Mon, 31 Aug 2015 17:20:08 +0200
8134031: Incorrect JIT compilation of complex code with inlining and escape analysis
Summary: Bad rewiring of memory edges when we split unique types during EA
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2005, 2015, 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
208 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
209 // processing, calls to CI to resolve symbols (types, fields, methods)
210 // referenced in bytecode. During symbol resolution VM may throw
211 // an exception which CI cleans and converts to compilation failure.
212 if (C->failing()) return false;
214 // 2. Finish Graph construction by propagating references to all
215 // java objects through graph.
216 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
217 java_objects_worklist, oop_fields_worklist)) {
218 // All objects escaped or hit time or iterations limits.
219 _collecting = false;
220 return false;
221 }
223 // 3. Adjust scalar_replaceable state of nonescaping objects and push
224 // scalar replaceable allocations on alloc_worklist for processing
225 // in split_unique_types().
226 int non_escaped_length = non_escaped_worklist.length();
227 for (int next = 0; next < non_escaped_length; next++) {
228 JavaObjectNode* ptn = non_escaped_worklist.at(next);
229 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
230 Node* n = ptn->ideal_node();
231 if (n->is_Allocate()) {
232 n->as_Allocate()->_is_non_escaping = noescape;
233 }
234 if (n->is_CallStaticJava()) {
235 n->as_CallStaticJava()->_is_non_escaping = noescape;
236 }
237 if (noescape && ptn->scalar_replaceable()) {
238 adjust_scalar_replaceable_state(ptn);
239 if (ptn->scalar_replaceable()) {
240 alloc_worklist.append(ptn->ideal_node());
241 }
242 }
243 }
245 #ifdef ASSERT
246 if (VerifyConnectionGraph) {
247 // Verify that graph is complete - no new edges could be added or needed.
248 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
249 java_objects_worklist, addp_worklist);
250 }
251 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
252 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
253 null_obj->edge_count() == 0 &&
254 !null_obj->arraycopy_src() &&
255 !null_obj->arraycopy_dst(), "sanity");
256 #endif
258 _collecting = false;
260 } // TracePhase t3("connectionGraph")
262 // 4. Optimize ideal graph based on EA information.
263 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
264 if (has_non_escaping_obj) {
265 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
266 }
268 #ifndef PRODUCT
269 if (PrintEscapeAnalysis) {
270 dump(ptnodes_worklist); // Dump ConnectionGraph
271 }
272 #endif
274 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
275 #ifdef ASSERT
276 if (VerifyConnectionGraph) {
277 int alloc_length = alloc_worklist.length();
278 for (int next = 0; next < alloc_length; ++next) {
279 Node* n = alloc_worklist.at(next);
280 PointsToNode* ptn = ptnode_adr(n->_idx);
281 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
282 }
283 }
284 #endif
286 // 5. Separate memory graph for scalar replaceable allcations.
287 if (has_scalar_replaceable_candidates &&
288 C->AliasLevel() >= 3 && EliminateAllocations) {
289 // Now use the escape information to create unique types for
290 // scalar replaceable objects.
291 split_unique_types(alloc_worklist);
292 if (C->failing()) return false;
293 C->print_method(PHASE_AFTER_EA, 2);
295 #ifdef ASSERT
296 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
297 tty->print("=== No allocations eliminated for ");
298 C->method()->print_short_name();
299 if(!EliminateAllocations) {
300 tty->print(" since EliminateAllocations is off ===");
301 } else if(!has_scalar_replaceable_candidates) {
302 tty->print(" since there are no scalar replaceable candidates ===");
303 } else if(C->AliasLevel() < 3) {
304 tty->print(" since AliasLevel < 3 ===");
305 }
306 tty->cr();
307 #endif
308 }
309 return has_non_escaping_obj;
310 }
312 // Utility function for nodes that load an object
313 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
314 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
315 // ThreadLocal has RawPtr type.
