Fri, 24 Oct 2014 10:28:19 -0700
8041984: CompilerThread seems to occupy all CPU in a very rare situation
Summary: Add new timeout checks to EA.
Reviewed-by: iveresov, drchase
1 /*
2 * Copyright (c) 2005, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.hpp"
30 #include "opto/c2compiler.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
38 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
39 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
40 _in_worklist(C->comp_arena()),
41 _next_pidx(0),
42 _collecting(true),
43 _verify(false),
44 _compile(C),
45 _igvn(igvn),
46 _node_map(C->comp_arena()) {
47 // Add unknown java object.
48 add_java_object(C->top(), PointsToNode::GlobalEscape);
49 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
50 // Add ConP(#NULL) and ConN(#NULL) nodes.
51 Node* oop_null = igvn->zerocon(T_OBJECT);
52 assert(oop_null->_idx < nodes_size(), "should be created already");
53 add_java_object(oop_null, PointsToNode::NoEscape);
54 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
55 if (UseCompressedOops) {
56 Node* noop_null = igvn->zerocon(T_NARROWOOP);
57 assert(noop_null->_idx < nodes_size(), "should be created already");
58 map_ideal_node(noop_null, null_obj);
59 }
60 _pcmp_neq = NULL; // Should be initialized
61 _pcmp_eq = NULL;
62 }
64 bool ConnectionGraph::has_candidates(Compile *C) {
65 // EA brings benefits only when the code has allocations and/or locks which
66 // are represented by ideal Macro nodes.
67 int cnt = C->macro_count();
68 for (int i = 0; i < cnt; i++) {
69 Node *n = C->macro_node(i);
70 if (n->is_Allocate())
71 return true;
72 if (n->is_Lock()) {
73 Node* obj = n->as_Lock()->obj_node()->uncast();
74 if (!(obj->is_Parm() || obj->is_Con()))
75 return true;
76 }
77 if (n->is_CallStaticJava() &&
78 n->as_CallStaticJava()->is_boxing_method()) {
79 return true;
80 }
81 }
82 return false;
83 }
85 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
86 Compile::TracePhase t2("escapeAnalysis", &Phase::_t_escapeAnalysis, true);
87 ResourceMark rm;
89 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
90 // to create space for them in ConnectionGraph::_nodes[].
91 Node* oop_null = igvn->zerocon(T_OBJECT);
92 Node* noop_null = igvn->zerocon(T_NARROWOOP);
93 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
94 // Perform escape analysis
95 if (congraph->compute_escape()) {
96 // There are non escaping objects.
97 C->set_congraph(congraph);
98 }
99 // Cleanup.
100 if (oop_null->outcnt() == 0)
101 igvn->hash_delete(oop_null);
102 if (noop_null->outcnt() == 0)
103 igvn->hash_delete(noop_null);
104 }
106 bool ConnectionGraph::compute_escape() {
107 Compile* C = _compile;
108 PhaseGVN* igvn = _igvn;
110 // Worklists used by EA.
111 Unique_Node_List delayed_worklist;
112 GrowableArray<Node*> alloc_worklist;
113 GrowableArray<Node*> ptr_cmp_worklist;
114 GrowableArray<Node*> storestore_worklist;
115 GrowableArray<PointsToNode*> ptnodes_worklist;
116 GrowableArray<JavaObjectNode*> java_objects_worklist;
117 GrowableArray<JavaObjectNode*> non_escaped_worklist;
118 GrowableArray<FieldNode*> oop_fields_worklist;
119 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
121 { Compile::TracePhase t3("connectionGraph", &Phase::_t_connectionGraph, true);
123 // 1. Populate Connection Graph (CG) with PointsTo nodes.
124 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
125 // Initialize worklist
126 if (C->root() != NULL) {
127 ideal_nodes.push(C->root());
128 }
129 // Processed ideal nodes are unique on ideal_nodes list
130 // but several ideal nodes are mapped to the phantom_obj.
131 // To avoid duplicated entries on the following worklists
132 // add the phantom_obj only once to them.
133 ptnodes_worklist.append(phantom_obj);
134 java_objects_worklist.append(phantom_obj);
135 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
136 Node* n = ideal_nodes.at(next);
137 // Create PointsTo nodes and add them to Connection Graph. Called
138 // only once per ideal node since ideal_nodes is Unique_Node list.
139 add_node_to_connection_graph(n, &delayed_worklist);
140 PointsToNode* ptn = ptnode_adr(n->_idx);
141 if (ptn != NULL && ptn != phantom_obj) {
142 ptnodes_worklist.append(ptn);
143 if (ptn->is_JavaObject()) {
144 java_objects_worklist.append(ptn->as_JavaObject());
145 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
146 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
147 // Only allocations and java static calls results are interesting.
148 non_escaped_worklist.append(ptn->as_JavaObject());
149 }
150 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
151 oop_fields_worklist.append(ptn->as_Field());
152 }
153 }
154 if (n->is_MergeMem()) {
155 // Collect all MergeMem nodes to add memory slices for
156 // scalar replaceable objects in split_unique_types().
157 _mergemem_worklist.append(n->as_MergeMem());
158 } else if (OptimizePtrCompare && n->is_Cmp() &&
159 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
160 // Collect compare pointers nodes.
161 ptr_cmp_worklist.append(n);
162 } else if (n->is_MemBarStoreStore()) {
163 // Collect all MemBarStoreStore nodes so that depending on the
164 // escape status of the associated Allocate node some of them
165 // may be eliminated.
166 storestore_worklist.append(n);
167 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
168 (n->req() > MemBarNode::Precedent)) {
169 record_for_optimizer(n);
170 #ifdef ASSERT
171 } else if (n->is_AddP()) {
172 // Collect address nodes for graph verification.
173 addp_worklist.append(n);
174 #endif
175 }
176 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
177 Node* m = n->fast_out(i); // Get user
178 ideal_nodes.push(m);
179 }
180 }
181 if (non_escaped_worklist.length() == 0) {
182 _collecting = false;
183 return false; // Nothing to do.
184 }
185 // Add final simple edges to graph.
186 while(delayed_worklist.size() > 0) {
187 Node* n = delayed_worklist.pop();
188 add_final_edges(n);
189 }
190 int ptnodes_length = ptnodes_worklist.length();
192 #ifdef ASSERT
193 if (VerifyConnectionGraph) {
194 // Verify that no new simple edges could be created and all
195 // local vars has edges.
196 _verify = true;
197 for (int next = 0; next < ptnodes_length; ++next) {
198 PointsToNode* ptn = ptnodes_worklist.at(next);
199 add_final_edges(ptn->ideal_node());
200 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
201 ptn->dump();
202 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
203 }
204 }
205 _verify = false;
206 }
207 #endif
209 // 2. Finish Graph construction by propagating references to all
210 // java objects through graph.
211 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
212 java_objects_worklist, oop_fields_worklist)) {
213 // All objects escaped or hit time or iterations limits.
214 _collecting = false;
215 return false;
216 }
218 // 3. Adjust scalar_replaceable state of nonescaping objects and push
219 // scalar replaceable allocations on alloc_worklist for processing
220 // in split_unique_types().
221 int non_escaped_length = non_escaped_worklist.length();
222 for (int next = 0; next < non_escaped_length; next++) {
223 JavaObjectNode* ptn = non_escaped_worklist.at(next);
224 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
225 Node* n = ptn->ideal_node();
226 if (n->is_Allocate()) {
227 n->as_Allocate()->_is_non_escaping = noescape;
228 }
229 if (n->is_CallStaticJava()) {
230 n->as_CallStaticJava()->_is_non_escaping = noescape;
231 }
232 if (noescape && ptn->scalar_replaceable()) {
233 adjust_scalar_replaceable_state(ptn);
234 if (ptn->scalar_replaceable()) {
235 alloc_worklist.append(ptn->ideal_node());
236 }
237 }
238 }
240 #ifdef ASSERT
241 if (VerifyConnectionGraph) {
242 // Verify that graph is complete - no new edges could be added or needed.
243 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
244 java_objects_worklist, addp_worklist);
245 }
246 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
247 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
248 null_obj->edge_count() == 0 &&
249 !null_obj->arraycopy_src() &&
250 !null_obj->arraycopy_dst(), "sanity");
251 #endif
253 _collecting = false;
255 } // TracePhase t3("connectionGraph")
257 // 4. Optimize ideal graph based on EA information.
258 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
259 if (has_non_escaping_obj) {
260 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
261 }
263 #ifndef PRODUCT
264 if (PrintEscapeAnalysis) {
265 dump(ptnodes_worklist); // Dump ConnectionGraph
266 }
267 #endif
269 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
270 #ifdef ASSERT
271 if (VerifyConnectionGraph) {
272 int alloc_length = alloc_worklist.length();
273 for (int next = 0; next < alloc_length; ++next) {
274 Node* n = alloc_worklist.at(next);
275 PointsToNode* ptn = ptnode_adr(n->_idx);
276 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
277 }
278 }
279 #endif
281 // 5. Separate memory graph for scalar replaceable allcations.
282 if (has_scalar_replaceable_candidates &&
283 C->AliasLevel() >= 3 && EliminateAllocations) {
284 // Now use the escape information to create unique types for
285 // scalar replaceable objects.
286 split_unique_types(alloc_worklist);
287 if (C->failing()) return false;
288 C->print_method(PHASE_AFTER_EA, 2);
290 #ifdef ASSERT
291 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
292 tty->print("=== No allocations eliminated for ");
293 C->method()->print_short_name();
294 if(!EliminateAllocations) {
295 tty->print(" since EliminateAllocations is off ===");
296 } else if(!has_scalar_replaceable_candidates) {
297 tty->print(" since there are no scalar replaceable candidates ===");
298 } else if(C->AliasLevel() < 3) {
299 tty->print(" since AliasLevel < 3 ===");
300 }
301 tty->cr();
302 #endif
303 }
304 return has_non_escaping_obj;
305 }
307 // Utility function for nodes that load an object
308 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
309 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
310 // ThreadLocal has RawPtr type.
311 const Type* t = _igvn->type(n);
312 if (t->make_ptr() != NULL) {
313 Node* adr = n->in(MemNode::Address);
314 #ifdef ASSERT
315 if (!adr->is_AddP()) {
316 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
317 } else {
318 assert((ptnode_adr(adr->_idx) == NULL ||
319 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
320 }
321 #endif
322 add_local_var_and_edge(n, PointsToNode::NoEscape,
323 adr, delayed_worklist);
324 }
325 }
327 // Populate Connection Graph with PointsTo nodes and create simple
328 // connection graph edges.
329 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
330 assert(!_verify, "this method sould not be called for verification");
331 PhaseGVN* igvn = _igvn;
332 uint n_idx = n->_idx;
333 PointsToNode* n_ptn = ptnode_adr(n_idx);
334 if (n_ptn != NULL)
335 return; // No need to redefine PointsTo node during first iteration.
337 if (n->is_Call()) {
338 // Arguments to allocation and locking don't escape.
339 if (n->is_AbstractLock()) {
340 // Put Lock and Unlock nodes on IGVN worklist to process them during
341 // first IGVN optimization when escape information is still available.
342 record_for_optimizer(n);
343 } else if (n->is_Allocate()) {
344 add_call_node(n->as_Call());
345 record_for_optimizer(n);
346 } else {
347 if (n->is_CallStaticJava()) {
348 const char* name = n->as_CallStaticJava()->_name;
349 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
350 return; // Skip uncommon traps
351 }
352 // Don't mark as processed since call's arguments have to be processed.
353 delayed_worklist->push(n);
354 // Check if a call returns an object.
355 if ((n->as_Call()->returns_pointer() &&
356 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) ||
357 (n->is_CallStaticJava() &&
358 n->as_CallStaticJava()->is_boxing_method())) {
359 add_call_node(n->as_Call());
360 }
361 }
362 return;
363 }
364 // Put this check here to process call arguments since some call nodes
365 // point to phantom_obj.
366 if (n_ptn == phantom_obj || n_ptn == null_obj)
367 return; // Skip predefined nodes.
369 int opcode = n->Opcode();
370 switch (opcode) {
371 case Op_AddP: {
372 Node* base = get_addp_base(n);
373 PointsToNode* ptn_base = ptnode_adr(base->_idx);
374 // Field nodes are created for all field types. They are used in
375 // adjust_scalar_replaceable_state() and split_unique_types().
376 // Note, non-oop fields will have only base edges in Connection
377 // Graph because such fields are not used for oop loads and stores.
378 int offset = address_offset(n, igvn);
379 add_field(n, PointsToNode::NoEscape, offset);
380 if (ptn_base == NULL) {
381 delayed_worklist->push(n); // Process it later.
382 } else {
383 n_ptn = ptnode_adr(n_idx);
384 add_base(n_ptn->as_Field(), ptn_base);
385 }
386 break;
387 }
388 case Op_CastX2P: {
389 map_ideal_node(n, phantom_obj);
390 break;
391 }
392 case Op_CastPP:
393 case Op_CheckCastPP:
394 case Op_EncodeP:
395 case Op_DecodeN:
396 case Op_EncodePKlass:
397 case Op_DecodeNKlass: {
398 add_local_var_and_edge(n, PointsToNode::NoEscape,
399 n->in(1), delayed_worklist);
400 break;
401 }
402 case Op_CMoveP: {
403 add_local_var(n, PointsToNode::NoEscape);
404 // Do not add edges during first iteration because some could be
405 // not defined yet.
406 delayed_worklist->push(n);
407 break;
408 }
409 case Op_ConP:
410 case Op_ConN:
411 case Op_ConNKlass: {
412 // assume all oop constants globally escape except for null
413 PointsToNode::EscapeState es;
414 const Type* t = igvn->type(n);
415 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
416 es = PointsToNode::NoEscape;
417 } else {
418 es = PointsToNode::GlobalEscape;
419 }
420 add_java_object(n, es);
421 break;
422 }
423 case Op_CreateEx: {
424 // assume that all exception objects globally escape
425 map_ideal_node(n, phantom_obj);
426 break;
427 }
428 case Op_LoadKlass:
429 case Op_LoadNKlass: {
430 // Unknown class is loaded
431 map_ideal_node(n, phantom_obj);
432 break;
433 }
434 case Op_LoadP:
435 case Op_LoadN:
436 case Op_LoadPLocked: {
437 add_objload_to_connection_graph(n, delayed_worklist);
438 break;
439 }
440 case Op_Parm: {
441 map_ideal_node(n, phantom_obj);
442 break;
443 }
444 case Op_PartialSubtypeCheck: {
445 // Produces Null or notNull and is used in only in CmpP so
446 // phantom_obj could be used.
