Thu, 13 Mar 2008 16:31:32 -0700
6673473: (Escape Analysis) Add the instance's field information to PhiNode
Summary: Avoid an infinite generation of instance's field values Phi nodes.
Reviewed-by: 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 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
91 uint alias_idx = phase->C->get_alias_index(tp);
92 Node *mem = mmem;
93 #ifdef ASSERT
94 {
95 // Check that current type is consistent with the alias index used during graph construction
96 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
97 bool consistent = adr_check == NULL || adr_check->empty() ||
98 phase->C->must_alias(adr_check, alias_idx );
99 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
100 if( !consistent && adr_check != NULL && !adr_check->empty() &&
101 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
102 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
103 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
104 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
105 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
106 // don't assert if it is dead code.
107 consistent = true;
108 }
109 if( !consistent ) {
110 st->print("alias_idx==%d, adr_check==", alias_idx);
111 if( adr_check == NULL ) {
112 st->print("NULL");
113 } else {
114 adr_check->dump();
115 }
116 st->cr();
117 print_alias_types();
118 assert(consistent, "adr_check must match alias idx");
119 }
120 }
121 #endif
122 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
123 // means an array I have not precisely typed yet. Do not do any
124 // alias stuff with it any time soon.
125 const TypeOopPtr *tinst = tp->isa_oopptr();
126 if( tp->base() != Type::AnyPtr &&
127 !(tinst &&
128 tinst->klass()->is_java_lang_Object() &&
129 tinst->offset() == Type::OffsetBot) ) {
130 // compress paths and change unreachable cycles to TOP
131 // If not, we can update the input infinitely along a MergeMem cycle
132 // Equivalent code in PhiNode::Ideal
133 Node* m = phase->transform(mmem);
134 // If tranformed to a MergeMem, get the desired slice
135 // Otherwise the returned node represents memory for every slice
136 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
137 // Update input if it is progress over what we have now
138 }
139 return mem;
140 }
142 //--------------------------Ideal_common---------------------------------------
143 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
144 // Unhook non-raw memories from complete (macro-expanded) initializations.
145 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
146 // If our control input is a dead region, kill all below the region
147 Node *ctl = in(MemNode::Control);
148 if (ctl && remove_dead_region(phase, can_reshape))
149 return this;
151 // Ignore if memory is dead, or self-loop
152 Node *mem = in(MemNode::Memory);
153 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
154 assert( mem != this, "dead loop in MemNode::Ideal" );
156 Node *address = in(MemNode::Address);
157 const Type *t_adr = phase->type( address );
158 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
160 // Avoid independent memory operations
161 Node* old_mem = mem;
163 // The code which unhooks non-raw memories from complete (macro-expanded)
164 // initializations was removed. After macro-expansion all stores catched
165 // by Initialize node became raw stores and there is no information
166 // which memory slices they modify. So it is unsafe to move any memory
167 // operation above these stores. Also in most cases hooked non-raw memories
168 // were already unhooked by using information from detect_ptr_independence()
169 // and find_previous_store().
171 if (mem->is_MergeMem()) {
172 MergeMemNode* mmem = mem->as_MergeMem();
173 const TypePtr *tp = t_adr->is_ptr();
175 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
176 }
178 if (mem != old_mem) {
179 set_req(MemNode::Memory, mem);
180 return this;
181 }
183 // let the subclass continue analyzing...
184 return NULL;
185 }
187 // Helper function for proving some simple control dominations.
188 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'.
189 // Already assumes that 'dom' is available at 'sub', and that 'sub'
190 // is not a constant (dominated by the method's StartNode).
191 // Used by MemNode::find_previous_store to prove that the
192 // control input of a memory operation predates (dominates)
193 // an allocation it wants to look past.
194 bool MemNode::detect_dominating_control(Node* dom, Node* sub) {
195 if (dom == NULL) return false;
196 if (dom->is_Proj()) dom = dom->in(0);
197 if (dom->is_Start()) return true; // anything inside the method
198 if (dom->is_Root()) return true; // dom 'controls' a constant
199 int cnt = 20; // detect cycle or too much effort
200 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
201 if (--cnt < 0) return false; // in a cycle or too complex
202 if (sub == dom) return true;
203 if (sub->is_Start()) return false;
204 if (sub->is_Root()) return false;
205 Node* up = sub->in(0);
206 if (sub == up && sub->is_Region()) {
207 for (uint i = 1; i < sub->req(); i++) {
208 Node* in = sub->in(i);
209 if (in != NULL && !in->is_top() && in != sub) {
210 up = in; break; // take any path on the way up to 'dom'
211 }
212 }
213 }
214 if (sub == up) return false; // some kind of tight cycle
215 sub = up;
216 }
217 return false;
218 }
220 //---------------------detect_ptr_independence---------------------------------
221 // Used by MemNode::find_previous_store to prove that two base
222 // pointers are never equal.
223 // The pointers are accompanied by their associated allocations,
224 // if any, which have been previously discovered by the caller.
225 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
226 Node* p2, AllocateNode* a2,
227 PhaseTransform* phase) {
228 // Attempt to prove that these two pointers cannot be aliased.
229 // They may both manifestly be allocations, and they should differ.
230 // Or, if they are not both allocations, they can be distinct constants.
231 // Otherwise, one is an allocation and the other a pre-existing value.
232 if (a1 == NULL && a2 == NULL) { // neither an allocation
233 return (p1 != p2) && p1->is_Con() && p2->is_Con();
234 } else if (a1 != NULL && a2 != NULL) { // both allocations
235 return (a1 != a2);
236 } else if (a1 != NULL) { // one allocation a1
237 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
238 return detect_dominating_control(p2->in(0), a1->in(0));
239 } else { //(a2 != NULL) // one allocation a2
240 return detect_dominating_control(p1->in(0), a2->in(0));
241 }
242 return false;
243 }
246 // The logic for reordering loads and stores uses four steps:
247 // (a) Walk carefully past stores and initializations which we
248 // can prove are independent of this load.
249 // (b) Observe that the next memory state makes an exact match
250 // with self (load or store), and locate the relevant store.
251 // (c) Ensure that, if we were to wire self directly to the store,
252 // the optimizer would fold it up somehow.
253 // (d) Do the rewiring, and return, depending on some other part of
254 // the optimizer to fold up the load.
255 // This routine handles steps (a) and (b). Steps (c) and (d) are
256 // specific to loads and stores, so they are handled by the callers.
257 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
258 //
259 Node* MemNode::find_previous_store(PhaseTransform* phase) {
260 Node* ctrl = in(MemNode::Control);
261 Node* adr = in(MemNode::Address);
262 intptr_t offset = 0;
263 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
264 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
266 if (offset == Type::OffsetBot)
267 return NULL; // cannot unalias unless there are precise offsets
269 intptr_t size_in_bytes = memory_size();
271 Node* mem = in(MemNode::Memory); // start searching here...
273 int cnt = 50; // Cycle limiter
274 for (;;) { // While we can dance past unrelated stores...
275 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
277 if (mem->is_Store()) {
278 Node* st_adr = mem->in(MemNode::Address);
279 intptr_t st_offset = 0;
280 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
281 if (st_base == NULL)
282 break; // inscrutable pointer
283 if (st_offset != offset && st_offset != Type::OffsetBot) {
284 const int MAX_STORE = BytesPerLong;
285 if (st_offset >= offset + size_in_bytes ||
286 st_offset <= offset - MAX_STORE ||
287 st_offset <= offset - mem->as_Store()->memory_size()) {
288 // Success: The offsets are provably independent.
289 // (You may ask, why not just test st_offset != offset and be done?
290 // The answer is that stores of different sizes can co-exist
291 // in the same sequence of RawMem effects. We sometimes initialize
292 // a whole 'tile' of array elements with a single jint or jlong.)
293 mem = mem->in(MemNode::Memory);
294 continue; // (a) advance through independent store memory
295 }
296 }
297 if (st_base != base &&
298 detect_ptr_independence(base, alloc,
299 st_base,
300 AllocateNode::Ideal_allocation(st_base, phase),
301 phase)) {
302 // Success: The bases are provably independent.
303 mem = mem->in(MemNode::Memory);
304 continue; // (a) advance through independent store memory
305 }
307 // (b) At this point, if the bases or offsets do not agree, we lose,
308 // since we have not managed to prove 'this' and 'mem' independent.
309 if (st_base == base && st_offset == offset) {
310 return mem; // let caller handle steps (c), (d)
311 }
313 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
314 InitializeNode* st_init = mem->in(0)->as_Initialize();
315 AllocateNode* st_alloc = st_init->allocation();
316 if (st_alloc == NULL)
317 break; // something degenerated
318 bool known_identical = false;
319 bool known_independent = false;
320 if (alloc == st_alloc)
321 known_identical = true;
322 else if (alloc != NULL)
323 known_independent = true;
324 else if (ctrl != NULL &&
325 detect_dominating_control(ctrl, st_alloc->in(0)))
326 known_independent = true;
328 if (known_independent) {
329 // The bases are provably independent: Either they are
330 // manifestly distinct allocations, or else the control
331 // of this load dominates the store's allocation.
332 int alias_idx = phase->C->get_alias_index(adr_type());
333 if (alias_idx == Compile::AliasIdxRaw) {
334 mem = st_alloc->in(TypeFunc::Memory);
335 } else {
336 mem = st_init->memory(alias_idx);
337 }
338 continue; // (a) advance through independent store memory
339 }
341 // (b) at this point, if we are not looking at a store initializing
342 // the same allocation we are loading from, we lose.
343 if (known_identical) {
344 // From caller, can_see_stored_value will consult find_captured_store.
345 return mem; // let caller handle steps (c), (d)
346 }
348 }
350 // Unless there is an explicit 'continue', we must bail out here,
351 // because 'mem' is an inscrutable memory state (e.g., a call).
352 break;
353 }
355 return NULL; // bail out
356 }
358 //----------------------calculate_adr_type-------------------------------------
359 // Helper function. Notices when the given type of address hits top or bottom.
360 // Also, asserts a cross-check of the type against the expected address type.
361 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
362 if (t == Type::TOP) return NULL; // does not touch memory any more?
363 #ifdef PRODUCT
364 cross_check = NULL;
365 #else
366 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
367 #endif
368 const TypePtr* tp = t->isa_ptr();
369 if (tp == NULL) {
370 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
371 return TypePtr::BOTTOM; // touches lots of memory
372 } else {
373 #ifdef ASSERT
374 // %%%% [phh] We don't check the alias index if cross_check is
375 // TypeRawPtr::BOTTOM. Needs to be investigated.
376 if (cross_check != NULL &&
377 cross_check != TypePtr::BOTTOM &&
378 cross_check != TypeRawPtr::BOTTOM) {
379 // Recheck the alias index, to see if it has changed (due to a bug).
380 Compile* C = Compile::current();
381 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
382 "must stay in the original alias category");
383 // The type of the address must be contained in the adr_type,
384 // disregarding "null"-ness.
385 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
386 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
387 assert(cross_check->meet(tp_notnull) == cross_check,
388 "real address must not escape from expected memory type");
389 }
390 #endif
391 return tp;
392 }
393 }
395 //------------------------adr_phi_is_loop_invariant----------------------------
396 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
397 // loop is loop invariant. Make a quick traversal of Phi and associated
398 // CastPP nodes, looking to see if they are a closed group within the loop.
399 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
400 // The idea is that the phi-nest must boil down to only CastPP nodes
401 // with the same data. This implies that any path into the loop already
402 // includes such a CastPP, and so the original cast, whatever its input,
403 // must be covered by an equivalent cast, with an earlier control input.
404 ResourceMark rm;
406 // The loop entry input of the phi should be the unique dominating
407 // node for every Phi/CastPP in the loop.
408 Unique_Node_List closure;
409 closure.push(adr_phi->in(LoopNode::EntryControl));
411 // Add the phi node and the cast to the worklist.
412 Unique_Node_List worklist;
413 worklist.push(adr_phi);
414 if( cast != NULL ){
415 if( !cast->is_ConstraintCast() ) return false;
416 worklist.push(cast);
417 }
419 // Begin recursive walk of phi nodes.
420 while( worklist.size() ){
421 // Take a node off the worklist
422 Node *n = worklist.pop();
423 if( !closure.member(n) ){
424 // Add it to the closure.
425 closure.push(n);
426 // Make a sanity check to ensure we don't waste too much time here.
427 if( closure.size() > 20) return false;
428 // This node is OK if:
429 // - it is a cast of an identical value
430 // - or it is a phi node (then we add its inputs to the worklist)
431 // Otherwise, the node is not OK, and we presume the cast is not invariant
432 if( n->is_ConstraintCast() ){
433 worklist.push(n->in(1));
434 } else if( n->is_Phi() ) {
435 for( uint i = 1; i < n->req(); i++ ) {
436 worklist.push(n->in(i));
437 }
438 } else {
439 return false;
440 }
441 }
442 }
444 // Quit when the worklist is empty, and we've found no offending nodes.
445 return true;
446 }
448 //------------------------------Ideal_DU_postCCP-------------------------------
449 // Find any cast-away of null-ness and keep its control. Null cast-aways are
450 // going away in this pass and we need to make this memory op depend on the
451 // gating null check.
453 // I tried to leave the CastPP's in. This makes the graph more accurate in
454 // some sense; we get to keep around the knowledge that an oop is not-null
455 // after some test. Alas, the CastPP's interfere with GVN (some values are
456 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
457 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
458 // some of the more trivial cases in the optimizer. Removing more useless
459 // Phi's started allowing Loads to illegally float above null checks. I gave
460 // up on this approach. CNC 10/20/2000
461 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
462 Node *ctr = in(MemNode::Control);
463 Node *mem = in(MemNode::Memory);
464 Node *adr = in(MemNode::Address);
465 Node *skipped_cast = NULL;
466 // Need a null check? Regular static accesses do not because they are
467 // from constant addresses. Array ops are gated by the range check (which
468 // always includes a NULL check). Just check field ops.
469 if( !ctr ) {
470 // Scan upwards for the highest location we can place this memory op.
471 while( true ) {
472 switch( adr->Opcode() ) {
474 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
475 adr = adr->in(AddPNode::Base);
476 continue;
478 case Op_CastPP:
479 // If the CastPP is useless, just peek on through it.
480 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
481 // Remember the cast that we've peeked though. If we peek
482 // through more than one, then we end up remembering the highest
483 // one, that is, if in a loop, the one closest to the top.
484 skipped_cast = adr;
485 adr = adr->in(1);
486 continue;
487 }
488 // CastPP is going away in this pass! We need this memory op to be
489 // control-dependent on the test that is guarding the CastPP.
