Thu, 01 Aug 2013 17:25:10 -0700
Merge
1 /*
2 * Copyright (c) 2005, 2013, 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 ptn->set_escape_state(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_instance_klass() || // StressReflectiveCode
784 cik->as_instance_klass()->has_finalizer()) {
785 es = PointsToNode::GlobalEscape;
786 }
787 }
788 add_java_object(call, es);
789 PointsToNode* ptn = ptnode_adr(call_idx);
790 if (!scalar_replaceable && ptn->scalar_replaceable()) {
791 ptn->set_scalar_replaceable(false);
792 }
793 } else if (call->is_CallStaticJava()) {
794 // Call nodes could be different types:
795 //
796 // 1. CallDynamicJavaNode (what happened during call is unknown):
797 //
798 // - mapped to GlobalEscape JavaObject node if oop is returned;
799 //
800 // - all oop arguments are escaping globally;
801 //
802 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
803 //
804 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
805 //
806 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
807 // - mapped to NoEscape JavaObject node if non-escaping object allocated
808 // during call is returned;
809 // - mapped to ArgEscape LocalVar node pointed to object arguments
810 // which are returned and does not escape during call;
811 //
812 // - oop arguments escaping status is defined by bytecode analysis;
813 //
814 // For a static call, we know exactly what method is being called.
815 // Use bytecode estimator to record whether the call's return value escapes.
816 ciMethod* meth = call->as_CallJava()->method();
817 if (meth == NULL) {
818 const char* name = call->as_CallStaticJava()->_name;
819 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
820 // Returns a newly allocated unescaped object.
821 add_java_object(call, PointsToNode::NoEscape);
822 ptnode_adr(call_idx)->set_scalar_replaceable(false);
823 } else if (meth->is_boxing_method()) {
824 // Returns boxing object
825 PointsToNode::EscapeState es;
826 vmIntrinsics::ID intr = meth->intrinsic_id();
827 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
828 // It does not escape if object is always allocated.
829 es = PointsToNode::NoEscape;
830 } else {
831 // It escapes globally if object could be loaded from cache.
832 es = PointsToNode::GlobalEscape;
833 }
834 add_java_object(call, es);
835 } else {
836 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
837 call_analyzer->copy_dependencies(_compile->dependencies());
838 if (call_analyzer->is_return_allocated()) {
839 // Returns a newly allocated unescaped object, simply
840 // update dependency information.
841 // Mark it as NoEscape so that objects referenced by
842 // it's fields will be marked as NoEscape at least.
843 add_java_object(call, PointsToNode::NoEscape);
844 ptnode_adr(call_idx)->set_scalar_replaceable(false);
845 } else {
846 // Determine whether any arguments are returned.
847 const TypeTuple* d = call->tf()->domain();
848 bool ret_arg = false;
849 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
850 if (d->field_at(i)->isa_ptr() != NULL &&
851 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
852 ret_arg = true;
853 break;
854 }
855 }
856 if (ret_arg) {
857 add_local_var(call, PointsToNode::ArgEscape);
858 } else {
859 // Returns unknown object.
860 map_ideal_node(call, phantom_obj);
861 }
862 }
863 }
864 } else {
865 // An other type of call, assume the worst case:
866 // returned value is unknown and globally escapes.
867 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
868 map_ideal_node(call, phantom_obj);
869 }
870 }
872 void ConnectionGraph::process_call_arguments(CallNode *call) {
873 bool is_arraycopy = false;
874 switch (call->Opcode()) {
875 #ifdef ASSERT
876 case Op_Allocate:
877 case Op_AllocateArray:
878 case Op_Lock:
879 case Op_Unlock:
880 assert(false, "should be done already");
881 break;
882 #endif
883 case Op_CallLeafNoFP:
884 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
885 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
886 // fall through
887 case Op_CallLeaf: {
888 // Stub calls, objects do not escape but they are not scale replaceable.
889 // Adjust escape state for outgoing arguments.
890 const TypeTuple * d = call->tf()->domain();
891 bool src_has_oops = false;
892 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
893 const Type* at = d->field_at(i);
894 Node *arg = call->in(i);
895 const Type *aat = _igvn->type(arg);
896 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
897 continue;
898 if (arg->is_AddP()) {
899 //
900 // The inline_native_clone() case when the arraycopy stub is called
901 // after the allocation before Initialize and CheckCastPP nodes.
902 // Or normal arraycopy for object arrays case.
903 //
904 // Set AddP's base (Allocate) as not scalar replaceable since
905 // pointer to the base (with offset) is passed as argument.
906 //
907 arg = get_addp_base(arg);
908 }
909 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
910 assert(arg_ptn != NULL, "should be registered");
911 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
912 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
913 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
914 aat->isa_ptr() != NULL, "expecting an Ptr");
915 bool arg_has_oops = aat->isa_oopptr() &&
916 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
917 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
918 if (i == TypeFunc::Parms) {
919 src_has_oops = arg_has_oops;
920 }
921 //
922 // src or dst could be j.l.Object when other is basic type array:
923 //
924 // arraycopy(char[],0,Object*,0,size);
925 // arraycopy(Object*,0,char[],0,size);
926 //
927 // Don't add edges in such cases.
928 //
929 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
930 arg_has_oops && (i > TypeFunc::Parms);
931 #ifdef ASSERT
932 if (!(is_arraycopy ||
933 (call->as_CallLeaf()->_name != NULL &&
934 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
935 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
936 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
937 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
938 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
939 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
940 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0)
941 ))) {
942 call->dump();
943 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
944 }
945 #endif
946 // Always process arraycopy's destination object since
947 // we need to add all possible edges to references in
948 // source object.
949 if (arg_esc >= PointsToNode::ArgEscape &&
950 !arg_is_arraycopy_dest) {
951 continue;
952 }
953 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
954 if (arg_is_arraycopy_dest) {
955 Node* src = call->in(TypeFunc::Parms);
956 if (src->is_AddP()) {
957 src = get_addp_base(src);
958 }
959 PointsToNode* src_ptn = ptnode_adr(src->_idx);
960 assert(src_ptn != NULL, "should be registered");
961 if (arg_ptn != src_ptn) {
962 // Special arraycopy edge:
963 // A destination object's field can't have the source object
964 // as base since objects escape states are not related.
965 // Only escape state of destination object's fields affects
966 // escape state of fields in source object.
967 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
968 }
969 }
970 }
971 }
972 break;
973 }
974 case Op_CallStaticJava: {
975 // For a static call, we know exactly what method is being called.
976 // Use bytecode estimator to record the call's escape affects
977 #ifdef ASSERT
978 const char* name = call->as_CallStaticJava()->_name;
979 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
980 #endif
981 ciMethod* meth = call->as_CallJava()->method();
982 if ((meth != NULL) && meth->is_boxing_method()) {
983 break; // Boxing methods do not modify any oops.
984 }
985 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
986 // fall-through if not a Java method or no analyzer information
987 if (call_analyzer != NULL) {
988 PointsToNode* call_ptn = ptnode_adr(call->_idx);
989 const TypeTuple* d = call->tf()->domain();
990 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
991 const Type* at = d->field_at(i);
992 int k = i - TypeFunc::Parms;
993 Node* arg = call->in(i);
994 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
995 if (at->isa_ptr() != NULL &&
996 call_analyzer->is_arg_returned(k)) {
997 // The call returns arguments.
998 if (call_ptn != NULL) { // Is call's result used?
999 assert(call_ptn->is_LocalVar(), "node should be registered");
1000 assert(arg_ptn != NULL, "node should be registered");
1001 add_edge(call_ptn, arg_ptn);
1002 }
1003 }
1004 if (at->isa_oopptr() != NULL &&
1005 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1006 if (!call_analyzer->is_arg_stack(k)) {
1007 // The argument global escapes
1008 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1009 } else {
1010 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1011 if (!call_analyzer->is_arg_local(k)) {
1012 // The argument itself doesn't escape, but any fields might
1013 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1014 }
1015 }
1016 }
1017 }
1018 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1019 // The call returns arguments.
1020 assert(call_ptn->edge_count() > 0, "sanity");
1021 if (!call_analyzer->is_return_local()) {
1022 // Returns also unknown object.
1023 add_edge(call_ptn, phantom_obj);
1024 }
1025 }
1026 break;
1027 }
1028 }
1029 default: {
1030 // Fall-through here if not a Java method or no analyzer information
1031 // or some other type of call, assume the worst case: all arguments
1032 // globally escape.
1033 const TypeTuple* d = call->tf()->domain();
1034 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1035 const Type* at = d->field_at(i);
1036 if (at->isa_oopptr() != NULL) {
1037 Node* arg = call->in(i);
1038 if (arg->is_AddP()) {
1039 arg = get_addp_base(arg);
1040 }
1041 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1042 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1043 }
1044 }
1045 }
1046 }
1047 }
1050 // Finish Graph construction.
1051 bool ConnectionGraph::complete_connection_graph(
1052 GrowableArray<PointsToNode*>& ptnodes_worklist,
1053 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1054 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1055 GrowableArray<FieldNode*>& oop_fields_worklist) {
1056 // Normally only 1-3 passes needed to build Connection Graph depending
1057 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1058 // Set limit to 20 to catch situation when something did go wrong and
1059 // bailout Escape Analysis.
1060 // Also limit build time to 30 sec (60 in debug VM).
1061 #define CG_BUILD_ITER_LIMIT 20
1062 #ifdef ASSERT
1063 #define CG_BUILD_TIME_LIMIT 60.0
1064 #else
1065 #define CG_BUILD_TIME_LIMIT 30.0
1066 #endif
1068 // Propagate GlobalEscape and ArgEscape escape states and check that
1069 // we still have non-escaping objects. The method pushs on _worklist
1070 // Field nodes which reference phantom_object.
1071 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1072 return false; // Nothing to do.
1073 }
1074 // Now propagate references to all JavaObject nodes.
1075 int java_objects_length = java_objects_worklist.length();
1076 elapsedTimer time;
1077 int new_edges = 1;
1078 int iterations = 0;
1079 do {
1080 while ((new_edges > 0) &&
1081 (iterations++ < CG_BUILD_ITER_LIMIT) &&
1082 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1083 time.start();
1084 new_edges = 0;
1085 // Propagate references to phantom_object for nodes pushed on _worklist
1086 // by find_non_escaped_objects() and find_field_value().
1087 new_edges += add_java_object_edges(phantom_obj, false);
1088 for (int next = 0; next < java_objects_length; ++next) {
1089 JavaObjectNode* ptn = java_objects_worklist.at(next);
1090 new_edges += add_java_object_edges(ptn, true);
1091 }
1092 if (new_edges > 0) {
1093 // Update escape states on each iteration if graph was updated.
1094 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1095 return false; // Nothing to do.
1096 }
1097 }
1098 time.stop();
1099 }
1100 if ((iterations < CG_BUILD_ITER_LIMIT) &&
1101 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1102 time.start();
1103 // Find fields which have unknown value.
1104 int fields_length = oop_fields_worklist.length();
1105 for (int next = 0; next < fields_length; next++) {
1106 FieldNode* field = oop_fields_worklist.at(next);
1107 if (field->edge_count() == 0) {
1108 new_edges += find_field_value(field);
1109 // This code may added new edges to phantom_object.