447 map_ideal_node(n, phantom_obj); // Result is unknown
448 break;
449 }
450 case Op_Phi: {
451 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
452 // ThreadLocal has RawPtr type.
453 const Type* t = n->as_Phi()->type();
454 if (t->make_ptr() != NULL) {
455 add_local_var(n, PointsToNode::NoEscape);
456 // Do not add edges during first iteration because some could be
457 // not defined yet.
458 delayed_worklist->push(n);
459 }
460 break;
461 }
462 case Op_Proj: {
463 // we are only interested in the oop result projection from a call
464 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
465 n->in(0)->as_Call()->returns_pointer()) {
466 add_local_var_and_edge(n, PointsToNode::NoEscape,
467 n->in(0), delayed_worklist);
468 }
469 break;
470 }
471 case Op_Rethrow: // Exception object escapes
472 case Op_Return: {
473 if (n->req() > TypeFunc::Parms &&
474 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
475 // Treat Return value as LocalVar with GlobalEscape escape state.
476 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
477 n->in(TypeFunc::Parms), delayed_worklist);
478 }
479 break;
480 }
481 case Op_GetAndSetP:
482 case Op_GetAndSetN: {
483 add_objload_to_connection_graph(n, delayed_worklist);
484 // fallthrough
485 }
486 case Op_StoreP:
487 case Op_StoreN:
488 case Op_StoreNKlass:
489 case Op_StorePConditional:
490 case Op_CompareAndSwapP:
491 case Op_CompareAndSwapN: {
492 Node* adr = n->in(MemNode::Address);
493 const Type *adr_type = igvn->type(adr);
494 adr_type = adr_type->make_ptr();
495 if (adr_type == NULL) {
496 break; // skip dead nodes
497 }
498 if (adr_type->isa_oopptr() ||
499 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
500 (adr_type == TypeRawPtr::NOTNULL &&
501 adr->in(AddPNode::Address)->is_Proj() &&
502 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
503 delayed_worklist->push(n); // Process it later.
504 #ifdef ASSERT
505 assert(adr->is_AddP(), "expecting an AddP");
506 if (adr_type == TypeRawPtr::NOTNULL) {
507 // Verify a raw address for a store captured by Initialize node.
508 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
509 assert(offs != Type::OffsetBot, "offset must be a constant");
510 }
511 #endif
512 } else {
513 // Ignore copy the displaced header to the BoxNode (OSR compilation).
514 if (adr->is_BoxLock())
515 break;
516 // Stored value escapes in unsafe access.
517 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
518 // Pointer stores in G1 barriers looks like unsafe access.
519 // Ignore such stores to be able scalar replace non-escaping
520 // allocations.
521 if (UseG1GC && adr->is_AddP()) {
522 Node* base = get_addp_base(adr);
523 if (base->Opcode() == Op_LoadP &&
524 base->in(MemNode::Address)->is_AddP()) {
525 adr = base->in(MemNode::Address);
526 Node* tls = get_addp_base(adr);
527 if (tls->Opcode() == Op_ThreadLocal) {
528 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
529 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() +
530 PtrQueue::byte_offset_of_buf())) {
531 break; // G1 pre barier previous oop value store.
532 }
533 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() +
534 PtrQueue::byte_offset_of_buf())) {
535 break; // G1 post barier card address store.
536 }
537 }
538 }
539 }
540 delayed_worklist->push(n); // Process unsafe access later.
541 break;
542 }
543 #ifdef ASSERT
544 n->dump(1);
545 assert(false, "not unsafe or G1 barrier raw StoreP");
546 #endif
547 }
548 break;
549 }
550 case Op_AryEq:
551 case Op_StrComp:
552 case Op_StrEquals:
553 case Op_StrIndexOf:
554 case Op_EncodeISOArray: {
555 add_local_var(n, PointsToNode::ArgEscape);
556 delayed_worklist->push(n); // Process it later.
557 break;
558 }
559 case Op_ThreadLocal: {
560 add_java_object(n, PointsToNode::ArgEscape);
561 break;
562 }
563 default:
564 ; // Do nothing for nodes not related to EA.
565 }
566 return;
567 }
569 #ifdef ASSERT
570 #define ELSE_FAIL(name) \
571 /* Should not be called for not pointer type. */ \
572 n->dump(1); \
573 assert(false, name); \
574 break;
575 #else
576 #define ELSE_FAIL(name) \
577 break;
578 #endif
580 // Add final simple edges to graph.
581 void ConnectionGraph::add_final_edges(Node *n) {
582 PointsToNode* n_ptn = ptnode_adr(n->_idx);
583 #ifdef ASSERT
584 if (_verify && n_ptn->is_JavaObject())
585 return; // This method does not change graph for JavaObject.
586 #endif
588 if (n->is_Call()) {
589 process_call_arguments(n->as_Call());
590 return;
591 }
592 assert(n->is_Store() || n->is_LoadStore() ||
593 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
594 "node should be registered already");
595 int opcode = n->Opcode();
596 switch (opcode) {
597 case Op_AddP: {
598 Node* base = get_addp_base(n);
599 PointsToNode* ptn_base = ptnode_adr(base->_idx);
600 assert(ptn_base != NULL, "field's base should be registered");
601 add_base(n_ptn->as_Field(), ptn_base);
602 break;
603 }
604 case Op_CastPP:
605 case Op_CheckCastPP:
606 case Op_EncodeP:
607 case Op_DecodeN:
608 case Op_EncodePKlass:
609 case Op_DecodeNKlass: {
610 add_local_var_and_edge(n, PointsToNode::NoEscape,
611 n->in(1), NULL);
612 break;
613 }
614 case Op_CMoveP: {
615 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
616 Node* in = n->in(i);
617 if (in == NULL)
618 continue; // ignore NULL
619 Node* uncast_in = in->uncast();
620 if (uncast_in->is_top() || uncast_in == n)
621 continue; // ignore top or inputs which go back this node
622 PointsToNode* ptn = ptnode_adr(in->_idx);
623 assert(ptn != NULL, "node should be registered");
624 add_edge(n_ptn, ptn);
625 }
626 break;
627 }
628 case Op_LoadP:
629 case Op_LoadN:
630 case Op_LoadPLocked: {
631 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
632 // ThreadLocal has RawPtr type.
633 const Type* t = _igvn->type(n);
634 if (t->make_ptr() != NULL) {
635 Node* adr = n->in(MemNode::Address);
636 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
637 break;
638 }
639 ELSE_FAIL("Op_LoadP");
640 }
641 case Op_Phi: {
642 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
643 // ThreadLocal has RawPtr type.
644 const Type* t = n->as_Phi()->type();
645 if (t->make_ptr() != NULL) {
646 for (uint i = 1; i < n->req(); i++) {
647 Node* in = n->in(i);
648 if (in == NULL)
649 continue; // ignore NULL
650 Node* uncast_in = in->uncast();
651 if (uncast_in->is_top() || uncast_in == n)
652 continue; // ignore top or inputs which go back this node
653 PointsToNode* ptn = ptnode_adr(in->_idx);
654 assert(ptn != NULL, "node should be registered");
655 add_edge(n_ptn, ptn);
656 }
657 break;
658 }
659 ELSE_FAIL("Op_Phi");
660 }
661 case Op_Proj: {
662 // we are only interested in the oop result projection from a call
663 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
664 n->in(0)->as_Call()->returns_pointer()) {
665 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
666 break;
667 }
668 ELSE_FAIL("Op_Proj");
669 }
670 case Op_Rethrow: // Exception object escapes
671 case Op_Return: {
672 if (n->req() > TypeFunc::Parms &&
673 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
674 // Treat Return value as LocalVar with GlobalEscape escape state.
675 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
676 n->in(TypeFunc::Parms), NULL);
677 break;
678 }
679 ELSE_FAIL("Op_Return");
680 }
681 case Op_StoreP:
682 case Op_StoreN:
683 case Op_StoreNKlass:
684 case Op_StorePConditional:
685 case Op_CompareAndSwapP:
686 case Op_CompareAndSwapN:
687 case Op_GetAndSetP:
688 case Op_GetAndSetN: {
689 Node* adr = n->in(MemNode::Address);
690 const Type *adr_type = _igvn->type(adr);
691 adr_type = adr_type->make_ptr();
692 #ifdef ASSERT
693 if (adr_type == NULL) {
694 n->dump(1);
695 assert(adr_type != NULL, "dead node should not be on list");
696 break;
697 }
698 #endif
699 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) {
700 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
701 }
702 if (adr_type->isa_oopptr() ||
703 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
704 (adr_type == TypeRawPtr::NOTNULL &&
705 adr->in(AddPNode::Address)->is_Proj() &&
706 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
707 // Point Address to Value
708 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
709 assert(adr_ptn != NULL &&
710 adr_ptn->as_Field()->is_oop(), "node should be registered");
711 Node *val = n->in(MemNode::ValueIn);
712 PointsToNode* ptn = ptnode_adr(val->_idx);
713 assert(ptn != NULL, "node should be registered");
714 add_edge(adr_ptn, ptn);
715 break;
716 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
717 // Stored value escapes in unsafe access.
718 Node *val = n->in(MemNode::ValueIn);
719 PointsToNode* ptn = ptnode_adr(val->_idx);
720 assert(ptn != NULL, "node should be registered");
721 set_escape_state(ptn, PointsToNode::GlobalEscape);
722 // Add edge to object for unsafe access with offset.
723 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
724 assert(adr_ptn != NULL, "node should be registered");
725 if (adr_ptn->is_Field()) {
726 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
727 add_edge(adr_ptn, ptn);
728 }
729 break;
730 }
731 ELSE_FAIL("Op_StoreP");
732 }
733 case Op_AryEq:
734 case Op_StrComp:
735 case Op_StrEquals:
736 case Op_StrIndexOf:
737 case Op_EncodeISOArray: {
738 // char[] arrays passed to string intrinsic do not escape but
739 // they are not scalar replaceable. Adjust escape state for them.
740 // Start from in(2) edge since in(1) is memory edge.
741 for (uint i = 2; i < n->req(); i++) {
742 Node* adr = n->in(i);
743 const Type* at = _igvn->type(adr);
744 if (!adr->is_top() && at->isa_ptr()) {
745 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
746 at->isa_ptr() != NULL, "expecting a pointer");
747 if (adr->is_AddP()) {
748 adr = get_addp_base(adr);
749 }
750 PointsToNode* ptn = ptnode_adr(adr->_idx);
751 assert(ptn != NULL, "node should be registered");
752 add_edge(n_ptn, ptn);
753 }
754 }
755 break;
756 }
757 default: {
758 // This method should be called only for EA specific nodes which may
759 // miss some edges when they were created.
760 #ifdef ASSERT
761 n->dump(1);
762 #endif
763 guarantee(false, "unknown node");
764 }
765 }
766 return;
767 }
769 void ConnectionGraph::add_call_node(CallNode* call) {
770 assert(call->returns_pointer(), "only for call which returns pointer");
771 uint call_idx = call->_idx;
772 if (call->is_Allocate()) {
773 Node* k = call->in(AllocateNode::KlassNode);
774 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
775 assert(kt != NULL, "TypeKlassPtr required.");
776 ciKlass* cik = kt->klass();
777 PointsToNode::EscapeState es = PointsToNode::NoEscape;
778 bool scalar_replaceable = true;
779 if (call->is_AllocateArray()) {
780 if (!cik->is_array_klass()) { // StressReflectiveCode
781 es = PointsToNode::GlobalEscape;
782 } else {
783 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
784 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
785 // Not scalar replaceable if the length is not constant or too big.
786 scalar_replaceable = false;
787 }
788 }
789 } else { // Allocate instance
790 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
791 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
792 !cik->is_instance_klass() || // StressReflectiveCode
793 cik->as_instance_klass()->has_finalizer()) {
794 es = PointsToNode::GlobalEscape;
795 }
796 }
797 add_java_object(call, es);
798 PointsToNode* ptn = ptnode_adr(call_idx);
799 if (!scalar_replaceable && ptn->scalar_replaceable()) {
800 ptn->set_scalar_replaceable(false);
801 }
802 } else if (call->is_CallStaticJava()) {
803 // Call nodes could be different types:
804 //
805 // 1. CallDynamicJavaNode (what happened during call is unknown):
806 //
807 // - mapped to GlobalEscape JavaObject node if oop is returned;
808 //
809 // - all oop arguments are escaping globally;
810 //
811 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
812 //
813 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
814 //
815 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
816 // - mapped to NoEscape JavaObject node if non-escaping object allocated
817 // during call is returned;
818 // - mapped to ArgEscape LocalVar node pointed to object arguments
819 // which are returned and does not escape during call;
820 //
821 // - oop arguments escaping status is defined by bytecode analysis;
822 //
823 // For a static call, we know exactly what method is being called.
824 // Use bytecode estimator to record whether the call's return value escapes.
825 ciMethod* meth = call->as_CallJava()->method();
826 if (meth == NULL) {
827 const char* name = call->as_CallStaticJava()->_name;
828 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
829 // Returns a newly allocated unescaped object.
830 add_java_object(call, PointsToNode::NoEscape);
831 ptnode_adr(call_idx)->set_scalar_replaceable(false);
832 } else if (meth->is_boxing_method()) {
833 // Returns boxing object
834 PointsToNode::EscapeState es;
835 vmIntrinsics::ID intr = meth->intrinsic_id();
836 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
837 // It does not escape if object is always allocated.
838 es = PointsToNode::NoEscape;
839 } else {
840 // It escapes globally if object could be loaded from cache.
841 es = PointsToNode::GlobalEscape;
842 }
843 add_java_object(call, es);
844 } else {
845 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
846 call_analyzer->copy_dependencies(_compile->dependencies());
847 if (call_analyzer->is_return_allocated()) {
848 // Returns a newly allocated unescaped object, simply
849 // update dependency information.
850 // Mark it as NoEscape so that objects referenced by
851 // it's fields will be marked as NoEscape at least.
852 add_java_object(call, PointsToNode::NoEscape);
853 ptnode_adr(call_idx)->set_scalar_replaceable(false);
854 } else {
855 // Determine whether any arguments are returned.