490 ccp->hash_delete(this);
491 set_req(MemNode::Control, adr->in(0));
492 ccp->hash_insert(this);
493 return this;
495 case Op_Phi:
496 // Attempt to float above a Phi to some dominating point.
497 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
498 // If we've already peeked through a Cast (which could have set the
499 // control), we can't float above a Phi, because the skipped Cast
500 // may not be loop invariant.
501 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
502 adr = adr->in(1);
503 continue;
504 }
505 }
507 // Intentional fallthrough!
509 // No obvious dominating point. The mem op is pinned below the Phi
510 // by the Phi itself. If the Phi goes away (no true value is merged)
511 // then the mem op can float, but not indefinitely. It must be pinned
512 // behind the controls leading to the Phi.
513 case Op_CheckCastPP:
514 // These usually stick around to change address type, however a
515 // useless one can be elided and we still need to pick up a control edge
516 if (adr->in(0) == NULL) {
517 // This CheckCastPP node has NO control and is likely useless. But we
518 // need check further up the ancestor chain for a control input to keep
519 // the node in place. 4959717.
520 skipped_cast = adr;
521 adr = adr->in(1);
522 continue;
523 }
524 ccp->hash_delete(this);
525 set_req(MemNode::Control, adr->in(0));
526 ccp->hash_insert(this);
527 return this;
529 // List of "safe" opcodes; those that implicitly block the memory
530 // op below any null check.
531 case Op_CastX2P: // no null checks on native pointers
532 case Op_Parm: // 'this' pointer is not null
533 case Op_LoadP: // Loading from within a klass
534 case Op_LoadKlass: // Loading from within a klass
535 case Op_ConP: // Loading from a klass
536 case Op_CreateEx: // Sucking up the guts of an exception oop
537 case Op_Con: // Reading from TLS
538 case Op_CMoveP: // CMoveP is pinned
539 break; // No progress
541 case Op_Proj: // Direct call to an allocation routine
542 case Op_SCMemProj: // Memory state from store conditional ops
543 #ifdef ASSERT
544 {
545 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
546 const Node* call = adr->in(0);
547 if (call->is_CallStaticJava()) {
548 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
549 const TypeTuple *r = call_java->tf()->range();
550 assert(r->cnt() > TypeFunc::Parms, "must return value");
551 const Type* ret_type = r->field_at(TypeFunc::Parms);
552 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
553 // We further presume that this is one of
554 // new_instance_Java, new_array_Java, or
555 // the like, but do not assert for this.
556 } else if (call->is_Allocate()) {
557 // similar case to new_instance_Java, etc.
558 } else if (!call->is_CallLeaf()) {
559 // Projections from fetch_oop (OSR) are allowed as well.
560 ShouldNotReachHere();
561 }
562 }
563 #endif
564 break;
565 default:
566 ShouldNotReachHere();
567 }
568 break;
569 }
570 }
572 return NULL; // No progress
573 }
576 //=============================================================================
577 uint LoadNode::size_of() const { return sizeof(*this); }
578 uint LoadNode::cmp( const Node &n ) const
579 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
580 const Type *LoadNode::bottom_type() const { return _type; }
581 uint LoadNode::ideal_reg() const {
582 return Matcher::base2reg[_type->base()];
583 }
585 #ifndef PRODUCT
586 void LoadNode::dump_spec(outputStream *st) const {
587 MemNode::dump_spec(st);
588 if( !Verbose && !WizardMode ) {
589 // standard dump does this in Verbose and WizardMode
590 st->print(" #"); _type->dump_on(st);
591 }
592 }
593 #endif
596 //----------------------------LoadNode::make-----------------------------------
597 // Polymorphic factory method:
598 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
599 // sanity check the alias category against the created node type
600 assert(!(adr_type->isa_oopptr() &&
601 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
602 "use LoadKlassNode instead");
603 assert(!(adr_type->isa_aryptr() &&
604 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
605 "use LoadRangeNode instead");
606 switch (bt) {
607 case T_BOOLEAN:
608 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
609 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
610 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
611 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
612 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
613 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
614 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
615 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
616 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
617 }
618 ShouldNotReachHere();
619 return (LoadNode*)NULL;
620 }
622 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
623 bool require_atomic = true;
624 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
625 }
630 //------------------------------hash-------------------------------------------
631 uint LoadNode::hash() const {
632 // unroll addition of interesting fields
633 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
634 }
636 //---------------------------can_see_stored_value------------------------------
637 // This routine exists to make sure this set of tests is done the same
638 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
639 // will change the graph shape in a way which makes memory alive twice at the
640 // same time (uses the Oracle model of aliasing), then some
641 // LoadXNode::Identity will fold things back to the equivalence-class model
642 // of aliasing.
643 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
644 Node* ld_adr = in(MemNode::Address);
646 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
647 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
648 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
649 atp->field() != NULL && !atp->field()->is_volatile()) {
650 uint alias_idx = atp->index();
651 bool final = atp->field()->is_final();
652 Node* result = NULL;
653 Node* current = st;
654 // Skip through chains of MemBarNodes checking the MergeMems for
655 // new states for the slice of this load. Stop once any other
656 // kind of node is encountered. Loads from final memory can skip
657 // through any kind of MemBar but normal loads shouldn't skip
658 // through MemBarAcquire since the could allow them to move out of
659 // a synchronized region.
660 while (current->is_Proj()) {
661 int opc = current->in(0)->Opcode();
662 if ((final && opc == Op_MemBarAcquire) ||
663 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
664 Node* mem = current->in(0)->in(TypeFunc::Memory);
665 if (mem->is_MergeMem()) {
666 MergeMemNode* merge = mem->as_MergeMem();
667 Node* new_st = merge->memory_at(alias_idx);
668 if (new_st == merge->base_memory()) {
669 // Keep searching
670 current = merge->base_memory();
671 continue;
672 }
673 // Save the new memory state for the slice and fall through
674 // to exit.
675 result = new_st;
676 }
677 }
678 break;
679 }
680 if (result != NULL) {
681 st = result;
682 }
683 }
686 // Loop around twice in the case Load -> Initialize -> Store.
687 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
688 for (int trip = 0; trip <= 1; trip++) {
690 if (st->is_Store()) {
691 Node* st_adr = st->in(MemNode::Address);
692 if (!phase->eqv(st_adr, ld_adr)) {
693 // Try harder before giving up... Match raw and non-raw pointers.
694 intptr_t st_off = 0;
695 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
696 if (alloc == NULL) return NULL;
697 intptr_t ld_off = 0;
698 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
699 if (alloc != allo2) return NULL;
700 if (ld_off != st_off) return NULL;
701 // At this point we have proven something like this setup:
702 // A = Allocate(...)
703 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
704 // S = StoreQ(, AddP(, A.Parm , #Off), V)
705 // (Actually, we haven't yet proven the Q's are the same.)
706 // In other words, we are loading from a casted version of
707 // the same pointer-and-offset that we stored to.
708 // Thus, we are able to replace L by V.
709 }
710 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
711 if (store_Opcode() != st->Opcode())
712 return NULL;
713 return st->in(MemNode::ValueIn);
714 }
716 intptr_t offset = 0; // scratch
718 // A load from a freshly-created object always returns zero.
719 // (This can happen after LoadNode::Ideal resets the load's memory input
720 // to find_captured_store, which returned InitializeNode::zero_memory.)
721 if (st->is_Proj() && st->in(0)->is_Allocate() &&
722 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
723 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
724 // return a zero value for the load's basic type
725 // (This is one of the few places where a generic PhaseTransform
726 // can create new nodes. Think of it as lazily manifesting
727 // virtually pre-existing constants.)
728 return phase->zerocon(memory_type());
729 }
731 // A load from an initialization barrier can match a captured store.
732 if (st->is_Proj() && st->in(0)->is_Initialize()) {
733 InitializeNode* init = st->in(0)->as_Initialize();
734 AllocateNode* alloc = init->allocation();
735 if (alloc != NULL &&
736 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
737 // examine a captured store value
738 st = init->find_captured_store(offset, memory_size(), phase);
739 if (st != NULL)
740 continue; // take one more trip around
741 }
742 }
744 break;
745 }
747 return NULL;
748 }
750 //----------------------is_instance_field_load_with_local_phi------------------
751 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
752 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
753 in(MemNode::Address)->is_AddP() ) {
754 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
755 // Only instances.
756 if( t_oop != NULL && t_oop->is_instance_field() &&
757 t_oop->offset() != Type::OffsetBot &&
758 t_oop->offset() != Type::OffsetTop) {
759 return true;
760 }
761 }
762 return false;
763 }
765 //------------------------------Identity---------------------------------------
766 // Loads are identity if previous store is to same address
767 Node *LoadNode::Identity( PhaseTransform *phase ) {
768 // If the previous store-maker is the right kind of Store, and the store is
769 // to the same address, then we are equal to the value stored.
770 Node* mem = in(MemNode::Memory);
771 Node* value = can_see_stored_value(mem, phase);
772 if( value ) {
773 // byte, short & char stores truncate naturally.
774 // A load has to load the truncated value which requires
775 // some sort of masking operation and that requires an
776 // Ideal call instead of an Identity call.
777 if (memory_size() < BytesPerInt) {
778 // If the input to the store does not fit with the load's result type,
779 // it must be truncated via an Ideal call.
780 if (!phase->type(value)->higher_equal(phase->type(this)))
781 return this;
782 }
783 // (This works even when value is a Con, but LoadNode::Value
784 // usually runs first, producing the singleton type of the Con.)
785 return value;
786 }
788 // Search for an existing data phi which was generated before for the same
789 // instance's field to avoid infinite genertion of phis in a loop.
790 Node *region = mem->in(0);
791 if (is_instance_field_load_with_local_phi(region)) {
792 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
793 int this_index = phase->C->get_alias_index(addr_t);
794 int this_offset = addr_t->offset();
795 int this_id = addr_t->is_oopptr()->instance_id();
796 const Type* this_type = bottom_type();
797 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
798 Node* phi = region->fast_out(i);
799 if (phi->is_Phi() && phi != mem &&
800 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
801 return phi;
802 }
803 }
804 }
806 return this;
807 }
810 // Returns true if the AliasType refers to the field that holds the
811 // cached box array. Currently only handles the IntegerCache case.
812 static bool is_autobox_cache(Compile::AliasType* atp) {
813 if (atp != NULL && atp->field() != NULL) {
814 ciField* field = atp->field();
815 ciSymbol* klass = field->holder()->name();
816 if (field->name() == ciSymbol::cache_field_name() &&
817 field->holder()->uses_default_loader() &&
818 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
819 return true;
820 }
821 }
822 return false;
823 }
825 // Fetch the base value in the autobox array
826 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
827 if (atp != NULL && atp->field() != NULL) {
828 ciField* field = atp->field();
829 ciSymbol* klass = field->holder()->name();
830 if (field->name() == ciSymbol::cache_field_name() &&
831 field->holder()->uses_default_loader() &&
832 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
833 assert(field->is_constant(), "what?");
834 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
835 // Fetch the box object at the base of the array and get its value
836 ciInstance* box = array->obj_at(0)->as_instance();
837 ciInstanceKlass* ik = box->klass()->as_instance_klass();
838 if (ik->nof_nonstatic_fields() == 1) {
839 // This should be true nonstatic_field_at requires calling
840 // nof_nonstatic_fields so check it anyway
841 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
842 cache_offset = c.as_int();
843 }
844 return true;
845 }
846 }
847 return false;
848 }
850 // Returns true if the AliasType refers to the value field of an
851 // autobox object. Currently only handles Integer.
852 static bool is_autobox_object(Compile::AliasType* atp) {
853 if (atp != NULL && atp->field() != NULL) {
854 ciField* field = atp->field();
855 ciSymbol* klass = field->holder()->name();
856 if (field->name() == ciSymbol::value_name() &&
857 field->holder()->uses_default_loader() &&
858 klass == ciSymbol::java_lang_Integer()) {
859 return true;
860 }
861 }
862 return false;
863 }
866 // We're loading from an object which has autobox behaviour.
867 // If this object is result of a valueOf call we'll have a phi
868 // merging a newly allocated object and a load from the cache.
869 // We want to replace this load with the original incoming
870 // argument to the valueOf call.
871 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
872 Node* base = in(Address)->in(AddPNode::Base);
873 if (base->is_Phi() && base->req() == 3) {
874 AllocateNode* allocation = NULL;
875 int allocation_index = -1;
876 int load_index = -1;
877 for (uint i = 1; i < base->req(); i++) {
878 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
879 if (allocation != NULL) {
880 allocation_index = i;
881 load_index = 3 - allocation_index;
882 break;
883 }
884 }
885 LoadNode* load = NULL;
886 if (allocation != NULL && base->in(load_index)->is_Load()) {
887 load = base->in(load_index)->as_Load();
888 }
889 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
890 // Push the loads from the phi that comes from valueOf up
891 // through it to allow elimination of the loads and the recovery
892 // of the original value.
893 Node* mem_phi = in(Memory);
894 Node* offset = in(Address)->in(AddPNode::Offset);
896 Node* in1 = clone();
897 Node* in1_addr = in1->in(Address)->clone();
898 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
899 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
900 in1_addr->set_req(AddPNode::Offset, offset);
901 in1->set_req(0, base->in(allocation_index));
902 in1->set_req(Address, in1_addr);
903 in1->set_req(Memory, mem_phi->in(allocation_index));
905 Node* in2 = clone();
906 Node* in2_addr = in2->in(Address)->clone();
907 in2_addr->set_req(AddPNode::Base, base->in(load_index));
908 in2_addr->set_req(AddPNode::Address, base->in(load_index));
909 in2_addr->set_req(AddPNode::Offset, offset);
910 in2->set_req(0, base->in(load_index));
911 in2->set_req(Address, in2_addr);
912 in2->set_req(Memory, mem_phi->in(load_index));
914 in1_addr = phase->transform(in1_addr);
915 in1 = phase->transform(in1);
916 in2_addr = phase->transform(in2_addr);
917 in2 = phase->transform(in2);
919 PhiNode* result = PhiNode::make_blank(base->in(0), this);
920 result->set_req(allocation_index, in1);
921 result->set_req(load_index, in2);
922 return result;
923 }
924 } else if (base->is_Load()) {
925 // Eliminate the load of Integer.value for integers from the cache
926 // array by deriving the value from the index into the array.
927 // Capture the offset of the load and then reverse the computation.
