Fri, 01 Aug 2008 10:06:45 -0700
6731641: assert(m->adr_type() == mach->adr_type(),"matcher should not change adr type")
Summary: fixed few addP node type and narrow oop type problems.
Reviewed-by: rasbold, never
1 /*
2 * Copyright 1997-2008 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 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
37 const TypePtr *MemNode::adr_type() const {
38 Node* adr = in(Address);
39 const TypePtr* cross_check = NULL;
40 DEBUG_ONLY(cross_check = _adr_type);
41 return calculate_adr_type(adr->bottom_type(), cross_check);
42 }
44 #ifndef PRODUCT
45 void MemNode::dump_spec(outputStream *st) const {
46 if (in(Address) == NULL) return; // node is dead
47 #ifndef ASSERT
48 // fake the missing field
49 const TypePtr* _adr_type = NULL;
50 if (in(Address) != NULL)
51 _adr_type = in(Address)->bottom_type()->isa_ptr();
52 #endif
53 dump_adr_type(this, _adr_type, st);
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
61 st->print(" @");
62 if (adr_type == NULL) {
63 st->print("NULL");
64 } else {
65 adr_type->dump_on(st);
66 Compile* C = Compile::current();
67 Compile::AliasType* atp = NULL;
68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
69 if (atp == NULL)
70 st->print(", idx=?\?;");
71 else if (atp->index() == Compile::AliasIdxBot)
72 st->print(", idx=Bot;");
73 else if (atp->index() == Compile::AliasIdxTop)
74 st->print(", idx=Top;");
75 else if (atp->index() == Compile::AliasIdxRaw)
76 st->print(", idx=Raw;");
77 else {
78 ciField* field = atp->field();
79 if (field) {
80 st->print(", name=");
81 field->print_name_on(st);
82 }
83 st->print(", idx=%d;", atp->index());
84 }
85 }
86 }
88 extern void print_alias_types();
90 #endif
92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
94 if (tinst == NULL || !tinst->is_known_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
98 Node *prev = NULL;
99 Node *result = mchain;
100 while (prev != result) {
101 prev = result;
102 if (result == start_mem)
103 break; // hit one of our sentinals
104 // skip over a call which does not affect this memory slice
105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
106 Node *proj_in = result->in(0);
107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
108 break; // hit one of our sentinals
109 } else if (proj_in->is_Call()) {
110 CallNode *call = proj_in->as_Call();
111 if (!call->may_modify(t_adr, phase)) {
112 result = call->in(TypeFunc::Memory);
113 }
114 } else if (proj_in->is_Initialize()) {
115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
116 // Stop if this is the initialization for the object instance which
117 // which contains this memory slice, otherwise skip over it.
118 if (alloc != NULL && alloc->_idx != instance_id) {
119 result = proj_in->in(TypeFunc::Memory);
120 }
121 } else if (proj_in->is_MemBar()) {
122 result = proj_in->in(TypeFunc::Memory);
123 } else {
124 assert(false, "unexpected projection");
125 }
126 } else if (result->is_MergeMem()) {
127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
128 }
129 }
130 return result;
131 }
133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
136 PhaseIterGVN *igvn = phase->is_IterGVN();
137 Node *result = mchain;
138 result = optimize_simple_memory_chain(result, t_adr, phase);
139 if (is_instance && igvn != NULL && result->is_Phi()) {
140 PhiNode *mphi = result->as_Phi();
141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
142 const TypePtr *t = mphi->adr_type();
143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
145 t->is_oopptr()->cast_to_exactness(true)
146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
148 // clone the Phi with our address type
149 result = mphi->split_out_instance(t_adr, igvn);
150 } else {
151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
152 }
153 }
154 return result;
155 }
157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
158 uint alias_idx = phase->C->get_alias_index(tp);
159 Node *mem = mmem;
160 #ifdef ASSERT
161 {
162 // Check that current type is consistent with the alias index used during graph construction
163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
164 bool consistent = adr_check == NULL || adr_check->empty() ||
165 phase->C->must_alias(adr_check, alias_idx );
166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
173 // don't assert if it is dead code.
174 consistent = true;
175 }
176 if( !consistent ) {
177 st->print("alias_idx==%d, adr_check==", alias_idx);
178 if( adr_check == NULL ) {
179 st->print("NULL");
180 } else {
181 adr_check->dump();
182 }
183 st->cr();
184 print_alias_types();
185 assert(consistent, "adr_check must match alias idx");
186 }
187 }
188 #endif
189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
190 // means an array I have not precisely typed yet. Do not do any
191 // alias stuff with it any time soon.
192 const TypeOopPtr *tinst = tp->isa_oopptr();
193 if( tp->base() != Type::AnyPtr &&
194 !(tinst &&
195 tinst->klass()->is_java_lang_Object() &&
196 tinst->offset() == Type::OffsetBot) ) {
197 // compress paths and change unreachable cycles to TOP
198 // If not, we can update the input infinitely along a MergeMem cycle
199 // Equivalent code in PhiNode::Ideal
200 Node* m = phase->transform(mmem);
201 // If tranformed to a MergeMem, get the desired slice
202 // Otherwise the returned node represents memory for every slice
203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
204 // Update input if it is progress over what we have now
205 }
206 return mem;
207 }
209 //--------------------------Ideal_common---------------------------------------
210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
211 // Unhook non-raw memories from complete (macro-expanded) initializations.
212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
213 // If our control input is a dead region, kill all below the region
214 Node *ctl = in(MemNode::Control);
215 if (ctl && remove_dead_region(phase, can_reshape))
216 return this;
218 // Ignore if memory is dead, or self-loop
219 Node *mem = in(MemNode::Memory);
220 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
221 assert( mem != this, "dead loop in MemNode::Ideal" );
223 Node *address = in(MemNode::Address);
224 const Type *t_adr = phase->type( address );
225 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
227 // Avoid independent memory operations
228 Node* old_mem = mem;
230 // The code which unhooks non-raw memories from complete (macro-expanded)
231 // initializations was removed. After macro-expansion all stores catched
232 // by Initialize node became raw stores and there is no information
233 // which memory slices they modify. So it is unsafe to move any memory
234 // operation above these stores. Also in most cases hooked non-raw memories
235 // were already unhooked by using information from detect_ptr_independence()
236 // and find_previous_store().
238 if (mem->is_MergeMem()) {
239 MergeMemNode* mmem = mem->as_MergeMem();
240 const TypePtr *tp = t_adr->is_ptr();
242 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
243 }
245 if (mem != old_mem) {
246 set_req(MemNode::Memory, mem);
247 return this;
248 }
250 // let the subclass continue analyzing...
251 return NULL;
252 }
254 // Helper function for proving some simple control dominations.
255 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
256 // Already assumes that 'dom' is available at 'sub', and that 'sub'
257 // is not a constant (dominated by the method's StartNode).
258 // Used by MemNode::find_previous_store to prove that the
259 // control input of a memory operation predates (dominates)
260 // an allocation it wants to look past.
261 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
262 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
263 return false; // Conservative answer for dead code
265 // Check 'dom'. Skip Proj and CatchProj nodes.
266 dom = dom->find_exact_control(dom);
267 if (dom == NULL || dom->is_top())
268 return false; // Conservative answer for dead code
270 if (dom == sub) {
271 // For the case when, for example, 'sub' is Initialize and the original
272 // 'dom' is Proj node of the 'sub'.
273 return false;
274 }
276 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
277 return true;
279 // 'dom' dominates 'sub' if its control edge and control edges
280 // of all its inputs dominate or equal to sub's control edge.
282 // Currently 'sub' is either Allocate, Initialize or Start nodes.
283 // Or Region for the check in LoadNode::Ideal();
284 // 'sub' should have sub->in(0) != NULL.
285 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
286 sub->is_Region(), "expecting only these nodes");
288 // Get control edge of 'sub'.
289 Node* orig_sub = sub;
290 sub = sub->find_exact_control(sub->in(0));
291 if (sub == NULL || sub->is_top())
292 return false; // Conservative answer for dead code
294 assert(sub->is_CFG(), "expecting control");
296 if (sub == dom)
297 return true;
299 if (sub->is_Start() || sub->is_Root())
300 return false;
302 {
303 // Check all control edges of 'dom'.
305 ResourceMark rm;
306 Arena* arena = Thread::current()->resource_area();
307 Node_List nlist(arena);
308 Unique_Node_List dom_list(arena);
310 dom_list.push(dom);
311 bool only_dominating_controls = false;
313 for (uint next = 0; next < dom_list.size(); next++) {
314 Node* n = dom_list.at(next);
315 if (n == orig_sub)
316 return false; // One of dom's inputs dominated by sub.
317 if (!n->is_CFG() && n->pinned()) {
318 // Check only own control edge for pinned non-control nodes.
319 n = n->find_exact_control(n->in(0));
320 if (n == NULL || n->is_top())
321 return false; // Conservative answer for dead code
322 assert(n->is_CFG(), "expecting control");
323 dom_list.push(n);
324 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
325 only_dominating_controls = true;
326 } else if (n->is_CFG()) {
327 if (n->dominates(sub, nlist))
328 only_dominating_controls = true;
329 else
330 return false;
331 } else {
332 // First, own control edge.
333 Node* m = n->find_exact_control(n->in(0));
334 if (m != NULL) {
335 if (m->is_top())
336 return false; // Conservative answer for dead code
337 dom_list.push(m);
338 }
339 // Now, the rest of edges.
340 uint cnt = n->req();
341 for (uint i = 1; i < cnt; i++) {
342 m = n->find_exact_control(n->in(i));
343 if (m == NULL || m->is_top())
344 continue;
345 dom_list.push(m);
346 }
347 }
348 }
349 return only_dominating_controls;
350 }
351 }
353 //---------------------detect_ptr_independence---------------------------------
354 // Used by MemNode::find_previous_store to prove that two base
355 // pointers are never equal.
356 // The pointers are accompanied by their associated allocations,
357 // if any, which have been previously discovered by the caller.
358 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
359 Node* p2, AllocateNode* a2,
360 PhaseTransform* phase) {
361 // Attempt to prove that these two pointers cannot be aliased.
362 // They may both manifestly be allocations, and they should differ.
363 // Or, if they are not both allocations, they can be distinct constants.
364 // Otherwise, one is an allocation and the other a pre-existing value.
365 if (a1 == NULL && a2 == NULL) { // neither an allocation
366 return (p1 != p2) && p1->is_Con() && p2->is_Con();
367 } else if (a1 != NULL && a2 != NULL) { // both allocations
368 return (a1 != a2);
369 } else if (a1 != NULL) { // one allocation a1
370 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
371 return all_controls_dominate(p2, a1);
372 } else { //(a2 != NULL) // one allocation a2
373 return all_controls_dominate(p1, a2);
374 }
375 return false;
376 }
379 // The logic for reordering loads and stores uses four steps:
380 // (a) Walk carefully past stores and initializations which we
381 // can prove are independent of this load.
382 // (b) Observe that the next memory state makes an exact match
383 // with self (load or store), and locate the relevant store.
384 // (c) Ensure that, if we were to wire self directly to the store,
385 // the optimizer would fold it up somehow.
386 // (d) Do the rewiring, and return, depending on some other part of
387 // the optimizer to fold up the load.
388 // This routine handles steps (a) and (b). Steps (c) and (d) are
389 // specific to loads and stores, so they are handled by the callers.
390 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
391 //
392 Node* MemNode::find_previous_store(PhaseTransform* phase) {
393 Node* ctrl = in(MemNode::Control);
394 Node* adr = in(MemNode::Address);
395 intptr_t offset = 0;
396 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
397 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
399 if (offset == Type::OffsetBot)
400 return NULL; // cannot unalias unless there are precise offsets
402 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
404 intptr_t size_in_bytes = memory_size();
406 Node* mem = in(MemNode::Memory); // start searching here...
408 int cnt = 50; // Cycle limiter
409 for (;;) { // While we can dance past unrelated stores...
410 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
412 if (mem->is_Store()) {
413 Node* st_adr = mem->in(MemNode::Address);
414 intptr_t st_offset = 0;
415 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
416 if (st_base == NULL)
417 break; // inscrutable pointer
418 if (st_offset != offset && st_offset != Type::OffsetBot) {
419 const int MAX_STORE = BytesPerLong;
420 if (st_offset >= offset + size_in_bytes ||
421 st_offset <= offset - MAX_STORE ||
422 st_offset <= offset - mem->as_Store()->memory_size()) {
423 // Success: The offsets are provably independent.
424 // (You may ask, why not just test st_offset != offset and be done?
425 // The answer is that stores of different sizes can co-exist
426 // in the same sequence of RawMem effects. We sometimes initialize
427 // a whole 'tile' of array elements with a single jint or jlong.)
428 mem = mem->in(MemNode::Memory);
429 continue; // (a) advance through independent store memory
430 }
431 }
432 if (st_base != base &&
433 detect_ptr_independence(base, alloc,
434 st_base,
435 AllocateNode::Ideal_allocation(st_base, phase),
436 phase)) {
437 // Success: The bases are provably independent.
438 mem = mem->in(MemNode::Memory);
439 continue; // (a) advance through independent store memory
440 }
442 // (b) At this point, if the bases or offsets do not agree, we lose,
443 // since we have not managed to prove 'this' and 'mem' independent.
444 if (st_base == base && st_offset == offset) {
445 return mem; // let caller handle steps (c), (d)
446 }
448 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
449 InitializeNode* st_init = mem->in(0)->as_Initialize();
450 AllocateNode* st_alloc = st_init->allocation();
451 if (st_alloc == NULL)
452 break; // something degenerated
453 bool known_identical = false;
454 bool known_independent = false;
455 if (alloc == st_alloc)
456 known_identical = true;
457 else if (alloc != NULL)
458 known_independent = true;
459 else if (all_controls_dominate(this, st_alloc))
460 known_independent = true;
462 if (known_independent) {
463 // The bases are provably independent: Either they are
464 // manifestly distinct allocations, or else the control
465 // of this load dominates the store's allocation.
466 int alias_idx = phase->C->get_alias_index(adr_type());
467 if (alias_idx == Compile::AliasIdxRaw) {
468 mem = st_alloc->in(TypeFunc::Memory);
469 } else {
470 mem = st_init->memory(alias_idx);
471 }
472 continue; // (a) advance through independent store memory
473 }
475 // (b) at this point, if we are not looking at a store initializing
476 // the same allocation we are loading from, we lose.
477 if (known_identical) {
478 // From caller, can_see_stored_value will consult find_captured_store.
479 return mem; // let caller handle steps (c), (d)
480 }
482 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
483 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
484 if (mem->is_Proj() && mem->in(0)->is_Call()) {
485 CallNode *call = mem->in(0)->as_Call();
486 if (!call->may_modify(addr_t, phase)) {
487 mem = call->in(TypeFunc::Memory);
488 continue; // (a) advance through independent call memory
489 }
490 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
491 mem = mem->in(0)->in(TypeFunc::Memory);
492 continue; // (a) advance through independent MemBar memory
493 } else if (mem->is_MergeMem()) {
494 int alias_idx = phase->C->get_alias_index(adr_type());
495 mem = mem->as_MergeMem()->memory_at(alias_idx);
496 continue; // (a) advance through independent MergeMem memory
497 }
498 }
500 // Unless there is an explicit 'continue', we must bail out here,
501 // because 'mem' is an inscrutable memory state (e.g., a call).
502 break;
503 }
505 return NULL; // bail out
506 }
508 //----------------------calculate_adr_type-------------------------------------
509 // Helper function. Notices when the given type of address hits top or bottom.
510 // Also, asserts a cross-check of the type against the expected address type.
511 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
512 if (t == Type::TOP) return NULL; // does not touch memory any more?
513 #ifdef PRODUCT
514 cross_check = NULL;
515 #else
516 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
517 #endif
518 const TypePtr* tp = t->isa_ptr();
519 if (tp == NULL) {
520 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
521 return TypePtr::BOTTOM; // touches lots of memory
522 } else {
523 #ifdef ASSERT
524 // %%%% [phh] We don't check the alias index if cross_check is
525 // TypeRawPtr::BOTTOM. Needs to be investigated.
