Fri, 11 Jul 2014 19:51:36 -0400
8036588: VerifyFieldClosure fails instanceKlass:3133
Summary: Changed deopt live-pointer test to use returns-object instead of live-and-returns-object
Reviewed-by: iveresov, kvn, jrose
1 /*
2 * Copyright (c) 2005, 2014, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.hpp"
30 #include "opto/c2compiler.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
38 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
39 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
40 _collecting(true),
41 _verify(false),
42 _compile(C),
43 _igvn(igvn),
44 _node_map(C->comp_arena()) {
45 // Add unknown java object.
46 add_java_object(C->top(), PointsToNode::GlobalEscape);
47 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
48 // Add ConP(#NULL) and ConN(#NULL) nodes.
49 Node* oop_null = igvn->zerocon(T_OBJECT);
50 assert(oop_null->_idx < nodes_size(), "should be created already");
51 add_java_object(oop_null, PointsToNode::NoEscape);
52 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
53 if (UseCompressedOops) {
54 Node* noop_null = igvn->zerocon(T_NARROWOOP);
55 assert(noop_null->_idx < nodes_size(), "should be created already");
56 map_ideal_node(noop_null, null_obj);
57 }
58 _pcmp_neq = NULL; // Should be initialized
59 _pcmp_eq = NULL;
60 }
62 bool ConnectionGraph::has_candidates(Compile *C) {
63 // EA brings benefits only when the code has allocations and/or locks which
64 // are represented by ideal Macro nodes.
65 int cnt = C->macro_count();
66 for (int i = 0; i < cnt; i++) {
67 Node *n = C->macro_node(i);
68 if (n->is_Allocate())
69 return true;
70 if (n->is_Lock()) {
71 Node* obj = n->as_Lock()->obj_node()->uncast();
72 if (!(obj->is_Parm() || obj->is_Con()))
73 return true;
74 }
75 if (n->is_CallStaticJava() &&
76 n->as_CallStaticJava()->is_boxing_method()) {
77 return true;
78 }
79 }
80 return false;
81 }
83 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
84 Compile::TracePhase t2("escapeAnalysis", &Phase::_t_escapeAnalysis, true);
85 ResourceMark rm;
87 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
88 // to create space for them in ConnectionGraph::_nodes[].
89 Node* oop_null = igvn->zerocon(T_OBJECT);
90 Node* noop_null = igvn->zerocon(T_NARROWOOP);
91 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
92 // Perform escape analysis
93 if (congraph->compute_escape()) {
94 // There are non escaping objects.
95 C->set_congraph(congraph);
96 }
97 // Cleanup.
98 if (oop_null->outcnt() == 0)
99 igvn->hash_delete(oop_null);
100 if (noop_null->outcnt() == 0)
101 igvn->hash_delete(noop_null);
102 }
104 bool ConnectionGraph::compute_escape() {
105 Compile* C = _compile;
106 PhaseGVN* igvn = _igvn;
108 // Worklists used by EA.
109 Unique_Node_List delayed_worklist;
110 GrowableArray<Node*> alloc_worklist;
111 GrowableArray<Node*> ptr_cmp_worklist;
112 GrowableArray<Node*> storestore_worklist;
113 GrowableArray<PointsToNode*> ptnodes_worklist;
114 GrowableArray<JavaObjectNode*> java_objects_worklist;
115 GrowableArray<JavaObjectNode*> non_escaped_worklist;
116 GrowableArray<FieldNode*> oop_fields_worklist;
117 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
119 { Compile::TracePhase t3("connectionGraph", &Phase::_t_connectionGraph, true);
121 // 1. Populate Connection Graph (CG) with PointsTo nodes.
122 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
123 // Initialize worklist
124 if (C->root() != NULL) {
125 ideal_nodes.push(C->root());
126 }
127 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
128 Node* n = ideal_nodes.at(next);
129 // Create PointsTo nodes and add them to Connection Graph. Called
130 // only once per ideal node since ideal_nodes is Unique_Node list.
131 add_node_to_connection_graph(n, &delayed_worklist);
132 PointsToNode* ptn = ptnode_adr(n->_idx);
133 if (ptn != NULL) {
134 ptnodes_worklist.append(ptn);
135 if (ptn->is_JavaObject()) {
136 java_objects_worklist.append(ptn->as_JavaObject());
137 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
138 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
139 // Only allocations and java static calls results are interesting.
140 non_escaped_worklist.append(ptn->as_JavaObject());
141 }
142 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
143 oop_fields_worklist.append(ptn->as_Field());
144 }
145 }
146 if (n->is_MergeMem()) {
147 // Collect all MergeMem nodes to add memory slices for
148 // scalar replaceable objects in split_unique_types().
149 _mergemem_worklist.append(n->as_MergeMem());
150 } else if (OptimizePtrCompare && n->is_Cmp() &&
151 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
152 // Collect compare pointers nodes.
153 ptr_cmp_worklist.append(n);
154 } else if (n->is_MemBarStoreStore()) {
155 // Collect all MemBarStoreStore nodes so that depending on the
156 // escape status of the associated Allocate node some of them
157 // may be eliminated.
158 storestore_worklist.append(n);
159 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
160 (n->req() > MemBarNode::Precedent)) {
161 record_for_optimizer(n);
162 #ifdef ASSERT
163 } else if (n->is_AddP()) {
164 // Collect address nodes for graph verification.
165 addp_worklist.append(n);
166 #endif
167 }
168 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
169 Node* m = n->fast_out(i); // Get user
170 ideal_nodes.push(m);
171 }
172 }
173 if (non_escaped_worklist.length() == 0) {
174 _collecting = false;
175 return false; // Nothing to do.
176 }
177 // Add final simple edges to graph.
178 while(delayed_worklist.size() > 0) {
179 Node* n = delayed_worklist.pop();
180 add_final_edges(n);
181 }
182 int ptnodes_length = ptnodes_worklist.length();
184 #ifdef ASSERT
185 if (VerifyConnectionGraph) {
186 // Verify that no new simple edges could be created and all
187 // local vars has edges.
188 _verify = true;
189 for (int next = 0; next < ptnodes_length; ++next) {
190 PointsToNode* ptn = ptnodes_worklist.at(next);
191 add_final_edges(ptn->ideal_node());
192 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
193 ptn->dump();
194 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
195 }
196 }
197 _verify = false;
198 }
199 #endif
201 // 2. Finish Graph construction by propagating references to all
202 // java objects through graph.
203 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
204 java_objects_worklist, oop_fields_worklist)) {
205 // All objects escaped or hit time or iterations limits.
206 _collecting = false;
207 return false;
208 }
210 // 3. Adjust scalar_replaceable state of nonescaping objects and push
211 // scalar replaceable allocations on alloc_worklist for processing
212 // in split_unique_types().
213 int non_escaped_length = non_escaped_worklist.length();
214 for (int next = 0; next < non_escaped_length; next++) {
215 JavaObjectNode* ptn = non_escaped_worklist.at(next);
216 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
217 Node* n = ptn->ideal_node();
218 if (n->is_Allocate()) {
219 n->as_Allocate()->_is_non_escaping = noescape;
220 }
221 if (n->is_CallStaticJava()) {
222 n->as_CallStaticJava()->_is_non_escaping = noescape;
223 }
224 if (noescape && ptn->scalar_replaceable()) {
225 adjust_scalar_replaceable_state(ptn);
226 if (ptn->scalar_replaceable()) {
227 alloc_worklist.append(ptn->ideal_node());
228 }
229 }
230 }
232 #ifdef ASSERT
233 if (VerifyConnectionGraph) {
234 // Verify that graph is complete - no new edges could be added or needed.
235 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
236 java_objects_worklist, addp_worklist);
237 }
238 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
239 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
240 null_obj->edge_count() == 0 &&
241 !null_obj->arraycopy_src() &&
242 !null_obj->arraycopy_dst(), "sanity");
243 #endif
245 _collecting = false;
247 } // TracePhase t3("connectionGraph")
249 // 4. Optimize ideal graph based on EA information.
250 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
251 if (has_non_escaping_obj) {
252 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
253 }
255 #ifndef PRODUCT
256 if (PrintEscapeAnalysis) {
257 dump(ptnodes_worklist); // Dump ConnectionGraph
258 }
259 #endif
261 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
262 #ifdef ASSERT
263 if (VerifyConnectionGraph) {
264 int alloc_length = alloc_worklist.length();
265 for (int next = 0; next < alloc_length; ++next) {
266 Node* n = alloc_worklist.at(next);
267 PointsToNode* ptn = ptnode_adr(n->_idx);
268 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
269 }
270 }
271 #endif
273 // 5. Separate memory graph for scalar replaceable allcations.
274 if (has_scalar_replaceable_candidates &&
275 C->AliasLevel() >= 3 && EliminateAllocations) {
276 // Now use the escape information to create unique types for
277 // scalar replaceable objects.
278 split_unique_types(alloc_worklist);
279 if (C->failing()) return false;
280 C->print_method(PHASE_AFTER_EA, 2);
282 #ifdef ASSERT
283 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
284 tty->print("=== No allocations eliminated for ");
285 C->method()->print_short_name();
286 if(!EliminateAllocations) {
287 tty->print(" since EliminateAllocations is off ===");
288 } else if(!has_scalar_replaceable_candidates) {
289 tty->print(" since there are no scalar replaceable candidates ===");
290 } else if(C->AliasLevel() < 3) {
291 tty->print(" since AliasLevel < 3 ===");
292 }
293 tty->cr();
294 #endif
295 }
296 return has_non_escaping_obj;
297 }
299 // Utility function for nodes that load an object
300 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
301 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
302 // ThreadLocal has RawPtr type.
303 const Type* t = _igvn->type(n);
304 if (t->make_ptr() != NULL) {
305 Node* adr = n->in(MemNode::Address);
306 #ifdef ASSERT
307 if (!adr->is_AddP()) {
308 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
309 } else {
310 assert((ptnode_adr(adr->_idx) == NULL ||
311 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
312 }
313 #endif
314 add_local_var_and_edge(n, PointsToNode::NoEscape,
315 adr, delayed_worklist);
316 }
317 }
319 // Populate Connection Graph with PointsTo nodes and create simple
320 // connection graph edges.
321 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
322 assert(!_verify, "this method sould not be called for verification");
323 PhaseGVN* igvn = _igvn;
324 uint n_idx = n->_idx;
325 PointsToNode* n_ptn = ptnode_adr(n_idx);
326 if (n_ptn != NULL)
327 return; // No need to redefine PointsTo node during first iteration.
329 if (n->is_Call()) {
330 // Arguments to allocation and locking don't escape.
331 if (n->is_AbstractLock()) {
332 // Put Lock and Unlock nodes on IGVN worklist to process them during
333 // first IGVN optimization when escape information is still available.
334 record_for_optimizer(n);
335 } else if (n->is_Allocate()) {
336 add_call_node(n->as_Call());
337 record_for_optimizer(n);
338 } else {
339 if (n->is_CallStaticJava()) {
340 const char* name = n->as_CallStaticJava()->_name;
341 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
342 return; // Skip uncommon traps
343 }
344 // Don't mark as processed since call's arguments have to be processed.
345 delayed_worklist->push(n);
346 // Check if a call returns an object.
347 if ((n->as_Call()->returns_pointer() &&
348 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) ||
349 (n->is_CallStaticJava() &&
350 n->as_CallStaticJava()->is_boxing_method())) {
351 add_call_node(n->as_Call());
352 }
353 }
354 return;
355 }
356 // Put this check here to process call arguments since some call nodes
357 // point to phantom_obj.
358 if (n_ptn == phantom_obj || n_ptn == null_obj)
359 return; // Skip predefined nodes.
361 int opcode = n->Opcode();
362 switch (opcode) {
363 case Op_AddP: {
364 Node* base = get_addp_base(n);
365 PointsToNode* ptn_base = ptnode_adr(base->_idx);
366 // Field nodes are created for all field types. They are used in
367 // adjust_scalar_replaceable_state() and split_unique_types().
368 // Note, non-oop fields will have only base edges in Connection
369 // Graph because such fields are not used for oop loads and stores.
370 int offset = address_offset(n, igvn);
371 add_field(n, PointsToNode::NoEscape, offset);
372 if (ptn_base == NULL) {
373 delayed_worklist->push(n); // Process it later.
374 } else {
375 n_ptn = ptnode_adr(n_idx);
376 add_base(n_ptn->as_Field(), ptn_base);
377 }
378 break;
379 }
380 case Op_CastX2P: {
381 map_ideal_node(n, phantom_obj);
382 break;
383 }
384 case Op_CastPP:
385 case Op_CheckCastPP:
386 case Op_EncodeP:
387 case Op_DecodeN:
388 case Op_EncodePKlass:
389 case Op_DecodeNKlass: {
390 add_local_var_and_edge(n, PointsToNode::NoEscape,
391 n->in(1), delayed_worklist);
392 break;
393 }
394 case Op_CMoveP: {
395 add_local_var(n, PointsToNode::NoEscape);
396 // Do not add edges during first iteration because some could be
397 // not defined yet.
398 delayed_worklist->push(n);
399 break;
400 }
401 case Op_ConP:
402 case Op_ConN:
403 case Op_ConNKlass: {
404 // assume all oop constants globally escape except for null
405 PointsToNode::EscapeState es;
406 const Type* t = igvn->type(n);
407 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
408 es = PointsToNode::NoEscape;
409 } else {
410 es = PointsToNode::GlobalEscape;
411 }
412 add_java_object(n, es);
413 break;
414 }
415 case Op_CreateEx: {
416 // assume that all exception objects globally escape
417 add_java_object(n, PointsToNode::GlobalEscape);
418 break;
419 }
420 case Op_LoadKlass:
421 case Op_LoadNKlass: {
422 // Unknown class is loaded
423 map_ideal_node(n, phantom_obj);
424 break;
425 }
426 case Op_LoadP:
427 case Op_LoadN:
428 case Op_LoadPLocked: {
429 add_objload_to_connection_graph(n, delayed_worklist);
430 break;
431 }
432 case Op_Parm: {
433 map_ideal_node(n, phantom_obj);
434 break;
435 }
436 case Op_PartialSubtypeCheck: {
437 // Produces Null or notNull and is used in only in CmpP so
438 // phantom_obj could be used.
439 map_ideal_node(n, phantom_obj); // Result is unknown
440 break;
441 }
442 case Op_Phi: {
443 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
444 // ThreadLocal has RawPtr type.