856 const TypeTuple* d = call->tf()->domain();
857 bool ret_arg = false;
858 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
859 if (d->field_at(i)->isa_ptr() != NULL &&
860 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
861 ret_arg = true;
862 break;
863 }
864 }
865 if (ret_arg) {
866 add_local_var(call, PointsToNode::ArgEscape);
867 } else {
868 // Returns unknown object.
869 map_ideal_node(call, phantom_obj);
870 }
871 }
872 }
873 } else {
874 // An other type of call, assume the worst case:
875 // returned value is unknown and globally escapes.
876 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
877 map_ideal_node(call, phantom_obj);
878 }
879 }
881 void ConnectionGraph::process_call_arguments(CallNode *call) {
882 bool is_arraycopy = false;
883 switch (call->Opcode()) {
884 #ifdef ASSERT
885 case Op_Allocate:
886 case Op_AllocateArray:
887 case Op_Lock:
888 case Op_Unlock:
889 assert(false, "should be done already");
890 break;
891 #endif
892 case Op_CallLeafNoFP:
893 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
894 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
895 // fall through
896 case Op_CallLeaf: {
897 // Stub calls, objects do not escape but they are not scale replaceable.
898 // Adjust escape state for outgoing arguments.
899 const TypeTuple * d = call->tf()->domain();
900 bool src_has_oops = false;
901 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
902 const Type* at = d->field_at(i);
903 Node *arg = call->in(i);
904 const Type *aat = _igvn->type(arg);
905 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
906 continue;
907 if (arg->is_AddP()) {
908 //
909 // The inline_native_clone() case when the arraycopy stub is called
910 // after the allocation before Initialize and CheckCastPP nodes.
911 // Or normal arraycopy for object arrays case.
912 //
913 // Set AddP's base (Allocate) as not scalar replaceable since
914 // pointer to the base (with offset) is passed as argument.
915 //
916 arg = get_addp_base(arg);
917 }
918 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
919 assert(arg_ptn != NULL, "should be registered");
920 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
921 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
922 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
923 aat->isa_ptr() != NULL, "expecting an Ptr");
924 bool arg_has_oops = aat->isa_oopptr() &&
925 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
926 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
927 if (i == TypeFunc::Parms) {
928 src_has_oops = arg_has_oops;
929 }
930 //
931 // src or dst could be j.l.Object when other is basic type array:
932 //
933 // arraycopy(char[],0,Object*,0,size);
934 // arraycopy(Object*,0,char[],0,size);
935 //
936 // Don't add edges in such cases.
937 //
938 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
939 arg_has_oops && (i > TypeFunc::Parms);
940 #ifdef ASSERT
941 if (!(is_arraycopy ||
942 (call->as_CallLeaf()->_name != NULL &&
943 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
944 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
945 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
946 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
947 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
948 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
949 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
950 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
951 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
952 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
953 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
954 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
955 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
956 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0)
957 ))) {
958 call->dump();
959 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
960 }
961 #endif
962 // Always process arraycopy's destination object since
963 // we need to add all possible edges to references in
964 // source object.
965 if (arg_esc >= PointsToNode::ArgEscape &&
966 !arg_is_arraycopy_dest) {
967 continue;
968 }
969 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
970 if (arg_is_arraycopy_dest) {
971 Node* src = call->in(TypeFunc::Parms);
972 if (src->is_AddP()) {
973 src = get_addp_base(src);
974 }
975 PointsToNode* src_ptn = ptnode_adr(src->_idx);
976 assert(src_ptn != NULL, "should be registered");
977 if (arg_ptn != src_ptn) {
978 // Special arraycopy edge:
979 // A destination object's field can't have the source object
980 // as base since objects escape states are not related.
981 // Only escape state of destination object's fields affects
982 // escape state of fields in source object.
983 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
984 }
985 }
986 }
987 }
988 break;
989 }
990 case Op_CallStaticJava: {
991 // For a static call, we know exactly what method is being called.
992 // Use bytecode estimator to record the call's escape affects
993 #ifdef ASSERT
994 const char* name = call->as_CallStaticJava()->_name;
995 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
996 #endif
997 ciMethod* meth = call->as_CallJava()->method();
998 if ((meth != NULL) && meth->is_boxing_method()) {
999 break; // Boxing methods do not modify any oops.
1000 }
1001 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1002 // fall-through if not a Java method or no analyzer information
1003 if (call_analyzer != NULL) {
1004 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1005 const TypeTuple* d = call->tf()->domain();
1006 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1007 const Type* at = d->field_at(i);
1008 int k = i - TypeFunc::Parms;
1009 Node* arg = call->in(i);
1010 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1011 if (at->isa_ptr() != NULL &&
1012 call_analyzer->is_arg_returned(k)) {
1013 // The call returns arguments.
1014 if (call_ptn != NULL) { // Is call's result used?
1015 assert(call_ptn->is_LocalVar(), "node should be registered");
1016 assert(arg_ptn != NULL, "node should be registered");
1017 add_edge(call_ptn, arg_ptn);
1018 }
1019 }
1020 if (at->isa_oopptr() != NULL &&
1021 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1022 if (!call_analyzer->is_arg_stack(k)) {
1023 // The argument global escapes
1024 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1025 } else {
1026 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1027 if (!call_analyzer->is_arg_local(k)) {
1028 // The argument itself doesn't escape, but any fields might
1029 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1030 }
1031 }
1032 }
1033 }
1034 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1035 // The call returns arguments.
1036 assert(call_ptn->edge_count() > 0, "sanity");
1037 if (!call_analyzer->is_return_local()) {
1038 // Returns also unknown object.
1039 add_edge(call_ptn, phantom_obj);
1040 }
1041 }
1042 break;
1043 }
1044 }
1045 default: {
1046 // Fall-through here if not a Java method or no analyzer information
1047 // or some other type of call, assume the worst case: all arguments
1048 // globally escape.
1049 const TypeTuple* d = call->tf()->domain();
1050 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1051 const Type* at = d->field_at(i);
1052 if (at->isa_oopptr() != NULL) {
1053 Node* arg = call->in(i);
1054 if (arg->is_AddP()) {
1055 arg = get_addp_base(arg);
1056 }
1057 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1058 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1059 }
1060 }
1061 }
1062 }
1063 }
1066 // Finish Graph construction.
1067 bool ConnectionGraph::complete_connection_graph(
1068 GrowableArray<PointsToNode*>& ptnodes_worklist,
1069 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1070 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1071 GrowableArray<FieldNode*>& oop_fields_worklist) {
1072 // Normally only 1-3 passes needed to build Connection Graph depending
1073 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1074 // Set limit to 20 to catch situation when something did go wrong and
1075 // bailout Escape Analysis.
1076 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1077 #define CG_BUILD_ITER_LIMIT 20
1079 // Propagate GlobalEscape and ArgEscape escape states and check that
1080 // we still have non-escaping objects. The method pushs on _worklist
1081 // Field nodes which reference phantom_object.
1082 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1083 return false; // Nothing to do.
1084 }
1085 // Now propagate references to all JavaObject nodes.
1086 int java_objects_length = java_objects_worklist.length();
1087 elapsedTimer time;
1088 bool timeout = false;
1089 int new_edges = 1;
1090 int iterations = 0;
1091 do {
1092 while ((new_edges > 0) &&
1093 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1094 double start_time = time.seconds();
1095 time.start();
1096 new_edges = 0;
1097 // Propagate references to phantom_object for nodes pushed on _worklist
1098 // by find_non_escaped_objects() and find_field_value().
1099 new_edges += add_java_object_edges(phantom_obj, false);
1100 for (int next = 0; next < java_objects_length; ++next) {
1101 JavaObjectNode* ptn = java_objects_worklist.at(next);
1102 new_edges += add_java_object_edges(ptn, true);
1104 #define SAMPLE_SIZE 4
1105 if ((next % SAMPLE_SIZE) == 0) {
1106 // Each 4 iterations calculate how much time it will take
1107 // to complete graph construction.
1108 time.stop();
1109 double stop_time = time.seconds();
1110 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1111 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1112 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1113 timeout = true;
1114 break; // Timeout
1115 }
1116 start_time = stop_time;
1117 time.start();
1118 }
1119 #undef SAMPLE_SIZE
1121 }
1122 if (timeout) break;
1123 if (new_edges > 0) {
1124 // Update escape states on each iteration if graph was updated.
1125 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1126 return false; // Nothing to do.
1127 }
1128 }
1129 time.stop();
1130 if (time.seconds() >= EscapeAnalysisTimeout) {
1131 timeout = true;
1132 break;
1133 }
1134 }
1135 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1136 time.start();
1137 // Find fields which have unknown value.
1138 int fields_length = oop_fields_worklist.length();
1139 for (int next = 0; next < fields_length; next++) {
1140 FieldNode* field = oop_fields_worklist.at(next);
1141 if (field->edge_count() == 0) {
1142 new_edges += find_field_value(field);
1143 // This code may added new edges to phantom_object.
1144 // Need an other cycle to propagate references to phantom_object.
1145 }
1146 }
1147 time.stop();
1148 if (time.seconds() >= EscapeAnalysisTimeout) {
1149 timeout = true;
1150 break;
1151 }
1152 } else {
1153 new_edges = 0; // Bailout
1154 }
1155 } while (new_edges > 0);
1157 // Bailout if passed limits.
1158 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1159 Compile* C = _compile;
1160 if (C->log() != NULL) {
1161 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1162 C->log()->text("%s", timeout ? "time" : "iterations");
1163 C->log()->end_elem(" limit'");
1164 }
1165 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1166 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1167 // Possible infinite build_connection_graph loop,
1168 // bailout (no changes to ideal graph were made).
1169 return false;
1170 }
1171 #ifdef ASSERT
1172 if (Verbose && PrintEscapeAnalysis) {
1173 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1174 iterations, nodes_size(), ptnodes_worklist.length());
1175 }
1176 #endif
1178 #undef CG_BUILD_ITER_LIMIT
1180 // Find fields initialized by NULL for non-escaping Allocations.
1181 int non_escaped_length = non_escaped_worklist.length();
1182 for (int next = 0; next < non_escaped_length; next++) {
1183 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1184 PointsToNode::EscapeState es = ptn->escape_state();
1185 assert(es <= PointsToNode::ArgEscape, "sanity");
1186 if (es == PointsToNode::NoEscape) {
1187 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1188 // Adding references to NULL object does not change escape states
1189 // since it does not escape. Also no fields are added to NULL object.
1190 add_java_object_edges(null_obj, false);
1191 }
1192 }
1193 Node* n = ptn->ideal_node();
1194 if (n->is_Allocate()) {
1195 // The object allocated by this Allocate node will never be
1196 // seen by an other thread. Mark it so that when it is
1197 // expanded no MemBarStoreStore is added.
1198 InitializeNode* ini = n->as_Allocate()->initialization();
1199 if (ini != NULL)
1200 ini->set_does_not_escape();
1201 }
1202 }
1203 return true; // Finished graph construction.
1204 }
1206 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1207 // and check that we still have non-escaping java objects.
1208 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1209 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1210 GrowableArray<PointsToNode*> escape_worklist;
1211 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1212 int ptnodes_length = ptnodes_worklist.length();
1213 for (int next = 0; next < ptnodes_length; ++next) {
1214 PointsToNode* ptn = ptnodes_worklist.at(next);
1215 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1216 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1217 escape_worklist.push(ptn);
1218 }
1219 }
1220 // Set escape states to referenced nodes (edges list).
1221 while (escape_worklist.length() > 0) {
1222 PointsToNode* ptn = escape_worklist.pop();
1223 PointsToNode::EscapeState es = ptn->escape_state();
1224 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1225 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1226 es >= PointsToNode::ArgEscape) {
1227 // GlobalEscape or ArgEscape state of field means it has unknown value.
1228 if (add_edge(ptn, phantom_obj)) {
1229 // New edge was added
1230 add_field_uses_to_worklist(ptn->as_Field());
1231 }
1232 }
1233 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1234 PointsToNode* e = i.get();
1235 if (e->is_Arraycopy()) {
1236 assert(ptn->arraycopy_dst(), "sanity");
1237 // Propagate only fields escape state through arraycopy edge.
1238 if (e->fields_escape_state() < field_es) {
1239 set_fields_escape_state(e, field_es);
1240 escape_worklist.push(e);
1241 }
1242 } else if (es >= field_es) {
1243 // fields_escape_state is also set to 'es' if it is less than 'es'.
1244 if (e->escape_state() < es) {
1245 set_escape_state(e, es);
1246 escape_worklist.push(e);
1247 }
1248 } else {
1249 // Propagate field escape state.
1250 bool es_changed = false;
1251 if (e->fields_escape_state() < field_es) {
1252 set_fields_escape_state(e, field_es);
1253 es_changed = true;
1254 }
1255 if ((e->escape_state() < field_es) &&
1256 e->is_Field() && ptn->is_JavaObject() &&
1257 e->as_Field()->is_oop()) {
1258 // Change escape state of referenced fileds.
1259 set_escape_state(e, field_es);
1260 es_changed = true;;
1261 } else if (e->escape_state() < es) {
1262 set_escape_state(e, es);
1263 es_changed = true;;
1264 }
1265 if (es_changed) {
1266 escape_worklist.push(e);
1267 }
1268 }
1269 }
1270 }
1271 // Remove escaped objects from non_escaped list.
1272 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1273 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1274 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1275 non_escaped_worklist.delete_at(next);
1276 }
1277 if (ptn->escape_state() == PointsToNode::NoEscape) {
1278 // Find fields in non-escaped allocations which have unknown value.
1279 find_init_values(ptn, phantom_obj, NULL);
1280 }
1281 }
1282 return (non_escaped_worklist.length() > 0);
1283 }
1285 // Add all references to JavaObject node by walking over all uses.
1286 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1287 int new_edges = 0;
1288 if (populate_worklist) {
1289 // Populate _worklist by uses of jobj's uses.
1290 for (UseIterator i(jobj); i.has_next(); i.next()) {
1291 PointsToNode* use = i.get();
1292 if (use->is_Arraycopy())
1293 continue;
1294 add_uses_to_worklist(use);
1295 if (use->is_Field() && use->as_Field()->is_oop()) {
1296 // Put on worklist all field's uses (loads) and
1297 // related field nodes (same base and offset).
1298 add_field_uses_to_worklist(use->as_Field());
1299 }
1300 }
1301 }
1302 for (int l = 0; l < _worklist.length(); l++) {
1303 PointsToNode* use = _worklist.at(l);
1304 if (PointsToNode::is_base_use(use)) {
1305 // Add reference from jobj to field and from field to jobj (field's base).