928 Node* load_base = base->in(Address)->in(AddPNode::Base);
929 if (load_base != NULL) {
930 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
931 intptr_t cache_offset;
932 int shift = -1;
933 Node* cache = NULL;
934 if (is_autobox_cache(atp)) {
935 shift = exact_log2(type2aelembytes(T_OBJECT));
936 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
937 }
938 if (cache != NULL && base->in(Address)->is_AddP()) {
939 Node* elements[4];
940 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
941 int cache_low;
942 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
943 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
944 // Add up all the offsets making of the address of the load
945 Node* result = elements[0];
946 for (int i = 1; i < count; i++) {
947 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
948 }
949 // Remove the constant offset from the address and then
950 // remove the scaling of the offset to recover the original index.
951 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
952 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
953 // Peel the shift off directly but wrap it in a dummy node
954 // since Ideal can't return existing nodes
955 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
956 } else {
957 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
958 }
959 #ifdef _LP64
960 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
961 #endif
962 return result;
963 }
964 }
965 }
966 }
967 return NULL;
968 }
971 //------------------------------Ideal------------------------------------------
972 // If the load is from Field memory and the pointer is non-null, we can
973 // zero out the control input.
974 // If the offset is constant and the base is an object allocation,
975 // try to hook me up to the exact initializing store.
976 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
977 Node* p = MemNode::Ideal_common(phase, can_reshape);
978 if (p) return (p == NodeSentinel) ? NULL : p;
980 Node* ctrl = in(MemNode::Control);
981 Node* address = in(MemNode::Address);
983 // Skip up past a SafePoint control. Cannot do this for Stores because
984 // pointer stores & cardmarks must stay on the same side of a SafePoint.
985 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
986 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
987 ctrl = ctrl->in(0);
988 set_req(MemNode::Control,ctrl);
989 }
991 // Check for useless control edge in some common special cases
992 if (in(MemNode::Control) != NULL) {
993 intptr_t ignore = 0;
994 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
995 if (base != NULL
996 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
997 && detect_dominating_control(base->in(0), phase->C->start())) {
998 // A method-invariant, non-null address (constant or 'this' argument).
999 set_req(MemNode::Control, NULL);
1000 }
1001 }
1003 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1004 Node* base = in(Address)->in(AddPNode::Base);
1005 if (base != NULL) {
1006 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1007 if (is_autobox_object(atp)) {
1008 Node* result = eliminate_autobox(phase);
1009 if (result != NULL) return result;
1010 }
1011 }
1012 }
1014 // Check for prior store with a different base or offset; make Load
1015 // independent. Skip through any number of them. Bail out if the stores
1016 // are in an endless dead cycle and report no progress. This is a key
1017 // transform for Reflection. However, if after skipping through the Stores
1018 // we can't then fold up against a prior store do NOT do the transform as
1019 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1020 // array memory alive twice: once for the hoisted Load and again after the
1021 // bypassed Store. This situation only works if EVERYBODY who does
1022 // anti-dependence work knows how to bypass. I.e. we need all
1023 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1024 // the alias index stuff. So instead, peek through Stores and IFF we can
1025 // fold up, do so.
1026 Node* prev_mem = find_previous_store(phase);
1027 // Steps (a), (b): Walk past independent stores to find an exact match.
1028 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1029 // (c) See if we can fold up on the spot, but don't fold up here.
1030 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1031 // just return a prior value, which is done by Identity calls.
1032 if (can_see_stored_value(prev_mem, phase)) {
1033 // Make ready for step (d):
1034 set_req(MemNode::Memory, prev_mem);
1035 return this;
1036 }
1037 }
1039 return NULL; // No further progress
1040 }
1042 // Helper to recognize certain Klass fields which are invariant across
1043 // some group of array types (e.g., int[] or all T[] where T < Object).
1044 const Type*
1045 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1046 ciKlass* klass) const {
1047 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1048 // The field is Klass::_modifier_flags. Return its (constant) value.
1049 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1050 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1051 return TypeInt::make(klass->modifier_flags());
1052 }
1053 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1054 // The field is Klass::_access_flags. Return its (constant) value.
1055 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1056 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1057 return TypeInt::make(klass->access_flags());
1058 }
1059 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1060 // The field is Klass::_layout_helper. Return its constant value if known.
1061 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1062 return TypeInt::make(klass->layout_helper());
1063 }
1065 // No match.
1066 return NULL;
1067 }
1069 //------------------------------Value-----------------------------------------
1070 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1071 // Either input is TOP ==> the result is TOP
1072 Node* mem = in(MemNode::Memory);
1073 const Type *t1 = phase->type(mem);
1074 if (t1 == Type::TOP) return Type::TOP;
1075 Node* adr = in(MemNode::Address);
1076 const TypePtr* tp = phase->type(adr)->isa_ptr();
1077 if (tp == NULL || tp->empty()) return Type::TOP;
1078 int off = tp->offset();
1079 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1081 // Try to guess loaded type from pointer type
1082 if (tp->base() == Type::AryPtr) {
1083 const Type *t = tp->is_aryptr()->elem();
1084 // Don't do this for integer types. There is only potential profit if
1085 // the element type t is lower than _type; that is, for int types, if _type is
1086 // more restrictive than t. This only happens here if one is short and the other
1087 // char (both 16 bits), and in those cases we've made an intentional decision
1088 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1089 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1090 //
1091 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1092 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1093 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1094 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1095 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1096 // In fact, that could have been the original type of p1, and p1 could have
1097 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1098 // expression (LShiftL quux 3) independently optimized to the constant 8.
1099 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1100 && Opcode() != Op_LoadKlass) {
1101 // t might actually be lower than _type, if _type is a unique
1102 // concrete subclass of abstract class t.
1103 // Make sure the reference is not into the header, by comparing
1104 // the offset against the offset of the start of the array's data.
1105 // Different array types begin at slightly different offsets (12 vs. 16).
1106 // We choose T_BYTE as an example base type that is least restrictive
1107 // as to alignment, which will therefore produce the smallest
1108 // possible base offset.
1109 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1110 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1111 const Type* jt = t->join(_type);
1112 // In any case, do not allow the join, per se, to empty out the type.
1113 if (jt->empty() && !t->empty()) {
1114 // This can happen if a interface-typed array narrows to a class type.
1115 jt = _type;
1116 }
1118 if (EliminateAutoBox) {
1119 // The pointers in the autobox arrays are always non-null
1120 Node* base = in(Address)->in(AddPNode::Base);
1121 if (base != NULL) {
1122 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1123 if (is_autobox_cache(atp)) {
1124 return jt->join(TypePtr::NOTNULL)->is_ptr();
1125 }
1126 }
1127 }
1128 return jt;
1129 }
1130 }
1131 } else if (tp->base() == Type::InstPtr) {
1132 assert( off != Type::OffsetBot ||
1133 // arrays can be cast to Objects
1134 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1135 // unsafe field access may not have a constant offset
1136 phase->C->has_unsafe_access(),
1137 "Field accesses must be precise" );
1138 // For oop loads, we expect the _type to be precise
1139 } else if (tp->base() == Type::KlassPtr) {
1140 assert( off != Type::OffsetBot ||
1141 // arrays can be cast to Objects
1142 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1143 // also allow array-loading from the primary supertype
1144 // array during subtype checks
1145 Opcode() == Op_LoadKlass,
1146 "Field accesses must be precise" );
1147 // For klass/static loads, we expect the _type to be precise
1148 }
1150 const TypeKlassPtr *tkls = tp->isa_klassptr();
1151 if (tkls != NULL && !StressReflectiveCode) {
1152 ciKlass* klass = tkls->klass();
1153 if (klass->is_loaded() && tkls->klass_is_exact()) {
1154 // We are loading a field from a Klass metaobject whose identity
1155 // is known at compile time (the type is "exact" or "precise").
1156 // Check for fields we know are maintained as constants by the VM.
1157 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1158 // The field is Klass::_super_check_offset. Return its (constant) value.
1159 // (Folds up type checking code.)
1160 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1161 return TypeInt::make(klass->super_check_offset());
1162 }
1163 // Compute index into primary_supers array
1164 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1165 // Check for overflowing; use unsigned compare to handle the negative case.
1166 if( depth < ciKlass::primary_super_limit() ) {
1167 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1168 // (Folds up type checking code.)
1169 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1170 ciKlass *ss = klass->super_of_depth(depth);
1171 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1172 }
1173 const Type* aift = load_array_final_field(tkls, klass);
1174 if (aift != NULL) return aift;
1175 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1176 && klass->is_array_klass()) {
1177 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1178 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1179 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1180 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1181 }
1182 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1183 // The field is Klass::_java_mirror. Return its (constant) value.
1184 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1185 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1186 return TypeInstPtr::make(klass->java_mirror());
1187 }
1188 }
1190 // We can still check if we are loading from the primary_supers array at a
1191 // shallow enough depth. Even though the klass is not exact, entries less
1192 // than or equal to its super depth are correct.
1193 if (klass->is_loaded() ) {
1194 ciType *inner = klass->klass();
1195 while( inner->is_obj_array_klass() )
1196 inner = inner->as_obj_array_klass()->base_element_type();
1197 if( inner->is_instance_klass() &&
1198 !inner->as_instance_klass()->flags().is_interface() ) {
1199 // Compute index into primary_supers array
1200 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1201 // Check for overflowing; use unsigned compare to handle the negative case.
1202 if( depth < ciKlass::primary_super_limit() &&
1203 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1204 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1205 // (Folds up type checking code.)
1206 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1207 ciKlass *ss = klass->super_of_depth(depth);
1208 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1209 }
1210 }
1211 }
1213 // If the type is enough to determine that the thing is not an array,
1214 // we can give the layout_helper a positive interval type.
1215 // This will help short-circuit some reflective code.
1216 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1217 && !klass->is_array_klass() // not directly typed as an array
1218 && !klass->is_interface() // specifically not Serializable & Cloneable
1219 && !klass->is_java_lang_Object() // not the supertype of all T[]
1220 ) {
1221 // Note: When interfaces are reliable, we can narrow the interface
1222 // test to (klass != Serializable && klass != Cloneable).
1223 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1224 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1225 // The key property of this type is that it folds up tests
1226 // for array-ness, since it proves that the layout_helper is positive.
1227 // Thus, a generic value like the basic object layout helper works fine.
1228 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1229 }
1230 }
1232 // If we are loading from a freshly-allocated object, produce a zero,
1233 // if the load is provably beyond the header of the object.
1234 // (Also allow a variable load from a fresh array to produce zero.)
1235 if (ReduceFieldZeroing) {
1236 Node* value = can_see_stored_value(mem,phase);
1237 if (value != NULL && value->is_Con())
1238 return value->bottom_type();
1239 }
1241 const TypeOopPtr *tinst = tp->isa_oopptr();
1242 if (tinst != NULL && tinst->is_instance_field()) {
1243 // If we have an instance type and our memory input is the
1244 // programs's initial memory state, there is no matching store,
1245 // so just return a zero of the appropriate type
1246 Node *mem = in(MemNode::Memory);
1247 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1248 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1249 return Type::get_zero_type(_type->basic_type());
1250 }
1251 }
1252 return _type;
1253 }
1255 //------------------------------match_edge-------------------------------------
1256 // Do we Match on this edge index or not? Match only the address.
1257 uint LoadNode::match_edge(uint idx) const {
1258 return idx == MemNode::Address;
1259 }
1261 //--------------------------LoadBNode::Ideal--------------------------------------
1262 //
1263 // If the previous store is to the same address as this load,
1264 // and the value stored was larger than a byte, replace this load
1265 // with the value stored truncated to a byte. If no truncation is
1266 // needed, the replacement is done in LoadNode::Identity().
1267 //
1268 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1269 Node* mem = in(MemNode::Memory);
1270 Node* value = can_see_stored_value(mem,phase);
1271 if( value && !phase->type(value)->higher_equal( _type ) ) {
1272 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1273 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1274 }
1275 // Identity call will handle the case where truncation is not needed.
1276 return LoadNode::Ideal(phase, can_reshape);
1277 }
1279 //--------------------------LoadCNode::Ideal--------------------------------------
1280 //
1281 // If the previous store is to the same address as this load,
1282 // and the value stored was larger than a char, replace this load
1283 // with the value stored truncated to a char. If no truncation is
1284 // needed, the replacement is done in LoadNode::Identity().
1285 //
1286 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1287 Node* mem = in(MemNode::Memory);
1288 Node* value = can_see_stored_value(mem,phase);
1289 if( value && !phase->type(value)->higher_equal( _type ) )
1290 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1291 // Identity call will handle the case where truncation is not needed.
1292 return LoadNode::Ideal(phase, can_reshape);
1293 }
1295 //--------------------------LoadSNode::Ideal--------------------------------------
1296 //
1297 // If the previous store is to the same address as this load,
1298 // and the value stored was larger than a short, replace this load
1299 // with the value stored truncated to a short. If no truncation is
1300 // needed, the replacement is done in LoadNode::Identity().
1301 //
1302 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1303 Node* mem = in(MemNode::Memory);
1304 Node* value = can_see_stored_value(mem,phase);
1305 if( value && !phase->type(value)->higher_equal( _type ) ) {
1306 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1307 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1308 }
1309 // Identity call will handle the case where truncation is not needed.
1310 return LoadNode::Ideal(phase, can_reshape);
1311 }
1313 //=============================================================================
1314 //------------------------------Value------------------------------------------
1315 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1316 // Either input is TOP ==> the result is TOP
1317 const Type *t1 = phase->type( in(MemNode::Memory) );
1318 if (t1 == Type::TOP) return Type::TOP;
1319 Node *adr = in(MemNode::Address);
1320 const Type *t2 = phase->type( adr );
1321 if (t2 == Type::TOP) return Type::TOP;
1322 const TypePtr *tp = t2->is_ptr();
1323 if (TypePtr::above_centerline(tp->ptr()) ||
1324 tp->ptr() == TypePtr::Null) return Type::TOP;
1326 // Return a more precise klass, if possible
1327 const TypeInstPtr *tinst = tp->isa_instptr();
1328 if (tinst != NULL) {
1329 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1330 int offset = tinst->offset();
1331 if (ik == phase->C->env()->Class_klass()
1332 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1333 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1334 // We are loading a special hidden field from a Class mirror object,
1335 // the field which points to the VM's Klass metaobject.
1336 ciType* t = tinst->java_mirror_type();
1337 // java_mirror_type returns non-null for compile-time Class constants.
1338 if (t != NULL) {
1339 // constant oop => constant klass
1340 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1341 return TypeKlassPtr::make(ciArrayKlass::make(t));
1342 }
1343 if (!t->is_klass()) {
1344 // a primitive Class (e.g., int.class) has NULL for a klass field
1345 return TypePtr::NULL_PTR;
1346 }
1347 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1348 return TypeKlassPtr::make(t->as_klass());
1349 }
1350 // non-constant mirror, so we can't tell what's going on
1351 }
1352 if( !ik->is_loaded() )
1353 return _type; // Bail out if not loaded
1354 if (offset == oopDesc::klass_offset_in_bytes()) {
1355 if (tinst->klass_is_exact()) {
1356 return TypeKlassPtr::make(ik);
1357 }
1358 // See if we can become precise: no subklasses and no interface
1359 // (Note: We need to support verified interfaces.)