445 const Type* t = n->as_Phi()->type();
446 if (t->make_ptr() != NULL) {
447 add_local_var(n, PointsToNode::NoEscape);
448 // Do not add edges during first iteration because some could be
449 // not defined yet.
450 delayed_worklist->push(n);
451 }
452 break;
453 }
454 case Op_Proj: {
455 // we are only interested in the oop result projection from a call
456 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
457 n->in(0)->as_Call()->returns_pointer()) {
458 add_local_var_and_edge(n, PointsToNode::NoEscape,
459 n->in(0), delayed_worklist);
460 }
461 break;
462 }
463 case Op_Rethrow: // Exception object escapes
464 case Op_Return: {
465 if (n->req() > TypeFunc::Parms &&
466 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
467 // Treat Return value as LocalVar with GlobalEscape escape state.
468 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
469 n->in(TypeFunc::Parms), delayed_worklist);
470 }
471 break;
472 }
473 case Op_GetAndSetP:
474 case Op_GetAndSetN: {
475 add_objload_to_connection_graph(n, delayed_worklist);
476 // fallthrough
477 }
478 case Op_StoreP:
479 case Op_StoreN:
480 case Op_StoreNKlass:
481 case Op_StorePConditional:
482 case Op_CompareAndSwapP:
483 case Op_CompareAndSwapN: {
484 Node* adr = n->in(MemNode::Address);
485 const Type *adr_type = igvn->type(adr);
486 adr_type = adr_type->make_ptr();
487 if (adr_type == NULL) {
488 break; // skip dead nodes
489 }
490 if (adr_type->isa_oopptr() ||
491 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
492 (adr_type == TypeRawPtr::NOTNULL &&
493 adr->in(AddPNode::Address)->is_Proj() &&
494 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
495 delayed_worklist->push(n); // Process it later.
496 #ifdef ASSERT
497 assert(adr->is_AddP(), "expecting an AddP");
498 if (adr_type == TypeRawPtr::NOTNULL) {
499 // Verify a raw address for a store captured by Initialize node.
500 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
501 assert(offs != Type::OffsetBot, "offset must be a constant");
502 }
503 #endif
504 } else {
505 // Ignore copy the displaced header to the BoxNode (OSR compilation).
506 if (adr->is_BoxLock())
507 break;
508 // Stored value escapes in unsafe access.
509 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
510 // Pointer stores in G1 barriers looks like unsafe access.
511 // Ignore such stores to be able scalar replace non-escaping
512 // allocations.
513 if (UseG1GC && adr->is_AddP()) {
514 Node* base = get_addp_base(adr);
515 if (base->Opcode() == Op_LoadP &&
516 base->in(MemNode::Address)->is_AddP()) {
517 adr = base->in(MemNode::Address);
518 Node* tls = get_addp_base(adr);
519 if (tls->Opcode() == Op_ThreadLocal) {
520 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
521 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() +
522 PtrQueue::byte_offset_of_buf())) {
523 break; // G1 pre barier previous oop value store.
524 }
525 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() +
526 PtrQueue::byte_offset_of_buf())) {
527 break; // G1 post barier card address store.
528 }
529 }
530 }
531 }
532 delayed_worklist->push(n); // Process unsafe access later.
533 break;
534 }
535 #ifdef ASSERT
536 n->dump(1);
537 assert(false, "not unsafe or G1 barrier raw StoreP");
538 #endif
539 }
540 break;
541 }
542 case Op_AryEq:
543 case Op_StrComp:
544 case Op_StrEquals:
545 case Op_StrIndexOf:
546 case Op_EncodeISOArray: {
547 add_local_var(n, PointsToNode::ArgEscape);
548 delayed_worklist->push(n); // Process it later.
549 break;
550 }
551 case Op_ThreadLocal: {
552 add_java_object(n, PointsToNode::ArgEscape);
553 break;
554 }
555 default:
556 ; // Do nothing for nodes not related to EA.
557 }
558 return;
559 }
561 #ifdef ASSERT
562 #define ELSE_FAIL(name) \
563 /* Should not be called for not pointer type. */ \
564 n->dump(1); \
565 assert(false, name); \
566 break;
567 #else
568 #define ELSE_FAIL(name) \
569 break;
570 #endif
572 // Add final simple edges to graph.
573 void ConnectionGraph::add_final_edges(Node *n) {
574 PointsToNode* n_ptn = ptnode_adr(n->_idx);
575 #ifdef ASSERT
576 if (_verify && n_ptn->is_JavaObject())
577 return; // This method does not change graph for JavaObject.
578 #endif
580 if (n->is_Call()) {
581 process_call_arguments(n->as_Call());
582 return;
583 }
584 assert(n->is_Store() || n->is_LoadStore() ||
585 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
586 "node should be registered already");
587 int opcode = n->Opcode();
588 switch (opcode) {
589 case Op_AddP: {
590 Node* base = get_addp_base(n);
591 PointsToNode* ptn_base = ptnode_adr(base->_idx);
592 assert(ptn_base != NULL, "field's base should be registered");
593 add_base(n_ptn->as_Field(), ptn_base);
594 break;
595 }
596 case Op_CastPP:
597 case Op_CheckCastPP:
598 case Op_EncodeP:
599 case Op_DecodeN:
600 case Op_EncodePKlass:
601 case Op_DecodeNKlass: {
602 add_local_var_and_edge(n, PointsToNode::NoEscape,
603 n->in(1), NULL);
604 break;
605 }
606 case Op_CMoveP: {
607 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
608 Node* in = n->in(i);
609 if (in == NULL)
610 continue; // ignore NULL
611 Node* uncast_in = in->uncast();
612 if (uncast_in->is_top() || uncast_in == n)
613 continue; // ignore top or inputs which go back this node
614 PointsToNode* ptn = ptnode_adr(in->_idx);
615 assert(ptn != NULL, "node should be registered");
616 add_edge(n_ptn, ptn);
617 }
618 break;
619 }
620 case Op_LoadP:
621 case Op_LoadN:
622 case Op_LoadPLocked: {
623 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
624 // ThreadLocal has RawPtr type.
625 const Type* t = _igvn->type(n);
626 if (t->make_ptr() != NULL) {
627 Node* adr = n->in(MemNode::Address);
628 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
629 break;
630 }
631 ELSE_FAIL("Op_LoadP");
632 }
633 case Op_Phi: {
634 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
635 // ThreadLocal has RawPtr type.
636 const Type* t = n->as_Phi()->type();
637 if (t->make_ptr() != NULL) {
638 for (uint i = 1; i < n->req(); i++) {
639 Node* in = n->in(i);
640 if (in == NULL)
641 continue; // ignore NULL
642 Node* uncast_in = in->uncast();
643 if (uncast_in->is_top() || uncast_in == n)
644 continue; // ignore top or inputs which go back this node
645 PointsToNode* ptn = ptnode_adr(in->_idx);
646 assert(ptn != NULL, "node should be registered");
647 add_edge(n_ptn, ptn);
648 }
649 break;
650 }
651 ELSE_FAIL("Op_Phi");
652 }
653 case Op_Proj: {
654 // we are only interested in the oop result projection from a call
655 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
656 n->in(0)->as_Call()->returns_pointer()) {
657 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
658 break;
659 }
660 ELSE_FAIL("Op_Proj");
661 }
662 case Op_Rethrow: // Exception object escapes
663 case Op_Return: {
664 if (n->req() > TypeFunc::Parms &&
665 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
666 // Treat Return value as LocalVar with GlobalEscape escape state.
667 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
668 n->in(TypeFunc::Parms), NULL);
669 break;
670 }
671 ELSE_FAIL("Op_Return");
672 }
673 case Op_StoreP:
674 case Op_StoreN:
675 case Op_StoreNKlass:
676 case Op_StorePConditional:
677 case Op_CompareAndSwapP:
678 case Op_CompareAndSwapN:
679 case Op_GetAndSetP:
680 case Op_GetAndSetN: {
681 Node* adr = n->in(MemNode::Address);
682 const Type *adr_type = _igvn->type(adr);
683 adr_type = adr_type->make_ptr();
684 #ifdef ASSERT
685 if (adr_type == NULL) {
686 n->dump(1);
687 assert(adr_type != NULL, "dead node should not be on list");
688 break;
689 }
690 #endif
691 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) {
692 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
693 }
694 if (adr_type->isa_oopptr() ||
695 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
696 (adr_type == TypeRawPtr::NOTNULL &&
697 adr->in(AddPNode::Address)->is_Proj() &&
698 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
699 // Point Address to Value
700 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
701 assert(adr_ptn != NULL &&
702 adr_ptn->as_Field()->is_oop(), "node should be registered");
703 Node *val = n->in(MemNode::ValueIn);
704 PointsToNode* ptn = ptnode_adr(val->_idx);
705 assert(ptn != NULL, "node should be registered");
706 add_edge(adr_ptn, ptn);
707 break;
708 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
709 // Stored value escapes in unsafe access.
710 Node *val = n->in(MemNode::ValueIn);
711 PointsToNode* ptn = ptnode_adr(val->_idx);
712 assert(ptn != NULL, "node should be registered");
713 set_escape_state(ptn, PointsToNode::GlobalEscape);
714 // Add edge to object for unsafe access with offset.
715 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
716 assert(adr_ptn != NULL, "node should be registered");
717 if (adr_ptn->is_Field()) {
718 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
719 add_edge(adr_ptn, ptn);
720 }
721 break;
722 }
723 ELSE_FAIL("Op_StoreP");
724 }
725 case Op_AryEq:
726 case Op_StrComp:
727 case Op_StrEquals:
728 case Op_StrIndexOf:
729 case Op_EncodeISOArray: {
730 // char[] arrays passed to string intrinsic do not escape but
731 // they are not scalar replaceable. Adjust escape state for them.
732 // Start from in(2) edge since in(1) is memory edge.
733 for (uint i = 2; i < n->req(); i++) {
734 Node* adr = n->in(i);
735 const Type* at = _igvn->type(adr);
736 if (!adr->is_top() && at->isa_ptr()) {
737 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
738 at->isa_ptr() != NULL, "expecting a pointer");
739 if (adr->is_AddP()) {
740 adr = get_addp_base(adr);
741 }
742 PointsToNode* ptn = ptnode_adr(adr->_idx);
743 assert(ptn != NULL, "node should be registered");
744 add_edge(n_ptn, ptn);
745 }
746 }
747 break;
748 }
749 default: {
750 // This method should be called only for EA specific nodes which may
751 // miss some edges when they were created.
752 #ifdef ASSERT
753 n->dump(1);
754 #endif
755 guarantee(false, "unknown node");
756 }
757 }
758 return;
759 }
761 void ConnectionGraph::add_call_node(CallNode* call) {
762 assert(call->returns_pointer(), "only for call which returns pointer");
763 uint call_idx = call->_idx;
764 if (call->is_Allocate()) {
765 Node* k = call->in(AllocateNode::KlassNode);
766 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
767 assert(kt != NULL, "TypeKlassPtr required.");
768 ciKlass* cik = kt->klass();
769 PointsToNode::EscapeState es = PointsToNode::NoEscape;
770 bool scalar_replaceable = true;
771 if (call->is_AllocateArray()) {
772 if (!cik->is_array_klass()) { // StressReflectiveCode
773 es = PointsToNode::GlobalEscape;
774 } else {
775 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
776 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
777 // Not scalar replaceable if the length is not constant or too big.
778 scalar_replaceable = false;
779 }
780 }
781 } else { // Allocate instance
782 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
783 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
784 !cik->is_instance_klass() || // StressReflectiveCode
785 cik->as_instance_klass()->has_finalizer()) {
786 es = PointsToNode::GlobalEscape;
787 }
788 }
789 add_java_object(call, es);
790 PointsToNode* ptn = ptnode_adr(call_idx);
791 if (!scalar_replaceable && ptn->scalar_replaceable()) {
792 ptn->set_scalar_replaceable(false);
793 }
794 } else if (call->is_CallStaticJava()) {
795 // Call nodes could be different types:
796 //
797 // 1. CallDynamicJavaNode (what happened during call is unknown):
798 //
799 // - mapped to GlobalEscape JavaObject node if oop is returned;
800 //
801 // - all oop arguments are escaping globally;
802 //
803 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
804 //
805 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
806 //
807 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
808 // - mapped to NoEscape JavaObject node if non-escaping object allocated
809 // during call is returned;
810 // - mapped to ArgEscape LocalVar node pointed to object arguments
811 // which are returned and does not escape during call;
812 //
813 // - oop arguments escaping status is defined by bytecode analysis;
814 //
815 // For a static call, we know exactly what method is being called.
816 // Use bytecode estimator to record whether the call's return value escapes.
817 ciMethod* meth = call->as_CallJava()->method();
818 if (meth == NULL) {
819 const char* name = call->as_CallStaticJava()->_name;
820 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
821 // Returns a newly allocated unescaped object.
822 add_java_object(call, PointsToNode::NoEscape);
823 ptnode_adr(call_idx)->set_scalar_replaceable(false);
824 } else if (meth->is_boxing_method()) {
825 // Returns boxing object
826 PointsToNode::EscapeState es;
827 vmIntrinsics::ID intr = meth->intrinsic_id();
828 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
829 // It does not escape if object is always allocated.
830 es = PointsToNode::NoEscape;
831 } else {
832 // It escapes globally if object could be loaded from cache.
833 es = PointsToNode::GlobalEscape;
834 }
835 add_java_object(call, es);
836 } else {
837 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
838 call_analyzer->copy_dependencies(_compile->dependencies());
839 if (call_analyzer->is_return_allocated()) {
840 // Returns a newly allocated unescaped object, simply
841 // update dependency information.
842 // Mark it as NoEscape so that objects referenced by
843 // it's fields will be marked as NoEscape at least.
844 add_java_object(call, PointsToNode::NoEscape);
845 ptnode_adr(call_idx)->set_scalar_replaceable(false);
846 } else {
847 // Determine whether any arguments are returned.
848 const TypeTuple* d = call->tf()->domain();
849 bool ret_arg = false;
850 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
851 if (d->field_at(i)->isa_ptr() != NULL &&
852 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
853 ret_arg = true;
854 break;
855 }
856 }
857 if (ret_arg) {
858 add_local_var(call, PointsToNode::ArgEscape);
859 } else {
860 // Returns unknown object.