526 if (cross_check != NULL &&
527 cross_check != TypePtr::BOTTOM &&
528 cross_check != TypeRawPtr::BOTTOM) {
529 // Recheck the alias index, to see if it has changed (due to a bug).
530 Compile* C = Compile::current();
531 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
532 "must stay in the original alias category");
533 // The type of the address must be contained in the adr_type,
534 // disregarding "null"-ness.
535 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
536 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
537 assert(cross_check->meet(tp_notnull) == cross_check,
538 "real address must not escape from expected memory type");
539 }
540 #endif
541 return tp;
542 }
543 }
545 //------------------------adr_phi_is_loop_invariant----------------------------
546 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
547 // loop is loop invariant. Make a quick traversal of Phi and associated
548 // CastPP nodes, looking to see if they are a closed group within the loop.
549 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
550 // The idea is that the phi-nest must boil down to only CastPP nodes
551 // with the same data. This implies that any path into the loop already
552 // includes such a CastPP, and so the original cast, whatever its input,
553 // must be covered by an equivalent cast, with an earlier control input.
554 ResourceMark rm;
556 // The loop entry input of the phi should be the unique dominating
557 // node for every Phi/CastPP in the loop.
558 Unique_Node_List closure;
559 closure.push(adr_phi->in(LoopNode::EntryControl));
561 // Add the phi node and the cast to the worklist.
562 Unique_Node_List worklist;
563 worklist.push(adr_phi);
564 if( cast != NULL ){
565 if( !cast->is_ConstraintCast() ) return false;
566 worklist.push(cast);
567 }
569 // Begin recursive walk of phi nodes.
570 while( worklist.size() ){
571 // Take a node off the worklist
572 Node *n = worklist.pop();
573 if( !closure.member(n) ){
574 // Add it to the closure.
575 closure.push(n);
576 // Make a sanity check to ensure we don't waste too much time here.
577 if( closure.size() > 20) return false;
578 // This node is OK if:
579 // - it is a cast of an identical value
580 // - or it is a phi node (then we add its inputs to the worklist)
581 // Otherwise, the node is not OK, and we presume the cast is not invariant
582 if( n->is_ConstraintCast() ){
583 worklist.push(n->in(1));
584 } else if( n->is_Phi() ) {
585 for( uint i = 1; i < n->req(); i++ ) {
586 worklist.push(n->in(i));
587 }
588 } else {
589 return false;
590 }
591 }
592 }
594 // Quit when the worklist is empty, and we've found no offending nodes.
595 return true;
596 }
598 //------------------------------Ideal_DU_postCCP-------------------------------
599 // Find any cast-away of null-ness and keep its control. Null cast-aways are
600 // going away in this pass and we need to make this memory op depend on the
601 // gating null check.
602 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
603 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
604 }
606 // I tried to leave the CastPP's in. This makes the graph more accurate in
607 // some sense; we get to keep around the knowledge that an oop is not-null
608 // after some test. Alas, the CastPP's interfere with GVN (some values are
609 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
610 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
611 // some of the more trivial cases in the optimizer. Removing more useless
612 // Phi's started allowing Loads to illegally float above null checks. I gave
613 // up on this approach. CNC 10/20/2000
614 // This static method may be called not from MemNode (EncodePNode calls it).
615 // Only the control edge of the node 'n' might be updated.
616 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
617 Node *skipped_cast = NULL;
618 // Need a null check? Regular static accesses do not because they are
619 // from constant addresses. Array ops are gated by the range check (which
620 // always includes a NULL check). Just check field ops.
621 if( n->in(MemNode::Control) == NULL ) {
622 // Scan upwards for the highest location we can place this memory op.
623 while( true ) {
624 switch( adr->Opcode() ) {
626 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
627 adr = adr->in(AddPNode::Base);
628 continue;
630 case Op_DecodeN: // No change to NULL-ness, so peek thru
631 adr = adr->in(1);
632 continue;
634 case Op_CastPP:
635 // If the CastPP is useless, just peek on through it.
636 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
637 // Remember the cast that we've peeked though. If we peek
638 // through more than one, then we end up remembering the highest
639 // one, that is, if in a loop, the one closest to the top.
640 skipped_cast = adr;
641 adr = adr->in(1);
642 continue;
643 }
644 // CastPP is going away in this pass! We need this memory op to be
645 // control-dependent on the test that is guarding the CastPP.
646 ccp->hash_delete(n);
647 n->set_req(MemNode::Control, adr->in(0));
648 ccp->hash_insert(n);
649 return n;
651 case Op_Phi:
652 // Attempt to float above a Phi to some dominating point.
653 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
654 // If we've already peeked through a Cast (which could have set the
655 // control), we can't float above a Phi, because the skipped Cast
656 // may not be loop invariant.
657 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
658 adr = adr->in(1);
659 continue;
660 }
661 }
663 // Intentional fallthrough!
665 // No obvious dominating point. The mem op is pinned below the Phi
666 // by the Phi itself. If the Phi goes away (no true value is merged)
667 // then the mem op can float, but not indefinitely. It must be pinned
668 // behind the controls leading to the Phi.
669 case Op_CheckCastPP:
670 // These usually stick around to change address type, however a
671 // useless one can be elided and we still need to pick up a control edge
672 if (adr->in(0) == NULL) {
673 // This CheckCastPP node has NO control and is likely useless. But we
674 // need check further up the ancestor chain for a control input to keep
675 // the node in place. 4959717.
676 skipped_cast = adr;
677 adr = adr->in(1);
678 continue;
679 }
680 ccp->hash_delete(n);
681 n->set_req(MemNode::Control, adr->in(0));
682 ccp->hash_insert(n);
683 return n;
685 // List of "safe" opcodes; those that implicitly block the memory
686 // op below any null check.
687 case Op_CastX2P: // no null checks on native pointers
688 case Op_Parm: // 'this' pointer is not null
689 case Op_LoadP: // Loading from within a klass
690 case Op_LoadN: // Loading from within a klass
691 case Op_LoadKlass: // Loading from within a klass
692 case Op_LoadNKlass: // Loading from within a klass
693 case Op_ConP: // Loading from a klass
694 case Op_ConN: // Loading from a klass
695 case Op_CreateEx: // Sucking up the guts of an exception oop
696 case Op_Con: // Reading from TLS
697 case Op_CMoveP: // CMoveP is pinned
698 case Op_CMoveN: // CMoveN is pinned
699 break; // No progress
701 case Op_Proj: // Direct call to an allocation routine
702 case Op_SCMemProj: // Memory state from store conditional ops
703 #ifdef ASSERT
704 {
705 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
706 const Node* call = adr->in(0);
707 if (call->is_CallJava()) {
708 const CallJavaNode* call_java = call->as_CallJava();
709 const TypeTuple *r = call_java->tf()->range();
710 assert(r->cnt() > TypeFunc::Parms, "must return value");
711 const Type* ret_type = r->field_at(TypeFunc::Parms);
712 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
713 // We further presume that this is one of
714 // new_instance_Java, new_array_Java, or
715 // the like, but do not assert for this.
716 } else if (call->is_Allocate()) {
717 // similar case to new_instance_Java, etc.
718 } else if (!call->is_CallLeaf()) {
719 // Projections from fetch_oop (OSR) are allowed as well.
720 ShouldNotReachHere();
721 }
722 }
723 #endif
724 break;
725 default:
726 ShouldNotReachHere();
727 }
728 break;
729 }
730 }
732 return NULL; // No progress
733 }
736 //=============================================================================
737 uint LoadNode::size_of() const { return sizeof(*this); }
738 uint LoadNode::cmp( const Node &n ) const
739 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
740 const Type *LoadNode::bottom_type() const { return _type; }
741 uint LoadNode::ideal_reg() const {
742 return Matcher::base2reg[_type->base()];
743 }
745 #ifndef PRODUCT
746 void LoadNode::dump_spec(outputStream *st) const {
747 MemNode::dump_spec(st);
748 if( !Verbose && !WizardMode ) {
749 // standard dump does this in Verbose and WizardMode
750 st->print(" #"); _type->dump_on(st);
751 }
752 }
753 #endif
756 //----------------------------LoadNode::make-----------------------------------
757 // Polymorphic factory method:
758 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
759 Compile* C = gvn.C;
761 // sanity check the alias category against the created node type
762 assert(!(adr_type->isa_oopptr() &&
763 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
764 "use LoadKlassNode instead");
765 assert(!(adr_type->isa_aryptr() &&
766 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
767 "use LoadRangeNode instead");
768 switch (bt) {
769 case T_BOOLEAN:
770 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
771 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
772 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
773 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
774 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
775 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
776 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
777 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
778 case T_OBJECT:
779 #ifdef _LP64
780 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
781 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
782 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
783 } else
784 #endif
785 {
786 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
787 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
788 }
789 }
790 ShouldNotReachHere();
791 return (LoadNode*)NULL;
792 }
794 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
795 bool require_atomic = true;
796 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
797 }
802 //------------------------------hash-------------------------------------------
803 uint LoadNode::hash() const {
804 // unroll addition of interesting fields
805 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
806 }
808 //---------------------------can_see_stored_value------------------------------
809 // This routine exists to make sure this set of tests is done the same
810 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
811 // will change the graph shape in a way which makes memory alive twice at the
812 // same time (uses the Oracle model of aliasing), then some
813 // LoadXNode::Identity will fold things back to the equivalence-class model
814 // of aliasing.
815 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
816 Node* ld_adr = in(MemNode::Address);
818 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
819 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
820 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
821 atp->field() != NULL && !atp->field()->is_volatile()) {
822 uint alias_idx = atp->index();
823 bool final = atp->field()->is_final();
824 Node* result = NULL;
825 Node* current = st;
826 // Skip through chains of MemBarNodes checking the MergeMems for
827 // new states for the slice of this load. Stop once any other
828 // kind of node is encountered. Loads from final memory can skip
829 // through any kind of MemBar but normal loads shouldn't skip
830 // through MemBarAcquire since the could allow them to move out of
831 // a synchronized region.
832 while (current->is_Proj()) {
833 int opc = current->in(0)->Opcode();
834 if ((final && opc == Op_MemBarAcquire) ||
835 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
836 Node* mem = current->in(0)->in(TypeFunc::Memory);
837 if (mem->is_MergeMem()) {
838 MergeMemNode* merge = mem->as_MergeMem();
839 Node* new_st = merge->memory_at(alias_idx);
840 if (new_st == merge->base_memory()) {
841 // Keep searching
842 current = merge->base_memory();
843 continue;
844 }
845 // Save the new memory state for the slice and fall through
846 // to exit.
847 result = new_st;
848 }
849 }
850 break;
851 }
852 if (result != NULL) {
853 st = result;
854 }
855 }
858 // Loop around twice in the case Load -> Initialize -> Store.
859 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
860 for (int trip = 0; trip <= 1; trip++) {
862 if (st->is_Store()) {
863 Node* st_adr = st->in(MemNode::Address);
864 if (!phase->eqv(st_adr, ld_adr)) {
865 // Try harder before giving up... Match raw and non-raw pointers.
866 intptr_t st_off = 0;
867 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
868 if (alloc == NULL) return NULL;
869 intptr_t ld_off = 0;
870 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
871 if (alloc != allo2) return NULL;
872 if (ld_off != st_off) return NULL;
873 // At this point we have proven something like this setup:
874 // A = Allocate(...)
875 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
876 // S = StoreQ(, AddP(, A.Parm , #Off), V)
877 // (Actually, we haven't yet proven the Q's are the same.)
878 // In other words, we are loading from a casted version of
879 // the same pointer-and-offset that we stored to.
880 // Thus, we are able to replace L by V.
881 }
882 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
883 if (store_Opcode() != st->Opcode())
884 return NULL;
885 return st->in(MemNode::ValueIn);
886 }
888 intptr_t offset = 0; // scratch
890 // A load from a freshly-created object always returns zero.
891 // (This can happen after LoadNode::Ideal resets the load's memory input
892 // to find_captured_store, which returned InitializeNode::zero_memory.)
893 if (st->is_Proj() && st->in(0)->is_Allocate() &&
894 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
895 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
896 // return a zero value for the load's basic type
897 // (This is one of the few places where a generic PhaseTransform
898 // can create new nodes. Think of it as lazily manifesting
899 // virtually pre-existing constants.)
900 return phase->zerocon(memory_type());
901 }
903 // A load from an initialization barrier can match a captured store.
904 if (st->is_Proj() && st->in(0)->is_Initialize()) {
905 InitializeNode* init = st->in(0)->as_Initialize();
906 AllocateNode* alloc = init->allocation();
907 if (alloc != NULL &&
908 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
909 // examine a captured store value
910 st = init->find_captured_store(offset, memory_size(), phase);
911 if (st != NULL)
912 continue; // take one more trip around
913 }
914 }
916 break;
917 }
919 return NULL;
920 }
922 //----------------------is_instance_field_load_with_local_phi------------------
923 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
924 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
925 in(MemNode::Address)->is_AddP() ) {
926 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
927 // Only instances.
928 if( t_oop != NULL && t_oop->is_known_instance_field() &&
929 t_oop->offset() != Type::OffsetBot &&
930 t_oop->offset() != Type::OffsetTop) {
931 return true;
932 }
933 }
934 return false;
935 }
937 //------------------------------Identity---------------------------------------
938 // Loads are identity if previous store is to same address
939 Node *LoadNode::Identity( PhaseTransform *phase ) {
940 // If the previous store-maker is the right kind of Store, and the store is
941 // to the same address, then we are equal to the value stored.
942 Node* mem = in(MemNode::Memory);
943 Node* value = can_see_stored_value(mem, phase);
944 if( value ) {
945 // byte, short & char stores truncate naturally.
946 // A load has to load the truncated value which requires
947 // some sort of masking operation and that requires an
948 // Ideal call instead of an Identity call.
949 if (memory_size() < BytesPerInt) {
950 // If the input to the store does not fit with the load's result type,
951 // it must be truncated via an Ideal call.
952 if (!phase->type(value)->higher_equal(phase->type(this)))
953 return this;
954 }
955 // (This works even when value is a Con, but LoadNode::Value
956 // usually runs first, producing the singleton type of the Con.)
957 return value;
958 }
960 // Search for an existing data phi which was generated before for the same
961 // instance's field to avoid infinite genertion of phis in a loop.
962 Node *region = mem->in(0);
963 if (is_instance_field_load_with_local_phi(region)) {
964 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
965 int this_index = phase->C->get_alias_index(addr_t);
966 int this_offset = addr_t->offset();
967 int this_id = addr_t->is_oopptr()->instance_id();
968 const Type* this_type = bottom_type();
969 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
970 Node* phi = region->fast_out(i);
971 if (phi->is_Phi() && phi != mem &&
972 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
973 return phi;
974 }
975 }
976 }
978 return this;
979 }
982 // Returns true if the AliasType refers to the field that holds the
983 // cached box array. Currently only handles the IntegerCache case.
984 static bool is_autobox_cache(Compile::AliasType* atp) {
985 if (atp != NULL && atp->field() != NULL) {
986 ciField* field = atp->field();
987 ciSymbol* klass = field->holder()->name();
988 if (field->name() == ciSymbol::cache_field_name() &&
989 field->holder()->uses_default_loader() &&
990 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
991 return true;
992 }
993 }
994 return false;
995 }
997 // Fetch the base value in the autobox array
998 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
999 if (atp != NULL && atp->field() != NULL) {
1000 ciField* field = atp->field();
1001 ciSymbol* klass = field->holder()->name();
1002 if (field->name() == ciSymbol::cache_field_name() &&
1003 field->holder()->uses_default_loader() &&
1004 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1005 assert(field->is_constant(), "what?");
1006 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1007 // Fetch the box object at the base of the array and get its value
1008 ciInstance* box = array->obj_at(0)->as_instance();
1009 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1010 if (ik->nof_nonstatic_fields() == 1) {
1011 // This should be true nonstatic_field_at requires calling
1012 // nof_nonstatic_fields so check it anyway
1013 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1014 cache_offset = c.as_int();
1015 }
1016 return true;
1017 }
1018 }
1019 return false;
1020 }
1022 // Returns true if the AliasType refers to the value field of an
1023 // autobox object. Currently only handles Integer.