316 const Type* t = _igvn->type(n);
317 if (t->make_ptr() != NULL) {
318 Node* adr = n->in(MemNode::Address);
319 #ifdef ASSERT
320 if (!adr->is_AddP()) {
321 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
322 } else {
323 assert((ptnode_adr(adr->_idx) == NULL ||
324 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
325 }
326 #endif
327 add_local_var_and_edge(n, PointsToNode::NoEscape,
328 adr, delayed_worklist);
329 }
330 }
332 // Populate Connection Graph with PointsTo nodes and create simple
333 // connection graph edges.
334 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
335 assert(!_verify, "this method sould not be called for verification");
336 PhaseGVN* igvn = _igvn;
337 uint n_idx = n->_idx;
338 PointsToNode* n_ptn = ptnode_adr(n_idx);
339 if (n_ptn != NULL)
340 return; // No need to redefine PointsTo node during first iteration.
342 if (n->is_Call()) {
343 // Arguments to allocation and locking don't escape.
344 if (n->is_AbstractLock()) {
345 // Put Lock and Unlock nodes on IGVN worklist to process them during
346 // first IGVN optimization when escape information is still available.
347 record_for_optimizer(n);
348 } else if (n->is_Allocate()) {
349 add_call_node(n->as_Call());
350 record_for_optimizer(n);
351 } else {
352 if (n->is_CallStaticJava()) {
353 const char* name = n->as_CallStaticJava()->_name;
354 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
355 return; // Skip uncommon traps
356 }
357 // Don't mark as processed since call's arguments have to be processed.
358 delayed_worklist->push(n);
359 // Check if a call returns an object.
360 if ((n->as_Call()->returns_pointer() &&
361 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) ||
362 (n->is_CallStaticJava() &&
363 n->as_CallStaticJava()->is_boxing_method())) {
364 add_call_node(n->as_Call());
365 }
366 }
367 return;
368 }
369 // Put this check here to process call arguments since some call nodes
370 // point to phantom_obj.
371 if (n_ptn == phantom_obj || n_ptn == null_obj)
372 return; // Skip predefined nodes.
374 int opcode = n->Opcode();
375 switch (opcode) {
376 case Op_AddP: {
377 Node* base = get_addp_base(n);
378 PointsToNode* ptn_base = ptnode_adr(base->_idx);
379 // Field nodes are created for all field types. They are used in
380 // adjust_scalar_replaceable_state() and split_unique_types().
381 // Note, non-oop fields will have only base edges in Connection
382 // Graph because such fields are not used for oop loads and stores.
383 int offset = address_offset(n, igvn);
384 add_field(n, PointsToNode::NoEscape, offset);
385 if (ptn_base == NULL) {
386 delayed_worklist->push(n); // Process it later.
387 } else {
388 n_ptn = ptnode_adr(n_idx);
389 add_base(n_ptn->as_Field(), ptn_base);
390 }
391 break;
392 }
393 case Op_CastX2P: {
394 map_ideal_node(n, phantom_obj);
395 break;
396 }
397 case Op_CastPP:
398 case Op_CheckCastPP:
399 case Op_EncodeP:
400 case Op_DecodeN:
401 case Op_EncodePKlass:
402 case Op_DecodeNKlass: {
403 add_local_var_and_edge(n, PointsToNode::NoEscape,
404 n->in(1), delayed_worklist);
405 break;
406 }
407 case Op_CMoveP: {
408 add_local_var(n, PointsToNode::NoEscape);
409 // Do not add edges during first iteration because some could be
410 // not defined yet.