1110 // Need an other cycle to propagate references to phantom_object.
1111 }
1112 }
1113 time.stop();
1114 } else {
1115 new_edges = 0; // Bailout
1116 }
1117 } while (new_edges > 0);
1119 // Bailout if passed limits.
1120 if ((iterations >= CG_BUILD_ITER_LIMIT) ||
1121 (time.seconds() >= CG_BUILD_TIME_LIMIT)) {
1122 Compile* C = _compile;
1123 if (C->log() != NULL) {
1124 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1125 C->log()->text("%s", (iterations >= CG_BUILD_ITER_LIMIT) ? "iterations" : "time");
1126 C->log()->end_elem(" limit'");
1127 }
1128 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1129 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1130 // Possible infinite build_connection_graph loop,
1131 // bailout (no changes to ideal graph were made).
1132 return false;
1133 }
1134 #ifdef ASSERT
1135 if (Verbose && PrintEscapeAnalysis) {
1136 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1137 iterations, nodes_size(), ptnodes_worklist.length());
1138 }
1139 #endif
1141 #undef CG_BUILD_ITER_LIMIT
1142 #undef CG_BUILD_TIME_LIMIT
1144 // Find fields initialized by NULL for non-escaping Allocations.
1145 int non_escaped_length = non_escaped_worklist.length();
1146 for (int next = 0; next < non_escaped_length; next++) {
1147 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1148 PointsToNode::EscapeState es = ptn->escape_state();
1149 assert(es <= PointsToNode::ArgEscape, "sanity");
1150 if (es == PointsToNode::NoEscape) {
1151 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1152 // Adding references to NULL object does not change escape states
1153 // since it does not escape. Also no fields are added to NULL object.
1154 add_java_object_edges(null_obj, false);
1155 }
1156 }
1157 Node* n = ptn->ideal_node();
1158 if (n->is_Allocate()) {
1159 // The object allocated by this Allocate node will never be
1160 // seen by an other thread. Mark it so that when it is
1161 // expanded no MemBarStoreStore is added.
1162 InitializeNode* ini = n->as_Allocate()->initialization();
1163 if (ini != NULL)
1164 ini->set_does_not_escape();
1165 }
1166 }
1167 return true; // Finished graph construction.
1168 }
1170 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1171 // and check that we still have non-escaping java objects.
1172 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1173 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1174 GrowableArray<PointsToNode*> escape_worklist;
1175 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1176 int ptnodes_length = ptnodes_worklist.length();
1177 for (int next = 0; next < ptnodes_length; ++next) {
1178 PointsToNode* ptn = ptnodes_worklist.at(next);
1179 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1180 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1181 escape_worklist.push(ptn);
1182 }
1183 }
1184 // Set escape states to referenced nodes (edges list).
1185 while (escape_worklist.length() > 0) {
1186 PointsToNode* ptn = escape_worklist.pop();
1187 PointsToNode::EscapeState es = ptn->escape_state();
1188 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1189 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1190 es >= PointsToNode::ArgEscape) {
1191 // GlobalEscape or ArgEscape state of field means it has unknown value.
1192 if (add_edge(ptn, phantom_obj)) {
1193 // New edge was added
1194 add_field_uses_to_worklist(ptn->as_Field());
1195 }
1196 }
1197 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1198 PointsToNode* e = i.get();
1199 if (e->is_Arraycopy()) {
1200 assert(ptn->arraycopy_dst(), "sanity");
1201 // Propagate only fields escape state through arraycopy edge.
1202 if (e->fields_escape_state() < field_es) {
1203 set_fields_escape_state(e, field_es);
1204 escape_worklist.push(e);
1205 }
1206 } else if (es >= field_es) {
1207 // fields_escape_state is also set to 'es' if it is less than 'es'.
1208 if (e->escape_state() < es) {
1209 set_escape_state(e, es);
1210 escape_worklist.push(e);
1211 }
1212 } else {
1213 // Propagate field escape state.
1214 bool es_changed = false;
1215 if (e->fields_escape_state() < field_es) {
1216 set_fields_escape_state(e, field_es);
1217 es_changed = true;
1218 }
1219 if ((e->escape_state() < field_es) &&
1220 e->is_Field() && ptn->is_JavaObject() &&
1221 e->as_Field()->is_oop()) {
1222 // Change escape state of referenced fileds.
1223 set_escape_state(e, field_es);
1224 es_changed = true;;
1225 } else if (e->escape_state() < es) {
1226 set_escape_state(e, es);
1227 es_changed = true;;
1228 }
1229 if (es_changed) {
1230 escape_worklist.push(e);
1231 }
1232 }
1233 }
1234 }
1235 // Remove escaped objects from non_escaped list.
1236 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1237 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1238 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1239 non_escaped_worklist.delete_at(next);
1240 }
1241 if (ptn->escape_state() == PointsToNode::NoEscape) {
1242 // Find fields in non-escaped allocations which have unknown value.
1243 find_init_values(ptn, phantom_obj, NULL);
1244 }
1245 }
1246 return (non_escaped_worklist.length() > 0);
1247 }
1249 // Add all references to JavaObject node by walking over all uses.
1250 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1251 int new_edges = 0;
1252 if (populate_worklist) {
1253 // Populate _worklist by uses of jobj's uses.
1254 for (UseIterator i(jobj); i.has_next(); i.next()) {
1255 PointsToNode* use = i.get();
1256 if (use->is_Arraycopy())
1257 continue;
1258 add_uses_to_worklist(use);
1259 if (use->is_Field() && use->as_Field()->is_oop()) {
1260 // Put on worklist all field's uses (loads) and
1261 // related field nodes (same base and offset).
1262 add_field_uses_to_worklist(use->as_Field());
1263 }
1264 }
1265 }
1266 while(_worklist.length() > 0) {
1267 PointsToNode* use = _worklist.pop();
1268 if (PointsToNode::is_base_use(use)) {
1269 // Add reference from jobj to field and from field to jobj (field's base).
1270 use = PointsToNode::get_use_node(use)->as_Field();
1271 if (add_base(use->as_Field(), jobj)) {
1272 new_edges++;
1273 }
1274 continue;
1275 }
1276 assert(!use->is_JavaObject(), "sanity");
1277 if (use->is_Arraycopy()) {
1278 if (jobj == null_obj) // NULL object does not have field edges
1279 continue;
1280 // Added edge from Arraycopy node to arraycopy's source java object
1281 if (add_edge(use, jobj)) {
1282 jobj->set_arraycopy_src();
1283 new_edges++;
1284 }
1285 // and stop here.
1286 continue;
1287 }
1288 if (!add_edge(use, jobj))
1289 continue; // No new edge added, there was such edge already.
1290 new_edges++;
1291 if (use->is_LocalVar()) {
1292 add_uses_to_worklist(use);
1293 if (use->arraycopy_dst()) {
1294 for (EdgeIterator i(use); i.has_next(); i.next()) {
1295 PointsToNode* e = i.get();
1296 if (e->is_Arraycopy()) {
1297 if (jobj == null_obj) // NULL object does not have field edges
1298 continue;
1299 // Add edge from arraycopy's destination java object to Arraycopy node.
1300 if (add_edge(jobj, e)) {
1301 new_edges++;
1302 jobj->set_arraycopy_dst();
1303 }
1304 }
1305 }
1306 }
1307 } else {
1308 // Added new edge to stored in field values.
1309 // Put on worklist all field's uses (loads) and
1310 // related field nodes (same base and offset).
1311 add_field_uses_to_worklist(use->as_Field());
1312 }
1313 }
1314 return new_edges;
1315 }
1317 // Put on worklist all related field nodes.
1318 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1319 assert(field->is_oop(), "sanity");
1320 int offset = field->offset();
1321 add_uses_to_worklist(field);
1322 // Loop over all bases of this field and push on worklist Field nodes
1323 // with the same offset and base (since they may reference the same field).
1324 for (BaseIterator i(field); i.has_next(); i.next()) {
1325 PointsToNode* base = i.get();
1326 add_fields_to_worklist(field, base);
1327 // Check if the base was source object of arraycopy and go over arraycopy's
1328 // destination objects since values stored to a field of source object are
1329 // accessable by uses (loads) of fields of destination objects.
1330 if (base->arraycopy_src()) {
1331 for (UseIterator j(base); j.has_next(); j.next()) {
1332 PointsToNode* arycp = j.get();
1333 if (arycp->is_Arraycopy()) {
1334 for (UseIterator k(arycp); k.has_next(); k.next()) {
1335 PointsToNode* abase = k.get();
1336 if (abase->arraycopy_dst() && abase != base) {
1337 // Look for the same arracopy reference.
1338 add_fields_to_worklist(field, abase);
1339 }
1340 }
1341 }
1342 }
1343 }
1344 }
1345 }
1347 // Put on worklist all related field nodes.
1348 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1349 int offset = field->offset();
1350 if (base->is_LocalVar()) {
1351 for (UseIterator j(base); j.has_next(); j.next()) {
1352 PointsToNode* f = j.get();
1353 if (PointsToNode::is_base_use(f)) { // Field
1354 f = PointsToNode::get_use_node(f);
1355 if (f == field || !f->as_Field()->is_oop())
1356 continue;
1357 int offs = f->as_Field()->offset();
1358 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1359 add_to_worklist(f);
1360 }
1361 }
1362 }
1363 } else {
1364 assert(base->is_JavaObject(), "sanity");
1365 if (// Skip phantom_object since it is only used to indicate that
1366 // this field's content globally escapes.
1367 (base != phantom_obj) &&
1368 // NULL object node does not have fields.
1369 (base != null_obj)) {
1370 for (EdgeIterator i(base); i.has_next(); i.next()) {
1371 PointsToNode* f = i.get();
1372 // Skip arraycopy edge since store to destination object field
1373 // does not update value in source object field.
1374 if (f->is_Arraycopy()) {
1375 assert(base->arraycopy_dst(), "sanity");
1376 continue;
1377 }
1378 if (f == field || !f->as_Field()->is_oop())
1379 continue;
1380 int offs = f->as_Field()->offset();
1381 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1382 add_to_worklist(f);
1383 }
1384 }
1385 }
1386 }
1387 }
1389 // Find fields which have unknown value.
1390 int ConnectionGraph::find_field_value(FieldNode* field) {
1391 // Escaped fields should have init value already.
1392 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1393 int new_edges = 0;
1394 for (BaseIterator i(field); i.has_next(); i.next()) {
1395 PointsToNode* base = i.get();
1396 if (base->is_JavaObject()) {
1397 // Skip Allocate's fields which will be processed later.
1398 if (base->ideal_node()->is_Allocate())
1399 return 0;
1400 assert(base == null_obj, "only NULL ptr base expected here");
1401 }
1402 }
1403 if (add_edge(field, phantom_obj)) {
1404 // New edge was added
1405 new_edges++;
1406 add_field_uses_to_worklist(field);
1407 }
1408 return new_edges;
1409 }
1411 // Find fields initializing values for allocations.