1306 use = PointsToNode::get_use_node(use)->as_Field();
1307 if (add_base(use->as_Field(), jobj)) {
1308 new_edges++;
1309 }
1310 continue;
1311 }
1312 assert(!use->is_JavaObject(), "sanity");
1313 if (use->is_Arraycopy()) {
1314 if (jobj == null_obj) // NULL object does not have field edges
1315 continue;
1316 // Added edge from Arraycopy node to arraycopy's source java object
1317 if (add_edge(use, jobj)) {
1318 jobj->set_arraycopy_src();
1319 new_edges++;
1320 }
1321 // and stop here.
1322 continue;
1323 }
1324 if (!add_edge(use, jobj))
1325 continue; // No new edge added, there was such edge already.
1326 new_edges++;
1327 if (use->is_LocalVar()) {
1328 add_uses_to_worklist(use);
1329 if (use->arraycopy_dst()) {
1330 for (EdgeIterator i(use); i.has_next(); i.next()) {
1331 PointsToNode* e = i.get();
1332 if (e->is_Arraycopy()) {
1333 if (jobj == null_obj) // NULL object does not have field edges
1334 continue;
1335 // Add edge from arraycopy's destination java object to Arraycopy node.
1336 if (add_edge(jobj, e)) {
1337 new_edges++;
1338 jobj->set_arraycopy_dst();
1339 }
1340 }
1341 }
1342 }
1343 } else {
1344 // Added new edge to stored in field values.
1345 // Put on worklist all field's uses (loads) and
1346 // related field nodes (same base and offset).
1347 add_field_uses_to_worklist(use->as_Field());
1348 }
1349 }
1350 _worklist.clear();
1351 _in_worklist.Reset();
1352 return new_edges;
1353 }
1355 // Put on worklist all related field nodes.
1356 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1357 assert(field->is_oop(), "sanity");
1358 int offset = field->offset();
1359 add_uses_to_worklist(field);
1360 // Loop over all bases of this field and push on worklist Field nodes
1361 // with the same offset and base (since they may reference the same field).
1362 for (BaseIterator i(field); i.has_next(); i.next()) {
1363 PointsToNode* base = i.get();
1364 add_fields_to_worklist(field, base);
1365 // Check if the base was source object of arraycopy and go over arraycopy's
1366 // destination objects since values stored to a field of source object are
1367 // accessable by uses (loads) of fields of destination objects.
1368 if (base->arraycopy_src()) {
1369 for (UseIterator j(base); j.has_next(); j.next()) {
1370 PointsToNode* arycp = j.get();
1371 if (arycp->is_Arraycopy()) {
1372 for (UseIterator k(arycp); k.has_next(); k.next()) {
1373 PointsToNode* abase = k.get();
1374 if (abase->arraycopy_dst() && abase != base) {
1375 // Look for the same arracopy reference.
1376 add_fields_to_worklist(field, abase);
1377 }
1378 }
1379 }
1380 }
1381 }
1382 }
1383 }
1385 // Put on worklist all related field nodes.
1386 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1387 int offset = field->offset();
1388 if (base->is_LocalVar()) {
1389 for (UseIterator j(base); j.has_next(); j.next()) {
1390 PointsToNode* f = j.get();
1391 if (PointsToNode::is_base_use(f)) { // Field
1392 f = PointsToNode::get_use_node(f);
1393 if (f == field || !f->as_Field()->is_oop())
1394 continue;
1395 int offs = f->as_Field()->offset();
1396 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1397 add_to_worklist(f);
1398 }
1399 }
1400 }
1401 } else {
1402 assert(base->is_JavaObject(), "sanity");
1403 if (// Skip phantom_object since it is only used to indicate that
1404 // this field's content globally escapes.
1405 (base != phantom_obj) &&
1406 // NULL object node does not have fields.
1407 (base != null_obj)) {
1408 for (EdgeIterator i(base); i.has_next(); i.next()) {
1409 PointsToNode* f = i.get();
1410 // Skip arraycopy edge since store to destination object field
1411 // does not update value in source object field.
1412 if (f->is_Arraycopy()) {
1413 assert(base->arraycopy_dst(), "sanity");
1414 continue;
1415 }
1416 if (f == field || !f->as_Field()->is_oop())
1417 continue;
1418 int offs = f->as_Field()->offset();
1419 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1420 add_to_worklist(f);
1421 }
1422 }
1423 }
1424 }
1425 }
1427 // Find fields which have unknown value.
1428 int ConnectionGraph::find_field_value(FieldNode* field) {
1429 // Escaped fields should have init value already.
1430 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1431 int new_edges = 0;
1432 for (BaseIterator i(field); i.has_next(); i.next()) {
1433 PointsToNode* base = i.get();
1434 if (base->is_JavaObject()) {
1435 // Skip Allocate's fields which will be processed later.
1436 if (base->ideal_node()->is_Allocate())
1437 return 0;
1438 assert(base == null_obj, "only NULL ptr base expected here");
1439 }
1440 }
1441 if (add_edge(field, phantom_obj)) {
1442 // New edge was added
1443 new_edges++;
1444 add_field_uses_to_worklist(field);
1445 }
1446 return new_edges;
1447 }
1449 // Find fields initializing values for allocations.
1450 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1451 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1452 int new_edges = 0;
1453 Node* alloc = pta->ideal_node();
1454 if (init_val == phantom_obj) {
1455 // Do nothing for Allocate nodes since its fields values are "known".
1456 if (alloc->is_Allocate())
1457 return 0;
1458 assert(alloc->as_CallStaticJava(), "sanity");
1459 #ifdef ASSERT
1460 if (alloc->as_CallStaticJava()->method() == NULL) {
1461 const char* name = alloc->as_CallStaticJava()->_name;
1462 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1463 }
1464 #endif
1465 // Non-escaped allocation returned from Java or runtime call have
1466 // unknown values in fields.
1467 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1468 PointsToNode* field = i.get();
1469 if (field->is_Field() && field->as_Field()->is_oop()) {
1470 if (add_edge(field, phantom_obj)) {
1471 // New edge was added
1472 new_edges++;
1473 add_field_uses_to_worklist(field->as_Field());
1474 }
1475 }
1476 }
1477 return new_edges;
1478 }
1479 assert(init_val == null_obj, "sanity");
1480 // Do nothing for Call nodes since its fields values are unknown.
1481 if (!alloc->is_Allocate())
1482 return 0;
1484 InitializeNode* ini = alloc->as_Allocate()->initialization();
1485 Compile* C = _compile;
1486 bool visited_bottom_offset = false;
1487 GrowableArray<int> offsets_worklist;
1489 // Check if an oop field's initializing value is recorded and add
1490 // a corresponding NULL if field's value if it is not recorded.
1491 // Connection Graph does not record a default initialization by NULL
1492 // captured by Initialize node.
1493 //
1494 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1495 PointsToNode* field = i.get(); // Field (AddP)
1496 if (!field->is_Field() || !field->as_Field()->is_oop())
1497 continue; // Not oop field
1498 int offset = field->as_Field()->offset();
1499 if (offset == Type::OffsetBot) {
1500 if (!visited_bottom_offset) {
1501 // OffsetBot is used to reference array's element,
1502 // always add reference to NULL to all Field nodes since we don't
1503 // known which element is referenced.
1504 if (add_edge(field, null_obj)) {
1505 // New edge was added
1506 new_edges++;
1507 add_field_uses_to_worklist(field->as_Field());
1508 visited_bottom_offset = true;
1509 }
1510 }
1511 } else {
1512 // Check only oop fields.
1513 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1514 if (adr_type->isa_rawptr()) {
1515 #ifdef ASSERT
1516 // Raw pointers are used for initializing stores so skip it
1517 // since it should be recorded already
1518 Node* base = get_addp_base(field->ideal_node());
1519 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1520 (base->in(0) == alloc),"unexpected pointer type");
1521 #endif
1522 continue;
1523 }
1524 if (!offsets_worklist.contains(offset)) {
1525 offsets_worklist.append(offset);
1526 Node* value = NULL;
1527 if (ini != NULL) {
1528 // StoreP::memory_type() == T_ADDRESS
1529 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1530 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1531 // Make sure initializing store has the same type as this AddP.
1532 // This AddP may reference non existing field because it is on a
1533 // dead branch of bimorphic call which is not eliminated yet.
1534 if (store != NULL && store->is_Store() &&
1535 store->as_Store()->memory_type() == ft) {
1536 value = store->in(MemNode::ValueIn);
1537 #ifdef ASSERT
1538 if (VerifyConnectionGraph) {
1539 // Verify that AddP already points to all objects the value points to.
1540 PointsToNode* val = ptnode_adr(value->_idx);
1541 assert((val != NULL), "should be processed already");
1542 PointsToNode* missed_obj = NULL;
1543 if (val->is_JavaObject()) {
1544 if (!field->points_to(val->as_JavaObject())) {
1545 missed_obj = val;
1546 }
1547 } else {
1548 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1549 tty->print_cr("----------init store has invalid value -----");
1550 store->dump();
1551 val->dump();
1552 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1553 }
1554 for (EdgeIterator j(val); j.has_next(); j.next()) {
1555 PointsToNode* obj = j.get();
1556 if (obj->is_JavaObject()) {
1557 if (!field->points_to(obj->as_JavaObject())) {
1558 missed_obj = obj;
1559 break;
1560 }
1561 }
1562 }
1563 }
1564 if (missed_obj != NULL) {
1565 tty->print_cr("----------field---------------------------------");
1566 field->dump();
1567 tty->print_cr("----------missed referernce to object-----------");
1568 missed_obj->dump();
1569 tty->print_cr("----------object referernced by init store -----");
1570 store->dump();
1571 val->dump();
1572 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1573 }
1574 }
1575 #endif
1576 } else {
1577 // There could be initializing stores which follow allocation.
1578 // For example, a volatile field store is not collected
1579 // by Initialize node.
1580 //
1581 // Need to check for dependent loads to separate such stores from
1582 // stores which follow loads. For now, add initial value NULL so
1583 // that compare pointers optimization works correctly.
1584 }
1585 }
1586 if (value == NULL) {
1587 // A field's initializing value was not recorded. Add NULL.
1588 if (add_edge(field, null_obj)) {
1589 // New edge was added
1590 new_edges++;
1591 add_field_uses_to_worklist(field->as_Field());
1592 }
1593 }
1594 }
1595 }
1596 }
1597 return new_edges;
1598 }
1600 // Adjust scalar_replaceable state after Connection Graph is built.
1601 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1602 // Search for non-escaping objects which are not scalar replaceable
1603 // and mark them to propagate the state to referenced objects.
1605 // 1. An object is not scalar replaceable if the field into which it is
1606 // stored has unknown offset (stored into unknown element of an array).
1607 //
1608 for (UseIterator i(jobj); i.has_next(); i.next()) {
1609 PointsToNode* use = i.get();
1610 assert(!use->is_Arraycopy(), "sanity");
1611 if (use->is_Field()) {
1612 FieldNode* field = use->as_Field();
1613 assert(field->is_oop() && field->scalar_replaceable() &&
1614 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1615 if (field->offset() == Type::OffsetBot) {
1616 jobj->set_scalar_replaceable(false);
1617 return;
1618 }
1619 // 2. An object is not scalar replaceable if the field into which it is
1620 // stored has multiple bases one of which is null.
1621 if (field->base_count() > 1) {
1622 for (BaseIterator i(field); i.has_next(); i.next()) {
1623 PointsToNode* base = i.get();
1624 if (base == null_obj) {
1625 jobj->set_scalar_replaceable(false);
1626 return;
1627 }
1628 }
1629 }
1630 }
1631 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1632 // 3. An object is not scalar replaceable if it is merged with other objects.
1633 for (EdgeIterator j(use); j.has_next(); j.next()) {
1634 PointsToNode* ptn = j.get();
1635 if (ptn->is_JavaObject() && ptn != jobj) {
1636 // Mark all objects.
1637 jobj->set_scalar_replaceable(false);
1638 ptn->set_scalar_replaceable(false);
1639 }
1640 }
1641 if (!jobj->scalar_replaceable()) {
1642 return;
1643 }
1644 }
1646 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1647 // Non-escaping object node should point only to field nodes.
1648 FieldNode* field = j.get()->as_Field();
1649 int offset = field->as_Field()->offset();
1651 // 4. An object is not scalar replaceable if it has a field with unknown
1652 // offset (array's element is accessed in loop).
1653 if (offset == Type::OffsetBot) {
1654 jobj->set_scalar_replaceable(false);
1655 return;
1656 }
1657 // 5. Currently an object is not scalar replaceable if a LoadStore node
1658 // access its field since the field value is unknown after it.
1659 //
1660 Node* n = field->ideal_node();
1661 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1662 if (n->fast_out(i)->is_LoadStore()) {
1663 jobj->set_scalar_replaceable(false);
1664 return;
1665 }
1666 }
1668 // 6. Or the address may point to more then one object. This may produce
1669 // the false positive result (set not scalar replaceable)
1670 // since the flow-insensitive escape analysis can't separate
1671 // the case when stores overwrite the field's value from the case
1672 // when stores happened on different control branches.
1673 //
1674 // Note: it will disable scalar replacement in some cases:
1675 //
1676 // Point p[] = new Point[1];
1677 // p[0] = new Point(); // Will be not scalar replaced
1678 //
1679 // but it will save us from incorrect optimizations in next cases:
1680 //
1681 // Point p[] = new Point[1];
1682 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1683 //
1684 if (field->base_count() > 1) {
1685 for (BaseIterator i(field); i.has_next(); i.next()) {
1686 PointsToNode* base = i.get();
1687 // Don't take into account LocalVar nodes which
1688 // may point to only one object which should be also
1689 // this field's base by now.
1690 if (base->is_JavaObject() && base != jobj) {
1691 // Mark all bases.
1692 jobj->set_scalar_replaceable(false);
1693 base->set_scalar_replaceable(false);
1694 }
1695 }
1696 }
1697 }
1698 }
1700 #ifdef ASSERT
1701 void ConnectionGraph::verify_connection_graph(
1702 GrowableArray<PointsToNode*>& ptnodes_worklist,
1703 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1704 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1705 GrowableArray<Node*>& addp_worklist) {
1706 // Verify that graph is complete - no new edges could be added.