411 delayed_worklist->push(n);
412 break;
413 }
414 case Op_ConP:
415 case Op_ConN:
416 case Op_ConNKlass: {
417 // assume all oop constants globally escape except for null
418 PointsToNode::EscapeState es;
419 const Type* t = igvn->type(n);
420 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
421 es = PointsToNode::NoEscape;
422 } else {
423 es = PointsToNode::GlobalEscape;
424 }
425 add_java_object(n, es);
426 break;
427 }
428 case Op_CreateEx: {
429 // assume that all exception objects globally escape
430 map_ideal_node(n, phantom_obj);
431 break;
432 }
433 case Op_LoadKlass:
434 case Op_LoadNKlass: {
435 // Unknown class is loaded
436 map_ideal_node(n, phantom_obj);
437 break;
438 }
439 case Op_LoadP:
440 case Op_LoadN:
441 case Op_LoadPLocked: {
442 add_objload_to_connection_graph(n, delayed_worklist);
443 break;
444 }
445 case Op_Parm: {
446 map_ideal_node(n, phantom_obj);
447 break;
448 }
449 case Op_PartialSubtypeCheck: {
450 // Produces Null or notNull and is used in only in CmpP so
451 // phantom_obj could be used.
452 map_ideal_node(n, phantom_obj); // Result is unknown
453 break;
454 }
455 case Op_Phi: {
456 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
457 // ThreadLocal has RawPtr type.
458 const Type* t = n->as_Phi()->type();
459 if (t->make_ptr() != NULL) {
460 add_local_var(n, PointsToNode::NoEscape);
461 // Do not add edges during first iteration because some could be
462 // not defined yet.
463 delayed_worklist->push(n);
464 }
465 break;
466 }
467 case Op_Proj: {
468 // we are only interested in the oop result projection from a call
469 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
470 n->in(0)->as_Call()->returns_pointer()) {
471 add_local_var_and_edge(n, PointsToNode::NoEscape,
472 n->in(0), delayed_worklist);
473 }
474 break;
475 }
476 case Op_Rethrow: // Exception object escapes
477 case Op_Return: {
478 if (n->req() > TypeFunc::Parms &&
479 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
480 // Treat Return value as LocalVar with GlobalEscape escape state.
481 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
482 n->in(TypeFunc::Parms), delayed_worklist);
483 }
484 break;
485 }
486 case Op_GetAndSetP:
487 case Op_GetAndSetN: {
488 add_objload_to_connection_graph(n, delayed_worklist);
489 // fallthrough
490 }
491 case Op_StoreP:
492 case Op_StoreN:
493 case Op_StoreNKlass:
494 case Op_StorePConditional:
495 case Op_CompareAndSwapP:
496 case Op_CompareAndSwapN: {
497 Node* adr = n->in(MemNode::Address);
498 const Type *adr_type = igvn->type(adr);
499 adr_type = adr_type->make_ptr();
500 if (adr_type == NULL) {
501 break; // skip dead nodes
502 }
503 if (adr_type->isa_oopptr() ||
504 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
505 (adr_type == TypeRawPtr::NOTNULL &&
506 adr->in(AddPNode::Address)->is_Proj() &&
507 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
508 delayed_worklist->push(n); // Process it later.
509 #ifdef ASSERT
510 assert(adr->is_AddP(), "expecting an AddP");
511 if (adr_type == TypeRawPtr::NOTNULL) {
512 // Verify a raw address for a store captured by Initialize node.
513 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
514 assert(offs != Type::OffsetBot, "offset must be a constant");
515 }
516 #endif
517 } else {
518 // Ignore copy the displaced header to the BoxNode (OSR compilation).
519 if (adr->is_BoxLock())
520 break;
521 // Stored value escapes in unsafe access.
522 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
523 // Pointer stores in G1 barriers looks like unsafe access.
524 // Ignore such stores to be able scalar replace non-escaping
525 // allocations.
526 if (UseG1GC && adr->is_AddP()) {
527 Node* base = get_addp_base(adr);
528 if (base->Opcode() == Op_LoadP &&
529 base->in(MemNode::Address)->is_AddP()) {
530 adr = base->in(MemNode::Address);
531 Node* tls = get_addp_base(adr);
532 if (tls->Opcode() == Op_ThreadLocal) {
533 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
534 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() +
535 PtrQueue::byte_offset_of_buf())) {
536 break; // G1 pre barier previous oop value store.
537 }
538 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() +
539 PtrQueue::byte_offset_of_buf())) {
540 break; // G1 post barier card address store.
