------------------
3263 #ifndef PRODUCT
3264 void MergeMemNode::dump_spec(outputStream *st) const {
3265 st->print(" {");
3266 Node* base_mem = base_memory();
3267 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3268 Node* mem = memory_at(i);
3269 if (mem == base_mem) { st->print(" -"); continue; }
3270 st->print( " N%d:", mem->_idx );
3271 Compile::current()->get_adr_type(i)->dump_on(st);
3272 }
3273 st->print(" }");
3274 }
3275 #endif // !PRODUCT
3278 #ifdef ASSERT
3279 static bool might_be_same(Node* a, Node* b) {
3280 if (a == b) return true;
3281 if (!(a->is_Phi() || b->is_Phi())) return false;
3282 // phis shift around during optimization
3283 return true; // pretty stupid...
3284 }
3286 // verify a narrow slice (either incoming or outgoing)
3287 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3288 if (!VerifyAliases) return; // don't bother to verify unless requested
3289 if (is_error_reported()) return; // muzzle asserts when debugging an error
3290 if (Node::in_dump()) return; // muzzle asserts when printing
3291 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3292 assert(n != NULL, "");
3293 // Elide intervening MergeMem's
3294 while (n->is_MergeMem()) {
3295 n = n->as_MergeMem()->memory_at(alias_idx);
3296 }
3297 Compile* C = Compile::current();
3298 const TypePtr* n_adr_type = n->adr_type();
3299 if (n == m->empty_memory()) {
3300 // Implicit copy of base_memory()
3301 } else if (n_adr_type != TypePtr::BOTTOM) {
3302 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3303 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3304 } else {
3305 // A few places like make_runtime_call "know" that VM calls are narrow,
3306 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3307 bool expected_wide_mem = false;
3308 if (n == m->base_memory()) {
3309 expected_wide_mem = true;
3310 } else if (alias_idx == Compile::AliasIdxRaw ||
3311 n == m->memory_at(Compile::AliasIdxRaw)) {
3312 expected_wide_mem = true;
3313 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3314 // memory can "leak through" calls on channels that
3315 // are write-once. Allow this also.
3316 expected_wide_mem = true;
3317 }
3318 assert(expected_wide_mem, "expected narrow slice replacement");
3319 }
3320 }
3321 #else // !ASSERT
3322 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3323 #endif
3326 //-----------------------------memory_at---------------------------------------
3327 Node* MergeMemNode::memory_at(uint alias_idx) const {
3328 assert(alias_idx >= Compile::AliasIdxRaw ||
3329 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3330 "must avoid base_memory and AliasIdxTop");
3332 // Otherwise, it is a narrow slice.
3333 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3334 Compile *C = Compile::current();
3335 if (is_empty_memory(n)) {
3336 // the array is sparse; empty slots are the "top" node
3337 n = base_memory();
3338 assert(Node::in_dump()
3339 || n == NULL || n->bottom_type() == Type::TOP
3340 || n->adr_type() == TypePtr::BOTTOM
3341 || n->adr_type() == TypeRawPtr::BOTTOM
3342 || Compile::current()->AliasLevel() == 0,
3343 "must be a wide memory");
3344 // AliasLevel == 0 if we are organizing the memory states manually.
3345 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3346 } else {
3347 // make sure the stored slice is sane
3348 #ifdef ASSERT
3349 if (is_error_reported() || Node::in_dump()) {
3350 } else if (might_be_same(n, base_memory())) {
3351 // Give it a pass: It is a mostly harmless repetition of the base.
3352 // This can arise normally from node subsumption during optimization.
3353 } else {
3354 verify_memory_slice(this, alias_idx, n);
3355 }
3356 #endif
3357 }
3358 return n;
3359 }
3361 //---------------------------set_memory_at-------------------------------------
3362 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3363 verify_memory_slice(this, alias_idx, n);
3364 Node* empty_mem = empty_memory();
3365 if (n == base_memory()) n = empty_mem; // collapse default
3366 uint need_req = alias_idx+1;
3367 if (req() < need_req) {
3368 if (n == empty_mem) return; // already the default, so do not grow me
3369 // grow the sparse array
3370 do {
3371 add_req(empty_mem);
3372 } while (req() < need_req);
3373 }
3374 set_req( alias_idx, n );
3375 }
3379 //--------------------------iteration_setup------------------------------------
3380 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3381 if (other != NULL) {
3382 grow_to_match(other);
3383 // invariant: the finite support of mm2 is within mm->req()
3384 #ifdef ASSERT
3385 for (uint i = req(); i < other->req(); i++) {
3386 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3387 }
3388 #endif
3389 }
3390 // Replace spurious copies of base_memory by top.
3391 Node* base_mem = base_memory();
3392 if (base_mem != NULL && !base_mem->is_top()) {
3393 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3394 if (in(i) == base_mem)
3395 set_req(i, empty_memory());
3396 }
3397 }
3398 }
3400 //---------------------------grow_to_match-------------------------------------
3401 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3402 Node* empty_mem = empty_memory();
3403 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3404 // look for the finite support of the other memory
3405 for (uint i = other->req(); --i >= req(); ) {
3406 if (other->in(i) != empty_mem) {
3407 uint new_len = i+1;
3408 while (req() < new_len) add_req(empty_mem);
3409 break;
3410 }
3411 }
3412 }
3414 //---------------------------verify_sparse-------------------------------------
3415 #ifndef PRODUCT
3416 bool MergeMemNode::verify_sparse() const {
3417 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3418 Node* base_mem = base_memory();
3419 // The following can happen in degenerate cases, since empty==top.
3420 if (is_empty_memory(base_mem)) return true;
3421 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3422 assert(in(i) != NULL, "sane slice");
3423 if (in(i) == base_mem) return false; // should have been the sentinel value!
3424 }
3425 return true;
3426 }
3428 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3429 Node* n;
3430 n = mm->in(idx);
3431 if (mem == n) return true; // might be empty_memory()
3432 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3433 if (mem == n) return true;
3434 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3435 if (mem == n) return true;
3436 if (n == NULL) break;
3437 }
3438 return false;
3439 }
3440 #endif // !PRODUCT