1707 int java_objects_length = java_objects_worklist.length();
1708 int non_escaped_length = non_escaped_worklist.length();
1709 int new_edges = 0;
1710 for (int next = 0; next < java_objects_length; ++next) {
1711 JavaObjectNode* ptn = java_objects_worklist.at(next);
1712 new_edges += add_java_object_edges(ptn, true);
1713 }
1714 assert(new_edges == 0, "graph was not complete");
1715 // Verify that escape state is final.
1716 int length = non_escaped_worklist.length();
1717 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1718 assert((non_escaped_length == non_escaped_worklist.length()) &&
1719 (non_escaped_length == length) &&
1720 (_worklist.length() == 0), "escape state was not final");
1722 // Verify fields information.
1723 int addp_length = addp_worklist.length();
1724 for (int next = 0; next < addp_length; ++next ) {
1725 Node* n = addp_worklist.at(next);
1726 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1727 if (field->is_oop()) {
1728 // Verify that field has all bases
1729 Node* base = get_addp_base(n);
1730 PointsToNode* ptn = ptnode_adr(base->_idx);
1731 if (ptn->is_JavaObject()) {
1732 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1733 } else {
1734 assert(ptn->is_LocalVar(), "sanity");
1735 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1736 PointsToNode* e = i.get();
1737 if (e->is_JavaObject()) {
1738 assert(field->has_base(e->as_JavaObject()), "sanity");
1739 }
1740 }
1741 }
1742 // Verify that all fields have initializing values.
1743 if (field->edge_count() == 0) {
1744 tty->print_cr("----------field does not have references----------");
1745 field->dump();
1746 for (BaseIterator i(field); i.has_next(); i.next()) {
1747 PointsToNode* base = i.get();
1748 tty->print_cr("----------field has next base---------------------");
1749 base->dump();
1750 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1751 tty->print_cr("----------base has fields-------------------------");
1752 for (EdgeIterator j(base); j.has_next(); j.next()) {
1753 j.get()->dump();
1754 }
1755 tty->print_cr("----------base has references---------------------");
1756 for (UseIterator j(base); j.has_next(); j.next()) {
1757 j.get()->dump();
1758 }
1759 }
1760 }
1761 for (UseIterator i(field); i.has_next(); i.next()) {
1762 i.get()->dump();
1763 }
1764 assert(field->edge_count() > 0, "sanity");
1765 }
1766 }
1767 }
1768 }
1769 #endif
1771 // Optimize ideal graph.
1772 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1773 GrowableArray<Node*>& storestore_worklist) {
1774 Compile* C = _compile;
1775 PhaseIterGVN* igvn = _igvn;
1776 if (EliminateLocks) {
1777 // Mark locks before changing ideal graph.
1778 int cnt = C->macro_count();
1779 for( int i=0; i < cnt; i++ ) {
1780 Node *n = C->macro_node(i);
1781 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1782 AbstractLockNode* alock = n->as_AbstractLock();
1783 if (!alock->is_non_esc_obj()) {
1784 if (not_global_escape(alock->obj_node())) {
1785 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1786 // The lock could be marked eliminated by lock coarsening
1787 // code during first IGVN before EA. Replace coarsened flag
1788 // to eliminate all associated locks/unlocks.
1789 alock->set_non_esc_obj();
1790 }
1791 }
1792 }
1793 }
1794 }
1796 if (OptimizePtrCompare) {
1797 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1798 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1799 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1800 // Optimize objects compare.
1801 while (ptr_cmp_worklist.length() != 0) {
1802 Node *n = ptr_cmp_worklist.pop();
1803 Node *res = optimize_ptr_compare(n);
1804 if (res != NULL) {
1805 #ifndef PRODUCT
1806 if (PrintOptimizePtrCompare) {
1807 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"));
1808 if (Verbose) {
1809 n->dump(1);
1810 }
1811 }
1812 #endif
1813 igvn->replace_node(n, res);
1814 }
1815 }
1816 // cleanup
1817 if (_pcmp_neq->outcnt() == 0)
1818 igvn->hash_delete(_pcmp_neq);
1819 if (_pcmp_eq->outcnt() == 0)
1820 igvn->hash_delete(_pcmp_eq);
1821 }
1823 // For MemBarStoreStore nodes added in library_call.cpp, check
1824 // escape status of associated AllocateNode and optimize out
1825 // MemBarStoreStore node if the allocated object never escapes.
1826 while (storestore_worklist.length() != 0) {
1827 Node *n = storestore_worklist.pop();
1828 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1829 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1830 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1831 if (not_global_escape(alloc)) {
1832 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1833 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1834 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1835 igvn->register_new_node_with_optimizer(mb);
1836 igvn->replace_node(storestore, mb);
1837 }
1838 }
1839 }
1841 // Optimize objects compare.
1842 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1843 assert(OptimizePtrCompare, "sanity");
1844 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1845 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1846 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1847 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1848 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1849 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1851 // Check simple cases first.
1852 if (jobj1 != NULL) {
1853 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1854 if (jobj1 == jobj2) {
1855 // Comparing the same not escaping object.
1856 return _pcmp_eq;
1857 }
1858 Node* obj = jobj1->ideal_node();
1859 // Comparing not escaping allocation.
1860 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1861 !ptn2->points_to(jobj1)) {
1862 return _pcmp_neq; // This includes nullness check.
1863 }
1864 }
1865 }
1866 if (jobj2 != NULL) {
1867 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1868 Node* obj = jobj2->ideal_node();
1869 // Comparing not escaping allocation.
1870 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1871 !ptn1->points_to(jobj2)) {
1872 return _pcmp_neq; // This includes nullness check.
1873 }
1874 }
1875 }
1876 if (jobj1 != NULL && jobj1 != phantom_obj &&
1877 jobj2 != NULL && jobj2 != phantom_obj &&
1878 jobj1->ideal_node()->is_Con() &&
1879 jobj2->ideal_node()->is_Con()) {
1880 // Klass or String constants compare. Need to be careful with
1881 // compressed pointers - compare types of ConN and ConP instead of nodes.
1882 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1883 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1884 if (t1->make_ptr() == t2->make_ptr()) {
1885 return _pcmp_eq;
1886 } else {
1887 return _pcmp_neq;
1888 }
1889 }
1890 if (ptn1->meet(ptn2)) {
1891 return NULL; // Sets are not disjoint
1892 }
1894 // Sets are disjoint.
1895 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1896 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1897 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1898 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1899 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1900 set2_has_unknown_ptr && set1_has_null_ptr) {
1901 // Check nullness of unknown object.
1902 return NULL;
1903 }
1905 // Disjointness by itself is not sufficient since
1906 // alias analysis is not complete for escaped objects.
1907 // Disjoint sets are definitely unrelated only when
1908 // at least one set has only not escaping allocations.
1909 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1910 if (ptn1->non_escaping_allocation()) {
1911 return _pcmp_neq;
1912 }
1913 }
1914 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1915 if (ptn2->non_escaping_allocation()) {
1916 return _pcmp_neq;
1917 }
1918 }
1919 return NULL;
1920 }
1922 // Connection Graph constuction functions.
1924 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1925 PointsToNode* ptadr = _nodes.at(n->_idx);
1926 if (ptadr != NULL) {
1927 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1928 return;
1929 }
1930 Compile* C = _compile;
1931 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
1932 _nodes.at_put(n->_idx, ptadr);
1933 }
1935 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1936 PointsToNode* ptadr = _nodes.at(n->_idx);
1937 if (ptadr != NULL) {
1938 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1939 return;
1940 }
1941 Compile* C = _compile;
1942 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
1943 _nodes.at_put(n->_idx, ptadr);
1944 }
1946 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1947 PointsToNode* ptadr = _nodes.at(n->_idx);
1948 if (ptadr != NULL) {
1949 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1950 return;
1951 }
1952 bool unsafe = false;
1953 bool is_oop = is_oop_field(n, offset, &unsafe);
1954 if (unsafe) {
1955 es = PointsToNode::GlobalEscape;
1956 }
1957 Compile* C = _compile;
1958 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
1959 _nodes.at_put(n->_idx, field);
1960 }
1962 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1963 PointsToNode* src, PointsToNode* dst) {
1964 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1965 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1966 PointsToNode* ptadr = _nodes.at(n->_idx);
1967 if (ptadr != NULL) {
1968 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1969 return;
1970 }
1971 Compile* C = _compile;
1972 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
1973 _nodes.at_put(n->_idx, ptadr);
1974 // Add edge from arraycopy node to source object.
1975 (void)add_edge(ptadr, src);
1976 src->set_arraycopy_src();
1977 // Add edge from destination object to arraycopy node.
1978 (void)add_edge(dst, ptadr);
1979 dst->set_arraycopy_dst();
1980 }
1982 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1983 const Type* adr_type = n->as_AddP()->bottom_type();
1984 BasicType bt = T_INT;
1985 if (offset == Type::OffsetBot) {
1986 // Check only oop fields.
1987 if (!adr_type->isa_aryptr() ||
1988 (adr_type->isa_aryptr()->klass() == NULL) ||
1989 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1990 // OffsetBot is used to reference array's element. Ignore first AddP.
1991 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
1992 bt = T_OBJECT;
1993 }
1994 }
1995 } else if (offset != oopDesc::klass_offset_in_bytes()) {
1996 if (adr_type->isa_instptr()) {
1997 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
1998 if (field != NULL) {
1999 bt = field->layout_type();
2000 } else {
2001 // Check for unsafe oop field access
2002 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2003 int opcode = n->fast_out(i)->Opcode();
2004 if (opcode == Op_StoreP || opcode == Op_LoadP ||
2005 opcode == Op_StoreN || opcode == Op_LoadN) {
2006 bt = T_OBJECT;
2007 (*unsafe) = true;
2008 break;
2009 }
2010 }
2011 }
2012 } else if (adr_type->isa_aryptr()) {
2013 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2014 // Ignore array length load.
2015 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2016 // Ignore first AddP.
2017 } else {
2018 const Type* elemtype = adr_type->isa_aryptr()->elem();
2019 bt = elemtype->array_element_basic_type();
2020 }
2021 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2022 // Allocation initialization, ThreadLocal field access, unsafe access
2023 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2024 int opcode = n->fast_out(i)->Opcode();
2025 if (opcode == Op_StoreP || opcode == Op_LoadP ||
2026 opcode == Op_StoreN || opcode == Op_LoadN) {
2027 bt = T_OBJECT;
2028 break;
2029 }
2030 }
2031 }
2032 }
2033 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2034 }
2036 // Returns unique pointed java object or NULL.
2037 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2038 assert(!_collecting, "should not call when contructed graph");
2039 // If the node was created after the escape computation we can't answer.
2040 uint idx = n->_idx;
2041 if (idx >= nodes_size()) {
2042 return NULL;
2043 }
2044 PointsToNode* ptn = ptnode_adr(idx);
2045 if (ptn->is_JavaObject()) {
2046 return ptn->as_JavaObject();
2047 }
2048 assert(ptn->is_LocalVar(), "sanity");
2049 // Check all java objects it points to.
2050 JavaObjectNode* jobj = NULL;
2051 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2052 PointsToNode* e = i.get();
2053 if (e->is_JavaObject()) {
2054 if (jobj == NULL) {
2055 jobj = e->as_JavaObject();
2056 } else if (jobj != e) {
2057 return NULL;
2058 }
2059 }
2060 }
2061 return jobj;
2062 }
2064 // Return true if this node points only to non-escaping allocations.
2065 bool PointsToNode::non_escaping_allocation() {
2066 if (is_JavaObject()) {
2067 Node* n = ideal_node();
2068 if (n->is_Allocate() || n->is_CallStaticJava()) {
2069 return (escape_state() == PointsToNode::NoEscape);
2070 } else {
2071 return false;
2072 }
2073 }
2074 assert(is_LocalVar(), "sanity");
2075 // Check all java objects it points to.
2076 for (EdgeIterator i(this); i.has_next(); i.next()) {
2077 PointsToNode* e = i.get();
2078 if (e->is_JavaObject()) {
2079 Node* n = e->ideal_node();
2080 if ((e->escape_state() != PointsToNode::NoEscape) ||
2081 !(n->is_Allocate() || n->is_CallStaticJava())) {
2082 return false;
2083 }
2084 }
2085 }
2086 return true;
2087 }
2089 // Return true if we know the node does not escape globally.
2090 bool ConnectionGraph::not_global_escape(Node *n) {
2091 assert(!_collecting, "should not call during graph construction");
2092 // If the node was created after the escape computation we can't answer.
2093 uint idx = n->_idx;
2094 if (idx >= nodes_size()) {
2095 return false;
2096 }
2097 PointsToNode* ptn = ptnode_adr(idx);
2098 PointsToNode::EscapeState es = ptn->escape_state();
2099 // If we have already computed a value, return it.
2100 if (es >= PointsToNode::GlobalEscape)
2101 return false;
2102 if (ptn->is_JavaObject()) {
2103 return true; // (es < PointsToNode::GlobalEscape);
2104 }
2105 assert(ptn->is_LocalVar(), "sanity");
2106 // Check all java objects it points to.
2107 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2108 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2109 return false;
2110 }
2111 return true;
2112 }
2115 // Helper functions
2117 // Return true if this node points to specified node or nodes it points to.
2118 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2119 if (is_JavaObject()) {
2120 return (this == ptn);
2121 }
2122 assert(is_LocalVar() || is_Field(), "sanity");
2123 for (EdgeIterator i(this); i.has_next(); i.next()) {
2124 if (i.get() == ptn)
2125 return true;
2126 }
2127 return false;
2128 }
2130 // Return true if one node points to an other.
2131 bool PointsToNode::meet(PointsToNode* ptn) {
2132 if (this == ptn) {
2133 return true;
2134 } else if (ptn->is_JavaObject()) {
2135 return this->points_to(ptn->as_JavaObject());
2136 } else if (this->is_JavaObject()) {
2137 return ptn->points_to(this->as_JavaObject());
2138 }
2139 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2140 int ptn_count = ptn->edge_count();
2141 for (EdgeIterator i(this); i.has_next(); i.next()) {
2142 PointsToNode* this_e = i.get();
2143 for (int j = 0; j < ptn_count; j++) {
2144 if (this_e == ptn->edge(j))
2145 return true;
2146 }
2147 }
2148 return false;
2149 }
2151 #ifdef ASSERT
2152 // Return true if bases point to this java object.