1412 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1413 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1414 int new_edges = 0;
1415 Node* alloc = pta->ideal_node();
1416 if (init_val == phantom_obj) {
1417 // Do nothing for Allocate nodes since its fields values are "known".
1418 if (alloc->is_Allocate())
1419 return 0;
1420 assert(alloc->as_CallStaticJava(), "sanity");
1421 #ifdef ASSERT
1422 if (alloc->as_CallStaticJava()->method() == NULL) {
1423 const char* name = alloc->as_CallStaticJava()->_name;
1424 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1425 }
1426 #endif
1427 // Non-escaped allocation returned from Java or runtime call have
1428 // unknown values in fields.
1429 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1430 PointsToNode* field = i.get();
1431 if (field->is_Field() && field->as_Field()->is_oop()) {
1432 if (add_edge(field, phantom_obj)) {
1433 // New edge was added
1434 new_edges++;
1435 add_field_uses_to_worklist(field->as_Field());
1436 }
1437 }
1438 }
1439 return new_edges;
1440 }
1441 assert(init_val == null_obj, "sanity");
1442 // Do nothing for Call nodes since its fields values are unknown.
1443 if (!alloc->is_Allocate())
1444 return 0;
1446 InitializeNode* ini = alloc->as_Allocate()->initialization();
1447 Compile* C = _compile;
1448 bool visited_bottom_offset = false;
1449 GrowableArray<int> offsets_worklist;
1451 // Check if an oop field's initializing value is recorded and add
1452 // a corresponding NULL if field's value if it is not recorded.
1453 // Connection Graph does not record a default initialization by NULL
1454 // captured by Initialize node.
1455 //
1456 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1457 PointsToNode* field = i.get(); // Field (AddP)
1458 if (!field->is_Field() || !field->as_Field()->is_oop())
1459 continue; // Not oop field
1460 int offset = field->as_Field()->offset();
1461 if (offset == Type::OffsetBot) {
1462 if (!visited_bottom_offset) {
1463 // OffsetBot is used to reference array's element,
1464 // always add reference to NULL to all Field nodes since we don't
1465 // known which element is referenced.
1466 if (add_edge(field, null_obj)) {
1467 // New edge was added
1468 new_edges++;
1469 add_field_uses_to_worklist(field->as_Field());
1470 visited_bottom_offset = true;
1471 }
1472 }
1473 } else {
1474 // Check only oop fields.
1475 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1476 if (adr_type->isa_rawptr()) {
1477 #ifdef ASSERT
1478 // Raw pointers are used for initializing stores so skip it
1479 // since it should be recorded already
1480 Node* base = get_addp_base(field->ideal_node());
1481 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1482 (base->in(0) == alloc),"unexpected pointer type");
1483 #endif
1484 continue;
1485 }
1486 if (!offsets_worklist.contains(offset)) {
1487 offsets_worklist.append(offset);
1488 Node* value = NULL;
1489 if (ini != NULL) {
1490 // StoreP::memory_type() == T_ADDRESS
1491 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1492 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1493 // Make sure initializing store has the same type as this AddP.
1494 // This AddP may reference non existing field because it is on a
1495 // dead branch of bimorphic call which is not eliminated yet.
1496 if (store != NULL && store->is_Store() &&
1497 store->as_Store()->memory_type() == ft) {
1498 value = store->in(MemNode::ValueIn);
1499 #ifdef ASSERT
1500 if (VerifyConnectionGraph) {
1501 // Verify that AddP already points to all objects the value points to.
1502 PointsToNode* val = ptnode_adr(value->_idx);
1503 assert((val != NULL), "should be processed already");
1504 PointsToNode* missed_obj = NULL;
1505 if (val->is_JavaObject()) {
1506 if (!field->points_to(val->as_JavaObject())) {
1507 missed_obj = val;
1508 }
1509 } else {
1510 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1511 tty->print_cr("----------init store has invalid value -----");
1512 store->dump();
1513 val->dump();
1514 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1515 }
1516 for (EdgeIterator j(val); j.has_next(); j.next()) {
1517 PointsToNode* obj = j.get();
1518 if (obj->is_JavaObject()) {
1519 if (!field->points_to(obj->as_JavaObject())) {
1520 missed_obj = obj;
1521 break;
1522 }
1523 }
1524 }
1525 }
1526 if (missed_obj != NULL) {
1527 tty->print_cr("----------field---------------------------------");
1528 field->dump();
1529 tty->print_cr("----------missed referernce to object-----------");
1530 missed_obj->dump();
1531 tty->print_cr("----------object referernced by init store -----");
1532 store->dump();
1533 val->dump();
1534 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1535 }
1536 }
1537 #endif
1538 } else {
1539 // There could be initializing stores which follow allocation.
1540 // For example, a volatile field store is not collected
1541 // by Initialize node.
1542 //
1543 // Need to check for dependent loads to separate such stores from
1544 // stores which follow loads. For now, add initial value NULL so
1545 // that compare pointers optimization works correctly.
1546 }
1547 }
1548 if (value == NULL) {
1549 // A field's initializing value was not recorded. Add NULL.
1550 if (add_edge(field, null_obj)) {
1551 // New edge was added
1552 new_edges++;
1553 add_field_uses_to_worklist(field->as_Field());
1554 }
1555 }
1556 }
1557 }
1558 }
1559 return new_edges;
1560 }
1562 // Adjust scalar_replaceable state after Connection Graph is built.
1563 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1564 // Search for non-escaping objects which are not scalar replaceable
1565 // and mark them to propagate the state to referenced objects.
1567 // 1. An object is not scalar replaceable if the field into which it is
1568 // stored has unknown offset (stored into unknown element of an array).
1569 //
1570 for (UseIterator i(jobj); i.has_next(); i.next()) {
1571 PointsToNode* use = i.get();
1572 assert(!use->is_Arraycopy(), "sanity");
1573 if (use->is_Field()) {
1574 FieldNode* field = use->as_Field();
1575 assert(field->is_oop() && field->scalar_replaceable() &&
1576 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1577 if (field->offset() == Type::OffsetBot) {
1578 jobj->set_scalar_replaceable(false);
1579 return;
1580 }
1581 }
1582 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1583 // 2. An object is not scalar replaceable if it is merged with other objects.
1584 for (EdgeIterator j(use); j.has_next(); j.next()) {
1585 PointsToNode* ptn = j.get();
1586 if (ptn->is_JavaObject() && ptn != jobj) {
1587 // Mark all objects.
1588 jobj->set_scalar_replaceable(false);
1589 ptn->set_scalar_replaceable(false);
1590 }
1591 }
1592 if (!jobj->scalar_replaceable()) {
1593 return;
1594 }
1595 }
1597 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1598 // Non-escaping object node should point only to field nodes.
1599 FieldNode* field = j.get()->as_Field();
1600 int offset = field->as_Field()->offset();
1602 // 3. An object is not scalar replaceable if it has a field with unknown
1603 // offset (array's element is accessed in loop).
1604 if (offset == Type::OffsetBot) {
1605 jobj->set_scalar_replaceable(false);
1606 return;
1607 }
1608 // 4. Currently an object is not scalar replaceable if a LoadStore node
1609 // access its field since the field value is unknown after it.
1610 //
1611 Node* n = field->ideal_node();
1612 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1613 if (n->fast_out(i)->is_LoadStore()) {
1614 jobj->set_scalar_replaceable(false);
1615 return;
1616 }
1617 }
1619 // 5. Or the address may point to more then one object. This may produce
1620 // the false positive result (set not scalar replaceable)
1621 // since the flow-insensitive escape analysis can't separate
1622 // the case when stores overwrite the field's value from the case
1623 // when stores happened on different control branches.
1624 //
1625 // Note: it will disable scalar replacement in some cases:
1626 //
1627 // Point p[] = new Point[1];
1628 // p[0] = new Point(); // Will be not scalar replaced
1629 //
1630 // but it will save us from incorrect optimizations in next cases:
1631 //
1632 // Point p[] = new Point[1];
1633 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1634 //
1635 if (field->base_count() > 1) {
1636 for (BaseIterator i(field); i.has_next(); i.next()) {
1637 PointsToNode* base = i.get();
1638 // Don't take into account LocalVar nodes which
1639 // may point to only one object which should be also
1640 // this field's base by now.
1641 if (base->is_JavaObject() && base != jobj) {
1642 // Mark all bases.
1643 jobj->set_scalar_replaceable(false);
1644 base->set_scalar_replaceable(false);
1645 }
1646 }
1647 }
1648 }
1649 }
1651 #ifdef ASSERT
1652 void ConnectionGraph::verify_connection_graph(
1653 GrowableArray<PointsToNode*>& ptnodes_worklist,
1654 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1655 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1656 GrowableArray<Node*>& addp_worklist) {
1657 // Verify that graph is complete - no new edges could be added.
1658 int java_objects_length = java_objects_worklist.length();
1659 int non_escaped_length = non_escaped_worklist.length();
1660 int new_edges = 0;
1661 for (int next = 0; next < java_objects_length; ++next) {
1662 JavaObjectNode* ptn = java_objects_worklist.at(next);
1663 new_edges += add_java_object_edges(ptn, true);
1664 }
1665 assert(new_edges == 0, "graph was not complete");
1666 // Verify that escape state is final.
1667 int length = non_escaped_worklist.length();
1668 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1669 assert((non_escaped_length == non_escaped_worklist.length()) &&
1670 (non_escaped_length == length) &&
1671 (_worklist.length() == 0), "escape state was not final");
1673 // Verify fields information.
1674 int addp_length = addp_worklist.length();
1675 for (int next = 0; next < addp_length; ++next ) {
1676 Node* n = addp_worklist.at(next);
1677 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1678 if (field->is_oop()) {
1679 // Verify that field has all bases
1680 Node* base = get_addp_base(n);
1681 PointsToNode* ptn = ptnode_adr(base->_idx);
1682 if (ptn->is_JavaObject()) {
1683 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1684 } else {
1685 assert(ptn->is_LocalVar(), "sanity");
1686 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1687 PointsToNode* e = i.get();
1688 if (e->is_JavaObject()) {
1689 assert(field->has_base(e->as_JavaObject()), "sanity");
1690 }
1691 }
1692 }
1693 // Verify that all fields have initializing values.
1694 if (field->edge_count() == 0) {
1695 tty->print_cr("----------field does not have references----------");
1696 field->dump();
1697 for (BaseIterator i(field); i.has_next(); i.next()) {
1698 PointsToNode* base = i.get();
1699 tty->print_cr("----------field has next base---------------------");
1700 base->dump();
1701 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1702 tty->print_cr("----------base has fields-------------------------");
1703 for (EdgeIterator j(base); j.has_next(); j.next()) {
1704 j.get()->dump();
1705 }
1706 tty->print_cr("----------base has references---------------------");
1707 for (UseIterator j(base); j.has_next(); j.next()) {
1708 j.get()->dump();
1709 }
1710 }
1711 }
1712 for (UseIterator i(field); i.has_next(); i.next()) {
1713 i.get()->dump();
1714 }
1715 assert(field->edge_count() > 0, "sanity");
1716 }
1717 }
1718 }
1719 }
1720 #endif
1722 // Optimize ideal graph.