1360 if (!ik->is_interface() && !ik->has_subklass()) {
1361 //assert(!UseExactTypes, "this code should be useless with exact types");
1362 // Add a dependence; if any subclass added we need to recompile
1363 if (!ik->is_final()) {
1364 // %%% should use stronger assert_unique_concrete_subtype instead
1365 phase->C->dependencies()->assert_leaf_type(ik);
1366 }
1367 // Return precise klass
1368 return TypeKlassPtr::make(ik);
1369 }
1371 // Return root of possible klass
1372 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1373 }
1374 }
1376 // Check for loading klass from an array
1377 const TypeAryPtr *tary = tp->isa_aryptr();
1378 if( tary != NULL ) {
1379 ciKlass *tary_klass = tary->klass();
1380 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1381 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1382 if (tary->klass_is_exact()) {
1383 return TypeKlassPtr::make(tary_klass);
1384 }
1385 ciArrayKlass *ak = tary->klass()->as_array_klass();
1386 // If the klass is an object array, we defer the question to the
1387 // array component klass.
1388 if( ak->is_obj_array_klass() ) {
1389 assert( ak->is_loaded(), "" );
1390 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1391 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1392 ciInstanceKlass* ik = base_k->as_instance_klass();
1393 // See if we can become precise: no subklasses and no interface
1394 if (!ik->is_interface() && !ik->has_subklass()) {
1395 //assert(!UseExactTypes, "this code should be useless with exact types");
1396 // Add a dependence; if any subclass added we need to recompile
1397 if (!ik->is_final()) {
1398 phase->C->dependencies()->assert_leaf_type(ik);
1399 }
1400 // Return precise array klass
1401 return TypeKlassPtr::make(ak);
1402 }
1403 }
1404 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1405 } else { // Found a type-array?
1406 //assert(!UseExactTypes, "this code should be useless with exact types");
1407 assert( ak->is_type_array_klass(), "" );
1408 return TypeKlassPtr::make(ak); // These are always precise
1409 }
1410 }
1411 }
1413 // Check for loading klass from an array klass
1414 const TypeKlassPtr *tkls = tp->isa_klassptr();
1415 if (tkls != NULL && !StressReflectiveCode) {
1416 ciKlass* klass = tkls->klass();
1417 if( !klass->is_loaded() )
1418 return _type; // Bail out if not loaded
1419 if( klass->is_obj_array_klass() &&
1420 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1421 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1422 // // Always returning precise element type is incorrect,
1423 // // e.g., element type could be object and array may contain strings
1424 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1426 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1427 // according to the element type's subclassing.
1428 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1429 }
1430 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1431 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1432 ciKlass* sup = klass->as_instance_klass()->super();
1433 // The field is Klass::_super. Return its (constant) value.
1434 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1435 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1436 }
1437 }
1439 // Bailout case
1440 return LoadNode::Value(phase);
1441 }
1443 //------------------------------Identity---------------------------------------
1444 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1445 // Also feed through the klass in Allocate(...klass...)._klass.
1446 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1447 Node* x = LoadNode::Identity(phase);
1448 if (x != this) return x;
1450 // Take apart the address into an oop and and offset.
1451 // Return 'this' if we cannot.
1452 Node* adr = in(MemNode::Address);
1453 intptr_t offset = 0;
1454 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1455 if (base == NULL) return this;
1456 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1457 if (toop == NULL) return this;
1459 // We can fetch the klass directly through an AllocateNode.
1460 // This works even if the klass is not constant (clone or newArray).
1461 if (offset == oopDesc::klass_offset_in_bytes()) {
1462 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1463 if (allocated_klass != NULL) {
1464 return allocated_klass;
1465 }
1466 }
1468 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1469 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1470 // See inline_native_Class_query for occurrences of these patterns.
1471 // Java Example: x.getClass().isAssignableFrom(y)
1472 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1473 //
1474 // This improves reflective code, often making the Class
1475 // mirror go completely dead. (Current exception: Class
1476 // mirrors may appear in debug info, but we could clean them out by
1477 // introducing a new debug info operator for klassOop.java_mirror).
1478 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1479 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1480 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1481 // We are loading a special hidden field from a Class mirror,
1482 // the field which points to its Klass or arrayKlass metaobject.
1483 if (base->is_Load()) {
1484 Node* adr2 = base->in(MemNode::Address);
1485 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1486 if (tkls != NULL && !tkls->empty()
1487 && (tkls->klass()->is_instance_klass() ||
1488 tkls->klass()->is_array_klass())
1489 && adr2->is_AddP()
1490 ) {
1491 int mirror_field = Klass::java_mirror_offset_in_bytes();
1492 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1493 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1494 }
1495 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1496 return adr2->in(AddPNode::Base);
1497 }
1498 }
1499 }
1500 }
1502 return this;
1503 }
1505 //------------------------------Value-----------------------------------------
1506 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1507 // Either input is TOP ==> the result is TOP
1508 const Type *t1 = phase->type( in(MemNode::Memory) );
1509 if( t1 == Type::TOP ) return Type::TOP;
1510 Node *adr = in(MemNode::Address);
1511 const Type *t2 = phase->type( adr );
1512 if( t2 == Type::TOP ) return Type::TOP;
1513 const TypePtr *tp = t2->is_ptr();
1514 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1515 const TypeAryPtr *tap = tp->isa_aryptr();
1516 if( !tap ) return _type;
1517 return tap->size();
1518 }
1520 //------------------------------Identity---------------------------------------
1521 // Feed through the length in AllocateArray(...length...)._length.
1522 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1523 Node* x = LoadINode::Identity(phase);
1524 if (x != this) return x;
1526 // Take apart the address into an oop and and offset.
1527 // Return 'this' if we cannot.
1528 Node* adr = in(MemNode::Address);
1529 intptr_t offset = 0;
1530 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1531 if (base == NULL) return this;
1532 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1533 if (tary == NULL) return this;
1535 // We can fetch the length directly through an AllocateArrayNode.
1536 // This works even if the length is not constant (clone or newArray).
1537 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1538 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1539 if (allocated_length != NULL) {
1540 return allocated_length;
1541 }
1542 }
1544 return this;
1546 }
1547 //=============================================================================
1548 //---------------------------StoreNode::make-----------------------------------
1549 // Polymorphic factory method:
1550 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1551 switch (bt) {
1552 case T_BOOLEAN:
1553 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1554 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1555 case T_CHAR:
1556 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1557 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1558 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1559 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1560 case T_ADDRESS:
1561 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1562 }
1563 ShouldNotReachHere();
1564 return (StoreNode*)NULL;
1565 }
1567 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1568 bool require_atomic = true;
1569 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1570 }
1573 //--------------------------bottom_type----------------------------------------
1574 const Type *StoreNode::bottom_type() const {
1575 return Type::MEMORY;
1576 }
1578 //------------------------------hash-------------------------------------------
1579 uint StoreNode::hash() const {
1580 // unroll addition of interesting fields
1581 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1583 // Since they are not commoned, do not hash them:
1584 return NO_HASH;
1585 }
1587 //------------------------------Ideal------------------------------------------
1588 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1589 // When a store immediately follows a relevant allocation/initialization,
1590 // try to capture it into the initialization, or hoist it above.
1591 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1592 Node* p = MemNode::Ideal_common(phase, can_reshape);
1593 if (p) return (p == NodeSentinel) ? NULL : p;
1595 Node* mem = in(MemNode::Memory);
1596 Node* address = in(MemNode::Address);
1598 // Back-to-back stores to same address? Fold em up.
1599 // Generally unsafe if I have intervening uses...
1600 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1601 // Looking at a dead closed cycle of memory?
1602 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1604 assert(Opcode() == mem->Opcode() ||
1605 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1606 "no mismatched stores, except on raw memory");
1608 if (mem->outcnt() == 1 && // check for intervening uses
1609 mem->as_Store()->memory_size() <= this->memory_size()) {
1610 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1611 // For example, 'mem' might be the final state at a conditional return.
1612 // Or, 'mem' might be used by some node which is live at the same time
1613 // 'this' is live, which might be unschedulable. So, require exactly
1614 // ONE user, the 'this' store, until such time as we clone 'mem' for
1615 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1616 if (can_reshape) { // (%%% is this an anachronism?)
1617 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1618 phase->is_IterGVN());
1619 } else {
1620 // It's OK to do this in the parser, since DU info is always accurate,
1621 // and the parser always refers to nodes via SafePointNode maps.
1622 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1623 }
1624 return this;
1625 }
1626 }
1628 // Capture an unaliased, unconditional, simple store into an initializer.
1629 // Or, if it is independent of the allocation, hoist it above the allocation.
1630 if (ReduceFieldZeroing && /*can_reshape &&*/
1631 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1632 InitializeNode* init = mem->in(0)->as_Initialize();
1633 intptr_t offset = init->can_capture_store(this, phase);
1634 if (offset > 0) {
1635 Node* moved = init->capture_store(this, offset, phase);
1636 // If the InitializeNode captured me, it made a raw copy of me,
1637 // and I need to disappear.
1638 if (moved != NULL) {
1639 // %%% hack to ensure that Ideal returns a new node:
1640 mem = MergeMemNode::make(phase->C, mem);
1641 return mem; // fold me away
1642 }
1643 }
1644 }
1646 return NULL; // No further progress
1647 }
1649 //------------------------------Value-----------------------------------------
1650 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1651 // Either input is TOP ==> the result is TOP
1652 const Type *t1 = phase->type( in(MemNode::Memory) );
1653 if( t1 == Type::TOP ) return Type::TOP;
1654 const Type *t2 = phase->type( in(MemNode::Address) );
1655 if( t2 == Type::TOP ) return Type::TOP;
1656 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1657 if( t3 == Type::TOP ) return Type::TOP;
1658 return Type::MEMORY;
1659 }
1661 //------------------------------Identity---------------------------------------
1662 // Remove redundant stores:
1663 // Store(m, p, Load(m, p)) changes to m.
1664 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1665 Node *StoreNode::Identity( PhaseTransform *phase ) {
1666 Node* mem = in(MemNode::Memory);
1667 Node* adr = in(MemNode::Address);
1668 Node* val = in(MemNode::ValueIn);
1670 // Load then Store? Then the Store is useless
1671 if (val->is_Load() &&
1672 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1673 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1674 val->as_Load()->store_Opcode() == Opcode()) {
1675 return mem;
1676 }
1678 // Two stores in a row of the same value?
1679 if (mem->is_Store() &&
1680 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1681 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1682 mem->Opcode() == Opcode()) {
1683 return mem;
1684 }
1686 // Store of zero anywhere into a freshly-allocated object?
1687 // Then the store is useless.
1688 // (It must already have been captured by the InitializeNode.)
1689 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1690 // a newly allocated object is already all-zeroes everywhere
1691 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1692 return mem;
1693 }
1695 // the store may also apply to zero-bits in an earlier object
1696 Node* prev_mem = find_previous_store(phase);
1697 // Steps (a), (b): Walk past independent stores to find an exact match.
1698 if (prev_mem != NULL) {
1699 Node* prev_val = can_see_stored_value(prev_mem, phase);
1700 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1701 // prev_val and val might differ by a cast; it would be good
1702 // to keep the more informative of the two.
1703 return mem;
1704 }
1705 }
1706 }
1708 return this;
1709 }
1711 //------------------------------match_edge-------------------------------------
1712 // Do we Match on this edge index or not? Match only memory & value
1713 uint StoreNode::match_edge(uint idx) const {
1714 return idx == MemNode::Address || idx == MemNode::ValueIn;
1715 }
1717 //------------------------------cmp--------------------------------------------
1718 // Do not common stores up together. They generally have to be split
1719 // back up anyways, so do not bother.
1720 uint StoreNode::cmp( const Node &n ) const {
1721 return (&n == this); // Always fail except on self
1722 }
1724 //------------------------------Ideal_masked_input-----------------------------
1725 // Check for a useless mask before a partial-word store
1726 // (StoreB ... (AndI valIn conIa) )
1727 // If (conIa & mask == mask) this simplifies to
1728 // (StoreB ... (valIn) )
1729 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1730 Node *val = in(MemNode::ValueIn);
1731 if( val->Opcode() == Op_AndI ) {
1732 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1733 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1734 set_req(MemNode::ValueIn, val->in(1));
1735 return this;
1736 }
1737 }
1738 return NULL;
1739 }
1742 //------------------------------Ideal_sign_extended_input----------------------
1743 // Check for useless sign-extension before a partial-word store
1744 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1745 // If (conIL == conIR && conIR <= num_bits) this simplifies to
1746 // (StoreB ... (valIn) )
1747 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1748 Node *val = in(MemNode::ValueIn);
1749 if( val->Opcode() == Op_RShiftI ) {
1750 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1751 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
1752 Node *shl = val->in(1);
1753 if( shl->Opcode() == Op_LShiftI ) {
1754 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
1755 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
1756 set_req(MemNode::ValueIn, shl->in(1));
1757 return this;
1758 }
1759 }
1760 }
1761 }
1762 return NULL;
1763 }
1765 //------------------------------value_never_loaded-----------------------------------
1766 // Determine whether there are any possible loads of the value stored.
1767 // For simplicity, we actually check if there are any loads from the
1768 // address stored to, not just for loads of the value stored by this node.
1769 //
1770 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
1771 Node *adr = in(Address);
1772 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
1773 if (adr_oop == NULL)
1774 return false;
1775 if (!adr_oop->is_instance_field())
1776 return false; // if not a distinct instance, there may be aliases of the address
1777 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
1778 Node *use = adr->fast_out(i);
1779 int opc = use->Opcode();
1780 if (use->is_Load() || use->is_LoadStore()) {
1781 return false;
1782 }
1783 }
1784 return true;
1785 }
1787 //=============================================================================
1788 //------------------------------Ideal------------------------------------------
1789 // If the store is from an AND mask that leaves the low bits untouched, then
1790 // we can skip the AND operation. If the store is from a sign-extension
1791 // (a left shift, then right shift) we can skip both.