541 }
542 }
543 }
544 }
545 delayed_worklist->push(n); // Process unsafe access later.
546 break;
547 }
548 #ifdef ASSERT
549 n->dump(1);
550 assert(false, "not unsafe or G1 barrier raw StoreP");
551 #endif
552 }
553 break;
554 }
555 case Op_AryEq:
556 case Op_StrComp:
557 case Op_StrEquals:
558 case Op_StrIndexOf:
559 case Op_EncodeISOArray: {
560 add_local_var(n, PointsToNode::ArgEscape);
561 delayed_worklist->push(n); // Process it later.
562 break;
563 }
564 case Op_ThreadLocal: {
565 add_java_object(n, PointsToNode::ArgEscape);
566 break;
567 }
568 default:
569 ; // Do nothing for nodes not related to EA.
570 }
571 return;
572 }
574 #ifdef ASSERT
575 #define ELSE_FAIL(name) \
576 /* Should not be called for not pointer type. */ \
577 n->dump(1); \
578 assert(false, name); \
579 break;
580 #else
581 #define ELSE_FAIL(name) \
582 break;
583 #endif
585 // Add final simple edges to graph.
586 void ConnectionGraph::add_final_edges(Node *n) {
587 PointsToNode* n_ptn = ptnode_adr(n->_idx);
588 #ifdef ASSERT
589 if (_verify && n_ptn->is_JavaObject())
590 return; // This method does not change graph for JavaObject.
591 #endif
593 if (n->is_Call()) {
594 process_call_arguments(n->as_Call());
595 return;
596 }
597 assert(n->is_Store() || n->is_LoadStore() ||
598 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
599 "node should be registered already");
600 int opcode = n->Opcode();
601 switch (opcode) {
602 case Op_AddP: {
603 Node* base = get_addp_base(n);
604 PointsToNode* ptn_base = ptnode_adr(base->_idx);
605 assert(ptn_base != NULL, "field's base should be registered");
606 add_base(n_ptn->as_Field(), ptn_base);
607 break;
608 }
609 case Op_CastPP:
610 case Op_CheckCastPP:
611 case Op_EncodeP:
612 case Op_DecodeN:
613 case Op_EncodePKlass:
614 case Op_DecodeNKlass: {
615 add_local_var_and_edge(n, PointsToNode::NoEscape,
616 n->in(1), NULL);
617 break;
618 }
619 case Op_CMoveP: {
620 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
621 Node* in = n->in(i);
622 if (in == NULL)
623 continue; // ignore NULL
624 Node* uncast_in = in->uncast();
625 if (uncast_in->is_top() || uncast_in == n)
626 continue; // ignore top or inputs which go back this node
627 PointsToNode* ptn = ptnode_adr(in->_idx);
628 assert(ptn != NULL, "node should be registered");
629 add_edge(n_ptn, ptn);
630 }
631 break;
632 }
633 case Op_LoadP:
634 case Op_LoadN:
635 case Op_LoadPLocked: {
636 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
637 // ThreadLocal has RawPtr type.
638 const Type* t = _igvn->type(n);
639 if (t->make_ptr() != NULL) {
640 Node* adr = n->in(MemNode::Address);
641 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
642 break;
643 }
644 ELSE_FAIL("Op_LoadP");
645 }
646 case Op_Phi: {
647 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
648 // ThreadLocal has RawPtr type.
649 const Type* t = n->as_Phi()->type();
650 if (t->make_ptr() != NULL) {
651 for (uint i = 1; i < n->req(); i++) {
652 Node* in = n->in(i);
653 if (in == NULL)
654 continue; // ignore NULL
655 Node* uncast_in = in->uncast();
656 if (uncast_in->is_top() || uncast_in == n)
657 continue; // ignore top or inputs which go back this node
658 PointsToNode* ptn = ptnode_adr(in->_idx);
659 assert(ptn != NULL, "node should be registered");
660 add_edge(n_ptn, ptn);
661 }
662 break;
663 }
664 ELSE_FAIL("Op_Phi");
665 }
666 case Op_Proj: {
667 // we are only interested in the oop result projection from a call
668 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
669 n->in(0)->as_Call()->returns_pointer()) {
670 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
671 break;
672 }
673 ELSE_FAIL("Op_Proj");
674 }
675 case Op_Rethrow: // Exception object escapes
676 case Op_Return: {
677 if (n->req() > TypeFunc::Parms &&
678 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
679 // Treat Return value as LocalVar with GlobalEscape escape state.