1024 static bool is_autobox_object(Compile::AliasType* atp) {
1025 if (atp != NULL && atp->field() != NULL) {
1026 ciField* field = atp->field();
1027 ciSymbol* klass = field->holder()->name();
1028 if (field->name() == ciSymbol::value_name() &&
1029 field->holder()->uses_default_loader() &&
1030 klass == ciSymbol::java_lang_Integer()) {
1031 return true;
1032 }
1033 }
1034 return false;
1035 }
1038 // We're loading from an object which has autobox behaviour.
1039 // If this object is result of a valueOf call we'll have a phi
1040 // merging a newly allocated object and a load from the cache.
1041 // We want to replace this load with the original incoming
1042 // argument to the valueOf call.
1043 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1044 Node* base = in(Address)->in(AddPNode::Base);
1045 if (base->is_Phi() && base->req() == 3) {
1046 AllocateNode* allocation = NULL;
1047 int allocation_index = -1;
1048 int load_index = -1;
1049 for (uint i = 1; i < base->req(); i++) {
1050 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1051 if (allocation != NULL) {
1052 allocation_index = i;
1053 load_index = 3 - allocation_index;
1054 break;
1055 }
1056 }
1057 LoadNode* load = NULL;
1058 if (allocation != NULL && base->in(load_index)->is_Load()) {
1059 load = base->in(load_index)->as_Load();
1060 }
1061 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1062 // Push the loads from the phi that comes from valueOf up
1063 // through it to allow elimination of the loads and the recovery
1064 // of the original value.
1065 Node* mem_phi = in(Memory);
1066 Node* offset = in(Address)->in(AddPNode::Offset);
1068 Node* in1 = clone();
1069 Node* in1_addr = in1->in(Address)->clone();
1070 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1071 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1072 in1_addr->set_req(AddPNode::Offset, offset);
1073 in1->set_req(0, base->in(allocation_index));
1074 in1->set_req(Address, in1_addr);
1075 in1->set_req(Memory, mem_phi->in(allocation_index));
1077 Node* in2 = clone();
1078 Node* in2_addr = in2->in(Address)->clone();
1079 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1080 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1081 in2_addr->set_req(AddPNode::Offset, offset);
1082 in2->set_req(0, base->in(load_index));
1083 in2->set_req(Address, in2_addr);
1084 in2->set_req(Memory, mem_phi->in(load_index));
1086 in1_addr = phase->transform(in1_addr);
1087 in1 = phase->transform(in1);
1088 in2_addr = phase->transform(in2_addr);
1089 in2 = phase->transform(in2);
1091 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1092 result->set_req(allocation_index, in1);
1093 result->set_req(load_index, in2);
1094 return result;
1095 }
1096 } else if (base->is_Load()) {
1097 // Eliminate the load of Integer.value for integers from the cache
1098 // array by deriving the value from the index into the array.
1099 // Capture the offset of the load and then reverse the computation.
1100 Node* load_base = base->in(Address)->in(AddPNode::Base);
1101 if (load_base != NULL) {
1102 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1103 intptr_t cache_offset;
1104 int shift = -1;
1105 Node* cache = NULL;
1106 if (is_autobox_cache(atp)) {
1107 shift = exact_log2(type2aelembytes(T_OBJECT));
1108 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1109 }
1110 if (cache != NULL && base->in(Address)->is_AddP()) {
1111 Node* elements[4];
1112 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1113 int cache_low;
1114 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1115 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1116 // Add up all the offsets making of the address of the load
1117 Node* result = elements[0];
1118 for (int i = 1; i < count; i++) {
1119 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1120 }
1121 // Remove the constant offset from the address and then
1122 // remove the scaling of the offset to recover the original index.
1123 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1124 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1125 // Peel the shift off directly but wrap it in a dummy node
1126 // since Ideal can't return existing nodes
1127 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1128 } else {
1129 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1130 }
1131 #ifdef _LP64
1132 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1133 #endif
1134 return result;
1135 }
1136 }
1137 }
1138 }
1139 return NULL;
1140 }
1142 //------------------------------split_through_phi------------------------------
1143 // Split instance field load through Phi.
1144 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1145 Node* mem = in(MemNode::Memory);
1146 Node* address = in(MemNode::Address);
1147 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1148 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1150 assert(mem->is_Phi() && (t_oop != NULL) &&
1151 t_oop->is_known_instance_field(), "invalide conditions");
1153 Node *region = mem->in(0);
1154 if (region == NULL) {
1155 return NULL; // Wait stable graph
1156 }
1157 uint cnt = mem->req();
1158 for( uint i = 1; i < cnt; i++ ) {
1159 Node *in = mem->in(i);
1160 if( in == NULL ) {
1161 return NULL; // Wait stable graph
1162 }
1163 }
1164 // Check for loop invariant.
1165 if (cnt == 3) {
1166 for( uint i = 1; i < cnt; i++ ) {
1167 Node *in = mem->in(i);
1168 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1169 if (m == mem) {
1170 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1171 return this;
1172 }
1173 }
1174 }
1175 // Split through Phi (see original code in loopopts.cpp).
1176 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1178 // Do nothing here if Identity will find a value
1179 // (to avoid infinite chain of value phis generation).
1180 if ( !phase->eqv(this, this->Identity(phase)) )
1181 return NULL;
1183 // Skip the split if the region dominates some control edge of the address.
1184 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1185 return NULL;
1187 const Type* this_type = this->bottom_type();
1188 int this_index = phase->C->get_alias_index(addr_t);
1189 int this_offset = addr_t->offset();
1190 int this_iid = addr_t->is_oopptr()->instance_id();
1191 int wins = 0;
1192 PhaseIterGVN *igvn = phase->is_IterGVN();
1193 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1194 for( uint i = 1; i < region->req(); i++ ) {
1195 Node *x;
1196 Node* the_clone = NULL;
1197 if( region->in(i) == phase->C->top() ) {
1198 x = phase->C->top(); // Dead path? Use a dead data op
1199 } else {
1200 x = this->clone(); // Else clone up the data op
1201 the_clone = x; // Remember for possible deletion.
1202 // Alter data node to use pre-phi inputs
1203 if( this->in(0) == region ) {
1204 x->set_req( 0, region->in(i) );
1205 } else {
1206 x->set_req( 0, NULL );
1207 }
1208 for( uint j = 1; j < this->req(); j++ ) {
1209 Node *in = this->in(j);
1210 if( in->is_Phi() && in->in(0) == region )
1211 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1212 }
1213 }
1214 // Check for a 'win' on some paths
1215 const Type *t = x->Value(igvn);
1217 bool singleton = t->singleton();
1219 // See comments in PhaseIdealLoop::split_thru_phi().
1220 if( singleton && t == Type::TOP ) {
1221 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1222 }
1224 if( singleton ) {
1225 wins++;
1226 x = igvn->makecon(t);
1227 } else {
1228 // We now call Identity to try to simplify the cloned node.
1229 // Note that some Identity methods call phase->type(this).
1230 // Make sure that the type array is big enough for
1231 // our new node, even though we may throw the node away.
1232 // (This tweaking with igvn only works because x is a new node.)
1233 igvn->set_type(x, t);
1234 // If x is a TypeNode, capture any more-precise type permanently into Node
1235 // othewise it will be not updated during igvn->transform since
1236 // igvn->type(x) is set to x->Value() already.
1237 x->raise_bottom_type(t);
1238 Node *y = x->Identity(igvn);
1239 if( y != x ) {
1240 wins++;
1241 x = y;
1242 } else {
1243 y = igvn->hash_find(x);
1244 if( y ) {
1245 wins++;
1246 x = y;
1247 } else {
1248 // Else x is a new node we are keeping
1249 // We do not need register_new_node_with_optimizer
1250 // because set_type has already been called.
1251 igvn->_worklist.push(x);
1252 }
1253 }
1254 }
1255 if (x != the_clone && the_clone != NULL)
1256 igvn->remove_dead_node(the_clone);
1257 phi->set_req(i, x);
1258 }
1259 if( wins > 0 ) {
1260 // Record Phi
1261 igvn->register_new_node_with_optimizer(phi);
1262 return phi;
1263 }
1264 igvn->remove_dead_node(phi);
1265 return NULL;
1266 }
1268 //------------------------------Ideal------------------------------------------
1269 // If the load is from Field memory and the pointer is non-null, we can
1270 // zero out the control input.
1271 // If the offset is constant and the base is an object allocation,
1272 // try to hook me up to the exact initializing store.
1273 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1274 Node* p = MemNode::Ideal_common(phase, can_reshape);
1275 if (p) return (p == NodeSentinel) ? NULL : p;
1277 Node* ctrl = in(MemNode::Control);
1278 Node* address = in(MemNode::Address);
1280 // Skip up past a SafePoint control. Cannot do this for Stores because
1281 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1282 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1283 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1284 ctrl = ctrl->in(0);
1285 set_req(MemNode::Control,ctrl);
1286 }
1288 // Check for useless control edge in some common special cases
1289 if (in(MemNode::Control) != NULL) {
1290 intptr_t ignore = 0;
1291 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1292 if (base != NULL
1293 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1294 && all_controls_dominate(base, phase->C->start())) {
1295 // A method-invariant, non-null address (constant or 'this' argument).
1296 set_req(MemNode::Control, NULL);
1297 }
1298 }
1300 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1301 Node* base = in(Address)->in(AddPNode::Base);
1302 if (base != NULL) {
1303 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1304 if (is_autobox_object(atp)) {
1305 Node* result = eliminate_autobox(phase);
1306 if (result != NULL) return result;
1307 }
1308 }
1309 }
1311 Node* mem = in(MemNode::Memory);
1312 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1314 if (addr_t != NULL) {
1315 // try to optimize our memory input
1316 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1317 if (opt_mem != mem) {
1318 set_req(MemNode::Memory, opt_mem);
1319 return this;
1320 }
1321 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1322 if (can_reshape && opt_mem->is_Phi() &&
1323 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1324 // Split instance field load through Phi.
1325 Node* result = split_through_phi(phase);
1326 if (result != NULL) return result;
1327 }
1328 }
1330 // Check for prior store with a different base or offset; make Load
1331 // independent. Skip through any number of them. Bail out if the stores
1332 // are in an endless dead cycle and report no progress. This is a key
1333 // transform for Reflection. However, if after skipping through the Stores
1334 // we can't then fold up against a prior store do NOT do the transform as
1335 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1336 // array memory alive twice: once for the hoisted Load and again after the
1337 // bypassed Store. This situation only works if EVERYBODY who does
1338 // anti-dependence work knows how to bypass. I.e. we need all
1339 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1340 // the alias index stuff. So instead, peek through Stores and IFF we can
1341 // fold up, do so.
1342 Node* prev_mem = find_previous_store(phase);
1343 // Steps (a), (b): Walk past independent stores to find an exact match.
1344 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1345 // (c) See if we can fold up on the spot, but don't fold up here.
1346 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1347 // just return a prior value, which is done by Identity calls.
1348 if (can_see_stored_value(prev_mem, phase)) {
1349 // Make ready for step (d):
1350 set_req(MemNode::Memory, prev_mem);
1351 return this;
1352 }
1353 }
1355 return NULL; // No further progress
1356 }
1358 // Helper to recognize certain Klass fields which are invariant across
1359 // some group of array types (e.g., int[] or all T[] where T < Object).
1360 const Type*
1361 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1362 ciKlass* klass) const {
1363 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1364 // The field is Klass::_modifier_flags. Return its (constant) value.
1365 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1366 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1367 return TypeInt::make(klass->modifier_flags());
1368 }
1369 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1370 // The field is Klass::_access_flags. Return its (constant) value.
1371 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1372 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1373 return TypeInt::make(klass->access_flags());
1374 }
1375 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1376 // The field is Klass::_layout_helper. Return its constant value if known.
1377 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1378 return TypeInt::make(klass->layout_helper());
1379 }
1381 // No match.
1382 return NULL;
1383 }
1385 //------------------------------Value-----------------------------------------
1386 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1387 // Either input is TOP ==> the result is TOP
1388 Node* mem = in(MemNode::Memory);
1389 const Type *t1 = phase->type(mem);
1390 if (t1 == Type::TOP) return Type::TOP;
1391 Node* adr = in(MemNode::Address);
1392 const TypePtr* tp = phase->type(adr)->isa_ptr();
1393 if (tp == NULL || tp->empty()) return Type::TOP;
1394 int off = tp->offset();
1395 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1397 // Try to guess loaded type from pointer type
1398 if (tp->base() == Type::AryPtr) {
1399 const Type *t = tp->is_aryptr()->elem();
1400 // Don't do this for integer types. There is only potential profit if
1401 // the element type t is lower than _type; that is, for int types, if _type is
1402 // more restrictive than t. This only happens here if one is short and the other
1403 // char (both 16 bits), and in those cases we've made an intentional decision
1404 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1405 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1406 //
1407 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1408 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1409 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1410 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1411 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1412 // In fact, that could have been the original type of p1, and p1 could have
1413 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1414 // expression (LShiftL quux 3) independently optimized to the constant 8.
1415 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1416 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1417 // t might actually be lower than _type, if _type is a unique
1418 // concrete subclass of abstract class t.
1419 // Make sure the reference is not into the header, by comparing
1420 // the offset against the offset of the start of the array's data.
1421 // Different array types begin at slightly different offsets (12 vs. 16).
1422 // We choose T_BYTE as an example base type that is least restrictive
1423 // as to alignment, which will therefore produce the smallest
1424 // possible base offset.
1425 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1426 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1427 const Type* jt = t->join(_type);
1428 // In any case, do not allow the join, per se, to empty out the type.
1429 if (jt->empty() && !t->empty()) {
1430 // This can happen if a interface-typed array narrows to a class type.
1431 jt = _type;
1432 }
1434 if (EliminateAutoBox) {
1435 // The pointers in the autobox arrays are always non-null
1436 Node* base = in(Address)->in(AddPNode::Base);
1437 if (base != NULL) {
1438 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1439 if (is_autobox_cache(atp)) {
1440 return jt->join(TypePtr::NOTNULL)->is_ptr();
1441 }
1442 }
1443 }
1444 return jt;
1445 }
1446 }
1447 } else if (tp->base() == Type::InstPtr) {
1448 assert( off != Type::OffsetBot ||
1449 // arrays can be cast to Objects
1450 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1451 // unsafe field access may not have a constant offset
1452 phase->C->has_unsafe_access(),
1453 "Field accesses must be precise" );
1454 // For oop loads, we expect the _type to be precise
1455 } else if (tp->base() == Type::KlassPtr) {
1456 assert( off != Type::OffsetBot ||
1457 // arrays can be cast to Objects
1458 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1459 // also allow array-loading from the primary supertype
1460 // array during subtype checks
1461 Opcode() == Op_LoadKlass,
1462 "Field accesses must be precise" );
1463 // For klass/static loads, we expect the _type to be precise
1464 }
1466 const TypeKlassPtr *tkls = tp->isa_klassptr();
1467 if (tkls != NULL && !StressReflectiveCode) {
1468 ciKlass* klass = tkls->klass();
1469 if (klass->is_loaded() && tkls->klass_is_exact()) {
1470 // We are loading a field from a Klass metaobject whose identity
1471 // is known at compile time (the type is "exact" or "precise").
1472 // Check for fields we know are maintained as constants by the VM.
1473 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1474 // The field is Klass::_super_check_offset. Return its (constant) value.
1475 // (Folds up type checking code.)
1476 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1477 return TypeInt::make(klass->super_check_offset());
1478 }
1479 // Compute index into primary_supers array
1480 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1481 // Check for overflowing; use unsigned compare to handle the negative case.
1482 if( depth < ciKlass::primary_super_limit() ) {
1483 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1484 // (Folds up type checking code.)
1485 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1486 ciKlass *ss = klass->super_of_depth(depth);
1487 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1488 }
1489 const Type* aift = load_array_final_field(tkls, klass);
1490 if (aift != NULL) return aift;
1491 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1492 && klass->is_array_klass()) {
1493 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1494 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1495 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1496 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1497 }
1498 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1499 // The field is Klass::_java_mirror. Return its (constant) value.