1723 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1724 GrowableArray<Node*>& storestore_worklist) {
1725 Compile* C = _compile;
1726 PhaseIterGVN* igvn = _igvn;
1727 if (EliminateLocks) {
1728 // Mark locks before changing ideal graph.
1729 int cnt = C->macro_count();
1730 for( int i=0; i < cnt; i++ ) {
1731 Node *n = C->macro_node(i);
1732 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1733 AbstractLockNode* alock = n->as_AbstractLock();
1734 if (!alock->is_non_esc_obj()) {
1735 if (not_global_escape(alock->obj_node())) {
1736 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1737 // The lock could be marked eliminated by lock coarsening
1738 // code during first IGVN before EA. Replace coarsened flag
1739 // to eliminate all associated locks/unlocks.
1740 alock->set_non_esc_obj();
1741 }
1742 }
1743 }
1744 }
1745 }
1747 if (OptimizePtrCompare) {
1748 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1749 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1750 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1751 // Optimize objects compare.
1752 while (ptr_cmp_worklist.length() != 0) {
1753 Node *n = ptr_cmp_worklist.pop();
1754 Node *res = optimize_ptr_compare(n);
1755 if (res != NULL) {
1756 #ifndef PRODUCT
1757 if (PrintOptimizePtrCompare) {
1758 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"));
1759 if (Verbose) {
1760 n->dump(1);
1761 }
1762 }
1763 #endif
1764 igvn->replace_node(n, res);
1765 }
1766 }
1767 // cleanup
1768 if (_pcmp_neq->outcnt() == 0)
1769 igvn->hash_delete(_pcmp_neq);
1770 if (_pcmp_eq->outcnt() == 0)
1771 igvn->hash_delete(_pcmp_eq);
1772 }
1774 // For MemBarStoreStore nodes added in library_call.cpp, check
1775 // escape status of associated AllocateNode and optimize out
1776 // MemBarStoreStore node if the allocated object never escapes.
1777 while (storestore_worklist.length() != 0) {
1778 Node *n = storestore_worklist.pop();
1779 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1780 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1781 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1782 if (not_global_escape(alloc)) {
1783 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1784 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1785 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1786 igvn->register_new_node_with_optimizer(mb);
1787 igvn->replace_node(storestore, mb);
1788 }
1789 }
1790 }
1792 // Optimize objects compare.
1793 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1794 assert(OptimizePtrCompare, "sanity");
1795 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1796 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1797 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1798 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1799 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1800 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1802 // Check simple cases first.
1803 if (jobj1 != NULL) {
1804 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1805 if (jobj1 == jobj2) {
1806 // Comparing the same not escaping object.
1807 return _pcmp_eq;
1808 }
1809 Node* obj = jobj1->ideal_node();
1810 // Comparing not escaping allocation.
1811 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1812 !ptn2->points_to(jobj1)) {
1813 return _pcmp_neq; // This includes nullness check.
1814 }
1815 }
1816 }
1817 if (jobj2 != NULL) {
1818 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1819 Node* obj = jobj2->ideal_node();
1820 // Comparing not escaping allocation.
1821 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1822 !ptn1->points_to(jobj2)) {
1823 return _pcmp_neq; // This includes nullness check.
1824 }
1825 }
1826 }
1827 if (jobj1 != NULL && jobj1 != phantom_obj &&
1828 jobj2 != NULL && jobj2 != phantom_obj &&
1829 jobj1->ideal_node()->is_Con() &&
1830 jobj2->ideal_node()->is_Con()) {
1831 // Klass or String constants compare. Need to be careful with
1832 // compressed pointers - compare types of ConN and ConP instead of nodes.
1833 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1834 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1835 if (t1->make_ptr() == t2->make_ptr()) {
1836 return _pcmp_eq;
1837 } else {
1838 return _pcmp_neq;
1839 }
1840 }
1841 if (ptn1->meet(ptn2)) {
1842 return NULL; // Sets are not disjoint
1843 }
1845 // Sets are disjoint.
1846 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1847 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1848 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1849 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1850 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1851 set2_has_unknown_ptr && set1_has_null_ptr) {
1852 // Check nullness of unknown object.
1853 return NULL;
1854 }
1856 // Disjointness by itself is not sufficient since
1857 // alias analysis is not complete for escaped objects.
1858 // Disjoint sets are definitely unrelated only when
1859 // at least one set has only not escaping allocations.
1860 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1861 if (ptn1->non_escaping_allocation()) {
1862 return _pcmp_neq;
1863 }
1864 }
1865 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1866 if (ptn2->non_escaping_allocation()) {
1867 return _pcmp_neq;
1868 }
1869 }
1870 return NULL;
1871 }
1873 // Connection Graph constuction functions.
1875 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1876 PointsToNode* ptadr = _nodes.at(n->_idx);
1877 if (ptadr != NULL) {
1878 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1879 return;
1880 }
1881 Compile* C = _compile;
1882 ptadr = new (C->comp_arena()) LocalVarNode(C, n, es);
1883 _nodes.at_put(n->_idx, ptadr);
1884 }
1886 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1887 PointsToNode* ptadr = _nodes.at(n->_idx);
1888 if (ptadr != NULL) {
1889 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1890 return;
1891 }
1892 Compile* C = _compile;
1893 ptadr = new (C->comp_arena()) JavaObjectNode(C, n, es);
1894 _nodes.at_put(n->_idx, ptadr);
1895 }
1897 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1898 PointsToNode* ptadr = _nodes.at(n->_idx);
1899 if (ptadr != NULL) {
1900 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1901 return;
1902 }
1903 bool unsafe = false;
1904 bool is_oop = is_oop_field(n, offset, &unsafe);
1905 if (unsafe) {
1906 es = PointsToNode::GlobalEscape;
1907 }
1908 Compile* C = _compile;
1909 FieldNode* field = new (C->comp_arena()) FieldNode(C, n, es, offset, is_oop);
1910 _nodes.at_put(n->_idx, field);
1911 }
1913 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1914 PointsToNode* src, PointsToNode* dst) {
1915 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1916 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1917 PointsToNode* ptadr = _nodes.at(n->_idx);
1918 if (ptadr != NULL) {
1919 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1920 return;
1921 }
1922 Compile* C = _compile;
1923 ptadr = new (C->comp_arena()) ArraycopyNode(C, n, es);
1924 _nodes.at_put(n->_idx, ptadr);
1925 // Add edge from arraycopy node to source object.
1926 (void)add_edge(ptadr, src);
1927 src->set_arraycopy_src();
1928 // Add edge from destination object to arraycopy node.
1929 (void)add_edge(dst, ptadr);
1930 dst->set_arraycopy_dst();
1931 }
1933 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1934 const Type* adr_type = n->as_AddP()->bottom_type();
1935 BasicType bt = T_INT;
1936 if (offset == Type::OffsetBot) {
1937 // Check only oop fields.
1938 if (!adr_type->isa_aryptr() ||
1939 (adr_type->isa_aryptr()->klass() == NULL) ||
1940 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1941 // OffsetBot is used to reference array's element. Ignore first AddP.
1942 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
1943 bt = T_OBJECT;
1944 }
1945 }
1946 } else if (offset != oopDesc::klass_offset_in_bytes()) {
1947 if (adr_type->isa_instptr()) {
1948 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
1949 if (field != NULL) {
1950 bt = field->layout_type();
1951 } else {
1952 // Check for unsafe oop field access
1953 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1954 int opcode = n->fast_out(i)->Opcode();
1955 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1956 opcode == Op_StoreN || opcode == Op_LoadN) {
1957 bt = T_OBJECT;
1958 (*unsafe) = true;
1959 break;
1960 }
1961 }
1962 }
1963 } else if (adr_type->isa_aryptr()) {
1964 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1965 // Ignore array length load.
1966 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
1967 // Ignore first AddP.
1968 } else {
1969 const Type* elemtype = adr_type->isa_aryptr()->elem();
1970 bt = elemtype->array_element_basic_type();
1971 }
1972 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
1973 // Allocation initialization, ThreadLocal field access, unsafe access
1974 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1975 int opcode = n->fast_out(i)->Opcode();
1976 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1977 opcode == Op_StoreN || opcode == Op_LoadN) {
1978 bt = T_OBJECT;
1979 break;
1980 }
1981 }
1982 }
1983 }
1984 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
1985 }
1987 // Returns unique pointed java object or NULL.
1988 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
1989 assert(!_collecting, "should not call when contructed graph");
1990 // If the node was created after the escape computation we can't answer.
1991 uint idx = n->_idx;
1992 if (idx >= nodes_size()) {
1993 return NULL;
1994 }
1995 PointsToNode* ptn = ptnode_adr(idx);
1996 if (ptn->is_JavaObject()) {
1997 return ptn->as_JavaObject();
1998 }
1999 assert(ptn->is_LocalVar(), "sanity");
2000 // Check all java objects it points to.
2001 JavaObjectNode* jobj = NULL;
2002 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2003 PointsToNode* e = i.get();
2004 if (e->is_JavaObject()) {
2005 if (jobj == NULL) {
2006 jobj = e->as_JavaObject();
2007 } else if (jobj != e) {
2008 return NULL;
2009 }
2010 }
2011 }
2012 return jobj;
2013 }
2015 // Return true if this node points only to non-escaping allocations.
2016 bool PointsToNode::non_escaping_allocation() {
2017 if (is_JavaObject()) {
2018 Node* n = ideal_node();
2019 if (n->is_Allocate() || n->is_CallStaticJava()) {
2020 return (escape_state() == PointsToNode::NoEscape);
2021 } else {
2022 return false;
2023 }
2024 }
2025 assert(is_LocalVar(), "sanity");
2026 // Check all java objects it points to.
2027 for (EdgeIterator i(this); i.has_next(); i.next()) {
2028 PointsToNode* e = i.get();
2029 if (e->is_JavaObject()) {
2030 Node* n = e->ideal_node();
2031 if ((e->escape_state() != PointsToNode::NoEscape) ||
2032 !(n->is_Allocate() || n->is_CallStaticJava())) {
2033 return false;
2034 }
2035 }
2036 }
2037 return true;
2038 }
2040 // Return true if we know the node does not escape globally.
2041 bool ConnectionGraph::not_global_escape(Node *n) {
2042 assert(!_collecting, "should not call during graph construction");
2043 // If the node was created after the escape computation we can't answer.
2044 uint idx = n->_idx;
2045 if (idx >= nodes_size()) {
2046 return false;
2047 }
2048 PointsToNode* ptn = ptnode_adr(idx);
2049 PointsToNode::EscapeState es = ptn->escape_state();
2050 // If we have already computed a value, return it.
2051 if (es >= PointsToNode::GlobalEscape)
2052 return false;
2053 if (ptn->is_JavaObject()) {
2054 return true; // (es < PointsToNode::GlobalEscape);
2055 }
2056 assert(ptn->is_LocalVar(), "sanity");
2057 // Check all java objects it points to.
2058 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2059 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2060 return false;
2061 }
2062 return true;
2063 }
2066 // Helper functions
2068 // Return true if this node points to specified node or nodes it points to.