2153 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2154 for (BaseIterator i(this); i.has_next(); i.next()) {
2155 if (i.get() == jobj)
2156 return true;
2157 }
2158 return false;
2159 }
2160 #endif
2162 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2163 const Type *adr_type = phase->type(adr);
2164 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2165 adr->in(AddPNode::Address)->is_Proj() &&
2166 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2167 // We are computing a raw address for a store captured by an Initialize
2168 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2169 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2170 assert(offs != Type::OffsetBot ||
2171 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2172 "offset must be a constant or it is initialization of array");
2173 return offs;
2174 }
2175 const TypePtr *t_ptr = adr_type->isa_ptr();
2176 assert(t_ptr != NULL, "must be a pointer type");
2177 return t_ptr->offset();
2178 }
2180 Node* ConnectionGraph::get_addp_base(Node *addp) {
2181 assert(addp->is_AddP(), "must be AddP");
2182 //
2183 // AddP cases for Base and Address inputs:
2184 // case #1. Direct object's field reference:
2185 // Allocate
2186 // |
2187 // Proj #5 ( oop result )
2188 // |
2189 // CheckCastPP (cast to instance type)
2190 // | |
2191 // AddP ( base == address )
2192 //
2193 // case #2. Indirect object's field reference:
2194 // Phi
2195 // |
2196 // CastPP (cast to instance type)
2197 // | |
2198 // AddP ( base == address )
2199 //
2200 // case #3. Raw object's field reference for Initialize node:
2201 // Allocate
2202 // |
2203 // Proj #5 ( oop result )
2204 // top |
2205 // \ |
2206 // AddP ( base == top )
2207 //
2208 // case #4. Array's element reference:
2209 // {CheckCastPP | CastPP}
2210 // | | |
2211 // | AddP ( array's element offset )
2212 // | |
2213 // AddP ( array's offset )
2214 //
2215 // case #5. Raw object's field reference for arraycopy stub call:
2216 // The inline_native_clone() case when the arraycopy stub is called
2217 // after the allocation before Initialize and CheckCastPP nodes.
2218 // Allocate
2219 // |
2220 // Proj #5 ( oop result )
2221 // | |
2222 // AddP ( base == address )
2223 //
2224 // case #6. Constant Pool, ThreadLocal, CastX2P or
2225 // Raw object's field reference:
2226 // {ConP, ThreadLocal, CastX2P, raw Load}
2227 // top |
2228 // \ |
2229 // AddP ( base == top )
2230 //
2231 // case #7. Klass's field reference.
2232 // LoadKlass
2233 // | |
2234 // AddP ( base == address )
2235 //
2236 // case #8. narrow Klass's field reference.
2237 // LoadNKlass
2238 // |
2239 // DecodeN
2240 // | |
2241 // AddP ( base == address )
2242 //
2243 Node *base = addp->in(AddPNode::Base);
2244 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2245 base = addp->in(AddPNode::Address);
2246 while (base->is_AddP()) {
2247 // Case #6 (unsafe access) may have several chained AddP nodes.
2248 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2249 base = base->in(AddPNode::Address);
2250 }
2251 Node* uncast_base = base->uncast();
2252 int opcode = uncast_base->Opcode();
2253 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2254 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2255 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2256 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2257 }
2258 return base;
2259 }
2261 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2262 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2263 Node* addp2 = addp->raw_out(0);
2264 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2265 addp2->in(AddPNode::Base) == n &&
2266 addp2->in(AddPNode::Address) == addp) {
2267 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2268 //
2269 // Find array's offset to push it on worklist first and
2270 // as result process an array's element offset first (pushed second)
2271 // to avoid CastPP for the array's offset.
2272 // Otherwise the inserted CastPP (LocalVar) will point to what
2273 // the AddP (Field) points to. Which would be wrong since
2274 // the algorithm expects the CastPP has the same point as
2275 // as AddP's base CheckCastPP (LocalVar).
2276 //
2277 // ArrayAllocation
2278 // |
2279 // CheckCastPP
2280 // |
2281 // memProj (from ArrayAllocation CheckCastPP)
2282 // | ||
2283 // | || Int (element index)
2284 // | || | ConI (log(element size))
2285 // | || | /
2286 // | || LShift
2287 // | || /
2288 // | AddP (array's element offset)
2289 // | |
2290 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2291 // | / /
2292 // AddP (array's offset)
2293 // |
2294 // Load/Store (memory operation on array's element)
2295 //
2296 return addp2;
2297 }
2298 return NULL;
2299 }
2301 //
2302 // Adjust the type and inputs of an AddP which computes the
2303 // address of a field of an instance
2304 //
2305 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2306 PhaseGVN* igvn = _igvn;
2307 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2308 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2309 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2310 if (t == NULL) {
2311 // We are computing a raw address for a store captured by an Initialize
2312 // compute an appropriate address type (cases #3 and #5).
2313 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2314 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2315 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2316 assert(offs != Type::OffsetBot, "offset must be a constant");
2317 t = base_t->add_offset(offs)->is_oopptr();
2318 }
2319 int inst_id = base_t->instance_id();
2320 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2321 "old type must be non-instance or match new type");
2323 // The type 't' could be subclass of 'base_t'.
2324 // As result t->offset() could be large then base_t's size and it will
2325 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2326 // constructor verifies correctness of the offset.
2327 //
2328 // It could happened on subclass's branch (from the type profiling
2329 // inlining) which was not eliminated during parsing since the exactness
2330 // of the allocation type was not propagated to the subclass type check.
2331 //
2332 // Or the type 't' could be not related to 'base_t' at all.
2333 // It could happened when CHA type is different from MDO type on a dead path
2334 // (for example, from instanceof check) which is not collapsed during parsing.
2335 //
2336 // Do nothing for such AddP node and don't process its users since
2337 // this code branch will go away.
2338 //
2339 if (!t->is_known_instance() &&
2340 !base_t->klass()->is_subtype_of(t->klass())) {
2341 return false; // bail out
2342 }
2343 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2344 // Do NOT remove the next line: ensure a new alias index is allocated
2345 // for the instance type. Note: C++ will not remove it since the call
2346 // has side effect.
2347 int alias_idx = _compile->get_alias_index(tinst);
2348 igvn->set_type(addp, tinst);
2349 // record the allocation in the node map
2350 set_map(addp, get_map(base->_idx));
2351 // Set addp's Base and Address to 'base'.
2352 Node *abase = addp->in(AddPNode::Base);
2353 Node *adr = addp->in(AddPNode::Address);
2354 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2355 adr->in(0)->_idx == (uint)inst_id) {
2356 // Skip AddP cases #3 and #5.
2357 } else {
2358 assert(!abase->is_top(), "sanity"); // AddP case #3
2359 if (abase != base) {
2360 igvn->hash_delete(addp);
2361 addp->set_req(AddPNode::Base, base);
2362 if (abase == adr) {
2363 addp->set_req(AddPNode::Address, base);
2364 } else {
2365 // AddP case #4 (adr is array's element offset AddP node)
2366 #ifdef ASSERT
2367 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2368 assert(adr->is_AddP() && atype != NULL &&
2369 atype->instance_id() == inst_id, "array's element offset should be processed first");
2370 #endif
2371 }
2372 igvn->hash_insert(addp);
2373 }
2374 }
2375 // Put on IGVN worklist since at least addp's type was changed above.
2376 record_for_optimizer(addp);
2377 return true;
2378 }
2380 //
2381 // Create a new version of orig_phi if necessary. Returns either the newly
2382 // created phi or an existing phi. Sets create_new to indicate whether a new
2383 // phi was created. Cache the last newly created phi in the node map.
2384 //
2385 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2386 Compile *C = _compile;
2387 PhaseGVN* igvn = _igvn;
2388 new_created = false;
2389 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2390 // nothing to do if orig_phi is bottom memory or matches alias_idx
2391 if (phi_alias_idx == alias_idx) {
2392 return orig_phi;
2393 }
2394 // Have we recently created a Phi for this alias index?
2395 PhiNode *result = get_map_phi(orig_phi->_idx);
2396 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2397 return result;
2398 }
2399 // Previous check may fail when the same wide memory Phi was split into Phis
2400 // for different memory slices. Search all Phis for this region.
2401 if (result != NULL) {
2402 Node* region = orig_phi->in(0);
2403 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2404 Node* phi = region->fast_out(i);
2405 if (phi->is_Phi() &&
2406 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2407 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2408 return phi->as_Phi();
2409 }
2410 }
2411 }
2412 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) {
2413 if (C->do_escape_analysis() == true && !C->failing()) {
2414 // Retry compilation without escape analysis.
2415 // If this is the first failure, the sentinel string will "stick"
2416 // to the Compile object, and the C2Compiler will see it and retry.
2417 C->record_failure(C2Compiler::retry_no_escape_analysis());
2418 }
2419 return NULL;
2420 }
2421 orig_phi_worklist.append_if_missing(orig_phi);
2422 const TypePtr *atype = C->get_adr_type(alias_idx);
2423 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2424 C->copy_node_notes_to(result, orig_phi);
2425 igvn->set_type(result, result->bottom_type());
2426 record_for_optimizer(result);
2427 set_map(orig_phi, result);
2428 new_created = true;
2429 return result;
2430 }
2432 //
2433 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2434 // specified alias index.
2435 //
2436 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2437 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2438 Compile *C = _compile;
2439 PhaseGVN* igvn = _igvn;
2440 bool new_phi_created;
2441 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2442 if (!new_phi_created) {
2443 return result;
2444 }
2445 GrowableArray<PhiNode *> phi_list;
2446 GrowableArray<uint> cur_input;
2447 PhiNode *phi = orig_phi;
2448 uint idx = 1;
2449 bool finished = false;
2450 while(!finished) {
2451 while (idx < phi->req()) {
2452 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2453 if (mem != NULL && mem->is_Phi()) {
2454 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2455 if (new_phi_created) {
2456 // found an phi for which we created a new split, push current one on worklist and begin
2457 // processing new one
2458 phi_list.push(phi);
2459 cur_input.push(idx);
2460 phi = mem->as_Phi();
2461 result = newphi;
2462 idx = 1;
2463 continue;
2464 } else {
2465 mem = newphi;
2466 }
2467 }
2468 if (C->failing()) {
2469 return NULL;
2470 }
2471 result->set_req(idx++, mem);
2472 }
2473 #ifdef ASSERT
2474 // verify that the new Phi has an input for each input of the original
2475 assert( phi->req() == result->req(), "must have same number of inputs.");
2476 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2477 #endif
2478 // Check if all new phi's inputs have specified alias index.
2479 // Otherwise use old phi.
2480 for (uint i = 1; i < phi->req(); i++) {
2481 Node* in = result->in(i);
2482 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2483 }
2484 // we have finished processing a Phi, see if there are any more to do
2485 finished = (phi_list.length() == 0 );
2486 if (!finished) {
2487 phi = phi_list.pop();
2488 idx = cur_input.pop();
2489 PhiNode *prev_result = get_map_phi(phi->_idx);
2490 prev_result->set_req(idx++, result);
2491 result = prev_result;
2492 }
2493 }
2494 return result;
2495 }
2497 //
2498 // The next methods are derived from methods in MemNode.
2499 //
2500 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2501 Node *mem = mmem;
2502 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2503 // means an array I have not precisely typed yet. Do not do any
2504 // alias stuff with it any time soon.
2505 if (toop->base() != Type::AnyPtr &&
2506 !(toop->klass() != NULL &&
2507 toop->klass()->is_java_lang_Object() &&
2508 toop->offset() == Type::OffsetBot)) {
2509 mem = mmem->memory_at(alias_idx);
2510 // Update input if it is progress over what we have now
2511 }
2512 return mem;
2513 }
2515 //
2516 // Move memory users to their memory slices.
2517 //
2518 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2519 Compile* C = _compile;
2520 PhaseGVN* igvn = _igvn;
2521 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2522 assert(tp != NULL, "ptr type");
2523 int alias_idx = C->get_alias_index(tp);
2524 int general_idx = C->get_general_index(alias_idx);
2526 // Move users first
2527 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2528 Node* use = n->fast_out(i);
2529 if (use->is_MergeMem()) {
2530 MergeMemNode* mmem = use->as_MergeMem();
2531 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2532 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2533 continue; // Nothing to do
2534 }
2535 // Replace previous general reference to mem node.
2536 uint orig_uniq = C->unique();
2537 Node* m = find_inst_mem(n, general_idx, orig_phis);
2538 assert(orig_uniq == C->unique(), "no new nodes");
2539 mmem->set_memory_at(general_idx, m);
2540 --imax;
2541 --i;
2542 } else if (use->is_MemBar()) {
2543 assert(!use->is_Initialize(), "initializing stores should not be moved");
2544 if (use->req() > MemBarNode::Precedent &&
2545 use->in(MemBarNode::Precedent) == n) {
2546 // Don't move related membars.
2547 record_for_optimizer(use);
2548 continue;
2549 }
2550 tp = use->as_MemBar()->adr_type()->isa_ptr();
2551 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2552 alias_idx == general_idx) {
2553 continue; // Nothing to do
2554 }
2555 // Move to general memory slice.
2556 uint orig_uniq = C->unique();
2557 Node* m = find_inst_mem(n, general_idx, orig_phis);
2558 assert(orig_uniq == C->unique(), "no new nodes");
2559 igvn->hash_delete(use);
2560 imax -= use->replace_edge(n, m);
2561 igvn->hash_insert(use);
2562 record_for_optimizer(use);
2563 --i;
2564 #ifdef ASSERT
2565 } else if (use->is_Mem()) {
2566 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2567 // Don't move related cardmark.
2568 continue;
2569 }
2570 // Memory nodes should have new memory input.
2571 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2572 assert(tp != NULL, "ptr type");
2573 int idx = C->get_alias_index(tp);
2574 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2575 "Following memory nodes should have new memory input or be on the same memory slice");
2576 } else if (use->is_Phi()) {
2577 // Phi nodes should be split and moved already.
2578 tp = use->as_Phi()->adr_type()->isa_ptr();
2579 assert(tp != NULL, "ptr type");
2580 int idx = C->get_alias_index(tp);
2581 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2582 } else {
2583 use->dump();
2584 assert(false, "should not be here");
2585 #endif
2586 }
2587 }
2588 }
2590 //
2591 // Search memory chain of "mem" to find a MemNode whose address
2592 // is the specified alias index.