861 map_ideal_node(call, phantom_obj);
862 }
863 }
864 }
865 } else {
866 // An other type of call, assume the worst case:
867 // returned value is unknown and globally escapes.
868 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
869 map_ideal_node(call, phantom_obj);
870 }
871 }
873 void ConnectionGraph::process_call_arguments(CallNode *call) {
874 bool is_arraycopy = false;
875 switch (call->Opcode()) {
876 #ifdef ASSERT
877 case Op_Allocate:
878 case Op_AllocateArray:
879 case Op_Lock:
880 case Op_Unlock:
881 assert(false, "should be done already");
882 break;
883 #endif
884 case Op_CallLeafNoFP:
885 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
886 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
887 // fall through
888 case Op_CallLeaf: {
889 // Stub calls, objects do not escape but they are not scale replaceable.
890 // Adjust escape state for outgoing arguments.
891 const TypeTuple * d = call->tf()->domain();
892 bool src_has_oops = false;
893 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
894 const Type* at = d->field_at(i);
895 Node *arg = call->in(i);
896 const Type *aat = _igvn->type(arg);
897 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
898 continue;
899 if (arg->is_AddP()) {
900 //
901 // The inline_native_clone() case when the arraycopy stub is called
902 // after the allocation before Initialize and CheckCastPP nodes.
903 // Or normal arraycopy for object arrays case.
904 //
905 // Set AddP's base (Allocate) as not scalar replaceable since
906 // pointer to the base (with offset) is passed as argument.
907 //
908 arg = get_addp_base(arg);
909 }
910 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
911 assert(arg_ptn != NULL, "should be registered");
912 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
913 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
914 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
915 aat->isa_ptr() != NULL, "expecting an Ptr");
916 bool arg_has_oops = aat->isa_oopptr() &&
917 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
918 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
919 if (i == TypeFunc::Parms) {
920 src_has_oops = arg_has_oops;
921 }
922 //
923 // src or dst could be j.l.Object when other is basic type array:
924 //
925 // arraycopy(char[],0,Object*,0,size);
926 // arraycopy(Object*,0,char[],0,size);
927 //
928 // Don't add edges in such cases.
929 //
930 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
931 arg_has_oops && (i > TypeFunc::Parms);
932 #ifdef ASSERT
933 if (!(is_arraycopy ||
934 (call->as_CallLeaf()->_name != NULL &&
935 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
936 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
937 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
938 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
939 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
940 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
941 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
942 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
943 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
944 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
945 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
946 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
947 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0)
948 ))) {
949 call->dump();
950 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
951 }
952 #endif
953 // Always process arraycopy's destination object since
954 // we need to add all possible edges to references in
955 // source object.
956 if (arg_esc >= PointsToNode::ArgEscape &&
957 !arg_is_arraycopy_dest) {
958 continue;
959 }
960 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
961 if (arg_is_arraycopy_dest) {
962 Node* src = call->in(TypeFunc::Parms);
963 if (src->is_AddP()) {
964 src = get_addp_base(src);
965 }
966 PointsToNode* src_ptn = ptnode_adr(src->_idx);
967 assert(src_ptn != NULL, "should be registered");
968 if (arg_ptn != src_ptn) {
969 // Special arraycopy edge:
970 // A destination object's field can't have the source object
971 // as base since objects escape states are not related.
972 // Only escape state of destination object's fields affects
973 // escape state of fields in source object.
974 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
975 }
976 }
977 }
978 }
979 break;
980 }
981 case Op_CallStaticJava: {
982 // For a static call, we know exactly what method is being called.
983 // Use bytecode estimator to record the call's escape affects
984 #ifdef ASSERT
985 const char* name = call->as_CallStaticJava()->_name;
986 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
987 #endif
988 ciMethod* meth = call->as_CallJava()->method();
989 if ((meth != NULL) && meth->is_boxing_method()) {
990 break; // Boxing methods do not modify any oops.
991 }
992 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
993 // fall-through if not a Java method or no analyzer information
994 if (call_analyzer != NULL) {
995 PointsToNode* call_ptn = ptnode_adr(call->_idx);
996 const TypeTuple* d = call->tf()->domain();
997 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
998 const Type* at = d->field_at(i);
999 int k = i - TypeFunc::Parms;
1000 Node* arg = call->in(i);
1001 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1002 if (at->isa_ptr() != NULL &&
1003 call_analyzer->is_arg_returned(k)) {
1004 // The call returns arguments.
1005 if (call_ptn != NULL) { // Is call's result used?
1006 assert(call_ptn->is_LocalVar(), "node should be registered");
1007 assert(arg_ptn != NULL, "node should be registered");
1008 add_edge(call_ptn, arg_ptn);
1009 }
1010 }
1011 if (at->isa_oopptr() != NULL &&
1012 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1013 if (!call_analyzer->is_arg_stack(k)) {
1014 // The argument global escapes
1015 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1016 } else {
1017 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1018 if (!call_analyzer->is_arg_local(k)) {
1019 // The argument itself doesn't escape, but any fields might
1020 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1021 }
1022 }
1023 }
1024 }
1025 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1026 // The call returns arguments.
1027 assert(call_ptn->edge_count() > 0, "sanity");
1028 if (!call_analyzer->is_return_local()) {
1029 // Returns also unknown object.
1030 add_edge(call_ptn, phantom_obj);
1031 }
1032 }
1033 break;
1034 }
1035 }
1036 default: {
1037 // Fall-through here if not a Java method or no analyzer information
1038 // or some other type of call, assume the worst case: all arguments
1039 // globally escape.
1040 const TypeTuple* d = call->tf()->domain();
1041 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1042 const Type* at = d->field_at(i);
1043 if (at->isa_oopptr() != NULL) {
1044 Node* arg = call->in(i);
1045 if (arg->is_AddP()) {
1046 arg = get_addp_base(arg);
1047 }
1048 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1049 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1050 }
1051 }
1052 }
1053 }
1054 }
1057 // Finish Graph construction.
1058 bool ConnectionGraph::complete_connection_graph(
1059 GrowableArray<PointsToNode*>& ptnodes_worklist,
1060 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1061 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1062 GrowableArray<FieldNode*>& oop_fields_worklist) {
1063 // Normally only 1-3 passes needed to build Connection Graph depending
1064 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1065 // Set limit to 20 to catch situation when something did go wrong and
1066 // bailout Escape Analysis.
1067 // Also limit build time to 30 sec (60 in debug VM).
1068 #define CG_BUILD_ITER_LIMIT 20
1069 #ifdef ASSERT
1070 #define CG_BUILD_TIME_LIMIT 60.0
1071 #else
1072 #define CG_BUILD_TIME_LIMIT 30.0
1073 #endif
1075 // Propagate GlobalEscape and ArgEscape escape states and check that
1076 // we still have non-escaping objects. The method pushs on _worklist
1077 // Field nodes which reference phantom_object.
1078 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1079 return false; // Nothing to do.
1080 }
1081 // Now propagate references to all JavaObject nodes.
1082 int java_objects_length = java_objects_worklist.length();
1083 elapsedTimer time;
1084 int new_edges = 1;
1085 int iterations = 0;
1086 do {
1087 while ((new_edges > 0) &&
1088 (iterations++ < CG_BUILD_ITER_LIMIT) &&
1089 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1090 time.start();
1091 new_edges = 0;
1092 // Propagate references to phantom_object for nodes pushed on _worklist
1093 // by find_non_escaped_objects() and find_field_value().
1094 new_edges += add_java_object_edges(phantom_obj, false);
1095 for (int next = 0; next < java_objects_length; ++next) {
1096 JavaObjectNode* ptn = java_objects_worklist.at(next);
1097 new_edges += add_java_object_edges(ptn, true);
1098 }
1099 if (new_edges > 0) {
1100 // Update escape states on each iteration if graph was updated.
1101 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1102 return false; // Nothing to do.
1103 }
1104 }
1105 time.stop();
1106 }
1107 if ((iterations < CG_BUILD_ITER_LIMIT) &&
1108 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1109 time.start();
1110 // Find fields which have unknown value.
1111 int fields_length = oop_fields_worklist.length();
1112 for (int next = 0; next < fields_length; next++) {
1113 FieldNode* field = oop_fields_worklist.at(next);
1114 if (field->edge_count() == 0) {
1115 new_edges += find_field_value(field);
1116 // This code may added new edges to phantom_object.
1117 // Need an other cycle to propagate references to phantom_object.
1118 }
1119 }
1120 time.stop();
1121 } else {
1122 new_edges = 0; // Bailout
1123 }
1124 } while (new_edges > 0);
1126 // Bailout if passed limits.
1127 if ((iterations >= CG_BUILD_ITER_LIMIT) ||
1128 (time.seconds() >= CG_BUILD_TIME_LIMIT)) {
1129 Compile* C = _compile;
1130 if (C->log() != NULL) {
1131 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1132 C->log()->text("%s", (iterations >= CG_BUILD_ITER_LIMIT) ? "iterations" : "time");
1133 C->log()->end_elem(" limit'");
1134 }
1135 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1136 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1137 // Possible infinite build_connection_graph loop,
1138 // bailout (no changes to ideal graph were made).
1139 return false;
1140 }
1141 #ifdef ASSERT
1142 if (Verbose && PrintEscapeAnalysis) {
1143 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1144 iterations, nodes_size(), ptnodes_worklist.length());
1145 }
1146 #endif
1148 #undef CG_BUILD_ITER_LIMIT
1149 #undef CG_BUILD_TIME_LIMIT
1151 // Find fields initialized by NULL for non-escaping Allocations.
1152 int non_escaped_length = non_escaped_worklist.length();
1153 for (int next = 0; next < non_escaped_length; next++) {
1154 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1155 PointsToNode::EscapeState es = ptn->escape_state();
1156 assert(es <= PointsToNode::ArgEscape, "sanity");
1157 if (es == PointsToNode::NoEscape) {
1158 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1159 // Adding references to NULL object does not change escape states
1160 // since it does not escape. Also no fields are added to NULL object.
1161 add_java_object_edges(null_obj, false);
1162 }
1163 }
1164 Node* n = ptn->ideal_node();
1165 if (n->is_Allocate()) {
1166 // The object allocated by this Allocate node will never be
1167 // seen by an other thread. Mark it so that when it is
1168 // expanded no MemBarStoreStore is added.
1169 InitializeNode* ini = n->as_Allocate()->initialization();
1170 if (ini != NULL)
1171 ini->set_does_not_escape();
1172 }
1173 }
1174 return true; // Finished graph construction.
1175 }
1177 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1178 // and check that we still have non-escaping java objects.
1179 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1180 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1181 GrowableArray<PointsToNode*> escape_worklist;
1182 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1183 int ptnodes_length = ptnodes_worklist.length();
1184 for (int next = 0; next < ptnodes_length; ++next) {
1185 PointsToNode* ptn = ptnodes_worklist.at(next);
1186 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1187 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1188 escape_worklist.push(ptn);
1189 }
1190 }
1191 // Set escape states to referenced nodes (edges list).
1192 while (escape_worklist.length() > 0) {
1193 PointsToNode* ptn = escape_worklist.pop();
1194 PointsToNode::EscapeState es = ptn->escape_state();
1195 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1196 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1197 es >= PointsToNode::ArgEscape) {
1198 // GlobalEscape or ArgEscape state of field means it has unknown value.
1199 if (add_edge(ptn, phantom_obj)) {
1200 // New edge was added
1201 add_field_uses_to_worklist(ptn->as_Field());
1202 }
1203 }
1204 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1205 PointsToNode* e = i.get();
1206 if (e->is_Arraycopy()) {
1207 assert(ptn->arraycopy_dst(), "sanity");
1208 // Propagate only fields escape state through arraycopy edge.
1209 if (e->fields_escape_state() < field_es) {
1210 set_fields_escape_state(e, field_es);
1211 escape_worklist.push(e);
1212 }
1213 } else if (es >= field_es) {
1214 // fields_escape_state is also set to 'es' if it is less than 'es'.
1215 if (e->escape_state() < es) {
1216 set_escape_state(e, es);
1217 escape_worklist.push(e);
1218 }
1219 } else {
1220 // Propagate field escape state.
1221 bool es_changed = false;
1222 if (e->fields_escape_state() < field_es) {
1223 set_fields_escape_state(e, field_es);
1224 es_changed = true;
1225 }
1226 if ((e->escape_state() < field_es) &&
1227 e->is_Field() && ptn->is_JavaObject() &&
1228 e->as_Field()->is_oop()) {
1229 // Change escape state of referenced fileds.
1230 set_escape_state(e, field_es);
1231 es_changed = true;;
1232 } else if (e->escape_state() < es) {
1233 set_escape_state(e, es);
1234 es_changed = true;;
1235 }
1236 if (es_changed) {
1237 escape_worklist.push(e);
1238 }
1239 }
1240 }
1241 }
1242 // Remove escaped objects from non_escaped list.
1243 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1244 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1245 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1246 non_escaped_worklist.delete_at(next);
1247 }
1248 if (ptn->escape_state() == PointsToNode::NoEscape) {
1249 // Find fields in non-escaped allocations which have unknown value.
1250 find_init_values(ptn, phantom_obj, NULL);
1251 }
1252 }
1253 return (non_escaped_worklist.length() > 0);
1254 }
1256 // Add all references to JavaObject node by walking over all uses.
1257 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1258 int new_edges = 0;
1259 if (populate_worklist) {
1260 // Populate _worklist by uses of jobj's uses.
1261 for (UseIterator i(jobj); i.has_next(); i.next()) {
1262 PointsToNode* use = i.get();
1263 if (use->is_Arraycopy())
1264 continue;
1265 add_uses_to_worklist(use);
1266 if (use->is_Field() && use->as_Field()->is_oop()) {
1267 // Put on worklist all field's uses (loads) and
1268 // related field nodes (same base and offset).
1269 add_field_uses_to_worklist(use->as_Field());
1270 }
1271 }
1272 }
1273 while(_worklist.length() > 0) {
1274 PointsToNode* use = _worklist.pop();
1275 if (PointsToNode::is_base_use(use)) {
1276 // Add reference from jobj to field and from field to jobj (field's base).
1277 use = PointsToNode::get_use_node(use)->as_Field();
1278 if (add_base(use->as_Field(), jobj)) {
1279 new_edges++;
1280 }
1281 continue;
1282 }
1283 assert(!use->is_JavaObject(), "sanity");
1284 if (use->is_Arraycopy()) {
1285 if (jobj == null_obj) // NULL object does not have field edges
1286 continue;
1287 // Added edge from Arraycopy node to arraycopy's source java object
1288 if (add_edge(use, jobj)) {
1289 jobj->set_arraycopy_src();
1290 new_edges++;
1291 }
1292 // and stop here.