2069 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2070 if (is_JavaObject()) {
2071 return (this == ptn);
2072 }
2073 assert(is_LocalVar() || is_Field(), "sanity");
2074 for (EdgeIterator i(this); i.has_next(); i.next()) {
2075 if (i.get() == ptn)
2076 return true;
2077 }
2078 return false;
2079 }
2081 // Return true if one node points to an other.
2082 bool PointsToNode::meet(PointsToNode* ptn) {
2083 if (this == ptn) {
2084 return true;
2085 } else if (ptn->is_JavaObject()) {
2086 return this->points_to(ptn->as_JavaObject());
2087 } else if (this->is_JavaObject()) {
2088 return ptn->points_to(this->as_JavaObject());
2089 }
2090 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2091 int ptn_count = ptn->edge_count();
2092 for (EdgeIterator i(this); i.has_next(); i.next()) {
2093 PointsToNode* this_e = i.get();
2094 for (int j = 0; j < ptn_count; j++) {
2095 if (this_e == ptn->edge(j))
2096 return true;
2097 }
2098 }
2099 return false;
2100 }
2102 #ifdef ASSERT
2103 // Return true if bases point to this java object.
2104 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2105 for (BaseIterator i(this); i.has_next(); i.next()) {
2106 if (i.get() == jobj)
2107 return true;
2108 }
2109 return false;
2110 }
2111 #endif
2113 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2114 const Type *adr_type = phase->type(adr);
2115 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2116 adr->in(AddPNode::Address)->is_Proj() &&
2117 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2118 // We are computing a raw address for a store captured by an Initialize
2119 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2120 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2121 assert(offs != Type::OffsetBot ||
2122 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2123 "offset must be a constant or it is initialization of array");
2124 return offs;
2125 }
2126 const TypePtr *t_ptr = adr_type->isa_ptr();
2127 assert(t_ptr != NULL, "must be a pointer type");
2128 return t_ptr->offset();
2129 }
2131 Node* ConnectionGraph::get_addp_base(Node *addp) {
2132 assert(addp->is_AddP(), "must be AddP");
2133 //
2134 // AddP cases for Base and Address inputs:
2135 // case #1. Direct object's field reference:
2136 // Allocate
2137 // |
2138 // Proj #5 ( oop result )
2139 // |
2140 // CheckCastPP (cast to instance type)
2141 // | |
2142 // AddP ( base == address )
2143 //
2144 // case #2. Indirect object's field reference:
2145 // Phi
2146 // |
2147 // CastPP (cast to instance type)
2148 // | |
2149 // AddP ( base == address )
2150 //
2151 // case #3. Raw object's field reference for Initialize node:
2152 // Allocate
2153 // |
2154 // Proj #5 ( oop result )
2155 // top |
2156 // \ |
2157 // AddP ( base == top )
2158 //
2159 // case #4. Array's element reference:
2160 // {CheckCastPP | CastPP}
2161 // | | |
2162 // | AddP ( array's element offset )
2163 // | |
2164 // AddP ( array's offset )
2165 //
2166 // case #5. Raw object's field reference for arraycopy stub call:
2167 // The inline_native_clone() case when the arraycopy stub is called
2168 // after the allocation before Initialize and CheckCastPP nodes.
2169 // Allocate
2170 // |
2171 // Proj #5 ( oop result )
2172 // | |
2173 // AddP ( base == address )
2174 //
2175 // case #6. Constant Pool, ThreadLocal, CastX2P or
2176 // Raw object's field reference:
2177 // {ConP, ThreadLocal, CastX2P, raw Load}
2178 // top |
2179 // \ |
2180 // AddP ( base == top )
2181 //
2182 // case #7. Klass's field reference.
2183 // LoadKlass
2184 // | |
2185 // AddP ( base == address )
2186 //
2187 // case #8. narrow Klass's field reference.
2188 // LoadNKlass
2189 // |
2190 // DecodeN
2191 // | |
2192 // AddP ( base == address )
2193 //
2194 Node *base = addp->in(AddPNode::Base);
2195 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2196 base = addp->in(AddPNode::Address);
2197 while (base->is_AddP()) {
2198 // Case #6 (unsafe access) may have several chained AddP nodes.
2199 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2200 base = base->in(AddPNode::Address);
2201 }
2202 Node* uncast_base = base->uncast();
2203 int opcode = uncast_base->Opcode();
2204 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2205 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2206 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2207 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2208 }
2209 return base;
2210 }
2212 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2213 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2214 Node* addp2 = addp->raw_out(0);
2215 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2216 addp2->in(AddPNode::Base) == n &&
2217 addp2->in(AddPNode::Address) == addp) {
2218 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2219 //
2220 // Find array's offset to push it on worklist first and
2221 // as result process an array's element offset first (pushed second)
2222 // to avoid CastPP for the array's offset.
2223 // Otherwise the inserted CastPP (LocalVar) will point to what
2224 // the AddP (Field) points to. Which would be wrong since
2225 // the algorithm expects the CastPP has the same point as
2226 // as AddP's base CheckCastPP (LocalVar).
2227 //
2228 // ArrayAllocation
2229 // |
2230 // CheckCastPP
2231 // |
2232 // memProj (from ArrayAllocation CheckCastPP)
2233 // | ||
2234 // | || Int (element index)
2235 // | || | ConI (log(element size))
2236 // | || | /
2237 // | || LShift
2238 // | || /
2239 // | AddP (array's element offset)
2240 // | |
2241 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2242 // | / /
2243 // AddP (array's offset)
2244 // |
2245 // Load/Store (memory operation on array's element)
2246 //
2247 return addp2;
2248 }
2249 return NULL;
2250 }
2252 //
2253 // Adjust the type and inputs of an AddP which computes the
2254 // address of a field of an instance
2255 //
2256 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2257 PhaseGVN* igvn = _igvn;
2258 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2259 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2260 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2261 if (t == NULL) {
2262 // We are computing a raw address for a store captured by an Initialize
2263 // compute an appropriate address type (cases #3 and #5).
2264 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2265 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2266 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2267 assert(offs != Type::OffsetBot, "offset must be a constant");
2268 t = base_t->add_offset(offs)->is_oopptr();
2269 }
2270 int inst_id = base_t->instance_id();
2271 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2272 "old type must be non-instance or match new type");
2274 // The type 't' could be subclass of 'base_t'.
2275 // As result t->offset() could be large then base_t's size and it will
2276 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2277 // constructor verifies correctness of the offset.
2278 //
2279 // It could happened on subclass's branch (from the type profiling
2280 // inlining) which was not eliminated during parsing since the exactness
2281 // of the allocation type was not propagated to the subclass type check.
2282 //
2283 // Or the type 't' could be not related to 'base_t' at all.
2284 // It could happened when CHA type is different from MDO type on a dead path
2285 // (for example, from instanceof check) which is not collapsed during parsing.
2286 //
2287 // Do nothing for such AddP node and don't process its users since
2288 // this code branch will go away.
2289 //
2290 if (!t->is_known_instance() &&
2291 !base_t->klass()->is_subtype_of(t->klass())) {
2292 return false; // bail out
2293 }
2294 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2295 // Do NOT remove the next line: ensure a new alias index is allocated
2296 // for the instance type. Note: C++ will not remove it since the call
2297 // has side effect.
2298 int alias_idx = _compile->get_alias_index(tinst);
2299 igvn->set_type(addp, tinst);
2300 // record the allocation in the node map
2301 set_map(addp, get_map(base->_idx));
2302 // Set addp's Base and Address to 'base'.
2303 Node *abase = addp->in(AddPNode::Base);
2304 Node *adr = addp->in(AddPNode::Address);
2305 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2306 adr->in(0)->_idx == (uint)inst_id) {
2307 // Skip AddP cases #3 and #5.
2308 } else {
2309 assert(!abase->is_top(), "sanity"); // AddP case #3
2310 if (abase != base) {
2311 igvn->hash_delete(addp);
2312 addp->set_req(AddPNode::Base, base);
2313 if (abase == adr) {
2314 addp->set_req(AddPNode::Address, base);
2315 } else {
2316 // AddP case #4 (adr is array's element offset AddP node)
2317 #ifdef ASSERT
2318 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2319 assert(adr->is_AddP() && atype != NULL &&
2320 atype->instance_id() == inst_id, "array's element offset should be processed first");
2321 #endif
2322 }
2323 igvn->hash_insert(addp);
2324 }
2325 }
2326 // Put on IGVN worklist since at least addp's type was changed above.
2327 record_for_optimizer(addp);
2328 return true;
2329 }
2331 //
2332 // Create a new version of orig_phi if necessary. Returns either the newly
2333 // created phi or an existing phi. Sets create_new to indicate whether a new
2334 // phi was created. Cache the last newly created phi in the node map.
2335 //
2336 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2337 Compile *C = _compile;
2338 PhaseGVN* igvn = _igvn;
2339 new_created = false;
2340 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2341 // nothing to do if orig_phi is bottom memory or matches alias_idx
2342 if (phi_alias_idx == alias_idx) {
2343 return orig_phi;
2344 }
2345 // Have we recently created a Phi for this alias index?
2346 PhiNode *result = get_map_phi(orig_phi->_idx);
2347 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2348 return result;
2349 }
2350 // Previous check may fail when the same wide memory Phi was split into Phis
2351 // for different memory slices. Search all Phis for this region.
2352 if (result != NULL) {
2353 Node* region = orig_phi->in(0);
2354 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2355 Node* phi = region->fast_out(i);
2356 if (phi->is_Phi() &&
2357 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2358 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2359 return phi->as_Phi();
2360 }
2361 }
2362 }
2363 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) {
2364 if (C->do_escape_analysis() == true && !C->failing()) {
2365 // Retry compilation without escape analysis.
2366 // If this is the first failure, the sentinel string will "stick"
2367 // to the Compile object, and the C2Compiler will see it and retry.
2368 C->record_failure(C2Compiler::retry_no_escape_analysis());
2369 }
2370 return NULL;
2371 }
2372 orig_phi_worklist.append_if_missing(orig_phi);
2373 const TypePtr *atype = C->get_adr_type(alias_idx);
2374 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2375 C->copy_node_notes_to(result, orig_phi);
2376 igvn->set_type(result, result->bottom_type());
2377 record_for_optimizer(result);
2378 set_map(orig_phi, result);
2379 new_created = true;
2380 return result;
2381 }
2383 //
2384 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2385 // specified alias index.
2386 //
2387 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2388 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2389 Compile *C = _compile;
2390 PhaseGVN* igvn = _igvn;
2391 bool new_phi_created;
2392 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2393 if (!new_phi_created) {
2394 return result;
2395 }
2396 GrowableArray<PhiNode *> phi_list;
2397 GrowableArray<uint> cur_input;
2398 PhiNode *phi = orig_phi;
2399 uint idx = 1;
2400 bool finished = false;
2401 while(!finished) {
2402 while (idx < phi->req()) {
2403 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2404 if (mem != NULL && mem->is_Phi()) {
2405 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2406 if (new_phi_created) {
2407 // found an phi for which we created a new split, push current one on worklist and begin
2408 // processing new one
2409 phi_list.push(phi);
2410 cur_input.push(idx);
2411 phi = mem->as_Phi();
2412 result = newphi;
2413 idx = 1;
2414 continue;
2415 } else {
2416 mem = newphi;
2417 }
2418 }
2419 if (C->failing()) {
2420 return NULL;
2421 }
2422 result->set_req(idx++, mem);
2423 }
2424 #ifdef ASSERT
2425 // verify that the new Phi has an input for each input of the original
2426 assert( phi->req() == result->req(), "must have same number of inputs.");
2427 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2428 #endif
2429 // Check if all new phi's inputs have specified alias index.