2593 //
2594 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2595 if (orig_mem == NULL)
2596 return orig_mem;
2597 Compile* C = _compile;
2598 PhaseGVN* igvn = _igvn;
2599 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2600 bool is_instance = (toop != NULL) && toop->is_known_instance();
2601 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2602 Node *prev = NULL;
2603 Node *result = orig_mem;
2604 while (prev != result) {
2605 prev = result;
2606 if (result == start_mem)
2607 break; // hit one of our sentinels
2608 if (result->is_Mem()) {
2609 const Type *at = igvn->type(result->in(MemNode::Address));
2610 if (at == Type::TOP)
2611 break; // Dead
2612 assert (at->isa_ptr() != NULL, "pointer type required.");
2613 int idx = C->get_alias_index(at->is_ptr());
2614 if (idx == alias_idx)
2615 break; // Found
2616 if (!is_instance && (at->isa_oopptr() == NULL ||
2617 !at->is_oopptr()->is_known_instance())) {
2618 break; // Do not skip store to general memory slice.
2619 }
2620 result = result->in(MemNode::Memory);
2621 }
2622 if (!is_instance)
2623 continue; // don't search further for non-instance types
2624 // skip over a call which does not affect this memory slice
2625 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2626 Node *proj_in = result->in(0);
2627 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2628 break; // hit one of our sentinels
2629 } else if (proj_in->is_Call()) {
2630 CallNode *call = proj_in->as_Call();
2631 if (!call->may_modify(toop, igvn)) {
2632 result = call->in(TypeFunc::Memory);
2633 }
2634 } else if (proj_in->is_Initialize()) {
2635 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2636 // Stop if this is the initialization for the object instance which
2637 // which contains this memory slice, otherwise skip over it.
2638 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2639 result = proj_in->in(TypeFunc::Memory);
2640 }
2641 } else if (proj_in->is_MemBar()) {
2642 result = proj_in->in(TypeFunc::Memory);
2643 }
2644 } else if (result->is_MergeMem()) {
2645 MergeMemNode *mmem = result->as_MergeMem();
2646 result = step_through_mergemem(mmem, alias_idx, toop);
2647 if (result == mmem->base_memory()) {
2648 // Didn't find instance memory, search through general slice recursively.
2649 result = mmem->memory_at(C->get_general_index(alias_idx));
2650 result = find_inst_mem(result, alias_idx, orig_phis);
2651 if (C->failing()) {
2652 return NULL;
2653 }
2654 mmem->set_memory_at(alias_idx, result);
2655 }
2656 } else if (result->is_Phi() &&
2657 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2658 Node *un = result->as_Phi()->unique_input(igvn);
2659 if (un != NULL) {
2660 orig_phis.append_if_missing(result->as_Phi());
2661 result = un;
2662 } else {
2663 break;
2664 }
2665 } else if (result->is_ClearArray()) {
2666 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2667 // Can not bypass initialization of the instance
2668 // we are looking for.
2669 break;
2670 }
2671 // Otherwise skip it (the call updated 'result' value).
2672 } else if (result->Opcode() == Op_SCMemProj) {
2673 Node* mem = result->in(0);
2674 Node* adr = NULL;
2675 if (mem->is_LoadStore()) {
2676 adr = mem->in(MemNode::Address);
2677 } else {
2678 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2679 adr = mem->in(3); // Memory edge corresponds to destination array
2680 }
2681 const Type *at = igvn->type(adr);
2682 if (at != Type::TOP) {
2683 assert (at->isa_ptr() != NULL, "pointer type required.");
2684 int idx = C->get_alias_index(at->is_ptr());
2685 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2686 break;
2687 }
2688 result = mem->in(MemNode::Memory);
2689 }
2690 }
2691 if (result->is_Phi()) {
2692 PhiNode *mphi = result->as_Phi();
2693 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2694 const TypePtr *t = mphi->adr_type();
2695 if (!is_instance) {
2696 // Push all non-instance Phis on the orig_phis worklist to update inputs
2697 // during Phase 4 if needed.
2698 orig_phis.append_if_missing(mphi);
2699 } else if (C->get_alias_index(t) != alias_idx) {
2700 // Create a new Phi with the specified alias index type.
2701 result = split_memory_phi(mphi, alias_idx, orig_phis);
2702 }
2703 }
2704 // the result is either MemNode, PhiNode, InitializeNode.
2705 return result;
2706 }
2708 //
2709 // Convert the types of unescaped object to instance types where possible,
2710 // propagate the new type information through the graph, and update memory
2711 // edges and MergeMem inputs to reflect the new type.
2712 //
2713 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2714 // The processing is done in 4 phases:
2715 //
2716 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2717 // types for the CheckCastPP for allocations where possible.
2718 // Propagate the the new types through users as follows:
2719 // casts and Phi: push users on alloc_worklist
2720 // AddP: cast Base and Address inputs to the instance type
2721 // push any AddP users on alloc_worklist and push any memnode
2722 // users onto memnode_worklist.
2723 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2724 // search the Memory chain for a store with the appropriate type
2725 // address type. If a Phi is found, create a new version with
2726 // the appropriate memory slices from each of the Phi inputs.
2727 // For stores, process the users as follows:
2728 // MemNode: push on memnode_worklist
2729 // MergeMem: push on mergemem_worklist
2730 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2731 // moving the first node encountered of each instance type to the
2732 // the input corresponding to its alias index.
2733 // appropriate memory slice.
2734 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2735 //
2736 // In the following example, the CheckCastPP nodes are the cast of allocation
2737 // results and the allocation of node 29 is unescaped and eligible to be an
2738 // instance type.
2739 //
2740 // We start with:
2741 //
2742 // 7 Parm #memory
2743 // 10 ConI "12"
2744 // 19 CheckCastPP "Foo"
2745 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2746 // 29 CheckCastPP "Foo"
2747 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2748 //
2749 // 40 StoreP 25 7 20 ... alias_index=4
2750 // 50 StoreP 35 40 30 ... alias_index=4
2751 // 60 StoreP 45 50 20 ... alias_index=4
2752 // 70 LoadP _ 60 30 ... alias_index=4
2753 // 80 Phi 75 50 60 Memory alias_index=4
2754 // 90 LoadP _ 80 30 ... alias_index=4
2755 // 100 LoadP _ 80 20 ... alias_index=4
2756 //
2757 //
2758 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2759 // and creating a new alias index for node 30. This gives:
2760 //
2761 // 7 Parm #memory
2762 // 10 ConI "12"
2763 // 19 CheckCastPP "Foo"
2764 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2765 // 29 CheckCastPP "Foo" iid=24
2766 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2767 //
2768 // 40 StoreP 25 7 20 ... alias_index=4
2769 // 50 StoreP 35 40 30 ... alias_index=6
2770 // 60 StoreP 45 50 20 ... alias_index=4
2771 // 70 LoadP _ 60 30 ... alias_index=6
2772 // 80 Phi 75 50 60 Memory alias_index=4
2773 // 90 LoadP _ 80 30 ... alias_index=6
2774 // 100 LoadP _ 80 20 ... alias_index=4
2775 //
2776 // In phase 2, new memory inputs are computed for the loads and stores,
2777 // And a new version of the phi is created. In phase 4, the inputs to
2778 // node 80 are updated and then the memory nodes are updated with the
2779 // values computed in phase 2. This results in:
2780 //
2781 // 7 Parm #memory
2782 // 10 ConI "12"
2783 // 19 CheckCastPP "Foo"
2784 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2785 // 29 CheckCastPP "Foo" iid=24
2786 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2787 //
2788 // 40 StoreP 25 7 20 ... alias_index=4
2789 // 50 StoreP 35 7 30 ... alias_index=6
2790 // 60 StoreP 45 40 20 ... alias_index=4
2791 // 70 LoadP _ 50 30 ... alias_index=6
2792 // 80 Phi 75 40 60 Memory alias_index=4
2793 // 120 Phi 75 50 50 Memory alias_index=6
2794 // 90 LoadP _ 120 30 ... alias_index=6
2795 // 100 LoadP _ 80 20 ... alias_index=4
2796 //
2797 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2798 GrowableArray<Node *> memnode_worklist;
2799 GrowableArray<PhiNode *> orig_phis;
2800 PhaseIterGVN *igvn = _igvn;
2801 uint new_index_start = (uint) _compile->num_alias_types();
2802 Arena* arena = Thread::current()->resource_area();
2803 VectorSet visited(arena);
2804 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2805 uint unique_old = _compile->unique();
2807 // Phase 1: Process possible allocations from alloc_worklist.
2808 // Create instance types for the CheckCastPP for allocations where possible.
2809 //
2810 // (Note: don't forget to change the order of the second AddP node on
2811 // the alloc_worklist if the order of the worklist processing is changed,
2812 // see the comment in find_second_addp().)
2813 //
2814 while (alloc_worklist.length() != 0) {
2815 Node *n = alloc_worklist.pop();
2816 uint ni = n->_idx;
2817 if (n->is_Call()) {
2818 CallNode *alloc = n->as_Call();
2819 // copy escape information to call node
2820 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2821 PointsToNode::EscapeState es = ptn->escape_state();
2822 // We have an allocation or call which returns a Java object,
2823 // see if it is unescaped.
2824 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2825 continue;
2826 // Find CheckCastPP for the allocate or for the return value of a call
2827 n = alloc->result_cast();
2828 if (n == NULL) { // No uses except Initialize node
2829 if (alloc->is_Allocate()) {
2830 // Set the scalar_replaceable flag for allocation
2831 // so it could be eliminated if it has no uses.
2832 alloc->as_Allocate()->_is_scalar_replaceable = true;
2833 }
2834 if (alloc->is_CallStaticJava()) {
2835 // Set the scalar_replaceable flag for boxing method
2836 // so it could be eliminated if it has no uses.
2837 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2838 }
2839 continue;
2840 }
2841 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2842 assert(!alloc->is_Allocate(), "allocation should have unique type");
2843 continue;
2844 }
2846 // The inline code for Object.clone() casts the allocation result to
2847 // java.lang.Object and then to the actual type of the allocated
2848 // object. Detect this case and use the second cast.
2849 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2850 // the allocation result is cast to java.lang.Object and then
2851 // to the actual Array type.
2852 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2853 && (alloc->is_AllocateArray() ||
2854 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2855 Node *cast2 = NULL;
2856 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2857 Node *use = n->fast_out(i);
2858 if (use->is_CheckCastPP()) {
2859 cast2 = use;
2860 break;
2861 }
2862 }
2863 if (cast2 != NULL) {
2864 n = cast2;
2865 } else {
2866 // Non-scalar replaceable if the allocation type is unknown statically
2867 // (reflection allocation), the object can't be restored during
2868 // deoptimization without precise type.
2869 continue;
2870 }
2871 }
2873 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2874 if (t == NULL)
2875 continue; // not a TypeOopPtr
2876 if (!t->klass_is_exact())
2877 continue; // not an unique type
2879 if (alloc->is_Allocate()) {
2880 // Set the scalar_replaceable flag for allocation
2881 // so it could be eliminated.
2882 alloc->as_Allocate()->_is_scalar_replaceable = true;
2883 }
2884 if (alloc->is_CallStaticJava()) {
2885 // Set the scalar_replaceable flag for boxing method
2886 // so it could be eliminated.
2887 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2888 }
2889 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2890 // in order for an object to be scalar-replaceable, it must be:
2891 // - a direct allocation (not a call returning an object)
2892 // - non-escaping
2893 // - eligible to be a unique type
2894 // - not determined to be ineligible by escape analysis
2895 set_map(alloc, n);
2896 set_map(n, alloc);
2897 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
2898 igvn->hash_delete(n);
2899 igvn->set_type(n, tinst);
2900 n->raise_bottom_type(tinst);
2901 igvn->hash_insert(n);
2902 record_for_optimizer(n);
2903 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2905 // First, put on the worklist all Field edges from Connection Graph
2906 // which is more accurate then putting immediate users from Ideal Graph.
2907 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2908 PointsToNode* tgt = e.get();
2909 Node* use = tgt->ideal_node();
2910 assert(tgt->is_Field() && use->is_AddP(),
2911 "only AddP nodes are Field edges in CG");
2912 if (use->outcnt() > 0) { // Don't process dead nodes
2913 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2914 if (addp2 != NULL) {
2915 assert(alloc->is_AllocateArray(),"array allocation was expected");
2916 alloc_worklist.append_if_missing(addp2);
2917 }
2918 alloc_worklist.append_if_missing(use);
2919 }
2920 }
2922 // An allocation may have an Initialize which has raw stores. Scan
2923 // the users of the raw allocation result and push AddP users
2924 // on alloc_worklist.