1500 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1501 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1502 return TypeInstPtr::make(klass->java_mirror());
1503 }
1504 }
1506 // We can still check if we are loading from the primary_supers array at a
1507 // shallow enough depth. Even though the klass is not exact, entries less
1508 // than or equal to its super depth are correct.
1509 if (klass->is_loaded() ) {
1510 ciType *inner = klass->klass();
1511 while( inner->is_obj_array_klass() )
1512 inner = inner->as_obj_array_klass()->base_element_type();
1513 if( inner->is_instance_klass() &&
1514 !inner->as_instance_klass()->flags().is_interface() ) {
1515 // Compute index into primary_supers array
1516 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1517 // Check for overflowing; use unsigned compare to handle the negative case.
1518 if( depth < ciKlass::primary_super_limit() &&
1519 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1520 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1521 // (Folds up type checking code.)
1522 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1523 ciKlass *ss = klass->super_of_depth(depth);
1524 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1525 }
1526 }
1527 }
1529 // If the type is enough to determine that the thing is not an array,
1530 // we can give the layout_helper a positive interval type.
1531 // This will help short-circuit some reflective code.
1532 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1533 && !klass->is_array_klass() // not directly typed as an array
1534 && !klass->is_interface() // specifically not Serializable & Cloneable
1535 && !klass->is_java_lang_Object() // not the supertype of all T[]
1536 ) {
1537 // Note: When interfaces are reliable, we can narrow the interface
1538 // test to (klass != Serializable && klass != Cloneable).
1539 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1540 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1541 // The key property of this type is that it folds up tests
1542 // for array-ness, since it proves that the layout_helper is positive.
1543 // Thus, a generic value like the basic object layout helper works fine.
1544 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1545 }
1546 }
1548 // If we are loading from a freshly-allocated object, produce a zero,
1549 // if the load is provably beyond the header of the object.
1550 // (Also allow a variable load from a fresh array to produce zero.)
1551 if (ReduceFieldZeroing) {
1552 Node* value = can_see_stored_value(mem,phase);
1553 if (value != NULL && value->is_Con())
1554 return value->bottom_type();
1555 }
1557 const TypeOopPtr *tinst = tp->isa_oopptr();
1558 if (tinst != NULL && tinst->is_known_instance_field()) {
1559 // If we have an instance type and our memory input is the
1560 // programs's initial memory state, there is no matching store,
1561 // so just return a zero of the appropriate type
1562 Node *mem = in(MemNode::Memory);
1563 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1564 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1565 return Type::get_zero_type(_type->basic_type());
1566 }
1567 }
1568 return _type;
1569 }
1571 //------------------------------match_edge-------------------------------------
1572 // Do we Match on this edge index or not? Match only the address.
1573 uint LoadNode::match_edge(uint idx) const {
1574 return idx == MemNode::Address;
1575 }
1577 //--------------------------LoadBNode::Ideal--------------------------------------
1578 //
1579 // If the previous store is to the same address as this load,
1580 // and the value stored was larger than a byte, replace this load
1581 // with the value stored truncated to a byte. If no truncation is
1582 // needed, the replacement is done in LoadNode::Identity().
1583 //
1584 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1585 Node* mem = in(MemNode::Memory);
1586 Node* value = can_see_stored_value(mem,phase);
1587 if( value && !phase->type(value)->higher_equal( _type ) ) {
1588 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1589 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1590 }
1591 // Identity call will handle the case where truncation is not needed.
1592 return LoadNode::Ideal(phase, can_reshape);
1593 }
1595 //--------------------------LoadCNode::Ideal--------------------------------------
1596 //
1597 // If the previous store is to the same address as this load,
1598 // and the value stored was larger than a char, replace this load
1599 // with the value stored truncated to a char. If no truncation is
1600 // needed, the replacement is done in LoadNode::Identity().
1601 //
1602 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1603 Node* mem = in(MemNode::Memory);
1604 Node* value = can_see_stored_value(mem,phase);
1605 if( value && !phase->type(value)->higher_equal( _type ) )
1606 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1607 // Identity call will handle the case where truncation is not needed.
1608 return LoadNode::Ideal(phase, can_reshape);
1609 }
1611 //--------------------------LoadSNode::Ideal--------------------------------------
1612 //
1613 // If the previous store is to the same address as this load,
1614 // and the value stored was larger than a short, replace this load
1615 // with the value stored truncated to a short. If no truncation is
1616 // needed, the replacement is done in LoadNode::Identity().
1617 //
1618 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1619 Node* mem = in(MemNode::Memory);
1620 Node* value = can_see_stored_value(mem,phase);
1621 if( value && !phase->type(value)->higher_equal( _type ) ) {
1622 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1623 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1624 }
1625 // Identity call will handle the case where truncation is not needed.
1626 return LoadNode::Ideal(phase, can_reshape);
1627 }
1629 //=============================================================================
1630 //----------------------------LoadKlassNode::make------------------------------
1631 // Polymorphic factory method:
1632 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1633 Compile* C = gvn.C;
1634 Node *ctl = NULL;
1635 // sanity check the alias category against the created node type
1636 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1637 assert(adr_type != NULL, "expecting TypeOopPtr");
1638 #ifdef _LP64
1639 if (adr_type->is_ptr_to_narrowoop()) {
1640 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1641 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1642 }
1643 #endif
1644 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1645 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1646 }
1648 //------------------------------Value------------------------------------------
1649 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1650 return klass_value_common(phase);
1651 }
1653 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1654 // Either input is TOP ==> the result is TOP
1655 const Type *t1 = phase->type( in(MemNode::Memory) );
1656 if (t1 == Type::TOP) return Type::TOP;
1657 Node *adr = in(MemNode::Address);
1658 const Type *t2 = phase->type( adr );
1659 if (t2 == Type::TOP) return Type::TOP;
1660 const TypePtr *tp = t2->is_ptr();
1661 if (TypePtr::above_centerline(tp->ptr()) ||
1662 tp->ptr() == TypePtr::Null) return Type::TOP;
1664 // Return a more precise klass, if possible
1665 const TypeInstPtr *tinst = tp->isa_instptr();
1666 if (tinst != NULL) {
1667 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1668 int offset = tinst->offset();
1669 if (ik == phase->C->env()->Class_klass()
1670 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1671 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1672 // We are loading a special hidden field from a Class mirror object,
1673 // the field which points to the VM's Klass metaobject.
1674 ciType* t = tinst->java_mirror_type();
1675 // java_mirror_type returns non-null for compile-time Class constants.
1676 if (t != NULL) {
1677 // constant oop => constant klass
1678 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1679 return TypeKlassPtr::make(ciArrayKlass::make(t));
1680 }
1681 if (!t->is_klass()) {
1682 // a primitive Class (e.g., int.class) has NULL for a klass field
1683 return TypePtr::NULL_PTR;
1684 }
1685 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1686 return TypeKlassPtr::make(t->as_klass());
1687 }
1688 // non-constant mirror, so we can't tell what's going on
1689 }
1690 if( !ik->is_loaded() )
1691 return _type; // Bail out if not loaded
1692 if (offset == oopDesc::klass_offset_in_bytes()) {
1693 if (tinst->klass_is_exact()) {
1694 return TypeKlassPtr::make(ik);
1695 }
1696 // See if we can become precise: no subklasses and no interface
1697 // (Note: We need to support verified interfaces.)
1698 if (!ik->is_interface() && !ik->has_subklass()) {
1699 //assert(!UseExactTypes, "this code should be useless with exact types");
1700 // Add a dependence; if any subclass added we need to recompile
1701 if (!ik->is_final()) {
1702 // %%% should use stronger assert_unique_concrete_subtype instead
1703 phase->C->dependencies()->assert_leaf_type(ik);
1704 }
1705 // Return precise klass
1706 return TypeKlassPtr::make(ik);
1707 }
1709 // Return root of possible klass
1710 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1711 }
1712 }
1714 // Check for loading klass from an array
1715 const TypeAryPtr *tary = tp->isa_aryptr();
1716 if( tary != NULL ) {
1717 ciKlass *tary_klass = tary->klass();
1718 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1719 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1720 if (tary->klass_is_exact()) {
1721 return TypeKlassPtr::make(tary_klass);
1722 }
1723 ciArrayKlass *ak = tary->klass()->as_array_klass();
1724 // If the klass is an object array, we defer the question to the
1725 // array component klass.
1726 if( ak->is_obj_array_klass() ) {
1727 assert( ak->is_loaded(), "" );
1728 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1729 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1730 ciInstanceKlass* ik = base_k->as_instance_klass();
1731 // See if we can become precise: no subklasses and no interface
1732 if (!ik->is_interface() && !ik->has_subklass()) {
1733 //assert(!UseExactTypes, "this code should be useless with exact types");
1734 // Add a dependence; if any subclass added we need to recompile
1735 if (!ik->is_final()) {
1736 phase->C->dependencies()->assert_leaf_type(ik);
1737 }
1738 // Return precise array klass
1739 return TypeKlassPtr::make(ak);
1740 }
1741 }
1742 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1743 } else { // Found a type-array?
1744 //assert(!UseExactTypes, "this code should be useless with exact types");
1745 assert( ak->is_type_array_klass(), "" );
1746 return TypeKlassPtr::make(ak); // These are always precise
1747 }
1748 }
1749 }
1751 // Check for loading klass from an array klass
1752 const TypeKlassPtr *tkls = tp->isa_klassptr();
1753 if (tkls != NULL && !StressReflectiveCode) {
1754 ciKlass* klass = tkls->klass();
1755 if( !klass->is_loaded() )
1756 return _type; // Bail out if not loaded
1757 if( klass->is_obj_array_klass() &&
1758 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1759 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1760 // // Always returning precise element type is incorrect,
1761 // // e.g., element type could be object and array may contain strings
1762 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1764 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1765 // according to the element type's subclassing.
1766 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1767 }
1768 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1769 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1770 ciKlass* sup = klass->as_instance_klass()->super();
1771 // The field is Klass::_super. Return its (constant) value.
1772 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1773 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1774 }
1775 }
1777 // Bailout case
1778 return LoadNode::Value(phase);
1779 }
1781 //------------------------------Identity---------------------------------------
1782 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1783 // Also feed through the klass in Allocate(...klass...)._klass.
1784 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1785 return klass_identity_common(phase);
1786 }
1788 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1789 Node* x = LoadNode::Identity(phase);
1790 if (x != this) return x;
1792 // Take apart the address into an oop and and offset.
1793 // Return 'this' if we cannot.
1794 Node* adr = in(MemNode::Address);
1795 intptr_t offset = 0;
1796 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1797 if (base == NULL) return this;
1798 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1799 if (toop == NULL) return this;
1801 // We can fetch the klass directly through an AllocateNode.
1802 // This works even if the klass is not constant (clone or newArray).
1803 if (offset == oopDesc::klass_offset_in_bytes()) {
1804 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1805 if (allocated_klass != NULL) {
1806 return allocated_klass;
1807 }
1808 }
1810 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1811 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1812 // See inline_native_Class_query for occurrences of these patterns.
1813 // Java Example: x.getClass().isAssignableFrom(y)
1814 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1815 //
1816 // This improves reflective code, often making the Class
1817 // mirror go completely dead. (Current exception: Class
1818 // mirrors may appear in debug info, but we could clean them out by
1819 // introducing a new debug info operator for klassOop.java_mirror).
1820 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1821 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1822 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1823 // We are loading a special hidden field from a Class mirror,
1824 // the field which points to its Klass or arrayKlass metaobject.
1825 if (base->is_Load()) {
1826 Node* adr2 = base->in(MemNode::Address);
1827 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1828 if (tkls != NULL && !tkls->empty()
1829 && (tkls->klass()->is_instance_klass() ||
1830 tkls->klass()->is_array_klass())
1831 && adr2->is_AddP()
1832 ) {
1833 int mirror_field = Klass::java_mirror_offset_in_bytes();
1834 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1835 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1836 }
1837 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1838 return adr2->in(AddPNode::Base);
1839 }
1840 }
1841 }
1842 }
1844 return this;
1845 }
1848 //------------------------------Value------------------------------------------
1849 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1850 const Type *t = klass_value_common(phase);
1851 if (t == Type::TOP)
1852 return t;
1854 return t->make_narrowoop();
1855 }
1857 //------------------------------Identity---------------------------------------
1858 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1859 // Also feed through the klass in Allocate(...klass...)._klass.
1860 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1861 Node *x = klass_identity_common(phase);
1863 const Type *t = phase->type( x );
1864 if( t == Type::TOP ) return x;
1865 if( t->isa_narrowoop()) return x;
1867 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
1868 }
1870 //------------------------------Value-----------------------------------------
1871 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1872 // Either input is TOP ==> the result is TOP
1873 const Type *t1 = phase->type( in(MemNode::Memory) );
1874 if( t1 == Type::TOP ) return Type::TOP;
1875 Node *adr = in(MemNode::Address);
1876 const Type *t2 = phase->type( adr );
1877 if( t2 == Type::TOP ) return Type::TOP;
1878 const TypePtr *tp = t2->is_ptr();
1879 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1880 const TypeAryPtr *tap = tp->isa_aryptr();
1881 if( !tap ) return _type;
1882 return tap->size();
1883 }
1885 //------------------------------Identity---------------------------------------
1886 // Feed through the length in AllocateArray(...length...)._length.
1887 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1888 Node* x = LoadINode::Identity(phase);
1889 if (x != this) return x;
1891 // Take apart the address into an oop and and offset.
1892 // Return 'this' if we cannot.
1893 Node* adr = in(MemNode::Address);
1894 intptr_t offset = 0;
1895 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1896 if (base == NULL) return this;
1897 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1898 if (tary == NULL) return this;
1900 // We can fetch the length directly through an AllocateArrayNode.
1901 // This works even if the length is not constant (clone or newArray).
1902 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1903 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1904 if (allocated_length != NULL) {
1905 return allocated_length;
1906 }
1907 }
1909 return this;
1911 }
1912 //=============================================================================
1913 //---------------------------StoreNode::make-----------------------------------
1914 // Polymorphic factory method:
1915 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1916 Compile* C = gvn.C;
1918 switch (bt) {
1919 case T_BOOLEAN:
1920 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1921 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1922 case T_CHAR:
1923 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1924 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1925 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1926 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1927 case T_ADDRESS:
1928 case T_OBJECT:
1929 #ifdef _LP64
1930 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
1931 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1932 adr->bottom_type()->isa_rawptr())) {
1933 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
1934 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
1935 } else
1936 #endif
1937 {
1938 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1939 }
1940 }
1941 ShouldNotReachHere();
1942 return (StoreNode*)NULL;
1943 }
1945 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1946 bool require_atomic = true;
1947 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1948 }
1951 //--------------------------bottom_type----------------------------------------
1952 const Type *StoreNode::bottom_type() const {
1953 return Type::MEMORY;
1954 }
1956 //------------------------------hash-------------------------------------------
1957 uint StoreNode::hash() const {
1958 // unroll addition of interesting fields
1959 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1961 // Since they are not commoned, do not hash them:
1962 return NO_HASH;
1963 }
1965 //------------------------------Ideal------------------------------------------
1966 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1967 // When a store immediately follows a relevant allocation/initialization,
1968 // try to capture it into the initialization, or hoist it above.
1969 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1970 Node* p = MemNode::Ideal_common(phase, can_reshape);
1971 if (p) return (p == NodeSentinel) ? NULL : p;
1973 Node* mem = in(MemNode::Memory);
1974 Node* address = in(MemNode::Address);
1976 // Back-to-back stores to same address? Fold em up.
1977 // Generally unsafe if I have intervening uses...
1978 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1979 // Looking at a dead closed cycle of memory?
1980 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1982 assert(Opcode() == mem->Opcode() ||
1983 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1984 "no mismatched stores, except on raw memory");
1986 if (mem->outcnt() == 1 && // check for intervening uses
1987 mem->as_Store()->memory_size() <= this->memory_size()) {
1988 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1989 // For example, 'mem' might be the final state at a conditional return.
1990 // Or, 'mem' might be used by some node which is live at the same time
1991 // 'this' is live, which might be unschedulable. So, require exactly
1992 // ONE user, the 'this' store, until such time as we clone 'mem' for
1993 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1994 if (can_reshape) { // (%%% is this an anachronism?)