680 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
681 n->in(TypeFunc::Parms), NULL);
682 break;
683 }
684 ELSE_FAIL("Op_Return");
685 }
686 case Op_StoreP:
687 case Op_StoreN:
688 case Op_StoreNKlass:
689 case Op_StorePConditional:
690 case Op_CompareAndSwapP:
691 case Op_CompareAndSwapN:
692 case Op_GetAndSetP:
693 case Op_GetAndSetN: {
694 Node* adr = n->in(MemNode::Address);
695 const Type *adr_type = _igvn->type(adr);
696 adr_type = adr_type->make_ptr();
697 #ifdef ASSERT
698 if (adr_type == NULL) {
699 n->dump(1);
700 assert(adr_type != NULL, "dead node should not be on list");
701 break;
702 }
703 #endif
704 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) {
705 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
706 }
707 if (adr_type->isa_oopptr() ||
708 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
709 (adr_type == TypeRawPtr::NOTNULL &&
710 adr->in(AddPNode::Address)->is_Proj() &&
711 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
712 // Point Address to Value
713 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
714 assert(adr_ptn != NULL &&
715 adr_ptn->as_Field()->is_oop(), "node should be registered");
716 Node *val = n->in(MemNode::ValueIn);
717 PointsToNode* ptn = ptnode_adr(val->_idx);
718 assert(ptn != NULL, "node should be registered");
719 add_edge(adr_ptn, ptn);
720 break;
721 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
722 // Stored value escapes in unsafe access.
723 Node *val = n->in(MemNode::ValueIn);
724 PointsToNode* ptn = ptnode_adr(val->_idx);
725 assert(ptn != NULL, "node should be registered");
726 set_escape_state(ptn, PointsToNode::GlobalEscape);
727 // Add edge to object for unsafe access with offset.
728 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
729 assert(adr_ptn != NULL, "node should be registered");
730 if (adr_ptn->is_Field()) {
731 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
732 add_edge(adr_ptn, ptn);
733 }
734 break;
735 }
736 ELSE_FAIL("Op_StoreP");
737 }
738 case Op_AryEq:
739 case Op_StrComp:
740 case Op_StrEquals:
741 case Op_StrIndexOf:
742 case Op_EncodeISOArray: {
743 // char[] arrays passed to string intrinsic do not escape but
744 // they are not scalar replaceable. Adjust escape state for them.
745 // Start from in(2) edge since in(1) is memory edge.
746 for (uint i = 2; i < n->req(); i++) {
747 Node* adr = n->in(i);
748 const Type* at = _igvn->type(adr);
749 if (!adr->is_top() && at->isa_ptr()) {
750 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
751 at->isa_ptr() != NULL, "expecting a pointer");
752 if (adr->is_AddP()) {
753 adr = get_addp_base(adr);
754 }
755 PointsToNode* ptn = ptnode_adr(adr->_idx);
756 assert(ptn != NULL, "node should be registered");
757 add_edge(n_ptn, ptn);
758 }
759 }
760 break;
761 }
762 default: {
763 // This method should be called only for EA specific nodes which may
764 // miss some edges when they were created.