2430 // Otherwise use old phi.
2431 for (uint i = 1; i < phi->req(); i++) {
2432 Node* in = result->in(i);
2433 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2434 }
2435 // we have finished processing a Phi, see if there are any more to do
2436 finished = (phi_list.length() == 0 );
2437 if (!finished) {
2438 phi = phi_list.pop();
2439 idx = cur_input.pop();
2440 PhiNode *prev_result = get_map_phi(phi->_idx);
2441 prev_result->set_req(idx++, result);
2442 result = prev_result;
2443 }
2444 }
2445 return result;
2446 }
2448 //
2449 // The next methods are derived from methods in MemNode.
2450 //
2451 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2452 Node *mem = mmem;
2453 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2454 // means an array I have not precisely typed yet. Do not do any
2455 // alias stuff with it any time soon.
2456 if (toop->base() != Type::AnyPtr &&
2457 !(toop->klass() != NULL &&
2458 toop->klass()->is_java_lang_Object() &&
2459 toop->offset() == Type::OffsetBot)) {
2460 mem = mmem->memory_at(alias_idx);
2461 // Update input if it is progress over what we have now
2462 }
2463 return mem;
2464 }
2466 //
2467 // Move memory users to their memory slices.
2468 //
2469 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2470 Compile* C = _compile;
2471 PhaseGVN* igvn = _igvn;
2472 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2473 assert(tp != NULL, "ptr type");
2474 int alias_idx = C->get_alias_index(tp);
2475 int general_idx = C->get_general_index(alias_idx);
2477 // Move users first
2478 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2479 Node* use = n->fast_out(i);
2480 if (use->is_MergeMem()) {
2481 MergeMemNode* mmem = use->as_MergeMem();
2482 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2483 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2484 continue; // Nothing to do
2485 }
2486 // Replace previous general reference to mem node.
2487 uint orig_uniq = C->unique();
2488 Node* m = find_inst_mem(n, general_idx, orig_phis);
2489 assert(orig_uniq == C->unique(), "no new nodes");
2490 mmem->set_memory_at(general_idx, m);
2491 --imax;
2492 --i;
2493 } else if (use->is_MemBar()) {
2494 assert(!use->is_Initialize(), "initializing stores should not be moved");
2495 if (use->req() > MemBarNode::Precedent &&
2496 use->in(MemBarNode::Precedent) == n) {
2497 // Don't move related membars.
2498 record_for_optimizer(use);
2499 continue;
2500 }
2501 tp = use->as_MemBar()->adr_type()->isa_ptr();
2502 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2503 alias_idx == general_idx) {
2504 continue; // Nothing to do
2505 }
2506 // Move to general memory slice.
2507 uint orig_uniq = C->unique();
2508 Node* m = find_inst_mem(n, general_idx, orig_phis);
2509 assert(orig_uniq == C->unique(), "no new nodes");
2510 igvn->hash_delete(use);
2511 imax -= use->replace_edge(n, m);
2512 igvn->hash_insert(use);
2513 record_for_optimizer(use);
2514 --i;
2515 #ifdef ASSERT
2516 } else if (use->is_Mem()) {
2517 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2518 // Don't move related cardmark.
2519 continue;
2520 }
2521 // Memory nodes should have new memory input.
2522 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2523 assert(tp != NULL, "ptr type");
2524 int idx = C->get_alias_index(tp);
2525 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2526 "Following memory nodes should have new memory input or be on the same memory slice");
2527 } else if (use->is_Phi()) {
2528 // Phi nodes should be split and moved already.
2529 tp = use->as_Phi()->adr_type()->isa_ptr();
2530 assert(tp != NULL, "ptr type");
2531 int idx = C->get_alias_index(tp);
2532 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2533 } else {
2534 use->dump();
2535 assert(false, "should not be here");
2536 #endif
2537 }
2538 }
2539 }
2541 //
2542 // Search memory chain of "mem" to find a MemNode whose address
2543 // is the specified alias index.
2544 //
2545 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2546 if (orig_mem == NULL)
2547 return orig_mem;
2548 Compile* C = _compile;
2549 PhaseGVN* igvn = _igvn;
2550 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2551 bool is_instance = (toop != NULL) && toop->is_known_instance();
2552 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2553 Node *prev = NULL;
2554 Node *result = orig_mem;
2555 while (prev != result) {
2556 prev = result;
2557 if (result == start_mem)
2558 break; // hit one of our sentinels
2559 if (result->is_Mem()) {
2560 const Type *at = igvn->type(result->in(MemNode::Address));
2561 if (at == Type::TOP)
2562 break; // Dead
2563 assert (at->isa_ptr() != NULL, "pointer type required.");
2564 int idx = C->get_alias_index(at->is_ptr());
2565 if (idx == alias_idx)
2566 break; // Found
2567 if (!is_instance && (at->isa_oopptr() == NULL ||
2568 !at->is_oopptr()->is_known_instance())) {
2569 break; // Do not skip store to general memory slice.
2570 }
2571 result = result->in(MemNode::Memory);
2572 }
2573 if (!is_instance)
2574 continue; // don't search further for non-instance types
2575 // skip over a call which does not affect this memory slice
2576 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2577 Node *proj_in = result->in(0);
2578 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2579 break; // hit one of our sentinels
2580 } else if (proj_in->is_Call()) {
2581 CallNode *call = proj_in->as_Call();
2582 if (!call->may_modify(toop, igvn)) {
2583 result = call->in(TypeFunc::Memory);
2584 }
2585 } else if (proj_in->is_Initialize()) {
2586 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2587 // Stop if this is the initialization for the object instance which
2588 // which contains this memory slice, otherwise skip over it.
2589 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2590 result = proj_in->in(TypeFunc::Memory);
2591 }
2592 } else if (proj_in->is_MemBar()) {
2593 result = proj_in->in(TypeFunc::Memory);
2594 }
2595 } else if (result->is_MergeMem()) {
2596 MergeMemNode *mmem = result->as_MergeMem();
2597 result = step_through_mergemem(mmem, alias_idx, toop);
2598 if (result == mmem->base_memory()) {
2599 // Didn't find instance memory, search through general slice recursively.
2600 result = mmem->memory_at(C->get_general_index(alias_idx));
2601 result = find_inst_mem(result, alias_idx, orig_phis);
2602 if (C->failing()) {
2603 return NULL;
2604 }
2605 mmem->set_memory_at(alias_idx, result);
2606 }
2607 } else if (result->is_Phi() &&
2608 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2609 Node *un = result->as_Phi()->unique_input(igvn);
2610 if (un != NULL) {
2611 orig_phis.append_if_missing(result->as_Phi());
2612 result = un;
2613 } else {
2614 break;
2615 }
2616 } else if (result->is_ClearArray()) {
2617 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2618 // Can not bypass initialization of the instance
2619 // we are looking for.
2620 break;
2621 }
2622 // Otherwise skip it (the call updated 'result' value).
2623 } else if (result->Opcode() == Op_SCMemProj) {
2624 Node* mem = result->in(0);
2625 Node* adr = NULL;
2626 if (mem->is_LoadStore()) {
2627 adr = mem->in(MemNode::Address);
2628 } else {
2629 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2630 adr = mem->in(3); // Memory edge corresponds to destination array
2631 }
2632 const Type *at = igvn->type(adr);
2633 if (at != Type::TOP) {
2634 assert (at->isa_ptr() != NULL, "pointer type required.");
2635 int idx = C->get_alias_index(at->is_ptr());
2636 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2637 break;
2638 }
2639 result = mem->in(MemNode::Memory);
2640 }
2641 }
2642 if (result->is_Phi()) {
2643 PhiNode *mphi = result->as_Phi();
2644 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2645 const TypePtr *t = mphi->adr_type();
2646 if (!is_instance) {
2647 // Push all non-instance Phis on the orig_phis worklist to update inputs
2648 // during Phase 4 if needed.
2649 orig_phis.append_if_missing(mphi);
2650 } else if (C->get_alias_index(t) != alias_idx) {
2651 // Create a new Phi with the specified alias index type.
2652 result = split_memory_phi(mphi, alias_idx, orig_phis);
2653 }
2654 }
2655 // the result is either MemNode, PhiNode, InitializeNode.
2656 return result;
2657 }
2659 //
2660 // Convert the types of unescaped object to instance types where possible,
2661 // propagate the new type information through the graph, and update memory
2662 // edges and MergeMem inputs to reflect the new type.
2663 //
2664 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2665 // The processing is done in 4 phases:
2666 //
2667 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2668 // types for the CheckCastPP for allocations where possible.
2669 // Propagate the the new types through users as follows:
2670 // casts and Phi: push users on alloc_worklist
2671 // AddP: cast Base and Address inputs to the instance type
2672 // push any AddP users on alloc_worklist and push any memnode
2673 // users onto memnode_worklist.
2674 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2675 // search the Memory chain for a store with the appropriate type
2676 // address type. If a Phi is found, create a new version with
2677 // the appropriate memory slices from each of the Phi inputs.
2678 // For stores, process the users as follows:
2679 // MemNode: push on memnode_worklist
2680 // MergeMem: push on mergemem_worklist
2681 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2682 // moving the first node encountered of each instance type to the
2683 // the input corresponding to its alias index.
2684 // appropriate memory slice.
2685 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2686 //
2687 // In the following example, the CheckCastPP nodes are the cast of allocation
2688 // results and the allocation of node 29 is unescaped and eligible to be an
2689 // instance type.