1293 continue;
1294 }
1295 if (!add_edge(use, jobj))
1296 continue; // No new edge added, there was such edge already.
1297 new_edges++;
1298 if (use->is_LocalVar()) {
1299 add_uses_to_worklist(use);
1300 if (use->arraycopy_dst()) {
1301 for (EdgeIterator i(use); i.has_next(); i.next()) {
1302 PointsToNode* e = i.get();
1303 if (e->is_Arraycopy()) {
1304 if (jobj == null_obj) // NULL object does not have field edges
1305 continue;
1306 // Add edge from arraycopy's destination java object to Arraycopy node.
1307 if (add_edge(jobj, e)) {
1308 new_edges++;
1309 jobj->set_arraycopy_dst();
1310 }
1311 }
1312 }
1313 }
1314 } else {
1315 // Added new edge to stored in field values.
1316 // Put on worklist all field's uses (loads) and
1317 // related field nodes (same base and offset).
1318 add_field_uses_to_worklist(use->as_Field());
1319 }
1320 }
1321 return new_edges;
1322 }
1324 // Put on worklist all related field nodes.
1325 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1326 assert(field->is_oop(), "sanity");
1327 int offset = field->offset();
1328 add_uses_to_worklist(field);
1329 // Loop over all bases of this field and push on worklist Field nodes
1330 // with the same offset and base (since they may reference the same field).
1331 for (BaseIterator i(field); i.has_next(); i.next()) {
1332 PointsToNode* base = i.get();
1333 add_fields_to_worklist(field, base);
1334 // Check if the base was source object of arraycopy and go over arraycopy's
1335 // destination objects since values stored to a field of source object are
1336 // accessable by uses (loads) of fields of destination objects.
1337 if (base->arraycopy_src()) {
1338 for (UseIterator j(base); j.has_next(); j.next()) {
1339 PointsToNode* arycp = j.get();
1340 if (arycp->is_Arraycopy()) {
1341 for (UseIterator k(arycp); k.has_next(); k.next()) {
1342 PointsToNode* abase = k.get();
1343 if (abase->arraycopy_dst() && abase != base) {
1344 // Look for the same arracopy reference.
1345 add_fields_to_worklist(field, abase);
1346 }
1347 }
1348 }
1349 }
1350 }
1351 }
1352 }
1354 // Put on worklist all related field nodes.
1355 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1356 int offset = field->offset();
1357 if (base->is_LocalVar()) {
1358 for (UseIterator j(base); j.has_next(); j.next()) {
1359 PointsToNode* f = j.get();
1360 if (PointsToNode::is_base_use(f)) { // Field
1361 f = PointsToNode::get_use_node(f);
1362 if (f == field || !f->as_Field()->is_oop())
1363 continue;
1364 int offs = f->as_Field()->offset();
1365 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1366 add_to_worklist(f);
1367 }
1368 }
1369 }
1370 } else {
1371 assert(base->is_JavaObject(), "sanity");
1372 if (// Skip phantom_object since it is only used to indicate that
1373 // this field's content globally escapes.
1374 (base != phantom_obj) &&
1375 // NULL object node does not have fields.
1376 (base != null_obj)) {
1377 for (EdgeIterator i(base); i.has_next(); i.next()) {
1378 PointsToNode* f = i.get();
1379 // Skip arraycopy edge since store to destination object field
1380 // does not update value in source object field.
1381 if (f->is_Arraycopy()) {
1382 assert(base->arraycopy_dst(), "sanity");
1383 continue;
1384 }
1385 if (f == field || !f->as_Field()->is_oop())
1386 continue;
1387 int offs = f->as_Field()->offset();
1388 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1389 add_to_worklist(f);
1390 }
1391 }
1392 }
1393 }
1394 }
1396 // Find fields which have unknown value.
1397 int ConnectionGraph::find_field_value(FieldNode* field) {
1398 // Escaped fields should have init value already.
1399 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1400 int new_edges = 0;
1401 for (BaseIterator i(field); i.has_next(); i.next()) {
1402 PointsToNode* base = i.get();
1403 if (base->is_JavaObject()) {
1404 // Skip Allocate's fields which will be processed later.
1405 if (base->ideal_node()->is_Allocate())
1406 return 0;
1407 assert(base == null_obj, "only NULL ptr base expected here");
1408 }
1409 }
1410 if (add_edge(field, phantom_obj)) {
1411 // New edge was added
1412 new_edges++;
1413 add_field_uses_to_worklist(field);
1414 }
1415 return new_edges;
1416 }
1418 // Find fields initializing values for allocations.
1419 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1420 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1421 int new_edges = 0;
1422 Node* alloc = pta->ideal_node();
1423 if (init_val == phantom_obj) {
1424 // Do nothing for Allocate nodes since its fields values are "known".
1425 if (alloc->is_Allocate())
1426 return 0;
1427 assert(alloc->as_CallStaticJava(), "sanity");
1428 #ifdef ASSERT
1429 if (alloc->as_CallStaticJava()->method() == NULL) {
1430 const char* name = alloc->as_CallStaticJava()->_name;
1431 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1432 }
1433 #endif
1434 // Non-escaped allocation returned from Java or runtime call have
1435 // unknown values in fields.
1436 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1437 PointsToNode* field = i.get();
1438 if (field->is_Field() && field->as_Field()->is_oop()) {
1439 if (add_edge(field, phantom_obj)) {
1440 // New edge was added
1441 new_edges++;
1442 add_field_uses_to_worklist(field->as_Field());
1443 }
1444 }
1445 }
1446 return new_edges;
1447 }
1448 assert(init_val == null_obj, "sanity");
1449 // Do nothing for Call nodes since its fields values are unknown.
1450 if (!alloc->is_Allocate())
1451 return 0;
1453 InitializeNode* ini = alloc->as_Allocate()->initialization();
1454 Compile* C = _compile;
1455 bool visited_bottom_offset = false;
1456 GrowableArray<int> offsets_worklist;
1458 // Check if an oop field's initializing value is recorded and add
1459 // a corresponding NULL if field's value if it is not recorded.
1460 // Connection Graph does not record a default initialization by NULL
1461 // captured by Initialize node.
1462 //
1463 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1464 PointsToNode* field = i.get(); // Field (AddP)
1465 if (!field->is_Field() || !field->as_Field()->is_oop())
1466 continue; // Not oop field
1467 int offset = field->as_Field()->offset();
1468 if (offset == Type::OffsetBot) {
1469 if (!visited_bottom_offset) {
1470 // OffsetBot is used to reference array's element,
1471 // always add reference to NULL to all Field nodes since we don't
1472 // known which element is referenced.
1473 if (add_edge(field, null_obj)) {
1474 // New edge was added
1475 new_edges++;
1476 add_field_uses_to_worklist(field->as_Field());
1477 visited_bottom_offset = true;
1478 }
1479 }
1480 } else {
1481 // Check only oop fields.
1482 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1483 if (adr_type->isa_rawptr()) {
1484 #ifdef ASSERT
1485 // Raw pointers are used for initializing stores so skip it
1486 // since it should be recorded already
1487 Node* base = get_addp_base(field->ideal_node());
1488 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1489 (base->in(0) == alloc),"unexpected pointer type");
1490 #endif
1491 continue;
1492 }
1493 if (!offsets_worklist.contains(offset)) {
1494 offsets_worklist.append(offset);
1495 Node* value = NULL;
1496 if (ini != NULL) {
1497 // StoreP::memory_type() == T_ADDRESS
1498 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1499 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1500 // Make sure initializing store has the same type as this AddP.
1501 // This AddP may reference non existing field because it is on a
1502 // dead branch of bimorphic call which is not eliminated yet.
1503 if (store != NULL && store->is_Store() &&
1504 store->as_Store()->memory_type() == ft) {
1505 value = store->in(MemNode::ValueIn);
1506 #ifdef ASSERT
1507 if (VerifyConnectionGraph) {
1508 // Verify that AddP already points to all objects the value points to.
1509 PointsToNode* val = ptnode_adr(value->_idx);
1510 assert((val != NULL), "should be processed already");
1511 PointsToNode* missed_obj = NULL;
1512 if (val->is_JavaObject()) {
1513 if (!field->points_to(val->as_JavaObject())) {
1514 missed_obj = val;
1515 }
1516 } else {
1517 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1518 tty->print_cr("----------init store has invalid value -----");
1519 store->dump();
1520 val->dump();
1521 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1522 }
1523 for (EdgeIterator j(val); j.has_next(); j.next()) {
1524 PointsToNode* obj = j.get();
1525 if (obj->is_JavaObject()) {
1526 if (!field->points_to(obj->as_JavaObject())) {
1527 missed_obj = obj;
1528 break;
1529 }
1530 }
1531 }
1532 }
1533 if (missed_obj != NULL) {
1534 tty->print_cr("----------field---------------------------------");
1535 field->dump();
1536 tty->print_cr("----------missed referernce to object-----------");
1537 missed_obj->dump();
1538 tty->print_cr("----------object referernced by init store -----");
1539 store->dump();
1540 val->dump();
1541 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1542 }
1543 }
1544 #endif
1545 } else {
1546 // There could be initializing stores which follow allocation.
1547 // For example, a volatile field store is not collected
1548 // by Initialize node.
1549 //
1550 // Need to check for dependent loads to separate such stores from
1551 // stores which follow loads. For now, add initial value NULL so
1552 // that compare pointers optimization works correctly.
1553 }
1554 }
1555 if (value == NULL) {
1556 // A field's initializing value was not recorded. Add NULL.
1557 if (add_edge(field, null_obj)) {
1558 // New edge was added
1559 new_edges++;
1560 add_field_uses_to_worklist(field->as_Field());
1561 }
1562 }
1563 }
1564 }
1565 }
1566 return new_edges;
1567 }
1569 // Adjust scalar_replaceable state after Connection Graph is built.
1570 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1571 // Search for non-escaping objects which are not scalar replaceable
1572 // and mark them to propagate the state to referenced objects.
1574 // 1. An object is not scalar replaceable if the field into which it is
1575 // stored has unknown offset (stored into unknown element of an array).
1576 //
1577 for (UseIterator i(jobj); i.has_next(); i.next()) {
1578 PointsToNode* use = i.get();
1579 assert(!use->is_Arraycopy(), "sanity");
1580 if (use->is_Field()) {
1581 FieldNode* field = use->as_Field();
1582 assert(field->is_oop() && field->scalar_replaceable() &&
1583 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1584 if (field->offset() == Type::OffsetBot) {
1585 jobj->set_scalar_replaceable(false);
1586 return;
1587 }
1588 // 2. An object is not scalar replaceable if the field into which it is
1589 // stored has multiple bases one of which is null.
1590 if (field->base_count() > 1) {
1591 for (BaseIterator i(field); i.has_next(); i.next()) {
1592 PointsToNode* base = i.get();
1593 if (base == null_obj) {
1594 jobj->set_scalar_replaceable(false);
1595 return;
1596 }
1597 }
1598 }
1599 }
1600 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1601 // 3. An object is not scalar replaceable if it is merged with other objects.
1602 for (EdgeIterator j(use); j.has_next(); j.next()) {
1603 PointsToNode* ptn = j.get();
1604 if (ptn->is_JavaObject() && ptn != jobj) {
1605 // Mark all objects.
1606 jobj->set_scalar_replaceable(false);
1607 ptn->set_scalar_replaceable(false);
1608 }
1609 }
1610 if (!jobj->scalar_replaceable()) {
1611 return;
1612 }
1613 }
1615 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1616 // Non-escaping object node should point only to field nodes.
1617 FieldNode* field = j.get()->as_Field();
1618 int offset = field->as_Field()->offset();
1620 // 4. An object is not scalar replaceable if it has a field with unknown
1621 // offset (array's element is accessed in loop).
1622 if (offset == Type::OffsetBot) {
1623 jobj->set_scalar_replaceable(false);
1624 return;
1625 }
1626 // 5. Currently an object is not scalar replaceable if a LoadStore node
1627 // access its field since the field value is unknown after it.
1628 //
1629 Node* n = field->ideal_node();
1630 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1631 if (n->fast_out(i)->is_LoadStore()) {
1632 jobj->set_scalar_replaceable(false);
1633 return;
1634 }
1635 }
1637 // 6. Or the address may point to more then one object. This may produce
1638 // the false positive result (set not scalar replaceable)
1639 // since the flow-insensitive escape analysis can't separate
1640 // the case when stores overwrite the field's value from the case
1641 // when stores happened on different control branches.
1642 //
1643 // Note: it will disable scalar replacement in some cases:
1644 //
1645 // Point p[] = new Point[1];
1646 // p[0] = new Point(); // Will be not scalar replaced
1647 //
1648 // but it will save us from incorrect optimizations in next cases:
1649 //
1650 // Point p[] = new Point[1];
1651 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1652 //
1653 if (field->base_count() > 1) {
1654 for (BaseIterator i(field); i.has_next(); i.next()) {
1655 PointsToNode* base = i.get();
1656 // Don't take into account LocalVar nodes which
1657 // may point to only one object which should be also
1658 // this field's base by now.
1659 if (base->is_JavaObject() && base != jobj) {
1660 // Mark all bases.
1661 jobj->set_scalar_replaceable(false);
1662 base->set_scalar_replaceable(false);
1663 }
1664 }
1665 }
1666 }
1667 }
1669 #ifdef ASSERT
1670 void ConnectionGraph::verify_connection_graph(
1671 GrowableArray<PointsToNode*>& ptnodes_worklist,
1672 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1673 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1674 GrowableArray<Node*>& addp_worklist) {
1675 // Verify that graph is complete - no new edges could be added.
1676 int java_objects_length = java_objects_worklist.length();
1677 int non_escaped_length = non_escaped_worklist.length();
1678 int new_edges = 0;
1679 for (int next = 0; next < java_objects_length; ++next) {
1680 JavaObjectNode* ptn = java_objects_worklist.at(next);
1681 new_edges += add_java_object_edges(ptn, true);
1682 }
1683 assert(new_edges == 0, "graph was not complete");
1684 // Verify that escape state is final.