1792 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
1793 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
1794 if( progress != NULL ) return progress;
1796 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
1797 if( progress != NULL ) return progress;
1799 // Finally check the default case
1800 return StoreNode::Ideal(phase, can_reshape);
1801 }
1803 //=============================================================================
1804 //------------------------------Ideal------------------------------------------
1805 // If the store is from an AND mask that leaves the low bits untouched, then
1806 // we can skip the AND operation
1807 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
1808 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
1809 if( progress != NULL ) return progress;
1811 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
1812 if( progress != NULL ) return progress;
1814 // Finally check the default case
1815 return StoreNode::Ideal(phase, can_reshape);
1816 }
1818 //=============================================================================
1819 //------------------------------Identity---------------------------------------
1820 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
1821 // No need to card mark when storing a null ptr
1822 Node* my_store = in(MemNode::OopStore);
1823 if (my_store->is_Store()) {
1824 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
1825 if( t1 == TypePtr::NULL_PTR ) {
1826 return in(MemNode::Memory);
1827 }
1828 }
1829 return this;
1830 }
1832 //------------------------------Value-----------------------------------------
1833 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
1834 // Either input is TOP ==> the result is TOP
1835 const Type *t = phase->type( in(MemNode::Memory) );
1836 if( t == Type::TOP ) return Type::TOP;
1837 t = phase->type( in(MemNode::Address) );
1838 if( t == Type::TOP ) return Type::TOP;
1839 t = phase->type( in(MemNode::ValueIn) );
1840 if( t == Type::TOP ) return Type::TOP;
1841 // If extra input is TOP ==> the result is TOP
1842 t = phase->type( in(MemNode::OopStore) );
1843 if( t == Type::TOP ) return Type::TOP;
1845 return StoreNode::Value( phase );
1846 }
1849 //=============================================================================
1850 //----------------------------------SCMemProjNode------------------------------
1851 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
1852 {
1853 return bottom_type();
1854 }
1856 //=============================================================================
1857 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
1858 init_req(MemNode::Control, c );
1859 init_req(MemNode::Memory , mem);
1860 init_req(MemNode::Address, adr);
1861 init_req(MemNode::ValueIn, val);
1862 init_req( ExpectedIn, ex );
1863 init_class_id(Class_LoadStore);
1865 }
1867 //=============================================================================
1868 //-------------------------------adr_type--------------------------------------
1869 // Do we Match on this edge index or not? Do not match memory
1870 const TypePtr* ClearArrayNode::adr_type() const {
1871 Node *adr = in(3);
1872 return MemNode::calculate_adr_type(adr->bottom_type());
1873 }
1875 //------------------------------match_edge-------------------------------------
1876 // Do we Match on this edge index or not? Do not match memory
1877 uint ClearArrayNode::match_edge(uint idx) const {
1878 return idx > 1;
1879 }
1881 //------------------------------Identity---------------------------------------
1882 // Clearing a zero length array does nothing
1883 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
1884 return phase->type(in(2))->higher_equal(TypeInt::ZERO) ? in(1) : this;
1885 }
1887 //------------------------------Idealize---------------------------------------
1888 // Clearing a short array is faster with stores
1889 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
1890 const int unit = BytesPerLong;
1891 const TypeX* t = phase->type(in(2))->isa_intptr_t();
1892 if (!t) return NULL;
1893 if (!t->is_con()) return NULL;
1894 intptr_t raw_count = t->get_con();
1895 intptr_t size = raw_count;
1896 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
1897 // Clearing nothing uses the Identity call.
1898 // Negative clears are possible on dead ClearArrays
1899 // (see jck test stmt114.stmt11402.val).
1900 if (size <= 0 || size % unit != 0) return NULL;
1901 intptr_t count = size / unit;
1902 // Length too long; use fast hardware clear
1903 if (size > Matcher::init_array_short_size) return NULL;
1904 Node *mem = in(1);
1905 if( phase->type(mem)==Type::TOP ) return NULL;
1906 Node *adr = in(3);
1907 const Type* at = phase->type(adr);
1908 if( at==Type::TOP ) return NULL;
1909 const TypePtr* atp = at->isa_ptr();
1910 // adjust atp to be the correct array element address type
1911 if (atp == NULL) atp = TypePtr::BOTTOM;
1912 else atp = atp->add_offset(Type::OffsetBot);
1913 // Get base for derived pointer purposes
1914 if( adr->Opcode() != Op_AddP ) Unimplemented();
1915 Node *base = adr->in(1);
1917 Node *zero = phase->makecon(TypeLong::ZERO);
1918 Node *off = phase->MakeConX(BytesPerLong);
1919 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1920 count--;
1921 while( count-- ) {
1922 mem = phase->transform(mem);
1923 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
1924 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
1925 }
1926 return mem;
1927 }
1929 //----------------------------clear_memory-------------------------------------
1930 // Generate code to initialize object storage to zero.
1931 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1932 intptr_t start_offset,
1933 Node* end_offset,
1934 PhaseGVN* phase) {
1935 Compile* C = phase->C;
1936 intptr_t offset = start_offset;
1938 int unit = BytesPerLong;
1939 if ((offset % unit) != 0) {
1940 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
1941 adr = phase->transform(adr);
1942 const TypePtr* atp = TypeRawPtr::BOTTOM;
1943 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1944 mem = phase->transform(mem);
1945 offset += BytesPerInt;
1946 }
1947 assert((offset % unit) == 0, "");
1949 // Initialize the remaining stuff, if any, with a ClearArray.
1950 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
1951 }
1953 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1954 Node* start_offset,
1955 Node* end_offset,
1956 PhaseGVN* phase) {
1957 Compile* C = phase->C;
1958 int unit = BytesPerLong;
1959 Node* zbase = start_offset;
1960 Node* zend = end_offset;
1962 // Scale to the unit required by the CPU:
1963 if (!Matcher::init_array_count_is_in_bytes) {
1964 Node* shift = phase->intcon(exact_log2(unit));
1965 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
1966 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
1967 }
1969 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
1970 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
1972 // Bulk clear double-words
1973 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
1974 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
1975 return phase->transform(mem);
1976 }
1978 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1979 intptr_t start_offset,
1980 intptr_t end_offset,
1981 PhaseGVN* phase) {
1982 Compile* C = phase->C;
1983 assert((end_offset % BytesPerInt) == 0, "odd end offset");
1984 intptr_t done_offset = end_offset;
1985 if ((done_offset % BytesPerLong) != 0) {
1986 done_offset -= BytesPerInt;
1987 }
1988 if (done_offset > start_offset) {
1989 mem = clear_memory(ctl, mem, dest,
1990 start_offset, phase->MakeConX(done_offset), phase);
1991 }
1992 if (done_offset < end_offset) { // emit the final 32-bit store
1993 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
1994 adr = phase->transform(adr);
1995 const TypePtr* atp = TypeRawPtr::BOTTOM;
1996 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
1997 mem = phase->transform(mem);
1998 done_offset += BytesPerInt;
1999 }
2000 assert(done_offset == end_offset, "");
2001 return mem;
2002 }
2004 //=============================================================================
2005 // Do we match on this edge? No memory edges
2006 uint StrCompNode::match_edge(uint idx) const {
2007 return idx == 5 || idx == 6;
2008 }
2010 //------------------------------Ideal------------------------------------------
2011 // Return a node which is more "ideal" than the current node. Strip out
2012 // control copies
2013 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2014 return remove_dead_region(phase, can_reshape) ? this : NULL;
2015 }
2018 //=============================================================================
2019 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2020 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2021 _adr_type(C->get_adr_type(alias_idx))
2022 {
2023 init_class_id(Class_MemBar);
2024 Node* top = C->top();
2025 init_req(TypeFunc::I_O,top);
2026 init_req(TypeFunc::FramePtr,top);
2027 init_req(TypeFunc::ReturnAdr,top);
2028 if (precedent != NULL)
2029 init_req(TypeFunc::Parms, precedent);
2030 }
2032 //------------------------------cmp--------------------------------------------
2033 uint MemBarNode::hash() const { return NO_HASH; }
2034 uint MemBarNode::cmp( const Node &n ) const {
2035 return (&n == this); // Always fail except on self
2036 }
2038 //------------------------------make-------------------------------------------
2039 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2040 int len = Precedent + (pn == NULL? 0: 1);
2041 switch (opcode) {
2042 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2043 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2044 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2045 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2046 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2047 default: ShouldNotReachHere(); return NULL;
2048 }
2049 }
2051 //------------------------------Ideal------------------------------------------
2052 // Return a node which is more "ideal" than the current node. Strip out
2053 // control copies
2054 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2055 if (remove_dead_region(phase, can_reshape)) return this;
2056 return NULL;
2057 }
2059 //------------------------------Value------------------------------------------
2060 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2061 if( !in(0) ) return Type::TOP;
2062 if( phase->type(in(0)) == Type::TOP )
2063 return Type::TOP;
2064 return TypeTuple::MEMBAR;
2065 }
2067 //------------------------------match------------------------------------------
2068 // Construct projections for memory.
2069 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2070 switch (proj->_con) {
2071 case TypeFunc::Control:
2072 case TypeFunc::Memory:
2073 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2074 }
2075 ShouldNotReachHere();
2076 return NULL;
2077 }
2079 //===========================InitializeNode====================================
2080 // SUMMARY:
2081 // This node acts as a memory barrier on raw memory, after some raw stores.
2082 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2083 // The Initialize can 'capture' suitably constrained stores as raw inits.
2084 // It can coalesce related raw stores into larger units (called 'tiles').
2085 // It can avoid zeroing new storage for memory units which have raw inits.
2086 // At macro-expansion, it is marked 'complete', and does not optimize further.
2087 //
2088 // EXAMPLE:
2089 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2090 // ctl = incoming control; mem* = incoming memory
2091 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2092 // First allocate uninitialized memory and fill in the header:
2093 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2094 // ctl := alloc.Control; mem* := alloc.Memory*
2095 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2096 // Then initialize to zero the non-header parts of the raw memory block:
2097 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2098 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2099 // After the initialize node executes, the object is ready for service:
2100 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2101 // Suppose its body is immediately initialized as {1,2}:
2102 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2103 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2104 // mem.SLICE(#short[*]) := store2
2105 //
2106 // DETAILS:
2107 // An InitializeNode collects and isolates object initialization after
2108 // an AllocateNode and before the next possible safepoint. As a
2109 // memory barrier (MemBarNode), it keeps critical stores from drifting
2110 // down past any safepoint or any publication of the allocation.
2111 // Before this barrier, a newly-allocated object may have uninitialized bits.
2112 // After this barrier, it may be treated as a real oop, and GC is allowed.
2113 //
2114 // The semantics of the InitializeNode include an implicit zeroing of
2115 // the new object from object header to the end of the object.
2116 // (The object header and end are determined by the AllocateNode.)
2117 //
2118 // Certain stores may be added as direct inputs to the InitializeNode.
2119 // These stores must update raw memory, and they must be to addresses
2120 // derived from the raw address produced by AllocateNode, and with
2121 // a constant offset. They must be ordered by increasing offset.
2122 // The first one is at in(RawStores), the last at in(req()-1).
2123 // Unlike most memory operations, they are not linked in a chain,
2124 // but are displayed in parallel as users of the rawmem output of
2125 // the allocation.
2126 //
2127 // (See comments in InitializeNode::capture_store, which continue
2128 // the example given above.)
2129 //
2130 // When the associated Allocate is macro-expanded, the InitializeNode
2131 // may be rewritten to optimize collected stores. A ClearArrayNode
2132 // may also be created at that point to represent any required zeroing.
2133 // The InitializeNode is then marked 'complete', prohibiting further
2134 // capturing of nearby memory operations.
2135 //
2136 // During macro-expansion, all captured initializations which store
2137 // constant values of 32 bits or smaller are coalesced (if advantagous)
2138 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2139 // initialized in fewer memory operations. Memory words which are
2140 // covered by neither tiles nor non-constant stores are pre-zeroed
2141 // by explicit stores of zero. (The code shape happens to do all
2142 // zeroing first, then all other stores, with both sequences occurring
2143 // in order of ascending offsets.)
2144 //
2145 // Alternatively, code may be inserted between an AllocateNode and its
2146 // InitializeNode, to perform arbitrary initialization of the new object.
2147 // E.g., the object copying intrinsics insert complex data transfers here.
2148 // The initialization must then be marked as 'complete' disable the
2149 // built-in zeroing semantics and the collection of initializing stores.
2150 //
2151 // While an InitializeNode is incomplete, reads from the memory state
2152 // produced by it are optimizable if they match the control edge and
2153 // new oop address associated with the allocation/initialization.
2154 // They return a stored value (if the offset matches) or else zero.
2155 // A write to the memory state, if it matches control and address,
2156 // and if it is to a constant offset, may be 'captured' by the
2157 // InitializeNode. It is cloned as a raw memory operation and rewired
2158 // inside the initialization, to the raw oop produced by the allocation.
2159 // Operations on addresses which are provably distinct (e.g., to
2160 // other AllocateNodes) are allowed to bypass the initialization.
2161 //
2162 // The effect of all this is to consolidate object initialization
2163 // (both arrays and non-arrays, both piecewise and bulk) into a
2164 // single location, where it can be optimized as a unit.
2165 //
2166 // Only stores with an offset less than TrackedInitializationLimit words
2167 // will be considered for capture by an InitializeNode. This puts a
2168 // reasonable limit on the complexity of optimized initializations.
2170 //---------------------------InitializeNode------------------------------------
2171 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2172 : _is_complete(false),
2173 MemBarNode(C, adr_type, rawoop)
2174 {
2175 init_class_id(Class_Initialize);
2177 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2178 assert(in(RawAddress) == rawoop, "proper init");
2179 // Note: allocation() can be NULL, for secondary initialization barriers
2180 }
2182 // Since this node is not matched, it will be processed by the
2183 // register allocator. Declare that there are no constraints
2184 // on the allocation of the RawAddress edge.
2185 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2186 // This edge should be set to top, by the set_complete. But be conservative.
2187 if (idx == InitializeNode::RawAddress)
2188 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2189 return RegMask::Empty;
2190 }
2192 Node* InitializeNode::memory(uint alias_idx) {
2193 Node* mem = in(Memory);
2194 if (mem->is_MergeMem()) {
2195 return mem->as_MergeMem()->memory_at(alias_idx);
2196 } else {
2197 // incoming raw memory is not split
2198 return mem;
2199 }
2200 }
2202 bool InitializeNode::is_non_zero() {
2203 if (is_complete()) return false;
2204 remove_extra_zeroes();
2205 return (req() > RawStores);
2206 }
2208 void InitializeNode::set_complete(PhaseGVN* phase) {
2209 assert(!is_complete(), "caller responsibility");
2210 _is_complete = true;
2212 // After this node is complete, it contains a bunch of
2213 // raw-memory initializations. There is no need for
2214 // it to have anything to do with non-raw memory effects.