2690 //
2691 // We start with:
2692 //
2693 // 7 Parm #memory
2694 // 10 ConI "12"
2695 // 19 CheckCastPP "Foo"
2696 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2697 // 29 CheckCastPP "Foo"
2698 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2699 //
2700 // 40 StoreP 25 7 20 ... alias_index=4
2701 // 50 StoreP 35 40 30 ... alias_index=4
2702 // 60 StoreP 45 50 20 ... alias_index=4
2703 // 70 LoadP _ 60 30 ... alias_index=4
2704 // 80 Phi 75 50 60 Memory alias_index=4
2705 // 90 LoadP _ 80 30 ... alias_index=4
2706 // 100 LoadP _ 80 20 ... alias_index=4
2707 //
2708 //
2709 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2710 // and creating a new alias index for node 30. This gives:
2711 //
2712 // 7 Parm #memory
2713 // 10 ConI "12"
2714 // 19 CheckCastPP "Foo"
2715 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2716 // 29 CheckCastPP "Foo" iid=24
2717 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2718 //
2719 // 40 StoreP 25 7 20 ... alias_index=4
2720 // 50 StoreP 35 40 30 ... alias_index=6
2721 // 60 StoreP 45 50 20 ... alias_index=4
2722 // 70 LoadP _ 60 30 ... alias_index=6
2723 // 80 Phi 75 50 60 Memory alias_index=4
2724 // 90 LoadP _ 80 30 ... alias_index=6
2725 // 100 LoadP _ 80 20 ... alias_index=4
2726 //
2727 // In phase 2, new memory inputs are computed for the loads and stores,
2728 // And a new version of the phi is created. In phase 4, the inputs to
2729 // node 80 are updated and then the memory nodes are updated with the
2730 // values computed in phase 2. This results in:
2731 //
2732 // 7 Parm #memory
2733 // 10 ConI "12"
2734 // 19 CheckCastPP "Foo"
2735 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2736 // 29 CheckCastPP "Foo" iid=24
2737 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2738 //
2739 // 40 StoreP 25 7 20 ... alias_index=4
2740 // 50 StoreP 35 7 30 ... alias_index=6
2741 // 60 StoreP 45 40 20 ... alias_index=4
2742 // 70 LoadP _ 50 30 ... alias_index=6
2743 // 80 Phi 75 40 60 Memory alias_index=4
2744 // 120 Phi 75 50 50 Memory alias_index=6
2745 // 90 LoadP _ 120 30 ... alias_index=6
2746 // 100 LoadP _ 80 20 ... alias_index=4
2747 //
2748 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2749 GrowableArray<Node *> memnode_worklist;
2750 GrowableArray<PhiNode *> orig_phis;
2751 PhaseIterGVN *igvn = _igvn;
2752 uint new_index_start = (uint) _compile->num_alias_types();
2753 Arena* arena = Thread::current()->resource_area();
2754 VectorSet visited(arena);
2755 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2756 uint unique_old = _compile->unique();
2758 // Phase 1: Process possible allocations from alloc_worklist.
2759 // Create instance types for the CheckCastPP for allocations where possible.
2760 //
2761 // (Note: don't forget to change the order of the second AddP node on
2762 // the alloc_worklist if the order of the worklist processing is changed,
2763 // see the comment in find_second_addp().)
2764 //
2765 while (alloc_worklist.length() != 0) {
2766 Node *n = alloc_worklist.pop();
2767 uint ni = n->_idx;
2768 if (n->is_Call()) {
2769 CallNode *alloc = n->as_Call();
2770 // copy escape information to call node
2771 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2772 PointsToNode::EscapeState es = ptn->escape_state();
2773 // We have an allocation or call which returns a Java object,
2774 // see if it is unescaped.
2775 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2776 continue;
2777 // Find CheckCastPP for the allocate or for the return value of a call
2778 n = alloc->result_cast();
2779 if (n == NULL) { // No uses except Initialize node
2780 if (alloc->is_Allocate()) {
2781 // Set the scalar_replaceable flag for allocation
2782 // so it could be eliminated if it has no uses.
2783 alloc->as_Allocate()->_is_scalar_replaceable = true;
2784 }
2785 if (alloc->is_CallStaticJava()) {
2786 // Set the scalar_replaceable flag for boxing method
2787 // so it could be eliminated if it has no uses.
2788 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2789 }
2790 continue;
2791 }
2792 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2793 assert(!alloc->is_Allocate(), "allocation should have unique type");
2794 continue;
2795 }
2797 // The inline code for Object.clone() casts the allocation result to
2798 // java.lang.Object and then to the actual type of the allocated
2799 // object. Detect this case and use the second cast.
2800 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2801 // the allocation result is cast to java.lang.Object and then
2802 // to the actual Array type.
2803 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2804 && (alloc->is_AllocateArray() ||
2805 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2806 Node *cast2 = NULL;
2807 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2808 Node *use = n->fast_out(i);
2809 if (use->is_CheckCastPP()) {
2810 cast2 = use;
2811 break;
2812 }
2813 }
2814 if (cast2 != NULL) {
2815 n = cast2;
2816 } else {
2817 // Non-scalar replaceable if the allocation type is unknown statically
2818 // (reflection allocation), the object can't be restored during
2819 // deoptimization without precise type.
2820 continue;
2821 }
2822 }
2823 if (alloc->is_Allocate()) {
2824 // Set the scalar_replaceable flag for allocation
2825 // so it could be eliminated.
2826 alloc->as_Allocate()->_is_scalar_replaceable = true;
2827 }
2828 if (alloc->is_CallStaticJava()) {
2829 // Set the scalar_replaceable flag for boxing method
2830 // so it could be eliminated.
2831 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2832 }
2833 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2834 // in order for an object to be scalar-replaceable, it must be:
2835 // - a direct allocation (not a call returning an object)
2836 // - non-escaping
2837 // - eligible to be a unique type
2838 // - not determined to be ineligible by escape analysis
2839 set_map(alloc, n);
2840 set_map(n, alloc);
2841 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2842 if (t == NULL)
2843 continue; // not a TypeOopPtr
2844 const TypeOopPtr* tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
2845 igvn->hash_delete(n);
2846 igvn->set_type(n, tinst);
2847 n->raise_bottom_type(tinst);
2848 igvn->hash_insert(n);
2849 record_for_optimizer(n);
2850 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2852 // First, put on the worklist all Field edges from Connection Graph
2853 // which is more accurate then putting immediate users from Ideal Graph.
2854 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2855 PointsToNode* tgt = e.get();
2856 Node* use = tgt->ideal_node();
2857 assert(tgt->is_Field() && use->is_AddP(),
2858 "only AddP nodes are Field edges in CG");
2859 if (use->outcnt() > 0) { // Don't process dead nodes
2860 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2861 if (addp2 != NULL) {
2862 assert(alloc->is_AllocateArray(),"array allocation was expected");
2863 alloc_worklist.append_if_missing(addp2);
2864 }
2865 alloc_worklist.append_if_missing(use);
2866 }
2867 }
2869 // An allocation may have an Initialize which has raw stores. Scan
2870 // the users of the raw allocation result and push AddP users
2871 // on alloc_worklist.
2872 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2873 assert (raw_result != NULL, "must have an allocation result");
2874 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2875 Node *use = raw_result->fast_out(i);
2876 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2877 Node* addp2 = find_second_addp(use, raw_result);
2878 if (addp2 != NULL) {
2879 assert(alloc->is_AllocateArray(),"array allocation was expected");
2880 alloc_worklist.append_if_missing(addp2);
2881 }
2882 alloc_worklist.append_if_missing(use);
2883 } else if (use->is_MemBar()) {
2884 memnode_worklist.append_if_missing(use);
2885 }
2886 }
2887 }
2888 } else if (n->is_AddP()) {
2889 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2890 if (jobj == NULL || jobj == phantom_obj) {
2891 #ifdef ASSERT
2892 ptnode_adr(get_addp_base(n)->_idx)->dump();
2893 ptnode_adr(n->_idx)->dump();
2894 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2895 #endif
2896 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2897 return;
2898 }
2899 Node *base = get_map(jobj->idx()); // CheckCastPP node
2900 if (!split_AddP(n, base)) continue; // wrong type from dead path
2901 } else if (n->is_Phi() ||
2902 n->is_CheckCastPP() ||
2903 n->is_EncodeP() ||
2904 n->is_DecodeN() ||
2905 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2906 if (visited.test_set(n->_idx)) {
2907 assert(n->is_Phi(), "loops only through Phi's");
2908 continue; // already processed
2909 }
2910 JavaObjectNode* jobj = unique_java_object(n);
2911 if (jobj == NULL || jobj == phantom_obj) {
2912 #ifdef ASSERT
2913 ptnode_adr(n->_idx)->dump();
2914 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2915 #endif
2916 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2917 return;
2918 } else {
2919 Node *val = get_map(jobj->idx()); // CheckCastPP node
2920 TypeNode *tn = n->as_Type();
2921 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2922 assert(tinst != NULL && tinst->is_known_instance() &&
2923 tinst->instance_id() == jobj->idx() , "instance type expected.");
2925 const Type *tn_type = igvn->type(tn);
2926 const TypeOopPtr *tn_t;
2927 if (tn_type->isa_narrowoop()) {
2928 tn_t = tn_type->make_ptr()->isa_oopptr();
2929 } else {
2930 tn_t = tn_type->isa_oopptr();
2931 }
2932 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2933 if (tn_type->isa_narrowoop()) {
2934 tn_type = tinst->make_narrowoop();
2935 } else {
2936 tn_type = tinst;
2937 }
2938 igvn->hash_delete(tn);
2939 igvn->set_type(tn, tn_type);
2940 tn->set_type(tn_type);
2941 igvn->hash_insert(tn);
2942 record_for_optimizer(n);
2943 } else {
2944 assert(tn_type == TypePtr::NULL_PTR ||
2945 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
2946 "unexpected type");
2947 continue; // Skip dead path with different type
2948 }
2949 }
2950 } else {
2951 debug_only(n->dump();)
2952 assert(false, "EA: unexpected node");
2953 continue;
2954 }
2955 // push allocation's users on appropriate worklist
2956 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2957 Node *use = n->fast_out(i);
2958 if(use->is_Mem() && use->in(MemNode::Address) == n) {
2959 // Load/store to instance's field
2960 memnode_worklist.append_if_missing(use);
2961 } else if (use->is_MemBar()) {
2962 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
2963 memnode_worklist.append_if_missing(use);
2964 }
2965 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
2966 Node* addp2 = find_second_addp(use, n);
2967 if (addp2 != NULL) {
2968 alloc_worklist.append_if_missing(addp2);
2969 }
2970 alloc_worklist.append_if_missing(use);
2971 } else if (use->is_Phi() ||
2972 use->is_CheckCastPP() ||
2973 use->is_EncodeNarrowPtr() ||
2974 use->is_DecodeNarrowPtr() ||
2975 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
2976 alloc_worklist.append_if_missing(use);
2977 #ifdef ASSERT
2978 } else if (use->is_Mem()) {
2979 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
2980 } else if (use->is_MergeMem()) {
2981 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
2982 } else if (use->is_SafePoint()) {
2983 // Look for MergeMem nodes for calls which reference unique allocation
2984 // (through CheckCastPP nodes) even for debug info.
2985 Node* m = use->in(TypeFunc::Memory);
2986 if (m->is_MergeMem()) {
2987 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
2988 }
2989 } else if (use->Opcode() == Op_EncodeISOArray) {
2990 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
2991 // EncodeISOArray overwrites destination array
2992 memnode_worklist.append_if_missing(use);
2993 }
2994 } else {
2995 uint op = use->Opcode();
2996 if (!(op == Op_CmpP || op == Op_Conv2B ||
2997 op == Op_CastP2X || op == Op_StoreCM ||
2998 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
2999 op == Op_StrEquals || op == Op_StrIndexOf)) {
3000 n->dump();
3001 use->dump();
3002 assert(false, "EA: missing allocation reference path");
3003 }
3004 #endif
3005 }
3006 }
3008 }
3009 // New alias types were created in split_AddP().
3010 uint new_index_end = (uint) _compile->num_alias_types();
3011 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3013 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3014 // compute new values for Memory inputs (the Memory inputs are not
3015 // actually updated until phase 4.)