1685 int length = non_escaped_worklist.length();
1686 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1687 assert((non_escaped_length == non_escaped_worklist.length()) &&
1688 (non_escaped_length == length) &&
1689 (_worklist.length() == 0), "escape state was not final");
1691 // Verify fields information.
1692 int addp_length = addp_worklist.length();
1693 for (int next = 0; next < addp_length; ++next ) {
1694 Node* n = addp_worklist.at(next);
1695 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1696 if (field->is_oop()) {
1697 // Verify that field has all bases
1698 Node* base = get_addp_base(n);
1699 PointsToNode* ptn = ptnode_adr(base->_idx);
1700 if (ptn->is_JavaObject()) {
1701 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1702 } else {
1703 assert(ptn->is_LocalVar(), "sanity");
1704 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1705 PointsToNode* e = i.get();
1706 if (e->is_JavaObject()) {
1707 assert(field->has_base(e->as_JavaObject()), "sanity");
1708 }
1709 }
1710 }
1711 // Verify that all fields have initializing values.
1712 if (field->edge_count() == 0) {
1713 tty->print_cr("----------field does not have references----------");
1714 field->dump();
1715 for (BaseIterator i(field); i.has_next(); i.next()) {
1716 PointsToNode* base = i.get();
1717 tty->print_cr("----------field has next base---------------------");
1718 base->dump();
1719 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1720 tty->print_cr("----------base has fields-------------------------");
1721 for (EdgeIterator j(base); j.has_next(); j.next()) {
1722 j.get()->dump();
1723 }
1724 tty->print_cr("----------base has references---------------------");
1725 for (UseIterator j(base); j.has_next(); j.next()) {
1726 j.get()->dump();
1727 }
1728 }
1729 }
1730 for (UseIterator i(field); i.has_next(); i.next()) {
1731 i.get()->dump();
1732 }
1733 assert(field->edge_count() > 0, "sanity");
1734 }
1735 }
1736 }
1737 }
1738 #endif
1740 // Optimize ideal graph.
1741 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1742 GrowableArray<Node*>& storestore_worklist) {
1743 Compile* C = _compile;
1744 PhaseIterGVN* igvn = _igvn;
1745 if (EliminateLocks) {
1746 // Mark locks before changing ideal graph.
1747 int cnt = C->macro_count();
1748 for( int i=0; i < cnt; i++ ) {
1749 Node *n = C->macro_node(i);
1750 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1751 AbstractLockNode* alock = n->as_AbstractLock();
1752 if (!alock->is_non_esc_obj()) {
1753 if (not_global_escape(alock->obj_node())) {
1754 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1755 // The lock could be marked eliminated by lock coarsening
1756 // code during first IGVN before EA. Replace coarsened flag
1757 // to eliminate all associated locks/unlocks.
1758 alock->set_non_esc_obj();
1759 }
1760 }
1761 }
1762 }
1763 }
1765 if (OptimizePtrCompare) {
1766 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1767 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1768 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1769 // Optimize objects compare.
1770 while (ptr_cmp_worklist.length() != 0) {
1771 Node *n = ptr_cmp_worklist.pop();
1772 Node *res = optimize_ptr_compare(n);
1773 if (res != NULL) {
1774 #ifndef PRODUCT
1775 if (PrintOptimizePtrCompare) {
1776 tty->print_cr("++++ Replaced: %d %s(%d,%d) --> %s", n->_idx, (n->Opcode() == Op_CmpP ? "CmpP" : "CmpN"), n->in(1)->_idx, n->in(2)->_idx, (res == _pcmp_eq ? "EQ" : "NotEQ"));
1777 if (Verbose) {
1778 n->dump(1);
1779 }
1780 }
1781 #endif
1782 igvn->replace_node(n, res);
1783 }
1784 }
1785 // cleanup
1786 if (_pcmp_neq->outcnt() == 0)
1787 igvn->hash_delete(_pcmp_neq);
1788 if (_pcmp_eq->outcnt() == 0)
1789 igvn->hash_delete(_pcmp_eq);
1790 }
1792 // For MemBarStoreStore nodes added in library_call.cpp, check
1793 // escape status of associated AllocateNode and optimize out
1794 // MemBarStoreStore node if the allocated object never escapes.
1795 while (storestore_worklist.length() != 0) {
1796 Node *n = storestore_worklist.pop();
1797 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1798 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1799 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1800 if (not_global_escape(alloc)) {
1801 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1802 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1803 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1804 igvn->register_new_node_with_optimizer(mb);
1805 igvn->replace_node(storestore, mb);
1806 }
1807 }
1808 }
1810 // Optimize objects compare.
1811 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1812 assert(OptimizePtrCompare, "sanity");
1813 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1814 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1815 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1816 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1817 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1818 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1820 // Check simple cases first.
1821 if (jobj1 != NULL) {
1822 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1823 if (jobj1 == jobj2) {
1824 // Comparing the same not escaping object.
1825 return _pcmp_eq;
1826 }
1827 Node* obj = jobj1->ideal_node();
1828 // Comparing not escaping allocation.
1829 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1830 !ptn2->points_to(jobj1)) {
1831 return _pcmp_neq; // This includes nullness check.
1832 }
1833 }
1834 }
1835 if (jobj2 != NULL) {
1836 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1837 Node* obj = jobj2->ideal_node();
1838 // Comparing not escaping allocation.
1839 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1840 !ptn1->points_to(jobj2)) {
1841 return _pcmp_neq; // This includes nullness check.
1842 }
1843 }
1844 }
1845 if (jobj1 != NULL && jobj1 != phantom_obj &&
1846 jobj2 != NULL && jobj2 != phantom_obj &&
1847 jobj1->ideal_node()->is_Con() &&
1848 jobj2->ideal_node()->is_Con()) {
1849 // Klass or String constants compare. Need to be careful with
1850 // compressed pointers - compare types of ConN and ConP instead of nodes.
1851 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1852 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1853 if (t1->make_ptr() == t2->make_ptr()) {
1854 return _pcmp_eq;
1855 } else {
1856 return _pcmp_neq;
1857 }
1858 }
1859 if (ptn1->meet(ptn2)) {
1860 return NULL; // Sets are not disjoint
1861 }
1863 // Sets are disjoint.
1864 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1865 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1866 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1867 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1868 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1869 set2_has_unknown_ptr && set1_has_null_ptr) {
1870 // Check nullness of unknown object.
1871 return NULL;
1872 }
1874 // Disjointness by itself is not sufficient since
1875 // alias analysis is not complete for escaped objects.
1876 // Disjoint sets are definitely unrelated only when
1877 // at least one set has only not escaping allocations.
1878 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1879 if (ptn1->non_escaping_allocation()) {
1880 return _pcmp_neq;
1881 }
1882 }
1883 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1884 if (ptn2->non_escaping_allocation()) {
1885 return _pcmp_neq;
1886 }
1887 }
1888 return NULL;
1889 }
1891 // Connection Graph constuction functions.
1893 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1894 PointsToNode* ptadr = _nodes.at(n->_idx);
1895 if (ptadr != NULL) {
1896 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1897 return;
1898 }
1899 Compile* C = _compile;
1900 ptadr = new (C->comp_arena()) LocalVarNode(C, n, es);
1901 _nodes.at_put(n->_idx, ptadr);
1902 }
1904 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1905 PointsToNode* ptadr = _nodes.at(n->_idx);
1906 if (ptadr != NULL) {
1907 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1908 return;
1909 }
1910 Compile* C = _compile;
1911 ptadr = new (C->comp_arena()) JavaObjectNode(C, n, es);
1912 _nodes.at_put(n->_idx, ptadr);
1913 }
1915 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1916 PointsToNode* ptadr = _nodes.at(n->_idx);
1917 if (ptadr != NULL) {
1918 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1919 return;
1920 }
1921 bool unsafe = false;
1922 bool is_oop = is_oop_field(n, offset, &unsafe);
1923 if (unsafe) {
1924 es = PointsToNode::GlobalEscape;
1925 }
1926 Compile* C = _compile;
1927 FieldNode* field = new (C->comp_arena()) FieldNode(C, n, es, offset, is_oop);
1928 _nodes.at_put(n->_idx, field);
1929 }
1931 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1932 PointsToNode* src, PointsToNode* dst) {
1933 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1934 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1935 PointsToNode* ptadr = _nodes.at(n->_idx);
1936 if (ptadr != NULL) {
1937 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1938 return;
1939 }
1940 Compile* C = _compile;
1941 ptadr = new (C->comp_arena()) ArraycopyNode(C, n, es);
1942 _nodes.at_put(n->_idx, ptadr);
1943 // Add edge from arraycopy node to source object.
1944 (void)add_edge(ptadr, src);
1945 src->set_arraycopy_src();
1946 // Add edge from destination object to arraycopy node.
1947 (void)add_edge(dst, ptadr);
1948 dst->set_arraycopy_dst();
1949 }
1951 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1952 const Type* adr_type = n->as_AddP()->bottom_type();
1953 BasicType bt = T_INT;
1954 if (offset == Type::OffsetBot) {
1955 // Check only oop fields.
1956 if (!adr_type->isa_aryptr() ||
1957 (adr_type->isa_aryptr()->klass() == NULL) ||
1958 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1959 // OffsetBot is used to reference array's element. Ignore first AddP.
1960 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
1961 bt = T_OBJECT;
1962 }
1963 }
1964 } else if (offset != oopDesc::klass_offset_in_bytes()) {
1965 if (adr_type->isa_instptr()) {
1966 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
1967 if (field != NULL) {
1968 bt = field->layout_type();
1969 } else {
1970 // Check for unsafe oop field access
1971 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1972 int opcode = n->fast_out(i)->Opcode();
1973 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1974 opcode == Op_StoreN || opcode == Op_LoadN) {
1975 bt = T_OBJECT;
1976 (*unsafe) = true;
1977 break;
1978 }
1979 }
1980 }
1981 } else if (adr_type->isa_aryptr()) {
1982 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1983 // Ignore array length load.
1984 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
1985 // Ignore first AddP.
1986 } else {
1987 const Type* elemtype = adr_type->isa_aryptr()->elem();
1988 bt = elemtype->array_element_basic_type();
1989 }
1990 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
1991 // Allocation initialization, ThreadLocal field access, unsafe access
1992 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1993 int opcode = n->fast_out(i)->Opcode();
1994 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1995 opcode == Op_StoreN || opcode == Op_LoadN) {
1996 bt = T_OBJECT;
1997 break;
1998 }
1999 }
2000 }
2001 }
2002 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2003 }
2005 // Returns unique pointed java object or NULL.
2006 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2007 assert(!_collecting, "should not call when contructed graph");
2008 // If the node was created after the escape computation we can't answer.
2009 uint idx = n->_idx;
2010 if (idx >= nodes_size()) {
2011 return NULL;
2012 }
2013 PointsToNode* ptn = ptnode_adr(idx);
2014 if (ptn->is_JavaObject()) {
2015 return ptn->as_JavaObject();
2016 }
2017 assert(ptn->is_LocalVar(), "sanity");
2018 // Check all java objects it points to.
2019 JavaObjectNode* jobj = NULL;
2020 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2021 PointsToNode* e = i.get();
2022 if (e->is_JavaObject()) {
2023 if (jobj == NULL) {
2024 jobj = e->as_JavaObject();
2025 } else if (jobj != e) {
2026 return NULL;
2027 }
2028 }
2029 }
2030 return jobj;
2031 }
2033 // Return true if this node points only to non-escaping allocations.
2034 bool PointsToNode::non_escaping_allocation() {
2035 if (is_JavaObject()) {
2036 Node* n = ideal_node();
2037 if (n->is_Allocate() || n->is_CallStaticJava()) {
2038 return (escape_state() == PointsToNode::NoEscape);
2039 } else {
2040 return false;
2041 }
2042 }
2043 assert(is_LocalVar(), "sanity");
2044 // Check all java objects it points to.
2045 for (EdgeIterator i(this); i.has_next(); i.next()) {
2046 PointsToNode* e = i.get();
2047 if (e->is_JavaObject()) {
2048 Node* n = e->ideal_node();
2049 if ((e->escape_state() != PointsToNode::NoEscape) ||
2050 !(n->is_Allocate() || n->is_CallStaticJava())) {
2051 return false;
2052 }
2053 }
2054 }
2055 return true;
2056 }
2058 // Return true if we know the node does not escape globally.
2059 bool ConnectionGraph::not_global_escape(Node *n) {
2060 assert(!_collecting, "should not call during graph construction");
2061 // If the node was created after the escape computation we can't answer.
2062 uint idx = n->_idx;
2063 if (idx >= nodes_size()) {
2064 return false;
2065 }
2066 PointsToNode* ptn = ptnode_adr(idx);
2067 PointsToNode::EscapeState es = ptn->escape_state();
2068 // If we have already computed a value, return it.
2069 if (es >= PointsToNode::GlobalEscape)
2070 return false;
2071 if (ptn->is_JavaObject()) {
2072 return true; // (es < PointsToNode::GlobalEscape);
2073 }
2074 assert(ptn->is_LocalVar(), "sanity");
2075 // Check all java objects it points to.
2076 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2077 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2078 return false;
2079 }
2080 return true;
2081 }
2084 // Helper functions
2086 // Return true if this node points to specified node or nodes it points to.
2087 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2088 if (is_JavaObject()) {
2089 return (this == ptn);
2090 }
2091 assert(is_LocalVar() || is_Field(), "sanity");
2092 for (EdgeIterator i(this); i.has_next(); i.next()) {
2093 if (i.get() == ptn)
2094 return true;
2095 }
2096 return false;
2097 }
2099 // Return true if one node points to an other.
2100 bool PointsToNode::meet(PointsToNode* ptn) {
2101 if (this == ptn) {
2102 return true;
2103 } else if (ptn->is_JavaObject()) {
2104 return this->points_to(ptn->as_JavaObject());
2105 } else if (this->is_JavaObject()) {
2106 return ptn->points_to(this->as_JavaObject());
2107 }
2108 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2109 int ptn_count = ptn->edge_count();
2110 for (EdgeIterator i(this); i.has_next(); i.next()) {
2111 PointsToNode* this_e = i.get();
2112 for (int j = 0; j < ptn_count; j++) {
2113 if (this_e == ptn->edge(j))
2114 return true;
2115 }
2116 }
2117 return false;
2118 }
2120 #ifdef ASSERT
2121 // Return true if bases point to this java object.