2215 // Therefore, tell all non-raw users to re-optimize themselves,
2216 // after skipping the memory effects of this initialization.
2217 PhaseIterGVN* igvn = phase->is_IterGVN();
2218 if (igvn) igvn->add_users_to_worklist(this);
2219 }
2221 // convenience function
2222 // return false if the init contains any stores already
2223 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2224 InitializeNode* init = initialization();
2225 if (init == NULL || init->is_complete()) return false;
2226 init->remove_extra_zeroes();
2227 // for now, if this allocation has already collected any inits, bail:
2228 if (init->is_non_zero()) return false;
2229 init->set_complete(phase);
2230 return true;
2231 }
2233 void InitializeNode::remove_extra_zeroes() {
2234 if (req() == RawStores) return;
2235 Node* zmem = zero_memory();
2236 uint fill = RawStores;
2237 for (uint i = fill; i < req(); i++) {
2238 Node* n = in(i);
2239 if (n->is_top() || n == zmem) continue; // skip
2240 if (fill < i) set_req(fill, n); // compact
2241 ++fill;
2242 }
2243 // delete any empty spaces created:
2244 while (fill < req()) {
2245 del_req(fill);
2246 }
2247 }
2249 // Helper for remembering which stores go with which offsets.
2250 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2251 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2252 intptr_t offset = -1;
2253 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2254 phase, offset);
2255 if (base == NULL) return -1; // something is dead,
2256 if (offset < 0) return -1; // dead, dead
2257 return offset;
2258 }
2260 // Helper for proving that an initialization expression is
2261 // "simple enough" to be folded into an object initialization.
2262 // Attempts to prove that a store's initial value 'n' can be captured
2263 // within the initialization without creating a vicious cycle, such as:
2264 // { Foo p = new Foo(); p.next = p; }
2265 // True for constants and parameters and small combinations thereof.
2266 bool InitializeNode::detect_init_independence(Node* n,
2267 bool st_is_pinned,
2268 int& count) {
2269 if (n == NULL) return true; // (can this really happen?)
2270 if (n->is_Proj()) n = n->in(0);
2271 if (n == this) return false; // found a cycle
2272 if (n->is_Con()) return true;
2273 if (n->is_Start()) return true; // params, etc., are OK
2274 if (n->is_Root()) return true; // even better
2276 Node* ctl = n->in(0);
2277 if (ctl != NULL && !ctl->is_top()) {
2278 if (ctl->is_Proj()) ctl = ctl->in(0);
2279 if (ctl == this) return false;
2281 // If we already know that the enclosing memory op is pinned right after
2282 // the init, then any control flow that the store has picked up
2283 // must have preceded the init, or else be equal to the init.
2284 // Even after loop optimizations (which might change control edges)
2285 // a store is never pinned *before* the availability of its inputs.
2286 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
2287 return false; // failed to prove a good control
2289 }
2291 // Check data edges for possible dependencies on 'this'.
2292 if ((count += 1) > 20) return false; // complexity limit
2293 for (uint i = 1; i < n->req(); i++) {
2294 Node* m = n->in(i);
2295 if (m == NULL || m == n || m->is_top()) continue;
2296 uint first_i = n->find_edge(m);
2297 if (i != first_i) continue; // process duplicate edge just once
2298 if (!detect_init_independence(m, st_is_pinned, count)) {
2299 return false;
2300 }
2301 }
2303 return true;
2304 }
2306 // Here are all the checks a Store must pass before it can be moved into
2307 // an initialization. Returns zero if a check fails.
2308 // On success, returns the (constant) offset to which the store applies,
2309 // within the initialized memory.
2310 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2311 const int FAIL = 0;
2312 if (st->req() != MemNode::ValueIn + 1)
2313 return FAIL; // an inscrutable StoreNode (card mark?)
2314 Node* ctl = st->in(MemNode::Control);
2315 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2316 return FAIL; // must be unconditional after the initialization
2317 Node* mem = st->in(MemNode::Memory);
2318 if (!(mem->is_Proj() && mem->in(0) == this))
2319 return FAIL; // must not be preceded by other stores
2320 Node* adr = st->in(MemNode::Address);
2321 intptr_t offset;
2322 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2323 if (alloc == NULL)
2324 return FAIL; // inscrutable address
2325 if (alloc != allocation())
2326 return FAIL; // wrong allocation! (store needs to float up)
2327 Node* val = st->in(MemNode::ValueIn);
2328 int complexity_count = 0;
2329 if (!detect_init_independence(val, true, complexity_count))
2330 return FAIL; // stored value must be 'simple enough'
2332 return offset; // success
2333 }
2335 // Find the captured store in(i) which corresponds to the range
2336 // [start..start+size) in the initialized object.
2337 // If there is one, return its index i. If there isn't, return the
2338 // negative of the index where it should be inserted.
2339 // Return 0 if the queried range overlaps an initialization boundary
2340 // or if dead code is encountered.
2341 // If size_in_bytes is zero, do not bother with overlap checks.
2342 int InitializeNode::captured_store_insertion_point(intptr_t start,
2343 int size_in_bytes,
2344 PhaseTransform* phase) {
2345 const int FAIL = 0, MAX_STORE = BytesPerLong;
2347 if (is_complete())
2348 return FAIL; // arraycopy got here first; punt
2350 assert(allocation() != NULL, "must be present");
2352 // no negatives, no header fields:
2353 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
2354 if (start < (intptr_t) sizeof(arrayOopDesc) &&
2355 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2357 // after a certain size, we bail out on tracking all the stores:
2358 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2359 if (start >= ti_limit) return FAIL;
2361 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2362 if (i >= limit) return -(int)i; // not found; here is where to put it
2364 Node* st = in(i);
2365 intptr_t st_off = get_store_offset(st, phase);
2366 if (st_off < 0) {
2367 if (st != zero_memory()) {
2368 return FAIL; // bail out if there is dead garbage
2369 }
2370 } else if (st_off > start) {
2371 // ...we are done, since stores are ordered
2372 if (st_off < start + size_in_bytes) {
2373 return FAIL; // the next store overlaps
2374 }
2375 return -(int)i; // not found; here is where to put it
2376 } else if (st_off < start) {
2377 if (size_in_bytes != 0 &&
2378 start < st_off + MAX_STORE &&
2379 start < st_off + st->as_Store()->memory_size()) {
2380 return FAIL; // the previous store overlaps
2381 }
2382 } else {
2383 if (size_in_bytes != 0 &&
2384 st->as_Store()->memory_size() != size_in_bytes) {
2385 return FAIL; // mismatched store size
2386 }
2387 return i;
2388 }
2390 ++i;
2391 }
2392 }
2394 // Look for a captured store which initializes at the offset 'start'
2395 // with the given size. If there is no such store, and no other
2396 // initialization interferes, then return zero_memory (the memory
2397 // projection of the AllocateNode).
2398 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2399 PhaseTransform* phase) {
2400 assert(stores_are_sane(phase), "");
2401 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2402 if (i == 0) {
2403 return NULL; // something is dead
2404 } else if (i < 0) {
2405 return zero_memory(); // just primordial zero bits here
2406 } else {
2407 Node* st = in(i); // here is the store at this position
2408 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2409 return st;
2410 }
2411 }
2413 // Create, as a raw pointer, an address within my new object at 'offset'.
2414 Node* InitializeNode::make_raw_address(intptr_t offset,
2415 PhaseTransform* phase) {
2416 Node* addr = in(RawAddress);
2417 if (offset != 0) {
2418 Compile* C = phase->C;
2419 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2420 phase->MakeConX(offset)) );
2421 }
2422 return addr;
2423 }
2425 // Clone the given store, converting it into a raw store
2426 // initializing a field or element of my new object.
2427 // Caller is responsible for retiring the original store,
2428 // with subsume_node or the like.
2429 //
2430 // From the example above InitializeNode::InitializeNode,
2431 // here are the old stores to be captured:
2432 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2433 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2434 //
2435 // Here is the changed code; note the extra edges on init:
2436 // alloc = (Allocate ...)
2437 // rawoop = alloc.RawAddress
2438 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2439 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2440 // init = (Initialize alloc.Control alloc.Memory rawoop
2441 // rawstore1 rawstore2)
2442 //
2443 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2444 PhaseTransform* phase) {
2445 assert(stores_are_sane(phase), "");
2447 if (start < 0) return NULL;
2448 assert(can_capture_store(st, phase) == start, "sanity");
2450 Compile* C = phase->C;
2451 int size_in_bytes = st->memory_size();
2452 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2453 if (i == 0) return NULL; // bail out
2454 Node* prev_mem = NULL; // raw memory for the captured store
2455 if (i > 0) {
2456 prev_mem = in(i); // there is a pre-existing store under this one
2457 set_req(i, C->top()); // temporarily disconnect it
2458 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2459 } else {
2460 i = -i; // no pre-existing store
2461 prev_mem = zero_memory(); // a slice of the newly allocated object
2462 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2463 set_req(--i, C->top()); // reuse this edge; it has been folded away
2464 else
2465 ins_req(i, C->top()); // build a new edge
2466 }
2467 Node* new_st = st->clone();
2468 new_st->set_req(MemNode::Control, in(Control));
2469 new_st->set_req(MemNode::Memory, prev_mem);
2470 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2471 new_st = phase->transform(new_st);
2473 // At this point, new_st might have swallowed a pre-existing store
2474 // at the same offset, or perhaps new_st might have disappeared,
2475 // if it redundantly stored the same value (or zero to fresh memory).
2477 // In any case, wire it in:
2478 set_req(i, new_st);
2480 // The caller may now kill the old guy.
2481 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2482 assert(check_st == new_st || check_st == NULL, "must be findable");
2483 assert(!is_complete(), "");
2484 return new_st;
2485 }
2487 static bool store_constant(jlong* tiles, int num_tiles,
2488 intptr_t st_off, int st_size,
2489 jlong con) {
2490 if ((st_off & (st_size-1)) != 0)
2491 return false; // strange store offset (assume size==2**N)
2492 address addr = (address)tiles + st_off;
2493 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2494 switch (st_size) {
2495 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2496 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2497 case sizeof(jint): *(jint*) addr = (jint) con; break;
2498 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2499 default: return false; // strange store size (detect size!=2**N here)
2500 }
2501 return true; // return success to caller
2502 }
2504 // Coalesce subword constants into int constants and possibly
2505 // into long constants. The goal, if the CPU permits,
2506 // is to initialize the object with a small number of 64-bit tiles.
2507 // Also, convert floating-point constants to bit patterns.
2508 // Non-constants are not relevant to this pass.
2509 //
2510 // In terms of the running example on InitializeNode::InitializeNode
2511 // and InitializeNode::capture_store, here is the transformation
2512 // of rawstore1 and rawstore2 into rawstore12:
2513 // alloc = (Allocate ...)
2514 // rawoop = alloc.RawAddress
2515 // tile12 = 0x00010002
2516 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2517 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2518 //
2519 void
2520 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2521 Node* size_in_bytes,
2522 PhaseGVN* phase) {
2523 Compile* C = phase->C;
2525 assert(stores_are_sane(phase), "");
2526 // Note: After this pass, they are not completely sane,
2527 // since there may be some overlaps.
2529 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2531 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2532 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2533 size_limit = MIN2(size_limit, ti_limit);
2534 size_limit = align_size_up(size_limit, BytesPerLong);
2535 int num_tiles = size_limit / BytesPerLong;
2537 // allocate space for the tile map:
2538 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2539 jlong tiles_buf[small_len];
2540 Node* nodes_buf[small_len];
2541 jlong inits_buf[small_len];
2542 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2543 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2544 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2545 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2546 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2547 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2548 // tiles: exact bitwise model of all primitive constants
2549 // nodes: last constant-storing node subsumed into the tiles model
2550 // inits: which bytes (in each tile) are touched by any initializations
2552 //// Pass A: Fill in the tile model with any relevant stores.
2554 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2555 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2556 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2557 Node* zmem = zero_memory(); // initially zero memory state
2558 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2559 Node* st = in(i);
2560 intptr_t st_off = get_store_offset(st, phase);
2562 // Figure out the store's offset and constant value:
2563 if (st_off < header_size) continue; //skip (ignore header)
2564 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2565 int st_size = st->as_Store()->memory_size();
2566 if (st_off + st_size > size_limit) break;
2568 // Record which bytes are touched, whether by constant or not.
2569 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2570 continue; // skip (strange store size)
2572 const Type* val = phase->type(st->in(MemNode::ValueIn));
2573 if (!val->singleton()) continue; //skip (non-con store)
2574 BasicType type = val->basic_type();
2576 jlong con = 0;
2577 switch (type) {
2578 case T_INT: con = val->is_int()->get_con(); break;
2579 case T_LONG: con = val->is_long()->get_con(); break;
2580 case T_FLOAT: con = jint_cast(val->getf()); break;
2581 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2582 default: continue; //skip (odd store type)
2583 }
2585 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2586 st->Opcode() == Op_StoreL) {
2587 continue; // This StoreL is already optimal.
2588 }
2590 // Store down the constant.
2591 store_constant(tiles, num_tiles, st_off, st_size, con);
2593 intptr_t j = st_off >> LogBytesPerLong;
2595 if (type == T_INT && st_size == BytesPerInt
2596 && (st_off & BytesPerInt) == BytesPerInt) {
2597 jlong lcon = tiles[j];
2598 if (!Matcher::isSimpleConstant64(lcon) &&
2599 st->Opcode() == Op_StoreI) {
2600 // This StoreI is already optimal by itself.
2601 jint* intcon = (jint*) &tiles[j];
2602 intcon[1] = 0; // undo the store_constant()
2604 // If the previous store is also optimal by itself, back up and
2605 // undo the action of the previous loop iteration... if we can.
2606 // But if we can't, just let the previous half take care of itself.
2607 st = nodes[j];
2608 st_off -= BytesPerInt;
2609 con = intcon[0];
2610 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2611 assert(st_off >= header_size, "still ignoring header");
2612 assert(get_store_offset(st, phase) == st_off, "must be");
2613 assert(in(i-1) == zmem, "must be");
2614 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2615 assert(con == tcon->is_int()->get_con(), "must be");
2616 // Undo the effects of the previous loop trip, which swallowed st:
2617 intcon[0] = 0; // undo store_constant()
2618 set_req(i-1, st); // undo set_req(i, zmem)
2619 nodes[j] = NULL; // undo nodes[j] = st
2620 --old_subword; // undo ++old_subword
2621 }
2622 continue; // This StoreI is already optimal.
2623 }
2624 }
2626 // This store is not needed.
2627 set_req(i, zmem);
2628 nodes[j] = st; // record for the moment
2629 if (st_size < BytesPerLong) // something has changed
2630 ++old_subword; // includes int/float, but who's counting...