765 #ifdef ASSERT
766 n->dump(1);
767 #endif
768 guarantee(false, "unknown node");
769 }
770 }
771 return;
772 }
774 void ConnectionGraph::add_call_node(CallNode* call) {
775 assert(call->returns_pointer(), "only for call which returns pointer");
776 uint call_idx = call->_idx;
777 if (call->is_Allocate()) {
778 Node* k = call->in(AllocateNode::KlassNode);
779 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
780 assert(kt != NULL, "TypeKlassPtr required.");
781 ciKlass* cik = kt->klass();
782 PointsToNode::EscapeState es = PointsToNode::NoEscape;
783 bool scalar_replaceable = true;
784 if (call->is_AllocateArray()) {
785 if (!cik->is_array_klass()) { // StressReflectiveCode
786 es = PointsToNode::GlobalEscape;
787 } else {
788 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
789 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
790 // Not scalar replaceable if the length is not constant or too big.
791 scalar_replaceable = false;
792 }
793 }
794 } else { // Allocate instance
795 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
796 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
797 !cik->is_instance_klass() || // StressReflectiveCode
798 cik->as_instance_klass()->has_finalizer()) {
799 es = PointsToNode::GlobalEscape;
800 }
801 }
802 add_java_object(call, es);
803 PointsToNode* ptn = ptnode_adr(call_idx);
804 if (!scalar_replaceable && ptn->scalar_replaceable()) {
805 ptn->set_scalar_replaceable(false);
806 }
807 } else if (call->is_CallStaticJava()) {
808 // Call nodes could be different types:
809 //
810 // 1. CallDynamicJavaNode (what happened during call is unknown):
811 //
812 // - mapped to GlobalEscape JavaObject node if oop is returned;
813 //
814 // - all oop arguments are escaping globally;
815 //
816 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
817 //
818 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
819 //
820 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
821 // - mapped to NoEscape JavaObject node if non-escaping object allocated
822 // during call is returned;
823 // - mapped to ArgEscape LocalVar node pointed to object arguments
824 // which are returned and does not escape during call;
825 //
826 // - oop arguments escaping status is defined by bytecode analysis;
827 //
828 // For a static call, we know exactly what method is being called.
829 // Use bytecode estimator to record whether the call's return value escapes.
830 ciMethod* meth = call->as_CallJava()->method();
831 if (meth == NULL) {
832 const char* name = call->as_CallStaticJava()->_name;
833 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
834 // Returns a newly allocated unescaped object.
835 add_java_object(call, PointsToNode::NoEscape);
836 ptnode_adr(call_idx)->set_scalar_replaceable(false);
837 } else if (meth->is_boxing_method()) {
838 // Returns boxing object
839 PointsToNode::EscapeState es;
840 vmIntrinsics::ID intr = meth->intrinsic_id();
841 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
842 // It does not escape if object is always allocated.
843 es = PointsToNode::NoEscape;
844 } else {
845 // It escapes globally if object could be loaded from cache.
846 es = PointsToNode::GlobalEscape;
847 }
848 add_java_object(call, es);
849 } else {
850 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
851 call_analyzer->copy_dependencies(_compile->dependencies());
852 if (call_analyzer->is_return_allocated()) {
853 // Returns a newly allocated unescaped object, simply
854 // update dependency information.
855 // Mark it as NoEscape so that objects referenced by
856 // it's fields will be marked as NoEscape at least.
857 add_java_object(call, PointsToNode::NoEscape);
858 ptnode_adr(call_idx)->set_scalar_replaceable(false);
859 } else {
860 // Determine whether any arguments are returned.
861 const TypeTuple* d = call->tf()->domain();
862 bool ret_arg = false;
863 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
864 if (d->field_at(i)->isa_ptr() != NULL &&
865 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
866 ret_arg = true;
867 break;
868 }
869 }
870 if (ret_arg) {
871 add_local_var(call, PointsToNode::ArgEscape);
872 } else {
873 // Returns unknown object.
874 map_ideal_node(call, phantom_obj);
875 }
876 }
877 }
878 } else {
879 // An other type of call, assume the worst case:
880 // returned value is unknown and globally escapes.