1995 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1996 phase->is_IterGVN());
1997 } else {
1998 // It's OK to do this in the parser, since DU info is always accurate,
1999 // and the parser always refers to nodes via SafePointNode maps.
2000 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2001 }
2002 return this;
2003 }
2004 }
2006 // Capture an unaliased, unconditional, simple store into an initializer.
2007 // Or, if it is independent of the allocation, hoist it above the allocation.
2008 if (ReduceFieldZeroing && /*can_reshape &&*/
2009 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2010 InitializeNode* init = mem->in(0)->as_Initialize();
2011 intptr_t offset = init->can_capture_store(this, phase);
2012 if (offset > 0) {
2013 Node* moved = init->capture_store(this, offset, phase);
2014 // If the InitializeNode captured me, it made a raw copy of me,
2015 // and I need to disappear.
2016 if (moved != NULL) {
2017 // %%% hack to ensure that Ideal returns a new node:
2018 mem = MergeMemNode::make(phase->C, mem);
2019 return mem; // fold me away
2020 }
2021 }
2022 }
2024 return NULL; // No further progress
2025 }
2027 //------------------------------Value-----------------------------------------
2028 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2029 // Either input is TOP ==> the result is TOP
2030 const Type *t1 = phase->type( in(MemNode::Memory) );
2031 if( t1 == Type::TOP ) return Type::TOP;
2032 const Type *t2 = phase->type( in(MemNode::Address) );
2033 if( t2 == Type::TOP ) return Type::TOP;
2034 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2035 if( t3 == Type::TOP ) return Type::TOP;
2036 return Type::MEMORY;
2037 }
2039 //------------------------------Identity---------------------------------------
2040 // Remove redundant stores:
2041 // Store(m, p, Load(m, p)) changes to m.
2042 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2043 Node *StoreNode::Identity( PhaseTransform *phase ) {
2044 Node* mem = in(MemNode::Memory);
2045 Node* adr = in(MemNode::Address);
2046 Node* val = in(MemNode::ValueIn);
2048 // Load then Store? Then the Store is useless
2049 if (val->is_Load() &&
2050 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2051 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2052 val->as_Load()->store_Opcode() == Opcode()) {
2053 return mem;
2054 }
2056 // Two stores in a row of the same value?
2057 if (mem->is_Store() &&
2058 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2059 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2060 mem->Opcode() == Opcode()) {
2061 return mem;
2062 }
2064 // Store of zero anywhere into a freshly-allocated object?
2065 // Then the store is useless.
2066 // (It must already have been captured by the InitializeNode.)
2067 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2068 // a newly allocated object is already all-zeroes everywhere
2069 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2070 return mem;
2071 }
2073 // the store may also apply to zero-bits in an earlier object
2074 Node* prev_mem = find_previous_store(phase);
2075 // Steps (a), (b): Walk past independent stores to find an exact match.
2076 if (prev_mem != NULL) {
2077 Node* prev_val = can_see_stored_value(prev_mem, phase);
2078 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2079 // prev_val and val might differ by a cast; it would be good
2080 // to keep the more informative of the two.
2081 return mem;
2082 }
2083 }
2084 }
2086 return this;
2087 }
2089 //------------------------------match_edge-------------------------------------
2090 // Do we Match on this edge index or not? Match only memory & value
2091 uint StoreNode::match_edge(uint idx) const {
2092 return idx == MemNode::Address || idx == MemNode::ValueIn;
2093 }
2095 //------------------------------cmp--------------------------------------------
2096 // Do not common stores up together. They generally have to be split
2097 // back up anyways, so do not bother.
2098 uint StoreNode::cmp( const Node &n ) const {
2099 return (&n == this); // Always fail except on self
2100 }
2102 //------------------------------Ideal_masked_input-----------------------------
2103 // Check for a useless mask before a partial-word store
2104 // (StoreB ... (AndI valIn conIa) )
2105 // If (conIa & mask == mask) this simplifies to
2106 // (StoreB ... (valIn) )
2107 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2108 Node *val = in(MemNode::ValueIn);
2109 if( val->Opcode() == Op_AndI ) {
2110 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2111 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2112 set_req(MemNode::ValueIn, val->in(1));
2113 return this;
2114 }
2115 }
2116 return NULL;
2117 }
2120 //------------------------------Ideal_sign_extended_input----------------------
2121 // Check for useless sign-extension before a partial-word store
2122 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2123 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2124 // (StoreB ... (valIn) )
2125 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2126 Node *val = in(MemNode::ValueIn);
2127 if( val->Opcode() == Op_RShiftI ) {
2128 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2129 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2130 Node *shl = val->in(1);
2131 if( shl->Opcode() == Op_LShiftI ) {
2132 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2133 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2134 set_req(MemNode::ValueIn, shl->in(1));
2135 return this;
2136 }
2137 }
2138 }
2139 }
2140 return NULL;
2141 }
2143 //------------------------------value_never_loaded-----------------------------------
2144 // Determine whether there are any possible loads of the value stored.
2145 // For simplicity, we actually check if there are any loads from the
2146 // address stored to, not just for loads of the value stored by this node.
2147 //
2148 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2149 Node *adr = in(Address);
2150 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2151 if (adr_oop == NULL)
2152 return false;
2153 if (!adr_oop->is_known_instance_field())
2154 return false; // if not a distinct instance, there may be aliases of the address
2155 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2156 Node *use = adr->fast_out(i);
2157 int opc = use->Opcode();
2158 if (use->is_Load() || use->is_LoadStore()) {
2159 return false;
2160 }
2161 }
2162 return true;
2163 }
2165 //=============================================================================
2166 //------------------------------Ideal------------------------------------------
2167 // If the store is from an AND mask that leaves the low bits untouched, then
2168 // we can skip the AND operation. If the store is from a sign-extension
2169 // (a left shift, then right shift) we can skip both.
2170 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2171 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2172 if( progress != NULL ) return progress;
2174 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2175 if( progress != NULL ) return progress;
2177 // Finally check the default case
2178 return StoreNode::Ideal(phase, can_reshape);
2179 }
2181 //=============================================================================
2182 //------------------------------Ideal------------------------------------------
2183 // If the store is from an AND mask that leaves the low bits untouched, then
2184 // we can skip the AND operation
2185 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2186 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2187 if( progress != NULL ) return progress;
2189 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2190 if( progress != NULL ) return progress;
2192 // Finally check the default case
2193 return StoreNode::Ideal(phase, can_reshape);
2194 }
2196 //=============================================================================
2197 //------------------------------Identity---------------------------------------
2198 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2199 // No need to card mark when storing a null ptr
2200 Node* my_store = in(MemNode::OopStore);
2201 if (my_store->is_Store()) {
2202 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2203 if( t1 == TypePtr::NULL_PTR ) {
2204 return in(MemNode::Memory);
2205 }
2206 }
2207 return this;
2208 }
2210 //------------------------------Value-----------------------------------------
2211 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2212 // Either input is TOP ==> the result is TOP
2213 const Type *t = phase->type( in(MemNode::Memory) );
2214 if( t == Type::TOP ) return Type::TOP;
2215 t = phase->type( in(MemNode::Address) );
2216 if( t == Type::TOP ) return Type::TOP;
2217 t = phase->type( in(MemNode::ValueIn) );
2218 if( t == Type::TOP ) return Type::TOP;
2219 // If extra input is TOP ==> the result is TOP
2220 t = phase->type( in(MemNode::OopStore) );
2221 if( t == Type::TOP ) return Type::TOP;
2223 return StoreNode::Value( phase );
2224 }
2227 //=============================================================================
2228 //----------------------------------SCMemProjNode------------------------------
2229 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2230 {
2231 return bottom_type();
2232 }
2234 //=============================================================================
2235 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2236 init_req(MemNode::Control, c );
2237 init_req(MemNode::Memory , mem);
2238 init_req(MemNode::Address, adr);
2239 init_req(MemNode::ValueIn, val);
2240 init_req( ExpectedIn, ex );
2241 init_class_id(Class_LoadStore);
2243 }
2245 //=============================================================================
2246 //-------------------------------adr_type--------------------------------------
2247 // Do we Match on this edge index or not? Do not match memory
2248 const TypePtr* ClearArrayNode::adr_type() const {
2249 Node *adr = in(3);
2250 return MemNode::calculate_adr_type(adr->bottom_type());
2251 }
2253 //------------------------------match_edge-------------------------------------
2254 // Do we Match on this edge index or not? Do not match memory
2255 uint ClearArrayNode::match_edge(uint idx) const {
2256 return idx > 1;
2257 }
2259 //------------------------------Identity---------------------------------------
2260 // Clearing a zero length array does nothing
2261 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2262 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2263 }
2265 //------------------------------Idealize---------------------------------------
2266 // Clearing a short array is faster with stores
2267 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2268 const int unit = BytesPerLong;
2269 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2270 if (!t) return NULL;
2271 if (!t->is_con()) return NULL;
2272 intptr_t raw_count = t->get_con();
2273 intptr_t size = raw_count;
2274 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2275 // Clearing nothing uses the Identity call.
2276 // Negative clears are possible on dead ClearArrays
2277 // (see jck test stmt114.stmt11402.val).
2278 if (size <= 0 || size % unit != 0) return NULL;
2279 intptr_t count = size / unit;
2280 // Length too long; use fast hardware clear
2281 if (size > Matcher::init_array_short_size) return NULL;
2282 Node *mem = in(1);
2283 if( phase->type(mem)==Type::TOP ) return NULL;
2284 Node *adr = in(3);
2285 const Type* at = phase->type(adr);
2286 if( at==Type::TOP ) return NULL;
2287 const TypePtr* atp = at->isa_ptr();
2288 // adjust atp to be the correct array element address type
2289 if (atp == NULL) atp = TypePtr::BOTTOM;
2290 else atp = atp->add_offset(Type::OffsetBot);
2291 // Get base for derived pointer purposes
2292 if( adr->Opcode() != Op_AddP ) Unimplemented();
2293 Node *base = adr->in(1);
2295 Node *zero = phase->makecon(TypeLong::ZERO);
2296 Node *off = phase->MakeConX(BytesPerLong);
2297 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2298 count--;
2299 while( count-- ) {
2300 mem = phase->transform(mem);
2301 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2302 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2303 }
2304 return mem;
2305 }
2307 //----------------------------clear_memory-------------------------------------
2308 // Generate code to initialize object storage to zero.
2309 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2310 intptr_t start_offset,
2311 Node* end_offset,
2312 PhaseGVN* phase) {
2313 Compile* C = phase->C;
2314 intptr_t offset = start_offset;
2316 int unit = BytesPerLong;
2317 if ((offset % unit) != 0) {
2318 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2319 adr = phase->transform(adr);
2320 const TypePtr* atp = TypeRawPtr::BOTTOM;
2321 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2322 mem = phase->transform(mem);
2323 offset += BytesPerInt;
2324 }
2325 assert((offset % unit) == 0, "");
2327 // Initialize the remaining stuff, if any, with a ClearArray.
2328 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2329 }
2331 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2332 Node* start_offset,
2333 Node* end_offset,
2334 PhaseGVN* phase) {
2335 if (start_offset == end_offset) {
2336 // nothing to do
2337 return mem;
2338 }
2340 Compile* C = phase->C;
2341 int unit = BytesPerLong;
2342 Node* zbase = start_offset;
2343 Node* zend = end_offset;
2345 // Scale to the unit required by the CPU:
2346 if (!Matcher::init_array_count_is_in_bytes) {
2347 Node* shift = phase->intcon(exact_log2(unit));
2348 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2349 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2350 }
2352 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2353 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2355 // Bulk clear double-words
2356 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2357 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2358 return phase->transform(mem);
2359 }
2361 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2362 intptr_t start_offset,
2363 intptr_t end_offset,
2364 PhaseGVN* phase) {
2365 if (start_offset == end_offset) {
2366 // nothing to do
2367 return mem;
2368 }
2370 Compile* C = phase->C;
2371 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2372 intptr_t done_offset = end_offset;
2373 if ((done_offset % BytesPerLong) != 0) {
2374 done_offset -= BytesPerInt;
2375 }
2376 if (done_offset > start_offset) {
2377 mem = clear_memory(ctl, mem, dest,
2378 start_offset, phase->MakeConX(done_offset), phase);
2379 }
2380 if (done_offset < end_offset) { // emit the final 32-bit store
2381 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2382 adr = phase->transform(adr);
2383 const TypePtr* atp = TypeRawPtr::BOTTOM;
2384 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2385 mem = phase->transform(mem);
2386 done_offset += BytesPerInt;
2387 }
2388 assert(done_offset == end_offset, "");
2389 return mem;
2390 }
2392 //=============================================================================
2393 // Do we match on this edge? No memory edges
2394 uint StrCompNode::match_edge(uint idx) const {
2395 return idx == 5 || idx == 6;
2396 }
2398 //------------------------------Ideal------------------------------------------
2399 // Return a node which is more "ideal" than the current node. Strip out
2400 // control copies
2401 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2402 return remove_dead_region(phase, can_reshape) ? this : NULL;
2403 }
2405 //------------------------------Ideal------------------------------------------
2406 // Return a node which is more "ideal" than the current node. Strip out
2407 // control copies
2408 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2409 return remove_dead_region(phase, can_reshape) ? this : NULL;
2410 }
2413 //=============================================================================
2414 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2415 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2416 _adr_type(C->get_adr_type(alias_idx))
2417 {
2418 init_class_id(Class_MemBar);
2419 Node* top = C->top();
2420 init_req(TypeFunc::I_O,top);
2421 init_req(TypeFunc::FramePtr,top);
2422 init_req(TypeFunc::ReturnAdr,top);
2423 if (precedent != NULL)
2424 init_req(TypeFunc::Parms, precedent);
2425 }
2427 //------------------------------cmp--------------------------------------------
2428 uint MemBarNode::hash() const { return NO_HASH; }
2429 uint MemBarNode::cmp( const Node &n ) const {
2430 return (&n == this); // Always fail except on self
2431 }
2433 //------------------------------make-------------------------------------------
2434 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2435 int len = Precedent + (pn == NULL? 0: 1);
2436 switch (opcode) {
2437 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2438 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2439 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2440 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2441 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2442 default: ShouldNotReachHere(); return NULL;
2443 }
2444 }
2446 //------------------------------Ideal------------------------------------------
2447 // Return a node which is more "ideal" than the current node. Strip out
2448 // control copies
2449 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2450 if (remove_dead_region(phase, can_reshape)) return this;
2451 return NULL;
2452 }
2454 //------------------------------Value------------------------------------------
2455 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2456 if( !in(0) ) return Type::TOP;
2457 if( phase->type(in(0)) == Type::TOP )
2458 return Type::TOP;
2459 return TypeTuple::MEMBAR;
2460 }
2462 //------------------------------match------------------------------------------
2463 // Construct projections for memory.
2464 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2465 switch (proj->_con) {
2466 case TypeFunc::Control:
2467 case TypeFunc::Memory:
2468 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2469 }
2470 ShouldNotReachHere();
2471 return NULL;
2472 }
2474 //===========================InitializeNode====================================
2475 // SUMMARY:
2476 // This node acts as a memory barrier on raw memory, after some raw stores.
2477 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2478 // The Initialize can 'capture' suitably constrained stores as raw inits.
2479 // It can coalesce related raw stores into larger units (called 'tiles').
2480 // It can avoid zeroing new storage for memory units which have raw inits.
2481 // At macro-expansion, it is marked 'complete', and does not optimize further.
2482 //
2483 // EXAMPLE:
2484 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2485 // ctl = incoming control; mem* = incoming memory
2486 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2487 // First allocate uninitialized memory and fill in the header:
2488 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2489 // ctl := alloc.Control; mem* := alloc.Memory*
2490 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2491 // Then initialize to zero the non-header parts of the raw memory block:
2492 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2493 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2494 // After the initialize node executes, the object is ready for service:
2495 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2496 // Suppose its body is immediately initialized as {1,2}:
2497 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2498 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2499 // mem.SLICE(#short[*]) := store2
2500 //
2501 // DETAILS:
2502 // An InitializeNode collects and isolates object initialization after
2503 // an AllocateNode and before the next possible safepoint. As a
2504 // memory barrier (MemBarNode), it keeps critical stores from drifting
2505 // down past any safepoint or any publication of the allocation.