2631 else ++old_long;
2632 }
2634 if ((old_subword + old_long) == 0)
2635 return; // nothing more to do
2637 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2638 // Be sure to insert them before overlapping non-constant stores.
2639 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2640 for (int j = 0; j < num_tiles; j++) {
2641 jlong con = tiles[j];
2642 jlong init = inits[j];
2643 if (con == 0) continue;
2644 jint con0, con1; // split the constant, address-wise
2645 jint init0, init1; // split the init map, address-wise
2646 { union { jlong con; jint intcon[2]; } u;
2647 u.con = con;
2648 con0 = u.intcon[0];
2649 con1 = u.intcon[1];
2650 u.con = init;
2651 init0 = u.intcon[0];
2652 init1 = u.intcon[1];
2653 }
2655 Node* old = nodes[j];
2656 assert(old != NULL, "need the prior store");
2657 intptr_t offset = (j * BytesPerLong);
2659 bool split = !Matcher::isSimpleConstant64(con);
2661 if (offset < header_size) {
2662 assert(offset + BytesPerInt >= header_size, "second int counts");
2663 assert(*(jint*)&tiles[j] == 0, "junk in header");
2664 split = true; // only the second word counts
2665 // Example: int a[] = { 42 ... }
2666 } else if (con0 == 0 && init0 == -1) {
2667 split = true; // first word is covered by full inits
2668 // Example: int a[] = { ... foo(), 42 ... }
2669 } else if (con1 == 0 && init1 == -1) {
2670 split = true; // second word is covered by full inits
2671 // Example: int a[] = { ... 42, foo() ... }
2672 }
2674 // Here's a case where init0 is neither 0 nor -1:
2675 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2676 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2677 // In this case the tile is not split; it is (jlong)42.
2678 // The big tile is stored down, and then the foo() value is inserted.
2679 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2681 Node* ctl = old->in(MemNode::Control);
2682 Node* adr = make_raw_address(offset, phase);
2683 const TypePtr* atp = TypeRawPtr::BOTTOM;
2685 // One or two coalesced stores to plop down.
2686 Node* st[2];
2687 intptr_t off[2];
2688 int nst = 0;
2689 if (!split) {
2690 ++new_long;
2691 off[nst] = offset;
2692 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2693 phase->longcon(con), T_LONG);
2694 } else {
2695 // Omit either if it is a zero.
2696 if (con0 != 0) {
2697 ++new_int;
2698 off[nst] = offset;
2699 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2700 phase->intcon(con0), T_INT);
2701 }
2702 if (con1 != 0) {
2703 ++new_int;
2704 offset += BytesPerInt;
2705 adr = make_raw_address(offset, phase);
2706 off[nst] = offset;
2707 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2708 phase->intcon(con1), T_INT);
2709 }
2710 }
2712 // Insert second store first, then the first before the second.
2713 // Insert each one just before any overlapping non-constant stores.
2714 while (nst > 0) {
2715 Node* st1 = st[--nst];
2716 C->copy_node_notes_to(st1, old);
2717 st1 = phase->transform(st1);
2718 offset = off[nst];
2719 assert(offset >= header_size, "do not smash header");
2720 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
2721 guarantee(ins_idx != 0, "must re-insert constant store");
2722 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
2723 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
2724 set_req(--ins_idx, st1);
2725 else
2726 ins_req(ins_idx, st1);
2727 }
2728 }
2730 if (PrintCompilation && WizardMode)
2731 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
2732 old_subword, old_long, new_int, new_long);
2733 if (C->log() != NULL)
2734 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
2735 old_subword, old_long, new_int, new_long);
2737 // Clean up any remaining occurrences of zmem:
2738 remove_extra_zeroes();
2739 }
2741 // Explore forward from in(start) to find the first fully initialized
2742 // word, and return its offset. Skip groups of subword stores which
2743 // together initialize full words. If in(start) is itself part of a
2744 // fully initialized word, return the offset of in(start). If there
2745 // are no following full-word stores, or if something is fishy, return
2746 // a negative value.
2747 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
2748 int int_map = 0;
2749 intptr_t int_map_off = 0;
2750 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
2752 for (uint i = start, limit = req(); i < limit; i++) {
2753 Node* st = in(i);
2755 intptr_t st_off = get_store_offset(st, phase);
2756 if (st_off < 0) break; // return conservative answer
2758 int st_size = st->as_Store()->memory_size();
2759 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
2760 return st_off; // we found a complete word init
2761 }
2763 // update the map:
2765 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
2766 if (this_int_off != int_map_off) {
2767 // reset the map:
2768 int_map = 0;
2769 int_map_off = this_int_off;
2770 }
2772 int subword_off = st_off - this_int_off;
2773 int_map |= right_n_bits(st_size) << subword_off;
2774 if ((int_map & FULL_MAP) == FULL_MAP) {
2775 return this_int_off; // we found a complete word init
2776 }
2778 // Did this store hit or cross the word boundary?
2779 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
2780 if (next_int_off == this_int_off + BytesPerInt) {
2781 // We passed the current int, without fully initializing it.
2782 int_map_off = next_int_off;
2783 int_map >>= BytesPerInt;
2784 } else if (next_int_off > this_int_off + BytesPerInt) {
2785 // We passed the current and next int.
2786 return this_int_off + BytesPerInt;
2787 }
2788 }
2790 return -1;
2791 }
2794 // Called when the associated AllocateNode is expanded into CFG.
2795 // At this point, we may perform additional optimizations.
2796 // Linearize the stores by ascending offset, to make memory
2797 // activity as coherent as possible.
2798 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
2799 intptr_t header_size,
2800 Node* size_in_bytes,
2801 PhaseGVN* phase) {
2802 assert(!is_complete(), "not already complete");
2803 assert(stores_are_sane(phase), "");
2804 assert(allocation() != NULL, "must be present");
2806 remove_extra_zeroes();
2808 if (ReduceFieldZeroing || ReduceBulkZeroing)
2809 // reduce instruction count for common initialization patterns
2810 coalesce_subword_stores(header_size, size_in_bytes, phase);
2812 Node* zmem = zero_memory(); // initially zero memory state
2813 Node* inits = zmem; // accumulating a linearized chain of inits
2814 #ifdef ASSERT
2815 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
2816 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
2817 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
2818 #endif
2819 intptr_t zeroes_done = header_size;
2821 bool do_zeroing = true; // we might give up if inits are very sparse
2822 int big_init_gaps = 0; // how many large gaps have we seen?
2824 if (ZeroTLAB) do_zeroing = false;
2825 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
2827 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2828 Node* st = in(i);
2829 intptr_t st_off = get_store_offset(st, phase);
2830 if (st_off < 0)
2831 break; // unknown junk in the inits
2832 if (st->in(MemNode::Memory) != zmem)
2833 break; // complicated store chains somehow in list
2835 int st_size = st->as_Store()->memory_size();
2836 intptr_t next_init_off = st_off + st_size;
2838 if (do_zeroing && zeroes_done < next_init_off) {
2839 // See if this store needs a zero before it or under it.
2840 intptr_t zeroes_needed = st_off;
2842 if (st_size < BytesPerInt) {
2843 // Look for subword stores which only partially initialize words.
2844 // If we find some, we must lay down some word-level zeroes first,
2845 // underneath the subword stores.
2846 //
2847 // Examples:
2848 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
2849 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
2850 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
2851 //
2852 // Note: coalesce_subword_stores may have already done this,
2853 // if it was prompted by constant non-zero subword initializers.
2854 // But this case can still arise with non-constant stores.
2856 intptr_t next_full_store = find_next_fullword_store(i, phase);
2858 // In the examples above:
2859 // in(i) p q r s x y z
2860 // st_off 12 13 14 15 12 13 14
2861 // st_size 1 1 1 1 1 1 1
2862 // next_full_s. 12 16 16 16 16 16 16
2863 // z's_done 12 16 16 16 12 16 12
2864 // z's_needed 12 16 16 16 16 16 16
2865 // zsize 0 0 0 0 4 0 4
2866 if (next_full_store < 0) {
2867 // Conservative tack: Zero to end of current word.
2868 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
2869 } else {
2870 // Zero to beginning of next fully initialized word.
2871 // Or, don't zero at all, if we are already in that word.
2872 assert(next_full_store >= zeroes_needed, "must go forward");
2873 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
2874 zeroes_needed = next_full_store;
2875 }
2876 }
2878 if (zeroes_needed > zeroes_done) {
2879 intptr_t zsize = zeroes_needed - zeroes_done;
2880 // Do some incremental zeroing on rawmem, in parallel with inits.
2881 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2882 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2883 zeroes_done, zeroes_needed,
2884 phase);
2885 zeroes_done = zeroes_needed;
2886 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
2887 do_zeroing = false; // leave the hole, next time
2888 }
2889 }
2891 // Collect the store and move on:
2892 st->set_req(MemNode::Memory, inits);
2893 inits = st; // put it on the linearized chain
2894 set_req(i, zmem); // unhook from previous position
2896 if (zeroes_done == st_off)
2897 zeroes_done = next_init_off;
2899 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
2901 #ifdef ASSERT
2902 // Various order invariants. Weaker than stores_are_sane because
2903 // a large constant tile can be filled in by smaller non-constant stores.
2904 assert(st_off >= last_init_off, "inits do not reverse");
2905 last_init_off = st_off;
2906 const Type* val = NULL;
2907 if (st_size >= BytesPerInt &&
2908 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
2909 (int)val->basic_type() < (int)T_OBJECT) {
2910 assert(st_off >= last_tile_end, "tiles do not overlap");
2911 assert(st_off >= last_init_end, "tiles do not overwrite inits");
2912 last_tile_end = MAX2(last_tile_end, next_init_off);
2913 } else {
2914 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
2915 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
2916 assert(st_off >= last_init_end, "inits do not overlap");
2917 last_init_end = next_init_off; // it's a non-tile
2918 }
2919 #endif //ASSERT
2920 }
2922 remove_extra_zeroes(); // clear out all the zmems left over
2923 add_req(inits);
2925 if (!ZeroTLAB) {
2926 // If anything remains to be zeroed, zero it all now.
2927 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
2928 // if it is the last unused 4 bytes of an instance, forget about it
2929 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
2930 if (zeroes_done + BytesPerLong >= size_limit) {
2931 assert(allocation() != NULL, "");
2932 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
2933 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
2934 if (zeroes_done == k->layout_helper())
2935 zeroes_done = size_limit;
2936 }
2937 if (zeroes_done < size_limit) {
2938 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
2939 zeroes_done, size_in_bytes, phase);
2940 }
2941 }
2943 set_complete(phase);
2944 return rawmem;
2945 }
2948 #ifdef ASSERT
2949 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
2950 if (is_complete())
2951 return true; // stores could be anything at this point
2952 intptr_t last_off = sizeof(oopDesc);
2953 for (uint i = InitializeNode::RawStores; i < req(); i++) {
2954 Node* st = in(i);
2955 intptr_t st_off = get_store_offset(st, phase);
2956 if (st_off < 0) continue; // ignore dead garbage
2957 if (last_off > st_off) {
2958 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
2959 this->dump(2);
2960 assert(false, "ascending store offsets");
2961 return false;
2962 }
2963 last_off = st_off + st->as_Store()->memory_size();
2964 }
2965 return true;
2966 }
2967 #endif //ASSERT
2972 //============================MergeMemNode=====================================
2973 //
2974 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
2975 // contributing store or call operations. Each contributor provides the memory
2976 // state for a particular "alias type" (see Compile::alias_type). For example,
2977 // if a MergeMem has an input X for alias category #6, then any memory reference
2978 // to alias category #6 may use X as its memory state input, as an exact equivalent
2979 // to using the MergeMem as a whole.
2980 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
2981 //
2982 // (Here, the <N> notation gives the index of the relevant adr_type.)
2983 //
2984 // In one special case (and more cases in the future), alias categories overlap.
2985 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
2986 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
2987 // it is exactly equivalent to that state W:
2988 // MergeMem(<Bot>: W) <==> W
2989 //
2990 // Usually, the merge has more than one input. In that case, where inputs
2991 // overlap (i.e., one is Bot), the narrower alias type determines the memory
2992 // state for that type, and the wider alias type (Bot) fills in everywhere else:
2993 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
2994 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
2995 //
2996 // A merge can take a "wide" memory state as one of its narrow inputs.
2997 // This simply means that the merge observes out only the relevant parts of
2998 // the wide input. That is, wide memory states arriving at narrow merge inputs
2999 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3000 //
3001 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3002 // and that memory slices "leak through":
3003 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3004 //
3005 // But, in such a cascade, repeated memory slices can "block the leak":
3006 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3007 //
3008 // In the last example, Y is not part of the combined memory state of the
3009 // outermost MergeMem. The system must, of course, prevent unschedulable
3010 // memory states from arising, so you can be sure that the state Y is somehow
3011 // a precursor to state Y'.
3012 //
3013 //
3014 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3015 // of each MergeMemNode array are exactly the numerical alias indexes, including
3016 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3017 // Compile::alias_type (and kin) produce and manage these indexes.
3018 //
3019 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3020 // (Note that this provides quick access to the top node inside MergeMem methods,
3021 // without the need to reach out via TLS to Compile::current.)
3022 //
3023 // As a consequence of what was just described, a MergeMem that represents a full
3024 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3025 // containing all alias categories.
3026 //
3027 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3028 //
3029 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3030 // a memory state for the alias type <N>, or else the top node, meaning that
3031 // there is no particular input for that alias type. Note that the length of
3032 // a MergeMem is variable, and may be extended at any time to accommodate new
3033 // memory states at larger alias indexes. When merges grow, they are of course
3034 // filled with "top" in the unused in() positions.
3035 //
3036 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3037 // (Top was chosen because it works smoothly with passes like GCM.)
3038 //
3039 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3040 // the type of random VM bits like TLS references.) Since it is always the
3041 // first non-Bot memory slice, some low-level loops use it to initialize an
3042 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3043 //
3044 //
3045 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3046 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3047 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3048 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3049 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3050 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3051 //
3052 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3053 // really that different from the other memory inputs. An abbreviation called
3054 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3055 //
3056 //
3057 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3058 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3059 // that "emerges though" the base memory will be marked as excluding the alias types
3060 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3061 //
3062 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3063 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3064 //
3065 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3066 // (It is currently unimplemented.) As you can see, the resulting merge is
3067 // actually a disjoint union of memory states, rather than an overlay.