2122 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2123 for (BaseIterator i(this); i.has_next(); i.next()) {
2124 if (i.get() == jobj)
2125 return true;
2126 }
2127 return false;
2128 }
2129 #endif
2131 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2132 const Type *adr_type = phase->type(adr);
2133 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2134 adr->in(AddPNode::Address)->is_Proj() &&
2135 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2136 // We are computing a raw address for a store captured by an Initialize
2137 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2138 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2139 assert(offs != Type::OffsetBot ||
2140 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2141 "offset must be a constant or it is initialization of array");
2142 return offs;
2143 }
2144 const TypePtr *t_ptr = adr_type->isa_ptr();
2145 assert(t_ptr != NULL, "must be a pointer type");
2146 return t_ptr->offset();
2147 }
2149 Node* ConnectionGraph::get_addp_base(Node *addp) {
2150 assert(addp->is_AddP(), "must be AddP");
2151 //
2152 // AddP cases for Base and Address inputs:
2153 // case #1. Direct object's field reference:
2154 // Allocate
2155 // |
2156 // Proj #5 ( oop result )
2157 // |
2158 // CheckCastPP (cast to instance type)
2159 // | |
2160 // AddP ( base == address )
2161 //
2162 // case #2. Indirect object's field reference:
2163 // Phi
2164 // |
2165 // CastPP (cast to instance type)
2166 // | |
2167 // AddP ( base == address )
2168 //
2169 // case #3. Raw object's field reference for Initialize node:
2170 // Allocate
2171 // |
2172 // Proj #5 ( oop result )
2173 // top |
2174 // \ |
2175 // AddP ( base == top )
2176 //
2177 // case #4. Array's element reference:
2178 // {CheckCastPP | CastPP}
2179 // | | |
2180 // | AddP ( array's element offset )
2181 // | |
2182 // AddP ( array's offset )
2183 //
2184 // case #5. Raw object's field reference for arraycopy stub call:
2185 // The inline_native_clone() case when the arraycopy stub is called
2186 // after the allocation before Initialize and CheckCastPP nodes.
2187 // Allocate
2188 // |
2189 // Proj #5 ( oop result )
2190 // | |
2191 // AddP ( base == address )
2192 //
2193 // case #6. Constant Pool, ThreadLocal, CastX2P or
2194 // Raw object's field reference:
2195 // {ConP, ThreadLocal, CastX2P, raw Load}
2196 // top |
2197 // \ |
2198 // AddP ( base == top )
2199 //
2200 // case #7. Klass's field reference.
2201 // LoadKlass
2202 // | |
2203 // AddP ( base == address )
2204 //
2205 // case #8. narrow Klass's field reference.
2206 // LoadNKlass
2207 // |
2208 // DecodeN
2209 // | |
2210 // AddP ( base == address )
2211 //
2212 Node *base = addp->in(AddPNode::Base);
2213 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2214 base = addp->in(AddPNode::Address);
2215 while (base->is_AddP()) {
2216 // Case #6 (unsafe access) may have several chained AddP nodes.
2217 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2218 base = base->in(AddPNode::Address);
2219 }
2220 Node* uncast_base = base->uncast();
2221 int opcode = uncast_base->Opcode();
2222 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2223 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2224 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2225 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2226 }
2227 return base;
2228 }
2230 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2231 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2232 Node* addp2 = addp->raw_out(0);
2233 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2234 addp2->in(AddPNode::Base) == n &&
2235 addp2->in(AddPNode::Address) == addp) {
2236 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2237 //
2238 // Find array's offset to push it on worklist first and
2239 // as result process an array's element offset first (pushed second)
2240 // to avoid CastPP for the array's offset.
2241 // Otherwise the inserted CastPP (LocalVar) will point to what
2242 // the AddP (Field) points to. Which would be wrong since
2243 // the algorithm expects the CastPP has the same point as
2244 // as AddP's base CheckCastPP (LocalVar).
2245 //
2246 // ArrayAllocation
2247 // |
2248 // CheckCastPP
2249 // |
2250 // memProj (from ArrayAllocation CheckCastPP)
2251 // | ||
2252 // | || Int (element index)
2253 // | || | ConI (log(element size))
2254 // | || | /
2255 // | || LShift
2256 // | || /
2257 // | AddP (array's element offset)
2258 // | |
2259 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2260 // | / /
2261 // AddP (array's offset)
2262 // |
2263 // Load/Store (memory operation on array's element)
2264 //
2265 return addp2;
2266 }
2267 return NULL;
2268 }
2270 //
2271 // Adjust the type and inputs of an AddP which computes the
2272 // address of a field of an instance
2273 //
2274 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2275 PhaseGVN* igvn = _igvn;
2276 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2277 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2278 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2279 if (t == NULL) {
2280 // We are computing a raw address for a store captured by an Initialize
2281 // compute an appropriate address type (cases #3 and #5).
2282 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2283 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2284 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2285 assert(offs != Type::OffsetBot, "offset must be a constant");
2286 t = base_t->add_offset(offs)->is_oopptr();
2287 }
2288 int inst_id = base_t->instance_id();
2289 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2290 "old type must be non-instance or match new type");
2292 // The type 't' could be subclass of 'base_t'.
2293 // As result t->offset() could be large then base_t's size and it will
2294 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2295 // constructor verifies correctness of the offset.
2296 //
2297 // It could happened on subclass's branch (from the type profiling
2298 // inlining) which was not eliminated during parsing since the exactness
2299 // of the allocation type was not propagated to the subclass type check.
2300 //
2301 // Or the type 't' could be not related to 'base_t' at all.
2302 // It could happened when CHA type is different from MDO type on a dead path
2303 // (for example, from instanceof check) which is not collapsed during parsing.
2304 //
2305 // Do nothing for such AddP node and don't process its users since
2306 // this code branch will go away.
2307 //
2308 if (!t->is_known_instance() &&
2309 !base_t->klass()->is_subtype_of(t->klass())) {
2310 return false; // bail out
2311 }
2312 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2313 // Do NOT remove the next line: ensure a new alias index is allocated
2314 // for the instance type. Note: C++ will not remove it since the call
2315 // has side effect.
2316 int alias_idx = _compile->get_alias_index(tinst);
2317 igvn->set_type(addp, tinst);
2318 // record the allocation in the node map
2319 set_map(addp, get_map(base->_idx));
2320 // Set addp's Base and Address to 'base'.
2321 Node *abase = addp->in(AddPNode::Base);
2322 Node *adr = addp->in(AddPNode::Address);
2323 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2324 adr->in(0)->_idx == (uint)inst_id) {
2325 // Skip AddP cases #3 and #5.
2326 } else {
2327 assert(!abase->is_top(), "sanity"); // AddP case #3
2328 if (abase != base) {
2329 igvn->hash_delete(addp);
2330 addp->set_req(AddPNode::Base, base);
2331 if (abase == adr) {
2332 addp->set_req(AddPNode::Address, base);
2333 } else {
2334 // AddP case #4 (adr is array's element offset AddP node)
2335 #ifdef ASSERT
2336 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2337 assert(adr->is_AddP() && atype != NULL &&
2338 atype->instance_id() == inst_id, "array's element offset should be processed first");
2339 #endif
2340 }
2341 igvn->hash_insert(addp);
2342 }
2343 }
2344 // Put on IGVN worklist since at least addp's type was changed above.
2345 record_for_optimizer(addp);
2346 return true;
2347 }
2349 //
2350 // Create a new version of orig_phi if necessary. Returns either the newly
2351 // created phi or an existing phi. Sets create_new to indicate whether a new
2352 // phi was created. Cache the last newly created phi in the node map.
2353 //
2354 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2355 Compile *C = _compile;
2356 PhaseGVN* igvn = _igvn;
2357 new_created = false;
2358 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2359 // nothing to do if orig_phi is bottom memory or matches alias_idx
2360 if (phi_alias_idx == alias_idx) {
2361 return orig_phi;
2362 }
2363 // Have we recently created a Phi for this alias index?
2364 PhiNode *result = get_map_phi(orig_phi->_idx);
2365 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2366 return result;
2367 }
2368 // Previous check may fail when the same wide memory Phi was split into Phis
2369 // for different memory slices. Search all Phis for this region.
2370 if (result != NULL) {
2371 Node* region = orig_phi->in(0);
2372 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2373 Node* phi = region->fast_out(i);
2374 if (phi->is_Phi() &&
2375 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2376 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2377 return phi->as_Phi();
2378 }
2379 }
2380 }
2381 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) {
2382 if (C->do_escape_analysis() == true && !C->failing()) {
2383 // Retry compilation without escape analysis.
2384 // If this is the first failure, the sentinel string will "stick"
2385 // to the Compile object, and the C2Compiler will see it and retry.
2386 C->record_failure(C2Compiler::retry_no_escape_analysis());
2387 }
2388 return NULL;
2389 }
2390 orig_phi_worklist.append_if_missing(orig_phi);
2391 const TypePtr *atype = C->get_adr_type(alias_idx);
2392 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2393 C->copy_node_notes_to(result, orig_phi);
2394 igvn->set_type(result, result->bottom_type());
2395 record_for_optimizer(result);
2396 set_map(orig_phi, result);
2397 new_created = true;
2398 return result;
2399 }
2401 //
2402 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2403 // specified alias index.
2404 //
2405 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2406 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2407 Compile *C = _compile;
2408 PhaseGVN* igvn = _igvn;
2409 bool new_phi_created;
2410 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2411 if (!new_phi_created) {
2412 return result;
2413 }
2414 GrowableArray<PhiNode *> phi_list;
2415 GrowableArray<uint> cur_input;
2416 PhiNode *phi = orig_phi;
2417 uint idx = 1;
2418 bool finished = false;
2419 while(!finished) {
2420 while (idx < phi->req()) {
2421 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2422 if (mem != NULL && mem->is_Phi()) {
2423 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2424 if (new_phi_created) {
2425 // found an phi for which we created a new split, push current one on worklist and begin
2426 // processing new one
2427 phi_list.push(phi);
2428 cur_input.push(idx);
2429 phi = mem->as_Phi();
2430 result = newphi;
2431 idx = 1;
2432 continue;
2433 } else {
2434 mem = newphi;
2435 }
2436 }
2437 if (C->failing()) {
2438 return NULL;
2439 }
2440 result->set_req(idx++, mem);
2441 }
2442 #ifdef ASSERT
2443 // verify that the new Phi has an input for each input of the original
2444 assert( phi->req() == result->req(), "must have same number of inputs.");
2445 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2446 #endif
2447 // Check if all new phi's inputs have specified alias index.
2448 // Otherwise use old phi.
2449 for (uint i = 1; i < phi->req(); i++) {
2450 Node* in = result->in(i);
2451 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2452 }
2453 // we have finished processing a Phi, see if there are any more to do
2454 finished = (phi_list.length() == 0 );
2455 if (!finished) {
2456 phi = phi_list.pop();
2457 idx = cur_input.pop();
2458 PhiNode *prev_result = get_map_phi(phi->_idx);
2459 prev_result->set_req(idx++, result);
2460 result = prev_result;
2461 }
2462 }
2463 return result;
2464 }
2466 //
2467 // The next methods are derived from methods in MemNode.
2468 //
2469 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2470 Node *mem = mmem;
2471 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2472 // means an array I have not precisely typed yet. Do not do any
2473 // alias stuff with it any time soon.
2474 if (toop->base() != Type::AnyPtr &&
2475 !(toop->klass() != NULL &&
2476 toop->klass()->is_java_lang_Object() &&
2477 toop->offset() == Type::OffsetBot)) {
2478 mem = mmem->memory_at(alias_idx);
2479 // Update input if it is progress over what we have now
2480 }
2481 return mem;
2482 }
2484 //
2485 // Move memory users to their memory slices.
2486 //
2487 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2488 Compile* C = _compile;
2489 PhaseGVN* igvn = _igvn;
2490 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2491 assert(tp != NULL, "ptr type");
2492 int alias_idx = C->get_alias_index(tp);
2493 int general_idx = C->get_general_index(alias_idx);
2495 // Move users first
2496 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2497 Node* use = n->fast_out(i);
2498 if (use->is_MergeMem()) {
2499 MergeMemNode* mmem = use->as_MergeMem();
2500 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2501 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2502 continue; // Nothing to do
2503 }
2504 // Replace previous general reference to mem node.
2505 uint orig_uniq = C->unique();
2506 Node* m = find_inst_mem(n, general_idx, orig_phis);
2507 assert(orig_uniq == C->unique(), "no new nodes");
2508 mmem->set_memory_at(general_idx, m);
2509 --imax;
2510 --i;
2511 } else if (use->is_MemBar()) {
2512 assert(!use->is_Initialize(), "initializing stores should not be moved");
2513 if (use->req() > MemBarNode::Precedent &&
2514 use->in(MemBarNode::Precedent) == n) {
2515 // Don't move related membars.
2516 record_for_optimizer(use);
2517 continue;
2518 }
2519 tp = use->as_MemBar()->adr_type()->isa_ptr();
2520 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2521 alias_idx == general_idx) {
2522 continue; // Nothing to do
2523 }
2524 // Move to general memory slice.
2525 uint orig_uniq = C->unique();
2526 Node* m = find_inst_mem(n, general_idx, orig_phis);
2527 assert(orig_uniq == C->unique(), "no new nodes");
2528 igvn->hash_delete(use);
2529 imax -= use->replace_edge(n, m);
2530 igvn->hash_insert(use);
2531 record_for_optimizer(use);
2532 --i;
2533 #ifdef ASSERT
2534 } else if (use->is_Mem()) {
2535 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2536 // Don't move related cardmark.
2537 continue;
2538 }
2539 // Memory nodes should have new memory input.
2540 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2541 assert(tp != NULL, "ptr type");
2542 int idx = C->get_alias_index(tp);
2543 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2544 "Following memory nodes should have new memory input or be on the same memory slice");
2545 } else if (use->is_Phi()) {
2546 // Phi nodes should be split and moved already.
2547 tp = use->as_Phi()->adr_type()->isa_ptr();
2548 assert(tp != NULL, "ptr type");
2549 int idx = C->get_alias_index(tp);
2550 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2551 } else {
2552 use->dump();
2553 assert(false, "should not be here");
2554 #endif
2555 }
2556 }
2557 }
2559 //
2560 // Search memory chain of "mem" to find a MemNode whose address
2561 // is the specified alias index.