3016 if (memnode_worklist.length() == 0)
3017 return; // nothing to do
3018 while (memnode_worklist.length() != 0) {
3019 Node *n = memnode_worklist.pop();
3020 if (visited.test_set(n->_idx))
3021 continue;
3022 if (n->is_Phi() || n->is_ClearArray()) {
3023 // we don't need to do anything, but the users must be pushed
3024 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3025 // we don't need to do anything, but the users must be pushed
3026 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3027 if (n == NULL)
3028 continue;
3029 } else if (n->Opcode() == Op_EncodeISOArray) {
3030 // get the memory projection
3031 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3032 Node *use = n->fast_out(i);
3033 if (use->Opcode() == Op_SCMemProj) {
3034 n = use;
3035 break;
3036 }
3037 }
3038 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3039 } else {
3040 assert(n->is_Mem(), "memory node required.");
3041 Node *addr = n->in(MemNode::Address);
3042 const Type *addr_t = igvn->type(addr);
3043 if (addr_t == Type::TOP)
3044 continue;
3045 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3046 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3047 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3048 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3049 if (_compile->failing()) {
3050 return;
3051 }
3052 if (mem != n->in(MemNode::Memory)) {
3053 // We delay the memory edge update since we need old one in
3054 // MergeMem code below when instances memory slices are separated.
3055 set_map(n, mem);
3056 }
3057 if (n->is_Load()) {
3058 continue; // don't push users
3059 } else if (n->is_LoadStore()) {
3060 // get the memory projection
3061 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3062 Node *use = n->fast_out(i);
3063 if (use->Opcode() == Op_SCMemProj) {
3064 n = use;
3065 break;
3066 }
3067 }
3068 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3069 }
3070 }
3071 // push user on appropriate worklist
3072 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3073 Node *use = n->fast_out(i);
3074 if (use->is_Phi() || use->is_ClearArray()) {
3075 memnode_worklist.append_if_missing(use);
3076 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3077 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3078 continue;
3079 memnode_worklist.append_if_missing(use);
3080 } else if (use->is_MemBar()) {
3081 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3082 memnode_worklist.append_if_missing(use);
3083 }
3084 #ifdef ASSERT
3085 } else if(use->is_Mem()) {
3086 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3087 } else if (use->is_MergeMem()) {
3088 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3089 } else if (use->Opcode() == Op_EncodeISOArray) {
3090 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3091 // EncodeISOArray overwrites destination array
3092 memnode_worklist.append_if_missing(use);
3093 }
3094 } else {
3095 uint op = use->Opcode();
3096 if (!(op == Op_StoreCM ||
3097 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3098 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3099 op == Op_AryEq || op == Op_StrComp ||
3100 op == Op_StrEquals || op == Op_StrIndexOf)) {
3101 n->dump();
3102 use->dump();
3103 assert(false, "EA: missing memory path");
3104 }
3105 #endif
3106 }
3107 }
3108 }
3110 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3111 // Walk each memory slice moving the first node encountered of each
3112 // instance type to the the input corresponding to its alias index.
3113 uint length = _mergemem_worklist.length();
3114 for( uint next = 0; next < length; ++next ) {
3115 MergeMemNode* nmm = _mergemem_worklist.at(next);
3116 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3117 // Note: we don't want to use MergeMemStream here because we only want to
3118 // scan inputs which exist at the start, not ones we add during processing.
3119 // Note 2: MergeMem may already contains instance memory slices added
3120 // during find_inst_mem() call when memory nodes were processed above.
3121 igvn->hash_delete(nmm);
3122 uint nslices = nmm->req();
3123 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3124 Node* mem = nmm->in(i);
3125 Node* cur = NULL;
3126 if (mem == NULL || mem->is_top())
3127 continue;
3128 // First, update mergemem by moving memory nodes to corresponding slices
3129 // if their type became more precise since this mergemem was created.
3130 while (mem->is_Mem()) {
3131 const Type *at = igvn->type(mem->in(MemNode::Address));
3132 if (at != Type::TOP) {
3133 assert (at->isa_ptr() != NULL, "pointer type required.");
3134 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3135 if (idx == i) {
3136 if (cur == NULL)
3137 cur = mem;
3138 } else {
3139 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3140 nmm->set_memory_at(idx, mem);
3141 }
3142 }
3143 }
3144 mem = mem->in(MemNode::Memory);
3145 }
3146 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3147 // Find any instance of the current type if we haven't encountered
3148 // already a memory slice of the instance along the memory chain.
3149 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3150 if((uint)_compile->get_general_index(ni) == i) {
3151 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3152 if (nmm->is_empty_memory(m)) {
3153 Node* result = find_inst_mem(mem, ni, orig_phis);
3154 if (_compile->failing()) {
3155 return;
3156 }
3157 nmm->set_memory_at(ni, result);
3158 }
3159 }
3160 }
3161 }
3162 // Find the rest of instances values
3163 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3164 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3165 Node* result = step_through_mergemem(nmm, ni, tinst);
3166 if (result == nmm->base_memory()) {
3167 // Didn't find instance memory, search through general slice recursively.
3168 result = nmm->memory_at(_compile->get_general_index(ni));
3169 result = find_inst_mem(result, ni, orig_phis);
3170 if (_compile->failing()) {
3171 return;
3172 }
3173 nmm->set_memory_at(ni, result);
3174 }
3175 }
3176 igvn->hash_insert(nmm);
3177 record_for_optimizer(nmm);
3178 }
3180 // Phase 4: Update the inputs of non-instance memory Phis and
3181 // the Memory input of memnodes
3182 // First update the inputs of any non-instance Phi's from
3183 // which we split out an instance Phi. Note we don't have
3184 // to recursively process Phi's encounted on the input memory
3185 // chains as is done in split_memory_phi() since they will
3186 // also be processed here.
3187 for (int j = 0; j < orig_phis.length(); j++) {
3188 PhiNode *phi = orig_phis.at(j);
3189 int alias_idx = _compile->get_alias_index(phi->adr_type());
3190 igvn->hash_delete(phi);
3191 for (uint i = 1; i < phi->req(); i++) {
3192 Node *mem = phi->in(i);
3193 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3194 if (_compile->failing()) {
3195 return;
3196 }
3197 if (mem != new_mem) {
3198 phi->set_req(i, new_mem);
3199 }
3200 }
3201 igvn->hash_insert(phi);
3202 record_for_optimizer(phi);
3203 }
3205 // Update the memory inputs of MemNodes with the value we computed
3206 // in Phase 2 and move stores memory users to corresponding memory slices.
3207 // Disable memory split verification code until the fix for 6984348.
3208 // Currently it produces false negative results since it does not cover all cases.
3209 #if 0 // ifdef ASSERT
3210 visited.Reset();
3211 Node_Stack old_mems(arena, _compile->unique() >> 2);
3212 #endif
3213 for (uint i = 0; i < ideal_nodes.size(); i++) {
3214 Node* n = ideal_nodes.at(i);
3215 Node* nmem = get_map(n->_idx);
3216 assert(nmem != NULL, "sanity");
3217 if (n->is_Mem()) {
3218 #if 0 // ifdef ASSERT
3219 Node* old_mem = n->in(MemNode::Memory);
3220 if (!visited.test_set(old_mem->_idx)) {
3221 old_mems.push(old_mem, old_mem->outcnt());
3222 }
3223 #endif
3224 assert(n->in(MemNode::Memory) != nmem, "sanity");
3225 if (!n->is_Load()) {
3226 // Move memory users of a store first.
3227 move_inst_mem(n, orig_phis);
3228 }
3229 // Now update memory input
3230 igvn->hash_delete(n);
3231 n->set_req(MemNode::Memory, nmem);
3232 igvn->hash_insert(n);
3233 record_for_optimizer(n);
3234 } else {
3235 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3236 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3237 }
3238 }
3239 #if 0 // ifdef ASSERT
3240 // Verify that memory was split correctly
3241 while (old_mems.is_nonempty()) {
3242 Node* old_mem = old_mems.node();
3243 uint old_cnt = old_mems.index();
3244 old_mems.pop();
3245 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3246 }
3247 #endif
3248 }
3250 #ifndef PRODUCT
3251 static const char *node_type_names[] = {
3252 "UnknownType",
3253 "JavaObject",
3254 "LocalVar",
3255 "Field",
3256 "Arraycopy"
3257 };
3259 static const char *esc_names[] = {
3260 "UnknownEscape",
3261 "NoEscape",
3262 "ArgEscape",
3263 "GlobalEscape"
3264 };
3266 void PointsToNode::dump(bool print_state) const {
3267 NodeType nt = node_type();
3268 tty->print("%s ", node_type_names[(int) nt]);
3269 if (print_state) {
3270 EscapeState es = escape_state();
3271 EscapeState fields_es = fields_escape_state();
3272 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3273 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3274 tty->print("NSR ");
3275 }
3276 if (is_Field()) {
3277 FieldNode* f = (FieldNode*)this;
3278 if (f->is_oop())
3279 tty->print("oop ");
3280 if (f->offset() > 0)
3281 tty->print("+%d ", f->offset());
3282 tty->print("(");
3283 for (BaseIterator i(f); i.has_next(); i.next()) {
3284 PointsToNode* b = i.get();
3285 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3286 }
3287 tty->print(" )");
3288 }
3289 tty->print("[");
3290 for (EdgeIterator i(this); i.has_next(); i.next()) {
3291 PointsToNode* e = i.get();
3292 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3293 }
3294 tty->print(" [");
3295 for (UseIterator i(this); i.has_next(); i.next()) {
3296 PointsToNode* u = i.get();
3297 bool is_base = false;
3298 if (PointsToNode::is_base_use(u)) {
3299 is_base = true;
3300 u = PointsToNode::get_use_node(u)->as_Field();
3301 }
3302 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3303 }
3304 tty->print(" ]] ");
3305 if (_node == NULL)
3306 tty->print_cr("<null>");
3307 else
3308 _node->dump();
3309 }
3311 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3312 bool first = true;
3313 int ptnodes_length = ptnodes_worklist.length();
3314 for (int i = 0; i < ptnodes_length; i++) {
3315 PointsToNode *ptn = ptnodes_worklist.at(i);
3316 if (ptn == NULL || !ptn->is_JavaObject())
3317 continue;
3318 PointsToNode::EscapeState es = ptn->escape_state();
3319 if ((es != PointsToNode::NoEscape) && !Verbose) {
3320 continue;
3321 }
3322 Node* n = ptn->ideal_node();
3323 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3324 n->as_CallStaticJava()->is_boxing_method())) {
3325 if (first) {
3326 tty->cr();
3327 tty->print("======== Connection graph for ");
3328 _compile->method()->print_short_name();
3329 tty->cr();
3330 first = false;
3331 }
3332 ptn->dump();
3333 // Print all locals and fields which reference this allocation
3334 for (UseIterator j(ptn); j.has_next(); j.next()) {
3335 PointsToNode* use = j.get();
3336 if (use->is_LocalVar()) {
3337 use->dump(Verbose);
3338 } else if (Verbose) {
3339 use->dump();
3340 }
3341 }
3342 tty->cr();
3343 }
3344 }
3345 }
3346 #endif