3068 //
3070 //------------------------------MergeMemNode-----------------------------------
3071 Node* MergeMemNode::make_empty_memory() {
3072 Node* empty_memory = (Node*) Compile::current()->top();
3073 assert(empty_memory->is_top(), "correct sentinel identity");
3074 return empty_memory;
3075 }
3077 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3078 init_class_id(Class_MergeMem);
3079 // all inputs are nullified in Node::Node(int)
3080 // set_input(0, NULL); // no control input
3082 // Initialize the edges uniformly to top, for starters.
3083 Node* empty_mem = make_empty_memory();
3084 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3085 init_req(i,empty_mem);
3086 }
3087 assert(empty_memory() == empty_mem, "");
3089 if( new_base != NULL && new_base->is_MergeMem() ) {
3090 MergeMemNode* mdef = new_base->as_MergeMem();
3091 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3092 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3093 mms.set_memory(mms.memory2());
3094 }
3095 assert(base_memory() == mdef->base_memory(), "");
3096 } else {
3097 set_base_memory(new_base);
3098 }
3099 }
3101 // Make a new, untransformed MergeMem with the same base as 'mem'.
3102 // If mem is itself a MergeMem, populate the result with the same edges.
3103 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3104 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3105 }
3107 //------------------------------cmp--------------------------------------------
3108 uint MergeMemNode::hash() const { return NO_HASH; }
3109 uint MergeMemNode::cmp( const Node &n ) const {
3110 return (&n == this); // Always fail except on self
3111 }
3113 //------------------------------Identity---------------------------------------
3114 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3115 // Identity if this merge point does not record any interesting memory
3116 // disambiguations.
3117 Node* base_mem = base_memory();
3118 Node* empty_mem = empty_memory();
3119 if (base_mem != empty_mem) { // Memory path is not dead?
3120 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3121 Node* mem = in(i);
3122 if (mem != empty_mem && mem != base_mem) {
3123 return this; // Many memory splits; no change
3124 }
3125 }
3126 }
3127 return base_mem; // No memory splits; ID on the one true input
3128 }
3130 //------------------------------Ideal------------------------------------------
3131 // This method is invoked recursively on chains of MergeMem nodes
3132 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3133 // Remove chain'd MergeMems
3134 //
3135 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3136 // relative to the "in(Bot)". Since we are patching both at the same time,
3137 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3138 // but rewrite each "in(i)" relative to the new "in(Bot)".
3139 Node *progress = NULL;
3142 Node* old_base = base_memory();
3143 Node* empty_mem = empty_memory();
3144 if (old_base == empty_mem)
3145 return NULL; // Dead memory path.
3147 MergeMemNode* old_mbase;
3148 if (old_base != NULL && old_base->is_MergeMem())
3149 old_mbase = old_base->as_MergeMem();
3150 else
3151 old_mbase = NULL;
3152 Node* new_base = old_base;
3154 // simplify stacked MergeMems in base memory
3155 if (old_mbase) new_base = old_mbase->base_memory();
3157 // the base memory might contribute new slices beyond my req()
3158 if (old_mbase) grow_to_match(old_mbase);
3160 // Look carefully at the base node if it is a phi.
3161 PhiNode* phi_base;
3162 if (new_base != NULL && new_base->is_Phi())
3163 phi_base = new_base->as_Phi();
3164 else
3165 phi_base = NULL;
3167 Node* phi_reg = NULL;
3168 uint phi_len = (uint)-1;
3169 if (phi_base != NULL && !phi_base->is_copy()) {
3170 // do not examine phi if degraded to a copy
3171 phi_reg = phi_base->region();
3172 phi_len = phi_base->req();
3173 // see if the phi is unfinished
3174 for (uint i = 1; i < phi_len; i++) {
3175 if (phi_base->in(i) == NULL) {
3176 // incomplete phi; do not look at it yet!
3177 phi_reg = NULL;
3178 phi_len = (uint)-1;
3179 break;
3180 }
3181 }
3182 }
3184 // Note: We do not call verify_sparse on entry, because inputs
3185 // can normalize to the base_memory via subsume_node or similar
3186 // mechanisms. This method repairs that damage.
3188 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3190 // Look at each slice.
3191 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3192 Node* old_in = in(i);
3193 // calculate the old memory value
3194 Node* old_mem = old_in;
3195 if (old_mem == empty_mem) old_mem = old_base;
3196 assert(old_mem == memory_at(i), "");
3198 // maybe update (reslice) the old memory value
3200 // simplify stacked MergeMems
3201 Node* new_mem = old_mem;
3202 MergeMemNode* old_mmem;
3203 if (old_mem != NULL && old_mem->is_MergeMem())
3204 old_mmem = old_mem->as_MergeMem();
3205 else
3206 old_mmem = NULL;
3207 if (old_mmem == this) {
3208 // This can happen if loops break up and safepoints disappear.
3209 // A merge of BotPtr (default) with a RawPtr memory derived from a
3210 // safepoint can be rewritten to a merge of the same BotPtr with
3211 // the BotPtr phi coming into the loop. If that phi disappears
3212 // also, we can end up with a self-loop of the mergemem.
3213 // In general, if loops degenerate and memory effects disappear,
3214 // a mergemem can be left looking at itself. This simply means
3215 // that the mergemem's default should be used, since there is
3216 // no longer any apparent effect on this slice.
3217 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3218 // from start. Update the input to TOP.
3219 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3220 }
3221 else if (old_mmem != NULL) {
3222 new_mem = old_mmem->memory_at(i);
3223 }
3224 // else preceeding memory was not a MergeMem
3226 // replace equivalent phis (unfortunately, they do not GVN together)
3227 if (new_mem != NULL && new_mem != new_base &&
3228 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3229 if (new_mem->is_Phi()) {
3230 PhiNode* phi_mem = new_mem->as_Phi();
3231 for (uint i = 1; i < phi_len; i++) {
3232 if (phi_base->in(i) != phi_mem->in(i)) {
3233 phi_mem = NULL;
3234 break;
3235 }
3236 }
3237 if (phi_mem != NULL) {
3238 // equivalent phi nodes; revert to the def
3239 new_mem = new_base;
3240 }
3241 }
3242 }
3244 // maybe store down a new value
3245 Node* new_in = new_mem;
3246 if (new_in == new_base) new_in = empty_mem;
3248 if (new_in != old_in) {
3249 // Warning: Do not combine this "if" with the previous "if"
3250 // A memory slice might have be be rewritten even if it is semantically
3251 // unchanged, if the base_memory value has changed.
3252 set_req(i, new_in);
3253 progress = this; // Report progress
3254 }
3255 }
3257 if (new_base != old_base) {
3258 set_req(Compile::AliasIdxBot, new_base);
3259 // Don't use set_base_memory(new_base), because we need to update du.
3260 assert(base_memory() == new_base, "");
3261 progress = this;
3262 }
3264 if( base_memory() == this ) {
3265 // a self cycle indicates this memory path is dead
3266 set_req(Compile::AliasIdxBot, empty_mem);
3267 }
3269 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3270 // Recursion must occur after the self cycle check above
3271 if( base_memory()->is_MergeMem() ) {
3272 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3273 Node *m = phase->transform(new_mbase); // Rollup any cycles
3274 if( m != NULL && (m->is_top() ||
3275 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3276 // propagate rollup of dead cycle to self
3277 set_req(Compile::AliasIdxBot, empty_mem);
3278 }
3279 }
3281 if( base_memory() == empty_mem ) {
3282 progress = this;
3283 // Cut inputs during Parse phase only.
3284 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3285 if( !can_reshape ) {
3286 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3287 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3288 }
3289 }
3290 }
3292 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3293 // Check if PhiNode::Ideal's "Split phis through memory merges"
3294 // transform should be attempted. Look for this->phi->this cycle.
3295 uint merge_width = req();
3296 if (merge_width > Compile::AliasIdxRaw) {
3297 PhiNode* phi = base_memory()->as_Phi();
3298 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3299 if (phi->in(i) == this) {
3300 phase->is_IterGVN()->_worklist.push(phi);
3301 break;
3302 }
3303 }
3304 }
3305 }
3307 assert(progress || verify_sparse(), "please, no dups of base");
3308 return progress;
3309 }
3311 //-------------------------set_base_memory-------------------------------------
3312 void MergeMemNode::set_base_memory(Node *new_base) {
3313 Node* empty_mem = empty_memory();
3314 set_req(Compile::AliasIdxBot, new_base);
3315 assert(memory_at(req()) == new_base, "must set default memory");
3316 // Clear out other occurrences of new_base:
3317 if (new_base != empty_mem) {
3318 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3319 if (in(i) == new_base) set_req(i, empty_mem);
3320 }
3321 }
3322 }
3324 //------------------------------out_RegMask------------------------------------
3325 const RegMask &MergeMemNode::out_RegMask() const {
3326 return RegMask::Empty;
3327 }
3329 //------------------------------dump_spec--------------------------------------
3330 #ifndef PRODUCT
3331 void MergeMemNode::dump_spec(outputStream *st) const {
3332 st->print(" {");
3333 Node* base_mem = base_memory();
3334 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3335 Node* mem = memory_at(i);
3336 if (mem == base_mem) { st->print(" -"); continue; }
3337 st->print( " N%d:", mem->_idx );
3338 Compile::current()->get_adr_type(i)->dump_on(st);
3339 }
3340 st->print(" }");
3341 }
3342 #endif // !PRODUCT
3345 #ifdef ASSERT
3346 static bool might_be_same(Node* a, Node* b) {
3347 if (a == b) return true;
3348 if (!(a->is_Phi() || b->is_Phi())) return false;
3349 // phis shift around during optimization
3350 return true; // pretty stupid...
3351 }
3353 // verify a narrow slice (either incoming or outgoing)
3354 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3355 if (!VerifyAliases) return; // don't bother to verify unless requested
3356 if (is_error_reported()) return; // muzzle asserts when debugging an error
3357 if (Node::in_dump()) return; // muzzle asserts when printing
3358 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3359 assert(n != NULL, "");
3360 // Elide intervening MergeMem's
3361 while (n->is_MergeMem()) {
3362 n = n->as_MergeMem()->memory_at(alias_idx);
3363 }
3364 Compile* C = Compile::current();
3365 const TypePtr* n_adr_type = n->adr_type();
3366 if (n == m->empty_memory()) {
3367 // Implicit copy of base_memory()
3368 } else if (n_adr_type != TypePtr::BOTTOM) {
3369 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3370 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3371 } else {
3372 // A few places like make_runtime_call "know" that VM calls are narrow,
3373 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3374 bool expected_wide_mem = false;
3375 if (n == m->base_memory()) {
3376 expected_wide_mem = true;
3377 } else if (alias_idx == Compile::AliasIdxRaw ||
3378 n == m->memory_at(Compile::AliasIdxRaw)) {
3379 expected_wide_mem = true;
3380 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3381 // memory can "leak through" calls on channels that
3382 // are write-once. Allow this also.
3383 expected_wide_mem = true;
3384 }
3385 assert(expected_wide_mem, "expected narrow slice replacement");
3386 }
3387 }
3388 #else // !ASSERT
3389 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3390 #endif
3393 //-----------------------------memory_at---------------------------------------
3394 Node* MergeMemNode::memory_at(uint alias_idx) const {
3395 assert(alias_idx >= Compile::AliasIdxRaw ||
3396 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3397 "must avoid base_memory and AliasIdxTop");
3399 // Otherwise, it is a narrow slice.
3400 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3401 Compile *C = Compile::current();
3402 if (is_empty_memory(n)) {
3403 // the array is sparse; empty slots are the "top" node
3404 n = base_memory();
3405 assert(Node::in_dump()
3406 || n == NULL || n->bottom_type() == Type::TOP
3407 || n->adr_type() == TypePtr::BOTTOM
3408 || n->adr_type() == TypeRawPtr::BOTTOM
3409 || Compile::current()->AliasLevel() == 0,
3410 "must be a wide memory");
3411 // AliasLevel == 0 if we are organizing the memory states manually.
3412 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3413 } else {
3414 // make sure the stored slice is sane
3415 #ifdef ASSERT
3416 if (is_error_reported() || Node::in_dump()) {
3417 } else if (might_be_same(n, base_memory())) {
3418 // Give it a pass: It is a mostly harmless repetition of the base.
3419 // This can arise normally from node subsumption during optimization.
3420 } else {
3421 verify_memory_slice(this, alias_idx, n);
3422 }
3423 #endif
3424 }
3425 return n;
3426 }
3428 //---------------------------set_memory_at-------------------------------------
3429 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3430 verify_memory_slice(this, alias_idx, n);
3431 Node* empty_mem = empty_memory();
3432 if (n == base_memory()) n = empty_mem; // collapse default
3433 uint need_req = alias_idx+1;
3434 if (req() < need_req) {
3435 if (n == empty_mem) return; // already the default, so do not grow me
3436 // grow the sparse array
3437 do {
3438 add_req(empty_mem);
3439 } while (req() < need_req);
3440 }
3441 set_req( alias_idx, n );
3442 }
3446 //--------------------------iteration_setup------------------------------------
3447 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3448 if (other != NULL) {
3449 grow_to_match(other);
3450 // invariant: the finite support of mm2 is within mm->req()
3451 #ifdef ASSERT
3452 for (uint i = req(); i < other->req(); i++) {
3453 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3454 }
3455 #endif
3456 }
3457 // Replace spurious copies of base_memory by top.
3458 Node* base_mem = base_memory();
3459 if (base_mem != NULL && !base_mem->is_top()) {
3460 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3461 if (in(i) == base_mem)
3462 set_req(i, empty_memory());
3463 }
3464 }
3465 }
3467 //---------------------------grow_to_match-------------------------------------
3468 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3469 Node* empty_mem = empty_memory();
3470 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3471 // look for the finite support of the other memory
3472 for (uint i = other->req(); --i >= req(); ) {
3473 if (other->in(i) != empty_mem) {
3474 uint new_len = i+1;
3475 while (req() < new_len) add_req(empty_mem);
3476 break;
3477 }
3478 }
3479 }
3481 //---------------------------verify_sparse-------------------------------------
3482 #ifndef PRODUCT
3483 bool MergeMemNode::verify_sparse() const {
3484 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3485 Node* base_mem = base_memory();
3486 // The following can happen in degenerate cases, since empty==top.
3487 if (is_empty_memory(base_mem)) return true;
3488 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3489 assert(in(i) != NULL, "sane slice");
3490 if (in(i) == base_mem) return false; // should have been the sentinel value!
3491 }
3492 return true;
3493 }
3495 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3496 Node* n;
3497 n = mm->in(idx);
3498 if (mem == n) return true; // might be empty_memory()
3499 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3500 if (mem == n) return true;
3501 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3502 if (mem == n) return true;
3503 if (n == NULL) break;
3504 }
3505 return false;
3506 }
3507 #endif // !PRODUCT