2562 //
2563 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2564 if (orig_mem == NULL)
2565 return orig_mem;
2566 Compile* C = _compile;
2567 PhaseGVN* igvn = _igvn;
2568 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2569 bool is_instance = (toop != NULL) && toop->is_known_instance();
2570 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2571 Node *prev = NULL;
2572 Node *result = orig_mem;
2573 while (prev != result) {
2574 prev = result;
2575 if (result == start_mem)
2576 break; // hit one of our sentinels
2577 if (result->is_Mem()) {
2578 const Type *at = igvn->type(result->in(MemNode::Address));
2579 if (at == Type::TOP)
2580 break; // Dead
2581 assert (at->isa_ptr() != NULL, "pointer type required.");
2582 int idx = C->get_alias_index(at->is_ptr());
2583 if (idx == alias_idx)
2584 break; // Found
2585 if (!is_instance && (at->isa_oopptr() == NULL ||
2586 !at->is_oopptr()->is_known_instance())) {
2587 break; // Do not skip store to general memory slice.
2588 }
2589 result = result->in(MemNode::Memory);
2590 }
2591 if (!is_instance)
2592 continue; // don't search further for non-instance types
2593 // skip over a call which does not affect this memory slice
2594 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2595 Node *proj_in = result->in(0);
2596 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2597 break; // hit one of our sentinels
2598 } else if (proj_in->is_Call()) {
2599 CallNode *call = proj_in->as_Call();
2600 if (!call->may_modify(toop, igvn)) {
2601 result = call->in(TypeFunc::Memory);
2602 }
2603 } else if (proj_in->is_Initialize()) {
2604 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2605 // Stop if this is the initialization for the object instance which
2606 // which contains this memory slice, otherwise skip over it.
2607 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2608 result = proj_in->in(TypeFunc::Memory);
2609 }
2610 } else if (proj_in->is_MemBar()) {
2611 result = proj_in->in(TypeFunc::Memory);
2612 }
2613 } else if (result->is_MergeMem()) {
2614 MergeMemNode *mmem = result->as_MergeMem();
2615 result = step_through_mergemem(mmem, alias_idx, toop);
2616 if (result == mmem->base_memory()) {
2617 // Didn't find instance memory, search through general slice recursively.
2618 result = mmem->memory_at(C->get_general_index(alias_idx));
2619 result = find_inst_mem(result, alias_idx, orig_phis);
2620 if (C->failing()) {
2621 return NULL;
2622 }
2623 mmem->set_memory_at(alias_idx, result);
2624 }
2625 } else if (result->is_Phi() &&
2626 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2627 Node *un = result->as_Phi()->unique_input(igvn);
2628 if (un != NULL) {
2629 orig_phis.append_if_missing(result->as_Phi());
2630 result = un;
2631 } else {
2632 break;
2633 }
2634 } else if (result->is_ClearArray()) {
2635 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2636 // Can not bypass initialization of the instance
2637 // we are looking for.
2638 break;
2639 }
2640 // Otherwise skip it (the call updated 'result' value).
2641 } else if (result->Opcode() == Op_SCMemProj) {
2642 Node* mem = result->in(0);
2643 Node* adr = NULL;
2644 if (mem->is_LoadStore()) {
2645 adr = mem->in(MemNode::Address);
2646 } else {
2647 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2648 adr = mem->in(3); // Memory edge corresponds to destination array
2649 }
2650 const Type *at = igvn->type(adr);
2651 if (at != Type::TOP) {
2652 assert (at->isa_ptr() != NULL, "pointer type required.");
2653 int idx = C->get_alias_index(at->is_ptr());
2654 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2655 break;
2656 }
2657 result = mem->in(MemNode::Memory);
2658 }
2659 }
2660 if (result->is_Phi()) {
2661 PhiNode *mphi = result->as_Phi();
2662 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2663 const TypePtr *t = mphi->adr_type();
2664 if (!is_instance) {
2665 // Push all non-instance Phis on the orig_phis worklist to update inputs
2666 // during Phase 4 if needed.
2667 orig_phis.append_if_missing(mphi);
2668 } else if (C->get_alias_index(t) != alias_idx) {
2669 // Create a new Phi with the specified alias index type.
2670 result = split_memory_phi(mphi, alias_idx, orig_phis);
2671 }
2672 }
2673 // the result is either MemNode, PhiNode, InitializeNode.
2674 return result;
2675 }
2677 //
2678 // Convert the types of unescaped object to instance types where possible,
2679 // propagate the new type information through the graph, and update memory
2680 // edges and MergeMem inputs to reflect the new type.
2681 //
2682 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2683 // The processing is done in 4 phases:
2684 //
2685 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2686 // types for the CheckCastPP for allocations where possible.
2687 // Propagate the the new types through users as follows:
2688 // casts and Phi: push users on alloc_worklist
2689 // AddP: cast Base and Address inputs to the instance type
2690 // push any AddP users on alloc_worklist and push any memnode
2691 // users onto memnode_worklist.
2692 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2693 // search the Memory chain for a store with the appropriate type
2694 // address type. If a Phi is found, create a new version with
2695 // the appropriate memory slices from each of the Phi inputs.
2696 // For stores, process the users as follows:
2697 // MemNode: push on memnode_worklist
2698 // MergeMem: push on mergemem_worklist
2699 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2700 // moving the first node encountered of each instance type to the
2701 // the input corresponding to its alias index.
2702 // appropriate memory slice.
2703 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2704 //
2705 // In the following example, the CheckCastPP nodes are the cast of allocation
2706 // results and the allocation of node 29 is unescaped and eligible to be an
2707 // instance type.
2708 //
2709 // We start with:
2710 //
2711 // 7 Parm #memory
2712 // 10 ConI "12"
2713 // 19 CheckCastPP "Foo"
2714 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2715 // 29 CheckCastPP "Foo"
2716 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2717 //
2718 // 40 StoreP 25 7 20 ... alias_index=4
2719 // 50 StoreP 35 40 30 ... alias_index=4
2720 // 60 StoreP 45 50 20 ... alias_index=4
2721 // 70 LoadP _ 60 30 ... alias_index=4
2722 // 80 Phi 75 50 60 Memory alias_index=4
2723 // 90 LoadP _ 80 30 ... alias_index=4
2724 // 100 LoadP _ 80 20 ... alias_index=4
2725 //
2726 //
2727 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2728 // and creating a new alias index for node 30. This gives:
2729 //
2730 // 7 Parm #memory
2731 // 10 ConI "12"
2732 // 19 CheckCastPP "Foo"
2733 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2734 // 29 CheckCastPP "Foo" iid=24
2735 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2736 //
2737 // 40 StoreP 25 7 20 ... alias_index=4
2738 // 50 StoreP 35 40 30 ... alias_index=6
2739 // 60 StoreP 45 50 20 ... alias_index=4
2740 // 70 LoadP _ 60 30 ... alias_index=6
2741 // 80 Phi 75 50 60 Memory alias_index=4
2742 // 90 LoadP _ 80 30 ... alias_index=6
2743 // 100 LoadP _ 80 20 ... alias_index=4
2744 //
2745 // In phase 2, new memory inputs are computed for the loads and stores,
2746 // And a new version of the phi is created. In phase 4, the inputs to
2747 // node 80 are updated and then the memory nodes are updated with the
2748 // values computed in phase 2. This results in:
2749 //
2750 // 7 Parm #memory
2751 // 10 ConI "12"
2752 // 19 CheckCastPP "Foo"
2753 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2754 // 29 CheckCastPP "Foo" iid=24
2755 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2756 //
2757 // 40 StoreP 25 7 20 ... alias_index=4
2758 // 50 StoreP 35 7 30 ... alias_index=6
2759 // 60 StoreP 45 40 20 ... alias_index=4
2760 // 70 LoadP _ 50 30 ... alias_index=6
2761 // 80 Phi 75 40 60 Memory alias_index=4
2762 // 120 Phi 75 50 50 Memory alias_index=6
2763 // 90 LoadP _ 120 30 ... alias_index=6
2764 // 100 LoadP _ 80 20 ... alias_index=4
2765 //
2766 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2767 GrowableArray<Node *> memnode_worklist;
2768 GrowableArray<PhiNode *> orig_phis;
2769 PhaseIterGVN *igvn = _igvn;
2770 uint new_index_start = (uint) _compile->num_alias_types();
2771 Arena* arena = Thread::current()->resource_area();
2772 VectorSet visited(arena);
2773 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2774 uint unique_old = _compile->unique();
2776 // Phase 1: Process possible allocations from alloc_worklist.
2777 // Create instance types for the CheckCastPP for allocations where possible.
2778 //
2779 // (Note: don't forget to change the order of the second AddP node on
2780 // the alloc_worklist if the order of the worklist processing is changed,
2781 // see the comment in find_second_addp().)
2782 //
2783 while (alloc_worklist.length() != 0) {
2784 Node *n = alloc_worklist.pop();
2785 uint ni = n->_idx;
2786 if (n->is_Call()) {
2787 CallNode *alloc = n->as_Call();
2788 // copy escape information to call node
2789 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2790 PointsToNode::EscapeState es = ptn->escape_state();
2791 // We have an allocation or call which returns a Java object,
2792 // see if it is unescaped.
2793 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2794 continue;
2795 // Find CheckCastPP for the allocate or for the return value of a call
2796 n = alloc->result_cast();
2797 if (n == NULL) { // No uses except Initialize node
2798 if (alloc->is_Allocate()) {
2799 // Set the scalar_replaceable flag for allocation
2800 // so it could be eliminated if it has no uses.
2801 alloc->as_Allocate()->_is_scalar_replaceable = true;
2802 }
2803 if (alloc->is_CallStaticJava()) {
2804 // Set the scalar_replaceable flag for boxing method
2805 // so it could be eliminated if it has no uses.
2806 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2807 }
2808 continue;
2809 }
2810 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2811 assert(!alloc->is_Allocate(), "allocation should have unique type");
2812 continue;
2813 }
2815 // The inline code for Object.clone() casts the allocation result to
2816 // java.lang.Object and then to the actual type of the allocated
2817 // object. Detect this case and use the second cast.
2818 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2819 // the allocation result is cast to java.lang.Object and then
2820 // to the actual Array type.
2821 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2822 && (alloc->is_AllocateArray() ||
2823 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2824 Node *cast2 = NULL;
2825 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2826 Node *use = n->fast_out(i);
2827 if (use->is_CheckCastPP()) {
2828 cast2 = use;
2829 break;
2830 }
2831 }
2832 if (cast2 != NULL) {
2833 n = cast2;
2834 } else {
2835 // Non-scalar replaceable if the allocation type is unknown statically
2836 // (reflection allocation), the object can't be restored during
2837 // deoptimization without precise type.
2838 continue;
2839 }
2840 }
2841 if (alloc->is_Allocate()) {
2842 // Set the scalar_replaceable flag for allocation
2843 // so it could be eliminated.
2844 alloc->as_Allocate()->_is_scalar_replaceable = true;
2845 }
2846 if (alloc->is_CallStaticJava()) {
2847 // Set the scalar_replaceable flag for boxing method
2848 // so it could be eliminated.
2849 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2850 }
2851 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2852 // in order for an object to be scalar-replaceable, it must be:
2853 // - a direct allocation (not a call returning an object)
2854 // - non-escaping
2855 // - eligible to be a unique type
2856 // - not determined to be ineligible by escape analysis
2857 set_map(alloc, n);
2858 set_map(n, alloc);
2859 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2860 if (t == NULL)
2861 continue; // not a TypeOopPtr
2862 const TypeOopPtr* tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
2863 igvn->hash_delete(n);
2864 igvn->set_type(n, tinst);
2865 n->raise_bottom_type(tinst);
2866 igvn->hash_insert(n);
2867 record_for_optimizer(n);
2868 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2870 // First, put on the worklist all Field edges from Connection Graph
2871 // which is more accurate then putting immediate users from Ideal Graph.
2872 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2873 PointsToNode* tgt = e.get();
2874 Node* use = tgt->ideal_node();
2875 assert(tgt->is_Field() && use->is_AddP(),
2876 "only AddP nodes are Field edges in CG");
2877 if (use->outcnt() > 0) { // Don't process dead nodes
2878 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2879 if (addp2 != NULL) {
2880 assert(alloc->is_AllocateArray(),"array allocation was expected");
2881 alloc_worklist.append_if_missing(addp2);
2882 }
2883 alloc_worklist.append_if_missing(use);
2884 }
2885 }
2887 // An allocation may have an Initialize which has raw stores. Scan
2888 // the users of the raw allocation result and push AddP users
2889 // on alloc_worklist.