2506 // Before this barrier, a newly-allocated object may have uninitialized bits.
2507 // After this barrier, it may be treated as a real oop, and GC is allowed.
2508 //
2509 // The semantics of the InitializeNode include an implicit zeroing of
2510 // the new object from object header to the end of the object.
2511 // (The object header and end are determined by the AllocateNode.)
2512 //
2513 // Certain stores may be added as direct inputs to the InitializeNode.
2514 // These stores must update raw memory, and they must be to addresses
2515 // derived from the raw address produced by AllocateNode, and with
2516 // a constant offset. They must be ordered by increasing offset.
2517 // The first one is at in(RawStores), the last at in(req()-1).
2518 // Unlike most memory operations, they are not linked in a chain,
2519 // but are displayed in parallel as users of the rawmem output of
2520 // the allocation.
2521 //
2522 // (See comments in InitializeNode::capture_store, which continue
2523 // the example given above.)
2524 //
2525 // When the associated Allocate is macro-expanded, the InitializeNode
2526 // may be rewritten to optimize collected stores. A ClearArrayNode
2527 // may also be created at that point to represent any required zeroing.
2528 // The InitializeNode is then marked 'complete', prohibiting further
2529 // capturing of nearby memory operations.
2530 //
2531 // During macro-expansion, all captured initializations which store
2532 // constant values of 32 bits or smaller are coalesced (if advantagous)
2533 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2534 // initialized in fewer memory operations. Memory words which are
2535 // covered by neither tiles nor non-constant stores are pre-zeroed
2536 // by explicit stores of zero. (The code shape happens to do all
2537 // zeroing first, then all other stores, with both sequences occurring
2538 // in order of ascending offsets.)
2539 //
2540 // Alternatively, code may be inserted between an AllocateNode and its
2541 // InitializeNode, to perform arbitrary initialization of the new object.
2542 // E.g., the object copying intrinsics insert complex data transfers here.
2543 // The initialization must then be marked as 'complete' disable the
2544 // built-in zeroing semantics and the collection of initializing stores.
2545 //
2546 // While an InitializeNode is incomplete, reads from the memory state
2547 // produced by it are optimizable if they match the control edge and
2548 // new oop address associated with the allocation/initialization.
2549 // They return a stored value (if the offset matches) or else zero.
2550 // A write to the memory state, if it matches control and address,
2551 // and if it is to a constant offset, may be 'captured' by the
2552 // InitializeNode. It is cloned as a raw memory operation and rewired
2553 // inside the initialization, to the raw oop produced by the allocation.
2554 // Operations on addresses which are provably distinct (e.g., to
2555 // other AllocateNodes) are allowed to bypass the initialization.
2556 //
2557 // The effect of all this is to consolidate object initialization
2558 // (both arrays and non-arrays, both piecewise and bulk) into a
2559 // single location, where it can be optimized as a unit.
2560 //
2561 // Only stores with an offset less than TrackedInitializationLimit words
2562 // will be considered for capture by an InitializeNode. This puts a
2563 // reasonable limit on the complexity of optimized initializations.
2565 //---------------------------InitializeNode------------------------------------
2566 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2567 : _is_complete(false),
2568 MemBarNode(C, adr_type, rawoop)
2569 {
2570 init_class_id(Class_Initialize);
2572 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2573 assert(in(RawAddress) == rawoop, "proper init");
2574 // Note: allocation() can be NULL, for secondary initialization barriers
2575 }
2577 // Since this node is not matched, it will be processed by the
2578 // register allocator. Declare that there are no constraints
2579 // on the allocation of the RawAddress edge.
2580 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2581 // This edge should be set to top, by the set_complete. But be conservative.
2582 if (idx == InitializeNode::RawAddress)
2583 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2584 return RegMask::Empty;
2585 }
2587 Node* InitializeNode::memory(uint alias_idx) {
2588 Node* mem = in(Memory);
2589 if (mem->is_MergeMem()) {
2590 return mem->as_MergeMem()->memory_at(alias_idx);
2591 } else {
2592 // incoming raw memory is not split
2593 return mem;
2594 }
2595 }
2597 bool InitializeNode::is_non_zero() {
2598 if (is_complete()) return false;
2599 remove_extra_zeroes();
2600 return (req() > RawStores);
2601 }
2603 void InitializeNode::set_complete(PhaseGVN* phase) {
2604 assert(!is_complete(), "caller responsibility");
2605 _is_complete = true;
2607 // After this node is complete, it contains a bunch of
2608 // raw-memory initializations. There is no need for
2609 // it to have anything to do with non-raw memory effects.
2610 // Therefore, tell all non-raw users to re-optimize themselves,
2611 // after skipping the memory effects of this initialization.
2612 PhaseIterGVN* igvn = phase->is_IterGVN();
2613 if (igvn) igvn->add_users_to_worklist(this);
2614 }
2616 // convenience function
2617 // return false if the init contains any stores already
2618 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2619 InitializeNode* init = initialization();
2620 if (init == NULL || init->is_complete()) return false;
2621 init->remove_extra_zeroes();
2622 // for now, if this allocation has already collected any inits, bail:
2623 if (init->is_non_zero()) return false;
2624 init->set_complete(phase);
2625 return true;
2626 }
2628 void InitializeNode::remove_extra_zeroes() {
2629 if (req() == RawStores) return;
2630 Node* zmem = zero_memory();
2631 uint fill = RawStores;
2632 for (uint i = fill; i < req(); i++) {
2633 Node* n = in(i);
2634 if (n->is_top() || n == zmem) continue; // skip
2635 if (fill < i) set_req(fill, n); // compact
2636 ++fill;
2637 }
2638 // delete any empty spaces created:
2639 while (fill < req()) {
2640 del_req(fill);
2641 }
2642 }
2644 // Helper for remembering which stores go with which offsets.
2645 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2646 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2647 intptr_t offset = -1;
2648 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2649 phase, offset);
2650 if (base == NULL) return -1; // something is dead,
2651 if (offset < 0) return -1; // dead, dead
2652 return offset;
2653 }
2655 // Helper for proving that an initialization expression is
2656 // "simple enough" to be folded into an object initialization.
2657 // Attempts to prove that a store's initial value 'n' can be captured
2658 // within the initialization without creating a vicious cycle, such as:
2659 // { Foo p = new Foo(); p.next = p; }
2660 // True for constants and parameters and small combinations thereof.
2661 bool InitializeNode::detect_init_independence(Node* n,
2662 bool st_is_pinned,
2663 int& count) {
2664 if (n == NULL) return true; // (can this really happen?)
2665 if (n->is_Proj()) n = n->in(0);
2666 if (n == this) return false; // found a cycle
2667 if (n->is_Con()) return true;
2668 if (n->is_Start()) return true; // params, etc., are OK
2669 if (n->is_Root()) return true; // even better
2671 Node* ctl = n->in(0);
2672 if (ctl != NULL && !ctl->is_top()) {
2673 if (ctl->is_Proj()) ctl = ctl->in(0);
2674 if (ctl == this) return false;
2676 // If we already know that the enclosing memory op is pinned right after
2677 // the init, then any control flow that the store has picked up
2678 // must have preceded the init, or else be equal to the init.
2679 // Even after loop optimizations (which might change control edges)
2680 // a store is never pinned *before* the availability of its inputs.
2681 if (!MemNode::all_controls_dominate(n, this))
2682 return false; // failed to prove a good control
2684 }
2686 // Check data edges for possible dependencies on 'this'.
2687 if ((count += 1) > 20) return false; // complexity limit
2688 for (uint i = 1; i < n->req(); i++) {
2689 Node* m = n->in(i);
2690 if (m == NULL || m == n || m->is_top()) continue;
2691 uint first_i = n->find_edge(m);
2692 if (i != first_i) continue; // process duplicate edge just once
2693 if (!detect_init_independence(m, st_is_pinned, count)) {
2694 return false;
2695 }
2696 }
2698 return true;
2699 }
2701 // Here are all the checks a Store must pass before it can be moved into
2702 // an initialization. Returns zero if a check fails.
2703 // On success, returns the (constant) offset to which the store applies,
2704 // within the initialized memory.
2705 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2706 const int FAIL = 0;
2707 if (st->req() != MemNode::ValueIn + 1)
2708 return FAIL; // an inscrutable StoreNode (card mark?)
2709 Node* ctl = st->in(MemNode::Control);
2710 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2711 return FAIL; // must be unconditional after the initialization
2712 Node* mem = st->in(MemNode::Memory);
2713 if (!(mem->is_Proj() && mem->in(0) == this))
2714 return FAIL; // must not be preceded by other stores
2715 Node* adr = st->in(MemNode::Address);
2716 intptr_t offset;
2717 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2718 if (alloc == NULL)
2719 return FAIL; // inscrutable address
2720 if (alloc != allocation())
2721 return FAIL; // wrong allocation! (store needs to float up)
2722 Node* val = st->in(MemNode::ValueIn);
2723 int complexity_count = 0;
2724 if (!detect_init_independence(val, true, complexity_count))
2725 return FAIL; // stored value must be 'simple enough'
2727 return offset; // success
2728 }
2730 // Find the captured store in(i) which corresponds to the range
2731 // [start..start+size) in the initialized object.
2732 // If there is one, return its index i. If there isn't, return the
2733 // negative of the index where it should be inserted.
2734 // Return 0 if the queried range overlaps an initialization boundary
2735 // or if dead code is encountered.
2736 // If size_in_bytes is zero, do not bother with overlap checks.
2737 int InitializeNode::captured_store_insertion_point(intptr_t start,
2738 int size_in_bytes,
2739 PhaseTransform* phase) {
2740 const int FAIL = 0, MAX_STORE = BytesPerLong;
2742 if (is_complete())
2743 return FAIL; // arraycopy got here first; punt
2745 assert(allocation() != NULL, "must be present");
2747 // no negatives, no header fields:
2748 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2750 // after a certain size, we bail out on tracking all the stores:
2751 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2752 if (start >= ti_limit) return FAIL;
2754 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2755 if (i >= limit) return -(int)i; // not found; here is where to put it
2757 Node* st = in(i);
2758 intptr_t st_off = get_store_offset(st, phase);
2759 if (st_off < 0) {
2760 if (st != zero_memory()) {
2761 return FAIL; // bail out if there is dead garbage
2762 }
2763 } else if (st_off > start) {
2764 // ...we are done, since stores are ordered
2765 if (st_off < start + size_in_bytes) {
2766 return FAIL; // the next store overlaps
2767 }
2768 return -(int)i; // not found; here is where to put it
2769 } else if (st_off < start) {
2770 if (size_in_bytes != 0 &&
2771 start < st_off + MAX_STORE &&
2772 start < st_off + st->as_Store()->memory_size()) {
2773 return FAIL; // the previous store overlaps
2774 }
2775 } else {
2776 if (size_in_bytes != 0 &&
2777 st->as_Store()->memory_size() != size_in_bytes) {
2778 return FAIL; // mismatched store size
2779 }
2780 return i;
2781 }
2783 ++i;
2784 }
2785 }
2787 // Look for a captured store which initializes at the offset 'start'
2788 // with the given size. If there is no such store, and no other
2789 // initialization interferes, then return zero_memory (the memory
2790 // projection of the AllocateNode).
2791 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2792 PhaseTransform* phase) {
2793 assert(stores_are_sane(phase), "");
2794 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2795 if (i == 0) {
2796 return NULL; // something is dead
2797 } else if (i < 0) {
2798 return zero_memory(); // just primordial zero bits here
2799 } else {
2800 Node* st = in(i); // here is the store at this position
2801 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2802 return st;
2803 }
2804 }
2806 // Create, as a raw pointer, an address within my new object at 'offset'.
2807 Node* InitializeNode::make_raw_address(intptr_t offset,
2808 PhaseTransform* phase) {
2809 Node* addr = in(RawAddress);
2810 if (offset != 0) {
2811 Compile* C = phase->C;
2812 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2813 phase->MakeConX(offset)) );
2814 }
2815 return addr;
2816 }
2818 // Clone the given store, converting it into a raw store
2819 // initializing a field or element of my new object.
2820 // Caller is responsible for retiring the original store,
2821 // with subsume_node or the like.
2822 //
2823 // From the example above InitializeNode::InitializeNode,
2824 // here are the old stores to be captured:
2825 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2826 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2827 //
2828 // Here is the changed code; note the extra edges on init:
2829 // alloc = (Allocate ...)
2830 // rawoop = alloc.RawAddress
2831 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2832 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2833 // init = (Initialize alloc.Control alloc.Memory rawoop
2834 // rawstore1 rawstore2)
2835 //
2836 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2837 PhaseTransform* phase) {
2838 assert(stores_are_sane(phase), "");
2840 if (start < 0) return NULL;
2841 assert(can_capture_store(st, phase) == start, "sanity");
2843 Compile* C = phase->C;
2844 int size_in_bytes = st->memory_size();
2845 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2846 if (i == 0) return NULL; // bail out
2847 Node* prev_mem = NULL; // raw memory for the captured store
2848 if (i > 0) {
2849 prev_mem = in(i); // there is a pre-existing store under this one
2850 set_req(i, C->top()); // temporarily disconnect it
2851 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2852 } else {
2853 i = -i; // no pre-existing store
2854 prev_mem = zero_memory(); // a slice of the newly allocated object
2855 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2856 set_req(--i, C->top()); // reuse this edge; it has been folded away
2857 else
2858 ins_req(i, C->top()); // build a new edge
2859 }
2860 Node* new_st = st->clone();
2861 new_st->set_req(MemNode::Control, in(Control));
2862 new_st->set_req(MemNode::Memory, prev_mem);
2863 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2864 new_st = phase->transform(new_st);
2866 // At this point, new_st might have swallowed a pre-existing store
2867 // at the same offset, or perhaps new_st might have disappeared,
2868 // if it redundantly stored the same value (or zero to fresh memory).
2870 // In any case, wire it in:
2871 set_req(i, new_st);
2873 // The caller may now kill the old guy.
2874 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2875 assert(check_st == new_st || check_st == NULL, "must be findable");
2876 assert(!is_complete(), "");
2877 return new_st;
2878 }
2880 static bool store_constant(jlong* tiles, int num_tiles,
2881 intptr_t st_off, int st_size,
2882 jlong con) {
2883 if ((st_off & (st_size-1)) != 0)
2884 return false; // strange store offset (assume size==2**N)
2885 address addr = (address)tiles + st_off;
2886 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2887 switch (st_size) {
2888 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2889 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2890 case sizeof(jint): *(jint*) addr = (jint) con; break;
2891 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2892 default: return false; // strange store size (detect size!=2**N here)
2893 }
2894 return true; // return success to caller
2895 }
2897 // Coalesce subword constants into int constants and possibly
2898 // into long constants. The goal, if the CPU permits,
2899 // is to initialize the object with a small number of 64-bit tiles.
2900 // Also, convert floating-point constants to bit patterns.
2901 // Non-constants are not relevant to this pass.
2902 //
2903 // In terms of the running example on InitializeNode::InitializeNode
2904 // and InitializeNode::capture_store, here is the transformation
2905 // of rawstore1 and rawstore2 into rawstore12:
2906 // alloc = (Allocate ...)
2907 // rawoop = alloc.RawAddress
2908 // tile12 = 0x00010002
2909 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2910 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2911 //
2912 void
2913 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2914 Node* size_in_bytes,
2915 PhaseGVN* phase) {
2916 Compile* C = phase->C;
2918 assert(stores_are_sane(phase), "");
2919 // Note: After this pass, they are not completely sane,
2920 // since there may be some overlaps.
2922 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2924 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2925 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2926 size_limit = MIN2(size_limit, ti_limit);
2927 size_limit = align_size_up(size_limit, BytesPerLong);
2928 int num_tiles = size_limit / BytesPerLong;
2930 // allocate space for the tile map:
2931 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2932 jlong tiles_buf[small_len];
2933 Node* nodes_buf[small_len];
2934 jlong inits_buf[small_len];
2935 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2936 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2937 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2938 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2939 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2940 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2941 // tiles: exact bitwise model of all primitive constants
2942 // nodes: last constant-storing node subsumed into the tiles model
2943 // inits: which bytes (in each tile) are touched by any initializations
2945 //// Pass A: Fill in the tile model with any relevant stores.