2925 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2926 assert (raw_result != NULL, "must have an allocation result");
2927 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2928 Node *use = raw_result->fast_out(i);
2929 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2930 Node* addp2 = find_second_addp(use, raw_result);
2931 if (addp2 != NULL) {
2932 assert(alloc->is_AllocateArray(),"array allocation was expected");
2933 alloc_worklist.append_if_missing(addp2);
2934 }
2935 alloc_worklist.append_if_missing(use);
2936 } else if (use->is_MemBar()) {
2937 memnode_worklist.append_if_missing(use);
2938 }
2939 }
2940 }
2941 } else if (n->is_AddP()) {
2942 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2943 if (jobj == NULL || jobj == phantom_obj) {
2944 #ifdef ASSERT
2945 ptnode_adr(get_addp_base(n)->_idx)->dump();
2946 ptnode_adr(n->_idx)->dump();
2947 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2948 #endif
2949 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2950 return;
2951 }
2952 Node *base = get_map(jobj->idx()); // CheckCastPP node
2953 if (!split_AddP(n, base)) continue; // wrong type from dead path
2954 } else if (n->is_Phi() ||
2955 n->is_CheckCastPP() ||
2956 n->is_EncodeP() ||
2957 n->is_DecodeN() ||
2958 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2959 if (visited.test_set(n->_idx)) {
2960 assert(n->is_Phi(), "loops only through Phi's");
2961 continue; // already processed
2962 }
2963 JavaObjectNode* jobj = unique_java_object(n);
2964 if (jobj == NULL || jobj == phantom_obj) {
2965 #ifdef ASSERT
2966 ptnode_adr(n->_idx)->dump();
2967 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2968 #endif
2969 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2970 return;
2971 } else {
2972 Node *val = get_map(jobj->idx()); // CheckCastPP node
2973 TypeNode *tn = n->as_Type();
2974 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2975 assert(tinst != NULL && tinst->is_known_instance() &&
2976 tinst->instance_id() == jobj->idx() , "instance type expected.");
2978 const Type *tn_type = igvn->type(tn);
2979 const TypeOopPtr *tn_t;
2980 if (tn_type->isa_narrowoop()) {
2981 tn_t = tn_type->make_ptr()->isa_oopptr();
2982 } else {
2983 tn_t = tn_type->isa_oopptr();
2984 }
2985 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2986 if (tn_type->isa_narrowoop()) {
2987 tn_type = tinst->make_narrowoop();
2988 } else {
2989 tn_type = tinst;
2990 }
2991 igvn->hash_delete(tn);
2992 igvn->set_type(tn, tn_type);
2993 tn->set_type(tn_type);
2994 igvn->hash_insert(tn);
2995 record_for_optimizer(n);
2996 } else {
2997 assert(tn_type == TypePtr::NULL_PTR ||
2998 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
2999 "unexpected type");
3000 continue; // Skip dead path with different type
3001 }
3002 }
3003 } else {
3004 debug_only(n->dump();)
3005 assert(false, "EA: unexpected node");
3006 continue;
3007 }
3008 // push allocation's users on appropriate worklist
3009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3010 Node *use = n->fast_out(i);
3011 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3012 // Load/store to instance's field
3013 memnode_worklist.append_if_missing(use);
3014 } else if (use->is_MemBar()) {
3015 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3016 memnode_worklist.append_if_missing(use);
3017 }
3018 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3019 Node* addp2 = find_second_addp(use, n);
3020 if (addp2 != NULL) {
3021 alloc_worklist.append_if_missing(addp2);
3022 }
3023 alloc_worklist.append_if_missing(use);
3024 } else if (use->is_Phi() ||
3025 use->is_CheckCastPP() ||
3026 use->is_EncodeNarrowPtr() ||
3027 use->is_DecodeNarrowPtr() ||
3028 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3029 alloc_worklist.append_if_missing(use);
3030 #ifdef ASSERT
3031 } else if (use->is_Mem()) {
3032 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3033 } else if (use->is_MergeMem()) {
3034 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3035 } else if (use->is_SafePoint()) {
3036 // Look for MergeMem nodes for calls which reference unique allocation
3037 // (through CheckCastPP nodes) even for debug info.
3038 Node* m = use->in(TypeFunc::Memory);
3039 if (m->is_MergeMem()) {
3040 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3041 }
3042 } else if (use->Opcode() == Op_EncodeISOArray) {
3043 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3044 // EncodeISOArray overwrites destination array
3045 memnode_worklist.append_if_missing(use);
3046 }
3047 } else {
3048 uint op = use->Opcode();
3049 if (!(op == Op_CmpP || op == Op_Conv2B ||
3050 op == Op_CastP2X || op == Op_StoreCM ||
3051 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
3052 op == Op_StrEquals || op == Op_StrIndexOf)) {
3053 n->dump();
3054 use->dump();
3055 assert(false, "EA: missing allocation reference path");
3056 }
3057 #endif
3058 }
3059 }
3061 }
3062 // New alias types were created in split_AddP().
3063 uint new_index_end = (uint) _compile->num_alias_types();
3064 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3066 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3067 // compute new values for Memory inputs (the Memory inputs are not
3068 // actually updated until phase 4.)
3069 if (memnode_worklist.length() == 0)
3070 return; // nothing to do
3071 while (memnode_worklist.length() != 0) {
3072 Node *n = memnode_worklist.pop();
3073 if (visited.test_set(n->_idx))
3074 continue;
3075 if (n->is_Phi() || n->is_ClearArray()) {
3076 // we don't need to do anything, but the users must be pushed
3077 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3078 // we don't need to do anything, but the users must be pushed
3079 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3080 if (n == NULL)
3081 continue;
3082 } else if (n->Opcode() == Op_EncodeISOArray) {
3083 // get the memory projection
3084 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3085 Node *use = n->fast_out(i);
3086 if (use->Opcode() == Op_SCMemProj) {
3087 n = use;
3088 break;
3089 }
3090 }
3091 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3092 } else {
3093 assert(n->is_Mem(), "memory node required.");
3094 Node *addr = n->in(MemNode::Address);
3095 const Type *addr_t = igvn->type(addr);
3096 if (addr_t == Type::TOP)
3097 continue;
3098 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3099 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3100 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3101 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3102 if (_compile->failing()) {
3103 return;
3104 }
3105 if (mem != n->in(MemNode::Memory)) {
3106 // We delay the memory edge update since we need old one in
3107 // MergeMem code below when instances memory slices are separated.
3108 set_map(n, mem);
3109 }
3110 if (n->is_Load()) {
3111 continue; // don't push users
3112 } else if (n->is_LoadStore()) {
3113 // get the memory projection
3114 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3115 Node *use = n->fast_out(i);
3116 if (use->Opcode() == Op_SCMemProj) {
3117 n = use;
3118 break;
3119 }
3120 }
3121 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3122 }
3123 }
3124 // push user on appropriate worklist
3125 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3126 Node *use = n->fast_out(i);
3127 if (use->is_Phi() || use->is_ClearArray()) {
3128 memnode_worklist.append_if_missing(use);
3129 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3130 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3131 continue;
3132 memnode_worklist.append_if_missing(use);
3133 } else if (use->is_MemBar()) {
3134 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3135 memnode_worklist.append_if_missing(use);
3136 }
3137 #ifdef ASSERT
3138 } else if(use->is_Mem()) {
3139 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3140 } else if (use->is_MergeMem()) {
3141 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3142 } else if (use->Opcode() == Op_EncodeISOArray) {
3143 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3144 // EncodeISOArray overwrites destination array
3145 memnode_worklist.append_if_missing(use);
3146 }
3147 } else {
3148 uint op = use->Opcode();
3149 if (!(op == Op_StoreCM ||
3150 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3151 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3152 op == Op_AryEq || op == Op_StrComp ||
3153 op == Op_StrEquals || op == Op_StrIndexOf)) {
3154 n->dump();
3155 use->dump();
3156 assert(false, "EA: missing memory path");
3157 }
3158 #endif
3159 }
3160 }
3161 }
3163 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3164 // Walk each memory slice moving the first node encountered of each
3165 // instance type to the the input corresponding to its alias index.
3166 uint length = _mergemem_worklist.length();
3167 for( uint next = 0; next < length; ++next ) {
3168 MergeMemNode* nmm = _mergemem_worklist.at(next);
3169 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3170 // Note: we don't want to use MergeMemStream here because we only want to
3171 // scan inputs which exist at the start, not ones we add during processing.
3172 // Note 2: MergeMem may already contains instance memory slices added
3173 // during find_inst_mem() call when memory nodes were processed above.
3174 igvn->hash_delete(nmm);
3175 uint nslices = nmm->req();
3176 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3177 Node* mem = nmm->in(i);
3178 Node* cur = NULL;
3179 if (mem == NULL || mem->is_top())
3180 continue;
3181 // First, update mergemem by moving memory nodes to corresponding slices
3182 // if their type became more precise since this mergemem was created.
3183 while (mem->is_Mem()) {
3184 const Type *at = igvn->type(mem->in(MemNode::Address));
3185 if (at != Type::TOP) {
3186 assert (at->isa_ptr() != NULL, "pointer type required.");
3187 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3188 if (idx == i) {
3189 if (cur == NULL)
3190 cur = mem;
3191 } else {
3192 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3193 nmm->set_memory_at(idx, mem);
3194 }
3195 }
3196 }
3197 mem = mem->in(MemNode::Memory);
3198 }
3199 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3200 // Find any instance of the current type if we haven't encountered
3201 // already a memory slice of the instance along the memory chain.
3202 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3203 if((uint)_compile->get_general_index(ni) == i) {
3204 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3205 if (nmm->is_empty_memory(m)) {
3206 Node* result = find_inst_mem(mem, ni, orig_phis);
3207 if (_compile->failing()) {
3208 return;
3209 }
3210 nmm->set_memory_at(ni, result);
3211 }
3212 }
3213 }
3214 }
3215 // Find the rest of instances values
3216 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3217 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3218 Node* result = step_through_mergemem(nmm, ni, tinst);
3219 if (result == nmm->base_memory()) {
3220 // Didn't find instance memory, search through general slice recursively.
3221 result = nmm->memory_at(_compile->get_general_index(ni));
3222 result = find_inst_mem(result, ni, orig_phis);
3223 if (_compile->failing()) {
3224 return;
3225 }
3226 nmm->set_memory_at(ni, result);
3227 }
3228 }
3229 igvn->hash_insert(nmm);
3230 record_for_optimizer(nmm);
3231 }
3233 // Phase 4: Update the inputs of non-instance memory Phis and
3234 // the Memory input of memnodes
3235 // First update the inputs of any non-instance Phi's from
3236 // which we split out an instance Phi. Note we don't have
3237 // to recursively process Phi's encounted on the input memory
3238 // chains as is done in split_memory_phi() since they will
3239 // also be processed here.
3240 for (int j = 0; j < orig_phis.length(); j++) {
3241 PhiNode *phi = orig_phis.at(j);
3242 int alias_idx = _compile->get_alias_index(phi->adr_type());
3243 igvn->hash_delete(phi);
3244 for (uint i = 1; i < phi->req(); i++) {
3245 Node *mem = phi->in(i);
3246 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3247 if (_compile->failing()) {
3248 return;
3249 }
3250 if (mem != new_mem) {
3251 phi->set_req(i, new_mem);
3252 }
3253 }
3254 igvn->hash_insert(phi);
3255 record_for_optimizer(phi);
3256 }
3258 // Update the memory inputs of MemNodes with the value we computed
3259 // in Phase 2 and move stores memory users to corresponding memory slices.
3260 // Disable memory split verification code until the fix for 6984348.
3261 // Currently it produces false negative results since it does not cover all cases.
3262 #if 0 // ifdef ASSERT
3263 visited.Reset();
3264 Node_Stack old_mems(arena, _compile->unique() >> 2);
3265 #endif
3266 for (uint i = 0; i < ideal_nodes.size(); i++) {
3267 Node* n = ideal_nodes.at(i);
3268 Node* nmem = get_map(n->_idx);
3269 assert(nmem != NULL, "sanity");
3270 if (n->is_Mem()) {
3271 #if 0 // ifdef ASSERT
3272 Node* old_mem = n->in(MemNode::Memory);
3273 if (!visited.test_set(old_mem->_idx)) {
3274 old_mems.push(old_mem, old_mem->outcnt());
3275 }
3276 #endif
3277 assert(n->in(MemNode::Memory) != nmem, "sanity");
3278 if (!n->is_Load()) {
3279 // Move memory users of a store first.
3280 move_inst_mem(n, orig_phis);
3281 }
3282 // Now update memory input
3283 igvn->hash_delete(n);
3284 n->set_req(MemNode::Memory, nmem);
3285 igvn->hash_insert(n);
3286 record_for_optimizer(n);
3287 } else {
3288 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3289 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3290 }
3291 }
3292 #if 0 // ifdef ASSERT
3293 // Verify that memory was split correctly
3294 while (old_mems.is_nonempty()) {
3295 Node* old_mem = old_mems.node();
3296 uint old_cnt = old_mems.index();
3297 old_mems.pop();
3298 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3299 }
3300 #endif
3301 }
3303 #ifndef PRODUCT
3304 static const char *node_type_names[] = {
3305 "UnknownType",
3306 "JavaObject",
3307 "LocalVar",
3308 "Field",
3309 "Arraycopy"
3310 };
3312 static const char *esc_names[] = {
3313 "UnknownEscape",
3314 "NoEscape",
3315 "ArgEscape",
3316 "GlobalEscape"
3317 };
3319 void PointsToNode::dump(bool print_state) const {
3320 NodeType nt = node_type();
3321 tty->print("%s ", node_type_names[(int) nt]);
3322 if (print_state) {
3323 EscapeState es = escape_state();
3324 EscapeState fields_es = fields_escape_state();
3325 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3326 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3327 tty->print("NSR ");
3328 }
3329 if (is_Field()) {
3330 FieldNode* f = (FieldNode*)this;
3331 if (f->is_oop())
3332 tty->print("oop ");
3333 if (f->offset() > 0)
3334 tty->print("+%d ", f->offset());
3335 tty->print("(");
3336 for (BaseIterator i(f); i.has_next(); i.next()) {
3337 PointsToNode* b = i.get();
3338 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3339 }
3340 tty->print(" )");
3341 }
3342 tty->print("[");
3343 for (EdgeIterator i(this); i.has_next(); i.next()) {
3344 PointsToNode* e = i.get();
3345 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3346 }
3347 tty->print(" [");
3348 for (UseIterator i(this); i.has_next(); i.next()) {
3349 PointsToNode* u = i.get();
3350 bool is_base = false;
3351 if (PointsToNode::is_base_use(u)) {
3352 is_base = true;
3353 u = PointsToNode::get_use_node(u)->as_Field();
3354 }
3355 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3356 }
3357 tty->print(" ]] ");
3358 if (_node == NULL)
3359 tty->print_cr("<null>");
3360 else
3361 _node->dump();
3362 }
3364 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3365 bool first = true;
3366 int ptnodes_length = ptnodes_worklist.length();
3367 for (int i = 0; i < ptnodes_length; i++) {
3368 PointsToNode *ptn = ptnodes_worklist.at(i);
3369 if (ptn == NULL || !ptn->is_JavaObject())
3370 continue;
3371 PointsToNode::EscapeState es = ptn->escape_state();
3372 if ((es != PointsToNode::NoEscape) && !Verbose) {
3373 continue;
3374 }
3375 Node* n = ptn->ideal_node();
3376 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3377 n->as_CallStaticJava()->is_boxing_method())) {
3378 if (first) {
3379 tty->cr();
3380 tty->print("======== Connection graph for ");
3381 _compile->method()->print_short_name();
3382 tty->cr();
3383 first = false;
3384 }
3385 ptn->dump();
3386 // Print all locals and fields which reference this allocation
3387 for (UseIterator j(ptn); j.has_next(); j.next()) {
3388 PointsToNode* use = j.get();
3389 if (use->is_LocalVar()) {
3390 use->dump(Verbose);
3391 } else if (Verbose) {
3392 use->dump();
3393 }
3394 }
3395 tty->cr();
3396 }
3397 }
3398 }
3399 #endif