2890 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2891 assert (raw_result != NULL, "must have an allocation result");
2892 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2893 Node *use = raw_result->fast_out(i);
2894 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2895 Node* addp2 = find_second_addp(use, raw_result);
2896 if (addp2 != NULL) {
2897 assert(alloc->is_AllocateArray(),"array allocation was expected");
2898 alloc_worklist.append_if_missing(addp2);
2899 }
2900 alloc_worklist.append_if_missing(use);
2901 } else if (use->is_MemBar()) {
2902 memnode_worklist.append_if_missing(use);
2903 }
2904 }
2905 }
2906 } else if (n->is_AddP()) {
2907 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2908 if (jobj == NULL || jobj == phantom_obj) {
2909 #ifdef ASSERT
2910 ptnode_adr(get_addp_base(n)->_idx)->dump();
2911 ptnode_adr(n->_idx)->dump();
2912 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2913 #endif
2914 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2915 return;
2916 }
2917 Node *base = get_map(jobj->idx()); // CheckCastPP node
2918 if (!split_AddP(n, base)) continue; // wrong type from dead path
2919 } else if (n->is_Phi() ||
2920 n->is_CheckCastPP() ||
2921 n->is_EncodeP() ||
2922 n->is_DecodeN() ||
2923 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2924 if (visited.test_set(n->_idx)) {
2925 assert(n->is_Phi(), "loops only through Phi's");
2926 continue; // already processed
2927 }
2928 JavaObjectNode* jobj = unique_java_object(n);
2929 if (jobj == NULL || jobj == phantom_obj) {
2930 #ifdef ASSERT
2931 ptnode_adr(n->_idx)->dump();
2932 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2933 #endif
2934 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2935 return;
2936 } else {
2937 Node *val = get_map(jobj->idx()); // CheckCastPP node
2938 TypeNode *tn = n->as_Type();
2939 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2940 assert(tinst != NULL && tinst->is_known_instance() &&
2941 tinst->instance_id() == jobj->idx() , "instance type expected.");
2943 const Type *tn_type = igvn->type(tn);
2944 const TypeOopPtr *tn_t;
2945 if (tn_type->isa_narrowoop()) {
2946 tn_t = tn_type->make_ptr()->isa_oopptr();
2947 } else {
2948 tn_t = tn_type->isa_oopptr();
2949 }
2950 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2951 if (tn_type->isa_narrowoop()) {
2952 tn_type = tinst->make_narrowoop();
2953 } else {
2954 tn_type = tinst;
2955 }
2956 igvn->hash_delete(tn);
2957 igvn->set_type(tn, tn_type);
2958 tn->set_type(tn_type);
2959 igvn->hash_insert(tn);
2960 record_for_optimizer(n);
2961 } else {
2962 assert(tn_type == TypePtr::NULL_PTR ||
2963 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
2964 "unexpected type");
2965 continue; // Skip dead path with different type
2966 }
2967 }
2968 } else {
2969 debug_only(n->dump();)
2970 assert(false, "EA: unexpected node");
2971 continue;
2972 }
2973 // push allocation's users on appropriate worklist
2974 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2975 Node *use = n->fast_out(i);
2976 if(use->is_Mem() && use->in(MemNode::Address) == n) {
2977 // Load/store to instance's field
2978 memnode_worklist.append_if_missing(use);
2979 } else if (use->is_MemBar()) {
2980 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
2981 memnode_worklist.append_if_missing(use);
2982 }
2983 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
2984 Node* addp2 = find_second_addp(use, n);
2985 if (addp2 != NULL) {
2986 alloc_worklist.append_if_missing(addp2);
2987 }
2988 alloc_worklist.append_if_missing(use);
2989 } else if (use->is_Phi() ||
2990 use->is_CheckCastPP() ||
2991 use->is_EncodeNarrowPtr() ||
2992 use->is_DecodeNarrowPtr() ||
2993 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
2994 alloc_worklist.append_if_missing(use);
2995 #ifdef ASSERT
2996 } else if (use->is_Mem()) {
2997 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
2998 } else if (use->is_MergeMem()) {
2999 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3000 } else if (use->is_SafePoint()) {
3001 // Look for MergeMem nodes for calls which reference unique allocation
3002 // (through CheckCastPP nodes) even for debug info.
3003 Node* m = use->in(TypeFunc::Memory);
3004 if (m->is_MergeMem()) {
3005 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3006 }
3007 } else if (use->Opcode() == Op_EncodeISOArray) {
3008 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3009 // EncodeISOArray overwrites destination array
3010 memnode_worklist.append_if_missing(use);
3011 }
3012 } else {
3013 uint op = use->Opcode();
3014 if (!(op == Op_CmpP || op == Op_Conv2B ||
3015 op == Op_CastP2X || op == Op_StoreCM ||
3016 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
3017 op == Op_StrEquals || op == Op_StrIndexOf)) {
3018 n->dump();
3019 use->dump();
3020 assert(false, "EA: missing allocation reference path");
3021 }
3022 #endif
3023 }
3024 }
3026 }
3027 // New alias types were created in split_AddP().
3028 uint new_index_end = (uint) _compile->num_alias_types();
3029 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3031 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3032 // compute new values for Memory inputs (the Memory inputs are not
3033 // actually updated until phase 4.)
3034 if (memnode_worklist.length() == 0)
3035 return; // nothing to do
3036 while (memnode_worklist.length() != 0) {
3037 Node *n = memnode_worklist.pop();
3038 if (visited.test_set(n->_idx))
3039 continue;
3040 if (n->is_Phi() || n->is_ClearArray()) {
3041 // we don't need to do anything, but the users must be pushed
3042 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3043 // we don't need to do anything, but the users must be pushed
3044 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3045 if (n == NULL)
3046 continue;
3047 } else if (n->Opcode() == Op_EncodeISOArray) {
3048 // get the memory projection
3049 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3050 Node *use = n->fast_out(i);
3051 if (use->Opcode() == Op_SCMemProj) {
3052 n = use;
3053 break;
3054 }
3055 }
3056 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3057 } else {
3058 assert(n->is_Mem(), "memory node required.");
3059 Node *addr = n->in(MemNode::Address);
3060 const Type *addr_t = igvn->type(addr);
3061 if (addr_t == Type::TOP)
3062 continue;
3063 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3064 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3065 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3066 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3067 if (_compile->failing()) {
3068 return;
3069 }
3070 if (mem != n->in(MemNode::Memory)) {
3071 // We delay the memory edge update since we need old one in
3072 // MergeMem code below when instances memory slices are separated.
3073 set_map(n, mem);
3074 }
3075 if (n->is_Load()) {
3076 continue; // don't push users
3077 } else if (n->is_LoadStore()) {
3078 // get the memory projection
3079 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3080 Node *use = n->fast_out(i);
3081 if (use->Opcode() == Op_SCMemProj) {
3082 n = use;
3083 break;
3084 }
3085 }
3086 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3087 }
3088 }
3089 // push user on appropriate worklist
3090 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3091 Node *use = n->fast_out(i);
3092 if (use->is_Phi() || use->is_ClearArray()) {
3093 memnode_worklist.append_if_missing(use);
3094 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3095 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3096 continue;
3097 memnode_worklist.append_if_missing(use);
3098 } else if (use->is_MemBar()) {
3099 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3100 memnode_worklist.append_if_missing(use);
3101 }
3102 #ifdef ASSERT
3103 } else if(use->is_Mem()) {
3104 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3105 } else if (use->is_MergeMem()) {
3106 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3107 } else if (use->Opcode() == Op_EncodeISOArray) {
3108 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3109 // EncodeISOArray overwrites destination array
3110 memnode_worklist.append_if_missing(use);
3111 }
3112 } else {
3113 uint op = use->Opcode();
3114 if (!(op == Op_StoreCM ||
3115 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3116 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3117 op == Op_AryEq || op == Op_StrComp ||
3118 op == Op_StrEquals || op == Op_StrIndexOf)) {
3119 n->dump();
3120 use->dump();
3121 assert(false, "EA: missing memory path");
3122 }
3123 #endif
3124 }
3125 }
3126 }
3128 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3129 // Walk each memory slice moving the first node encountered of each
3130 // instance type to the the input corresponding to its alias index.
3131 uint length = _mergemem_worklist.length();
3132 for( uint next = 0; next < length; ++next ) {
3133 MergeMemNode* nmm = _mergemem_worklist.at(next);
3134 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3135 // Note: we don't want to use MergeMemStream here because we only want to
3136 // scan inputs which exist at the start, not ones we add during processing.
3137 // Note 2: MergeMem may already contains instance memory slices added
3138 // during find_inst_mem() call when memory nodes were processed above.
3139 igvn->hash_delete(nmm);
3140 uint nslices = nmm->req();
3141 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3142 Node* mem = nmm->in(i);
3143 Node* cur = NULL;
3144 if (mem == NULL || mem->is_top())
3145 continue;
3146 // First, update mergemem by moving memory nodes to corresponding slices
3147 // if their type became more precise since this mergemem was created.
3148 while (mem->is_Mem()) {
3149 const Type *at = igvn->type(mem->in(MemNode::Address));
3150 if (at != Type::TOP) {
3151 assert (at->isa_ptr() != NULL, "pointer type required.");
3152 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3153 if (idx == i) {
3154 if (cur == NULL)
3155 cur = mem;
3156 } else {
3157 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3158 nmm->set_memory_at(idx, mem);
3159 }
3160 }
3161 }
3162 mem = mem->in(MemNode::Memory);
3163 }
3164 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3165 // Find any instance of the current type if we haven't encountered
3166 // already a memory slice of the instance along the memory chain.
3167 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3168 if((uint)_compile->get_general_index(ni) == i) {
3169 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3170 if (nmm->is_empty_memory(m)) {
3171 Node* result = find_inst_mem(mem, ni, orig_phis);
3172 if (_compile->failing()) {
3173 return;
3174 }
3175 nmm->set_memory_at(ni, result);
3176 }
3177 }
3178 }
3179 }
3180 // Find the rest of instances values
3181 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3182 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3183 Node* result = step_through_mergemem(nmm, ni, tinst);
3184 if (result == nmm->base_memory()) {
3185 // Didn't find instance memory, search through general slice recursively.
3186 result = nmm->memory_at(_compile->get_general_index(ni));
3187 result = find_inst_mem(result, ni, orig_phis);
3188 if (_compile->failing()) {
3189 return;
3190 }
3191 nmm->set_memory_at(ni, result);
3192 }
3193 }
3194 igvn->hash_insert(nmm);
3195 record_for_optimizer(nmm);
3196 }
3198 // Phase 4: Update the inputs of non-instance memory Phis and
3199 // the Memory input of memnodes
3200 // First update the inputs of any non-instance Phi's from
3201 // which we split out an instance Phi. Note we don't have
3202 // to recursively process Phi's encounted on the input memory
3203 // chains as is done in split_memory_phi() since they will
3204 // also be processed here.
3205 for (int j = 0; j < orig_phis.length(); j++) {
3206 PhiNode *phi = orig_phis.at(j);
3207 int alias_idx = _compile->get_alias_index(phi->adr_type());
3208 igvn->hash_delete(phi);
3209 for (uint i = 1; i < phi->req(); i++) {
3210 Node *mem = phi->in(i);
3211 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3212 if (_compile->failing()) {
3213 return;
3214 }
3215 if (mem != new_mem) {
3216 phi->set_req(i, new_mem);
3217 }
3218 }
3219 igvn->hash_insert(phi);
3220 record_for_optimizer(phi);
3221 }
3223 // Update the memory inputs of MemNodes with the value we computed
3224 // in Phase 2 and move stores memory users to corresponding memory slices.
3225 // Disable memory split verification code until the fix for 6984348.
3226 // Currently it produces false negative results since it does not cover all cases.
3227 #if 0 // ifdef ASSERT
3228 visited.Reset();
3229 Node_Stack old_mems(arena, _compile->unique() >> 2);
3230 #endif
3231 for (uint i = 0; i < ideal_nodes.size(); i++) {
3232 Node* n = ideal_nodes.at(i);
3233 Node* nmem = get_map(n->_idx);
3234 assert(nmem != NULL, "sanity");
3235 if (n->is_Mem()) {
3236 #if 0 // ifdef ASSERT
3237 Node* old_mem = n->in(MemNode::Memory);
3238 if (!visited.test_set(old_mem->_idx)) {
3239 old_mems.push(old_mem, old_mem->outcnt());
3240 }
3241 #endif
3242 assert(n->in(MemNode::Memory) != nmem, "sanity");
3243 if (!n->is_Load()) {
3244 // Move memory users of a store first.
3245 move_inst_mem(n, orig_phis);
3246 }
3247 // Now update memory input
3248 igvn->hash_delete(n);
3249 n->set_req(MemNode::Memory, nmem);
3250 igvn->hash_insert(n);
3251 record_for_optimizer(n);
3252 } else {
3253 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3254 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3255 }
3256 }
3257 #if 0 // ifdef ASSERT
3258 // Verify that memory was split correctly
3259 while (old_mems.is_nonempty()) {
3260 Node* old_mem = old_mems.node();
3261 uint old_cnt = old_mems.index();
3262 old_mems.pop();
3263 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3264 }
3265 #endif
3266 }
3268 #ifndef PRODUCT
3269 static const char *node_type_names[] = {
3270 "UnknownType",
3271 "JavaObject",
3272 "LocalVar",
3273 "Field",
3274 "Arraycopy"
3275 };
3277 static const char *esc_names[] = {
3278 "UnknownEscape",
3279 "NoEscape",
3280 "ArgEscape",
3281 "GlobalEscape"
3282 };
3284 void PointsToNode::dump(bool print_state) const {
3285 NodeType nt = node_type();
3286 tty->print("%s ", node_type_names[(int) nt]);
3287 if (print_state) {
3288 EscapeState es = escape_state();
3289 EscapeState fields_es = fields_escape_state();
3290 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3291 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3292 tty->print("NSR ");
3293 }
3294 if (is_Field()) {
3295 FieldNode* f = (FieldNode*)this;
3296 if (f->is_oop())
3297 tty->print("oop ");
3298 if (f->offset() > 0)
3299 tty->print("+%d ", f->offset());
3300 tty->print("(");
3301 for (BaseIterator i(f); i.has_next(); i.next()) {
3302 PointsToNode* b = i.get();
3303 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3304 }
3305 tty->print(" )");
3306 }
3307 tty->print("[");
3308 for (EdgeIterator i(this); i.has_next(); i.next()) {
3309 PointsToNode* e = i.get();
3310 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3311 }
3312 tty->print(" [");
3313 for (UseIterator i(this); i.has_next(); i.next()) {
3314 PointsToNode* u = i.get();
3315 bool is_base = false;
3316 if (PointsToNode::is_base_use(u)) {
3317 is_base = true;
3318 u = PointsToNode::get_use_node(u)->as_Field();
3319 }
3320 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3321 }
3322 tty->print(" ]] ");
3323 if (_node == NULL)
3324 tty->print_cr("<null>");
3325 else
3326 _node->dump();
3327 }
3329 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3330 bool first = true;
3331 int ptnodes_length = ptnodes_worklist.length();
3332 for (int i = 0; i < ptnodes_length; i++) {
3333 PointsToNode *ptn = ptnodes_worklist.at(i);
3334 if (ptn == NULL || !ptn->is_JavaObject())
3335 continue;
3336 PointsToNode::EscapeState es = ptn->escape_state();
3337 if ((es != PointsToNode::NoEscape) && !Verbose) {
3338 continue;
3339 }
3340 Node* n = ptn->ideal_node();
3341 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3342 n->as_CallStaticJava()->is_boxing_method())) {
3343 if (first) {
3344 tty->cr();
3345 tty->print("======== Connection graph for ");
3346 _compile->method()->print_short_name();
3347 tty->cr();
3348 first = false;
3349 }
3350 ptn->dump();
3351 // Print all locals and fields which reference this allocation
3352 for (UseIterator j(ptn); j.has_next(); j.next()) {
3353 PointsToNode* use = j.get();
3354 if (use->is_LocalVar()) {
3355 use->dump(Verbose);
3356 } else if (Verbose) {
3357 use->dump();
3358 }
3359 }
3360 tty->cr();
3361 }
3362 }
3363 }
3364 #endif