2947 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2948 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2949 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2950 Node* zmem = zero_memory(); // initially zero memory state
2951 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2952 Node* st = in(i);
2953 intptr_t st_off = get_store_offset(st, phase);
2955 // Figure out the store's offset and constant value:
2956 if (st_off < header_size) continue; //skip (ignore header)
2957 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2958 int st_size = st->as_Store()->memory_size();
2959 if (st_off + st_size > size_limit) break;
2961 // Record which bytes are touched, whether by constant or not.
2962 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2963 continue; // skip (strange store size)
2965 const Type* val = phase->type(st->in(MemNode::ValueIn));
2966 if (!val->singleton()) continue; //skip (non-con store)
2967 BasicType type = val->basic_type();
2969 jlong con = 0;
2970 switch (type) {
2971 case T_INT: con = val->is_int()->get_con(); break;
2972 case T_LONG: con = val->is_long()->get_con(); break;
2973 case T_FLOAT: con = jint_cast(val->getf()); break;
2974 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2975 default: continue; //skip (odd store type)
2976 }
2978 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2979 st->Opcode() == Op_StoreL) {
2980 continue; // This StoreL is already optimal.
2981 }
2983 // Store down the constant.
2984 store_constant(tiles, num_tiles, st_off, st_size, con);
2986 intptr_t j = st_off >> LogBytesPerLong;
2988 if (type == T_INT && st_size == BytesPerInt
2989 && (st_off & BytesPerInt) == BytesPerInt) {
2990 jlong lcon = tiles[j];
2991 if (!Matcher::isSimpleConstant64(lcon) &&
2992 st->Opcode() == Op_StoreI) {
2993 // This StoreI is already optimal by itself.
2994 jint* intcon = (jint*) &tiles[j];
2995 intcon[1] = 0; // undo the store_constant()
2997 // If the previous store is also optimal by itself, back up and
2998 // undo the action of the previous loop iteration... if we can.
2999 // But if we can't, just let the previous half take care of itself.
3000 st = nodes[j];
3001 st_off -= BytesPerInt;
3002 con = intcon[0];
3003 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3004 assert(st_off >= header_size, "still ignoring header");
3005 assert(get_store_offset(st, phase) == st_off, "must be");
3006 assert(in(i-1) == zmem, "must be");
3007 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3008 assert(con == tcon->is_int()->get_con(), "must be");
3009 // Undo the effects of the previous loop trip, which swallowed st:
3010 intcon[0] = 0; // undo store_constant()
3011 set_req(i-1, st); // undo set_req(i, zmem)
3012 nodes[j] = NULL; // undo nodes[j] = st
3013 --old_subword; // undo ++old_subword
3014 }
3015 continue; // This StoreI is already optimal.
3016 }
3017 }
3019 // This store is not needed.
3020 set_req(i, zmem);
3021 nodes[j] = st; // record for the moment
3022 if (st_size < BytesPerLong) // something has changed
3023 ++old_subword; // includes int/float, but who's counting...
3024 else ++old_long;
3025 }
3027 if ((old_subword + old_long) == 0)
3028 return; // nothing more to do
3030 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3031 // Be sure to insert them before overlapping non-constant stores.
3032 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3033 for (int j = 0; j < num_tiles; j++) {
3034 jlong con = tiles[j];
3035 jlong init = inits[j];
3036 if (con == 0) continue;
3037 jint con0, con1; // split the constant, address-wise
3038 jint init0, init1; // split the init map, address-wise
3039 { union { jlong con; jint intcon[2]; } u;
3040 u.con = con;
3041 con0 = u.intcon[0];
3042 con1 = u.intcon[1];
3043 u.con = init;
3044 init0 = u.intcon[0];
3045 init1 = u.intcon[1];
3046 }
3048 Node* old = nodes[j];
3049 assert(old != NULL, "need the prior store");
3050 intptr_t offset = (j * BytesPerLong);
3052 bool split = !Matcher::isSimpleConstant64(con);
3054 if (offset < header_size) {
3055 assert(offset + BytesPerInt >= header_size, "second int counts");
3056 assert(*(jint*)&tiles[j] == 0, "junk in header");
3057 split = true; // only the second word counts
3058 // Example: int a[] = { 42 ... }
3059 } else if (con0 == 0 && init0 == -1) {
3060 split = true; // first word is covered by full inits
3061 // Example: int a[] = { ... foo(), 42 ... }
3062 } else if (con1 == 0 && init1 == -1) {
3063 split = true; // second word is covered by full inits
3064 // Example: int a[] = { ... 42, foo() ... }
3065 }
3067 // Here's a case where init0 is neither 0 nor -1:
3068 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3069 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3070 // In this case the tile is not split; it is (jlong)42.
3071 // The big tile is stored down, and then the foo() value is inserted.
3072 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3074 Node* ctl = old->in(MemNode::Control);
3075 Node* adr = make_raw_address(offset, phase);
3076 const TypePtr* atp = TypeRawPtr::BOTTOM;
3078 // One or two coalesced stores to plop down.
3079 Node* st[2];
3080 intptr_t off[2];
3081 int nst = 0;
3082 if (!split) {
3083 ++new_long;
3084 off[nst] = offset;
3085 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3086 phase->longcon(con), T_LONG);
3087 } else {
3088 // Omit either if it is a zero.
3089 if (con0 != 0) {
3090 ++new_int;
3091 off[nst] = offset;
3092 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3093 phase->intcon(con0), T_INT);
3094 }
3095 if (con1 != 0) {
3096 ++new_int;
3097 offset += BytesPerInt;
3098 adr = make_raw_address(offset, phase);
3099 off[nst] = offset;
3100 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3101 phase->intcon(con1), T_INT);
3102 }
3103 }
3105 // Insert second store first, then the first before the second.
3106 // Insert each one just before any overlapping non-constant stores.
3107 while (nst > 0) {
3108 Node* st1 = st[--nst];
3109 C->copy_node_notes_to(st1, old);
3110 st1 = phase->transform(st1);
3111 offset = off[nst];
3112 assert(offset >= header_size, "do not smash header");
3113 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3114 guarantee(ins_idx != 0, "must re-insert constant store");
3115 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3116 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3117 set_req(--ins_idx, st1);
3118 else
3119 ins_req(ins_idx, st1);
3120 }
3121 }
3123 if (PrintCompilation && WizardMode)
3124 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3125 old_subword, old_long, new_int, new_long);
3126 if (C->log() != NULL)
3127 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3128 old_subword, old_long, new_int, new_long);
3130 // Clean up any remaining occurrences of zmem:
3131 remove_extra_zeroes();
3132 }
3134 // Explore forward from in(start) to find the first fully initialized
3135 // word, and return its offset. Skip groups of subword stores which
3136 // together initialize full words. If in(start) is itself part of a
3137 // fully initialized word, return the offset of in(start). If there
3138 // are no following full-word stores, or if something is fishy, return
3139 // a negative value.
3140 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3141 int int_map = 0;
3142 intptr_t int_map_off = 0;
3143 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3145 for (uint i = start, limit = req(); i < limit; i++) {
3146 Node* st = in(i);
3148 intptr_t st_off = get_store_offset(st, phase);
3149 if (st_off < 0) break; // return conservative answer
3151 int st_size = st->as_Store()->memory_size();
3152 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3153 return st_off; // we found a complete word init
3154 }
3156 // update the map:
3158 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3159 if (this_int_off != int_map_off) {
3160 // reset the map:
3161 int_map = 0;
3162 int_map_off = this_int_off;
3163 }
3165 int subword_off = st_off - this_int_off;
3166 int_map |= right_n_bits(st_size) << subword_off;
3167 if ((int_map & FULL_MAP) == FULL_MAP) {
3168 return this_int_off; // we found a complete word init
3169 }
3171 // Did this store hit or cross the word boundary?
3172 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3173 if (next_int_off == this_int_off + BytesPerInt) {
3174 // We passed the current int, without fully initializing it.
3175 int_map_off = next_int_off;
3176 int_map >>= BytesPerInt;
3177 } else if (next_int_off > this_int_off + BytesPerInt) {
3178 // We passed the current and next int.
3179 return this_int_off + BytesPerInt;
3180 }
3181 }
3183 return -1;
3184 }
3187 // Called when the associated AllocateNode is expanded into CFG.
3188 // At this point, we may perform additional optimizations.
3189 // Linearize the stores by ascending offset, to make memory
3190 // activity as coherent as possible.
3191 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3192 intptr_t header_size,
3193 Node* size_in_bytes,
3194 PhaseGVN* phase) {
3195 assert(!is_complete(), "not already complete");
3196 assert(stores_are_sane(phase), "");
3197 assert(allocation() != NULL, "must be present");
3199 remove_extra_zeroes();
3201 if (ReduceFieldZeroing || ReduceBulkZeroing)
3202 // reduce instruction count for common initialization patterns
3203 coalesce_subword_stores(header_size, size_in_bytes, phase);
3205 Node* zmem = zero_memory(); // initially zero memory state
3206 Node* inits = zmem; // accumulating a linearized chain of inits
3207 #ifdef ASSERT
3208 intptr_t first_offset = allocation()->minimum_header_size();
3209 intptr_t last_init_off = first_offset; // previous init offset
3210 intptr_t last_init_end = first_offset; // previous init offset+size
3211 intptr_t last_tile_end = first_offset; // previous tile offset+size
3212 #endif
3213 intptr_t zeroes_done = header_size;
3215 bool do_zeroing = true; // we might give up if inits are very sparse
3216 int big_init_gaps = 0; // how many large gaps have we seen?
3218 if (ZeroTLAB) do_zeroing = false;
3219 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3221 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3222 Node* st = in(i);
3223 intptr_t st_off = get_store_offset(st, phase);
3224 if (st_off < 0)
3225 break; // unknown junk in the inits
3226 if (st->in(MemNode::Memory) != zmem)
3227 break; // complicated store chains somehow in list
3229 int st_size = st->as_Store()->memory_size();
3230 intptr_t next_init_off = st_off + st_size;
3232 if (do_zeroing && zeroes_done < next_init_off) {
3233 // See if this store needs a zero before it or under it.
3234 intptr_t zeroes_needed = st_off;
3236 if (st_size < BytesPerInt) {
3237 // Look for subword stores which only partially initialize words.
3238 // If we find some, we must lay down some word-level zeroes first,
3239 // underneath the subword stores.
3240 //
3241 // Examples:
3242 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3243 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3244 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3245 //
3246 // Note: coalesce_subword_stores may have already done this,
3247 // if it was prompted by constant non-zero subword initializers.
3248 // But this case can still arise with non-constant stores.
3250 intptr_t next_full_store = find_next_fullword_store(i, phase);
3252 // In the examples above:
3253 // in(i) p q r s x y z
3254 // st_off 12 13 14 15 12 13 14
3255 // st_size 1 1 1 1 1 1 1
3256 // next_full_s. 12 16 16 16 16 16 16
3257 // z's_done 12 16 16 16 12 16 12
3258 // z's_needed 12 16 16 16 16 16 16
3259 // zsize 0 0 0 0 4 0 4
3260 if (next_full_store < 0) {
3261 // Conservative tack: Zero to end of current word.
3262 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3263 } else {
3264 // Zero to beginning of next fully initialized word.
3265 // Or, don't zero at all, if we are already in that word.
3266 assert(next_full_store >= zeroes_needed, "must go forward");
3267 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3268 zeroes_needed = next_full_store;
3269 }
3270 }
3272 if (zeroes_needed > zeroes_done) {
3273 intptr_t zsize = zeroes_needed - zeroes_done;
3274 // Do some incremental zeroing on rawmem, in parallel with inits.
3275 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3276 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3277 zeroes_done, zeroes_needed,
3278 phase);
3279 zeroes_done = zeroes_needed;
3280 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3281 do_zeroing = false; // leave the hole, next time
3282 }
3283 }
3285 // Collect the store and move on:
3286 st->set_req(MemNode::Memory, inits);
3287 inits = st; // put it on the linearized chain
3288 set_req(i, zmem); // unhook from previous position
3290 if (zeroes_done == st_off)
3291 zeroes_done = next_init_off;
3293 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3295 #ifdef ASSERT
3296 // Various order invariants. Weaker than stores_are_sane because
3297 // a large constant tile can be filled in by smaller non-constant stores.
3298 assert(st_off >= last_init_off, "inits do not reverse");
3299 last_init_off = st_off;
3300 const Type* val = NULL;
3301 if (st_size >= BytesPerInt &&
3302 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3303 (int)val->basic_type() < (int)T_OBJECT) {
3304 assert(st_off >= last_tile_end, "tiles do not overlap");
3305 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3306 last_tile_end = MAX2(last_tile_end, next_init_off);
3307 } else {
3308 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3309 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3310 assert(st_off >= last_init_end, "inits do not overlap");
3311 last_init_end = next_init_off; // it's a non-tile
3312 }
3313 #endif //ASSERT
3314 }
3316 remove_extra_zeroes(); // clear out all the zmems left over
3317 add_req(inits);
3319 if (!ZeroTLAB) {
3320 // If anything remains to be zeroed, zero it all now.
3321 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3322 // if it is the last unused 4 bytes of an instance, forget about it
3323 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3324 if (zeroes_done + BytesPerLong >= size_limit) {
3325 assert(allocation() != NULL, "");
3326 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3327 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3328 if (zeroes_done == k->layout_helper())
3329 zeroes_done = size_limit;
3330 }
3331 if (zeroes_done < size_limit) {
3332 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3333 zeroes_done, size_in_bytes, phase);
3334 }
3335 }
3337 set_complete(phase);
3338 return rawmem;
3339 }
3342 #ifdef ASSERT
3343 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3344 if (is_complete())
3345 return true; // stores could be anything at this point
3346 assert(allocation() != NULL, "must be present");
3347 intptr_t last_off = allocation()->minimum_header_size();
3348 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3349 Node* st = in(i);
3350 intptr_t st_off = get_store_offset(st, phase);
3351 if (st_off < 0) continue; // ignore dead garbage
3352 if (last_off > st_off) {
3353 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3354 this->dump(2);
3355 assert(false, "ascending store offsets");
3356 return false;
3357 }
3358 last_off = st_off + st->as_Store()->memory_size();
3359 }
3360 return true;
3361 }
3362 #endif //ASSERT
3367 //============================MergeMemNode=====================================
3368 //
3369 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3370 // contributing store or call operations. Each contributor provides the memory
3371 // state for a particular "alias type" (see Compile::alias_type). For example,
3372 // if a MergeMem has an input X for alias category #6, then any memory reference
3373 // to alias category #6 may use X as its memory state input, as an exact equivalent
3374 // to using the MergeMem as a whole.
3375 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3376 //
3377 // (Here, the <N> notation gives the index of the relevant adr_type.)
3378 //
3379 // In one special case (and more cases in the future), alias categories overlap.
3380 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3381 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3382 // it is exactly equivalent to that state W:
3383 // MergeMem(<Bot>: W) <==> W
3384 //
3385 // Usually, the merge has more than one input. In that case, where inputs
3386 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3387 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3388 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3389 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3390 //
3391 // A merge can take a "wide" memory state as one of its narrow inputs.
3392 // This simply means that the merge observes out only the relevant parts of
3393 // the wide input. That is, wide memory states arriving at narrow merge inputs
3394 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3395 //
3396 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3397 // and that memory slices "leak through":
3398 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3399 //
3400 // But, in such a cascade, repeated memory slices can "block the leak":
3401 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3402 //
3403 // In the last example, Y is not part of the combined memory state of the
3404 // outermost MergeMem. The system must, of course, prevent unschedulable
3405 // memory states from arising, so you can be sure that the state Y is somehow
3406 // a precursor to state Y'.
3407 //
3408 //
3409 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3410 // of each MergeMemNode array are exactly the numerical alias indexes, including
3411 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3412 // Compile::alias_type (and kin) produce and manage these indexes.
3413 //
3414 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3415 // (Note that this provides quick access to the top node inside MergeMem methods,
3416 // without the need to reach out via TLS to Compile::current.)
3417 //
3418 // As a consequence of what was just described, a MergeMem that represents a full
3419 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3420 // containing all alias categories.
3421 //
3422 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3423 //
3424 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3425 // a memory state for the alias type <N>, or else the top node, meaning that
3426 // there is no particular input for that alias type. Note that the length of
3427 // a MergeMem is variable, and may be extended at any time to accommodate new
3428 // memory states at larger alias indexes. When merges grow, they are of course
3429 // filled with "top" in the unused in() positions.
3430 //
3431 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3432 // (Top was chosen because it works smoothly with passes like GCM.)
3433 //
3434 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3435 // the type of random VM bits like TLS references.) Since it is always the
3436 // first non-Bot memory slice, some low-level loops use it to initialize an
3437 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3438 //
3439 //
3440 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3441 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3442 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3443 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3444 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3445 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3446 //
3447 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3448 // really that different from the other memory inputs. An abbreviation called
3449 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3450 //
3451 //
3452 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3453 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3454 // that "emerges though" the base memory will be marked as excluding the alias types
3455 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3456 //
3457 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3458 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3459 //
3460 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3461 // (It is currently unimplemented.) As you can see, the resulting merge is
3462 // actually a disjoint union of memory states, rather than an overlay.
3463 //
3465 //------------------------------MergeMemNode-----------------------------------
3466 Node* MergeMemNode::make_empty_memory() {
3467 Node* empty_memory = (Node*) Compile::current()->top();
3468 assert(empty_memory->is_top(), "correct sentinel identity");
3469 return empty_memory;
3470 }
3472 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3473 init_class_id(Class_MergeMem);
3474 // all inputs are nullified in Node::Node(int)
3475 // set_input(0, NULL); // no control input
3477 // Initialize the edges uniformly to top, for starters.
3478 Node* empty_mem = make_empty_memory();
3479 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3480 init_req(i,empty_mem);
3481 }
3482 assert(empty_memory() == empty_mem, "");
3484 if( new_base != NULL && new_base->is_MergeMem() ) {
3485 MergeMemNode* mdef = new_base->as_MergeMem();
3486 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3487 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3488 mms.set_memory(mms.memory2());
3489 }
3490 assert(base_memory() == mdef->base_memory(), "");
3491 } else {
3492 set_base_memory(new_base);
3493 }
3494 }
3496 // Make a new, untransformed MergeMem with the same base as 'mem'.
3497 // If mem is itself a MergeMem, populate the result with the same edges.
3498 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3499 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3500 }
3502 //------------------------------cmp--------------------------------------------
3503 uint MergeMemNode::hash() const { return NO_HASH; }
3504 uint MergeMemNode::cmp( const Node &n ) const {
3505 return (&n == this); // Always fail except on self
3506 }
3508 //------------------------------Identity---------------------------------------
3509 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3510 // Identity if this merge point does not record any interesting memory
3511 // disambiguations.
3512 Node* base_mem = base_memory();
3513 Node* empty_mem = empty_memory();
3514 if (base_mem != empty_mem) { // Memory path is not dead?
3515 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3516 Node* mem = in(i);
3517 if (mem != empty_mem && mem != base_mem) {
3518 return this; // Many memory splits; no change
3519 }
3520 }
3521 }
3522 return base_mem; // No memory splits; ID on the one true input
3523 }
3525 //------------------------------Ideal------------------------------------------
3526 // This method is invoked recursively on chains of MergeMem nodes
3527 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3528 // Remove chain'd MergeMems
3529 //
3530 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3531 // relative to the "in(Bot)". Since we are patching both at the same time,
3532 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3533 // but rewrite each "in(i)" relative to the new "in(Bot)".
3534 Node *progress = NULL;
3537 Node* old_base = base_memory();
3538 Node* empty_mem = empty_memory();
3539 if (old_base == empty_mem)
3540 return NULL; // Dead memory path.
3542 MergeMemNode* old_mbase;
3543 if (old_base != NULL && old_base->is_MergeMem())
3544 old_mbase = old_base->as_MergeMem();
3545 else
3546 old_mbase = NULL;
3547 Node* new_base = old_base;
3549 // simplify stacked MergeMems in base memory
3550 if (old_mbase) new_base = old_mbase->base_memory();
3552 // the base memory might contribute new slices beyond my req()
3553 if (old_mbase) grow_to_match(old_mbase);
3555 // Look carefully at the base node if it is a phi.
3556 PhiNode* phi_base;
3557 if (new_base != NULL && new_base->is_Phi())
3558 phi_base = new_base->as_Phi();
3559 else
3560 phi_base = NULL;
3562 Node* phi_reg = NULL;
3563 uint phi_len = (uint)-1;
3564 if (phi_base != NULL && !phi_base->is_copy()) {
3565 // do not examine phi if degraded to a copy
3566 phi_reg = phi_base->region();
3567 phi_len = phi_base->req();
3568 // see if the phi is unfinished
3569 for (uint i = 1; i < phi_len; i++) {
3570 if (phi_base->in(i) == NULL) {
3571 // incomplete phi; do not look at it yet!
3572 phi_reg = NULL;
3573 phi_len = (uint)-1;
3574 break;
3575 }
3576 }
3577 }
3579 // Note: We do not call verify_sparse on entry, because inputs
3580 // can normalize to the base_memory via subsume_node or similar
3581 // mechanisms. This method repairs that damage.
3583 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3585 // Look at each slice.
3586 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3587 Node* old_in = in(i);
3588 // calculate the old memory value
3589 Node* old_mem = old_in;
3590 if (old_mem == empty_mem) old_mem = old_base;
3591 assert(old_mem == memory_at(i), "");
3593 // maybe update (reslice) the old memory value
3595 // simplify stacked MergeMems
3596 Node* new_mem = old_mem;
3597 MergeMemNode* old_mmem;
3598 if (old_mem != NULL && old_mem->is_MergeMem())
3599 old_mmem = old_mem->as_MergeMem();
3600 else
3601 old_mmem = NULL;
3602 if (old_mmem == this) {
3603 // This can happen if loops break up and safepoints disappear.
3604 // A merge of BotPtr (default) with a RawPtr memory derived from a
3605 // safepoint can be rewritten to a merge of the same BotPtr with
3606 // the BotPtr phi coming into the loop. If that phi disappears
3607 // also, we can end up with a self-loop of the mergemem.
3608 // In general, if loops degenerate and memory effects disappear,
3609 // a mergemem can be left looking at itself. This simply means
3610 // that the mergemem's default should be used, since there is
3611 // no longer any apparent effect on this slice.
3612 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3613 // from start. Update the input to TOP.
3614 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3615 }
3616 else if (old_mmem != NULL) {
3617 new_mem = old_mmem->memory_at(i);
3618 }
3619 // else preceeding memory was not a MergeMem
3621 // replace equivalent phis (unfortunately, they do not GVN together)
3622 if (new_mem != NULL && new_mem != new_base &&
3623 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3624 if (new_mem->is_Phi()) {
3625 PhiNode* phi_mem = new_mem->as_Phi();
3626 for (uint i = 1; i < phi_len; i++) {
3627 if (phi_base->in(i) != phi_mem->in(i)) {
3628 phi_mem = NULL;
3629 break;
3630 }
3631 }
3632 if (phi_mem != NULL) {
3633 // equivalent phi nodes; revert to the def
3634 new_mem = new_base;
3635 }
3636 }
3637 }
3639 // maybe store down a new value
3640 Node* new_in = new_mem;
3641 if (new_in == new_base) new_in = empty_mem;
3643 if (new_in != old_in) {
3644 // Warning: Do not combine this "if" with the previous "if"
3645 // A memory slice might have be be rewritten even if it is semantically
3646 // unchanged, if the base_memory value has changed.
3647 set_req(i, new_in);
3648 progress = this; // Report progress
3649 }
3650 }
3652 if (new_base != old_base) {
3653 set_req(Compile::AliasIdxBot, new_base);
3654 // Don't use set_base_memory(new_base), because we need to update du.
3655 assert(base_memory() == new_base, "");
3656 progress = this;
3657 }
3659 if( base_memory() == this ) {
3660 // a self cycle indicates this memory path is dead
3661 set_req(Compile::AliasIdxBot, empty_mem);
3662 }
3664 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3665 // Recursion must occur after the self cycle check above
3666 if( base_memory()->is_MergeMem() ) {
3667 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3668 Node *m = phase->transform(new_mbase); // Rollup any cycles
3669 if( m != NULL && (m->is_top() ||
3670 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3671 // propagate rollup of dead cycle to self
3672 set_req(Compile::AliasIdxBot, empty_mem);
3673 }
3674 }
3676 if( base_memory() == empty_mem ) {
3677 progress = this;
3678 // Cut inputs during Parse phase only.
3679 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3680 if( !can_reshape ) {
3681 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3682 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3683 }
3684 }
3685 }
3687 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3688 // Check if PhiNode::Ideal's "Split phis through memory merges"
3689 // transform should be attempted. Look for this->phi->this cycle.
3690 uint merge_width = req();
3691 if (merge_width > Compile::AliasIdxRaw) {
3692 PhiNode* phi = base_memory()->as_Phi();
3693 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3694 if (phi->in(i) == this) {
3695 phase->is_IterGVN()->_worklist.push(phi);
3696 break;
3697 }
3698 }
3699 }
3700 }
3702 assert(progress || verify_sparse(), "please, no dups of base");
3703 return progress;
3704 }
3706 //-------------------------set_base_memory-------------------------------------
3707 void MergeMemNode::set_base_memory(Node *new_base) {
3708 Node* empty_mem = empty_memory();
3709 set_req(Compile::AliasIdxBot, new_base);
3710 assert(memory_at(req()) == new_base, "must set default memory");
3711 // Clear out other occurrences of new_base:
3712 if (new_base != empty_mem) {
3713 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3714 if (in(i) == new_base) set_req(i, empty_mem);
3715 }
3716 }
3717 }
3719 //------------------------------out_RegMask------------------------------------
3720 const RegMask &MergeMemNode::out_RegMask() const {
3721 return RegMask::Empty;
3722 }
3724 //------------------------------dump_spec--------------------------------------
3725 #ifndef PRODUCT
3726 void MergeMemNode::dump_spec(outputStream *st) const {
3727 st->print(" {");
3728 Node* base_mem = base_memory();
3729 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3730 Node* mem = memory_at(i);
3731 if (mem == base_mem) { st->print(" -"); continue; }
3732 st->print( " N%d:", mem->_idx );
3733 Compile::current()->get_adr_type(i)->dump_on(st);
3734 }
3735 st->print(" }");
3736 }
3737 #endif // !PRODUCT
3740 #ifdef ASSERT
3741 static bool might_be_same(Node* a, Node* b) {
3742 if (a == b) return true;
3743 if (!(a->is_Phi() || b->is_Phi())) return false;
3744 // phis shift around during optimization
3745 return true; // pretty stupid...
3746 }
3748 // verify a narrow slice (either incoming or outgoing)
3749 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3750 if (!VerifyAliases) return; // don't bother to verify unless requested
3751 if (is_error_reported()) return; // muzzle asserts when debugging an error
3752 if (Node::in_dump()) return; // muzzle asserts when printing
3753 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3754 assert(n != NULL, "");
3755 // Elide intervening MergeMem's
3756 while (n->is_MergeMem()) {
3757 n = n->as_MergeMem()->memory_at(alias_idx);
3758 }
3759 Compile* C = Compile::current();
3760 const TypePtr* n_adr_type = n->adr_type();
3761 if (n == m->empty_memory()) {
3762 // Implicit copy of base_memory()
3763 } else if (n_adr_type != TypePtr::BOTTOM) {
3764 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3765 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3766 } else {
3767 // A few places like make_runtime_call "know" that VM calls are narrow,
3768 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3769 bool expected_wide_mem = false;
3770 if (n == m->base_memory()) {
3771 expected_wide_mem = true;
3772 } else if (alias_idx == Compile::AliasIdxRaw ||
3773 n == m->memory_at(Compile::AliasIdxRaw)) {
3774 expected_wide_mem = true;
3775 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3776 // memory can "leak through" calls on channels that
3777 // are write-once. Allow this also.
3778 expected_wide_mem = true;
3779 }
3780 assert(expected_wide_mem, "expected narrow slice replacement");
3781 }
3782 }
3783 #else // !ASSERT
3784 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3785 #endif
3788 //-----------------------------memory_at---------------------------------------
3789 Node* MergeMemNode::memory_at(uint alias_idx) const {
3790 assert(alias_idx >= Compile::AliasIdxRaw ||
3791 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3792 "must avoid base_memory and AliasIdxTop");
3794 // Otherwise, it is a narrow slice.
3795 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3796 Compile *C = Compile::current();
3797 if (is_empty_memory(n)) {
3798 // the array is sparse; empty slots are the "top" node
3799 n = base_memory();
3800 assert(Node::in_dump()
3801 || n == NULL || n->bottom_type() == Type::TOP
3802 || n->adr_type() == TypePtr::BOTTOM
3803 || n->adr_type() == TypeRawPtr::BOTTOM
3804 || Compile::current()->AliasLevel() == 0,
3805 "must be a wide memory");
3806 // AliasLevel == 0 if we are organizing the memory states manually.
3807 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3808 } else {
3809 // make sure the stored slice is sane
3810 #ifdef ASSERT
3811 if (is_error_reported() || Node::in_dump()) {
3812 } else if (might_be_same(n, base_memory())) {
3813 // Give it a pass: It is a mostly harmless repetition of the base.
3814 // This can arise normally from node subsumption during optimization.
3815 } else {
3816 verify_memory_slice(this, alias_idx, n);
3817 }
3818 #endif
3819 }
3820 return n;
3821 }
3823 //---------------------------set_memory_at-------------------------------------
3824 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3825 verify_memory_slice(this, alias_idx, n);
3826 Node* empty_mem = empty_memory();
3827 if (n == base_memory()) n = empty_mem; // collapse default
3828 uint need_req = alias_idx+1;
3829 if (req() < need_req) {
3830 if (n == empty_mem) return; // already the default, so do not grow me
3831 // grow the sparse array
3832 do {
3833 add_req(empty_mem);
3834 } while (req() < need_req);
3835 }
3836 set_req( alias_idx, n );
3837 }
3841 //--------------------------iteration_setup------------------------------------
3842 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3843 if (other != NULL) {
3844 grow_to_match(other);
3845 // invariant: the finite support of mm2 is within mm->req()
3846 #ifdef ASSERT
3847 for (uint i = req(); i < other->req(); i++) {
3848 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3849 }
3850 #endif
3851 }
3852 // Replace spurious copies of base_memory by top.
3853 Node* base_mem = base_memory();
3854 if (base_mem != NULL && !base_mem->is_top()) {
3855 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3856 if (in(i) == base_mem)
3857 set_req(i, empty_memory());
3858 }
3859 }
3860 }
3862 //---------------------------grow_to_match-------------------------------------
3863 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3864 Node* empty_mem = empty_memory();
3865 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3866 // look for the finite support of the other memory
3867 for (uint i = other->req(); --i >= req(); ) {
3868 if (other->in(i) != empty_mem) {
3869 uint new_len = i+1;
3870 while (req() < new_len) add_req(empty_mem);
3871 break;
3872 }
3873 }
3874 }
3876 //---------------------------verify_sparse-------------------------------------
3877 #ifndef PRODUCT
3878 bool MergeMemNode::verify_sparse() const {
3879 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3880 Node* base_mem = base_memory();
3881 // The following can happen in degenerate cases, since empty==top.
3882 if (is_empty_memory(base_mem)) return true;
3883 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3884 assert(in(i) != NULL, "sane slice");
3885 if (in(i) == base_mem) return false; // should have been the sentinel value!
3886 }
3887 return true;
3888 }
3890 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3891 Node* n;
3892 n = mm->in(idx);
3893 if (mem == n) return true; // might be empty_memory()
3894 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3895 if (mem == n) return true;
3896 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3897 if (mem == n) return true;
3898 if (n == NULL) break;
3899 }
3900 return false;
3901 }
3902 #endif // !PRODUCT