881 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
882 map_ideal_node(call, phantom_obj);
883 }
884 }
886 void ConnectionGraph::process_call_arguments(CallNode *call) {
887 bool is_arraycopy = false;
888 switch (call->Opcode()) {
889 #ifdef ASSERT
890 case Op_Allocate:
891 case Op_AllocateArray:
892 case Op_Lock:
893 case Op_Unlock:
894 assert(false, "should be done already");
895 break;
896 #endif
897 case Op_CallLeafNoFP:
898 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
899 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
900 // fall through
901 case Op_CallLeaf: {
902 // Stub calls, objects do not escape but they are not scale replaceable.
903 // Adjust escape state for outgoing arguments.
904 const TypeTuple * d = call->tf()->domain();
905 bool src_has_oops = false;
906 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
907 const Type* at = d->field_at(i);
908 Node *arg = call->in(i);
909 const Type *aat = _igvn->type(arg);
910 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
911 continue;
912 if (arg->is_AddP()) {
913 //
914 // The inline_native_clone() case when the arraycopy stub is called
915 // after the allocation before Initialize and CheckCastPP nodes.
916 // Or normal arraycopy for object arrays case.
917 //
918 // Set AddP's base (Allocate) as not scalar replaceable since
919 // pointer to the base (with offset) is passed as argument.
920 //
921 arg = get_addp_base(arg);
922 }
923 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
924 assert(arg_ptn != NULL, "should be registered");
925 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
926 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
927 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
928 aat->isa_ptr() != NULL, "expecting an Ptr");
929 bool arg_has_oops = aat->isa_oopptr() &&
930 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
931 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
932 if (i == TypeFunc::Parms) {
933 src_has_oops = arg_has_oops;
934 }
935 //
936 // src or dst could be j.l.Object when other is basic type array:
937 //
938 // arraycopy(char[],0,Object*,0,size);
939 // arraycopy(Object*,0,char[],0,size);
940 //
941 // Don't add edges in such cases.
942 //
943 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
944 arg_has_oops && (i > TypeFunc::Parms);
945 #ifdef ASSERT
946 if (!(is_arraycopy ||
947 (call->as_CallLeaf()->_name != NULL &&
948 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
949 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
950 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
951 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
952 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
953 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
954 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
955 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
956 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
957 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
958 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
959 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
960 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
961 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0)
962 ))) {
963 call->dump();
964 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
965 }
966 #endif
967 // Always process arraycopy's destination object since
968 // we need to add all possible edges to references in
969 // source object.
970 if (arg_esc >= PointsToNode::ArgEscape &&
971 !arg_is_arraycopy_dest) {
972 continue;
973 }
974 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
975 if (arg_is_arraycopy_dest) {
976 Node* src = call->in(TypeFunc::Parms);
977 if (src->is_AddP()) {
978 src = get_addp_base(src);
979 }
980 PointsToNode* src_ptn = ptnode_adr(src->_idx);
981 assert(src_ptn != NULL, "should be registered");
982 if (arg_ptn != src_ptn) {
983 // Special arraycopy edge:
984 // A destination object's field can't have the source object
985 // as base since objects escape states are not related.
986 // Only escape state of destination object's fields affects
987 // escape state of fields in source object.
988 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
989 }
990 }
991 }
992 }
993 break;
994 }
995 case Op_CallStaticJava: {
996 // For a static call, we know exactly what method is being called.
997 // Use bytecode estimator to record the call's escape affects
998 #ifdef ASSERT
999 const char* name = call->as_CallStaticJava()->_name;
1000 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1001 #endif
1002 ciMethod* meth = call->as_CallJava()->method();
1003 if ((meth != NULL) && meth->is_boxing_method()) {
1004 break; // Boxing methods do not modify any oops.
1005 }
1006 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1007 // fall-through if not a Java method or no analyzer information
1008 if (call_analyzer != NULL) {
1009 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1010 const TypeTuple* d = call->tf()->domain();
1011 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1012 const Type* at = d->field_at(i);
1013 int k = i - TypeFunc::Parms;
1014 Node* arg = call->in(i);
1015 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1016 if (at->isa_ptr() != NULL &&
1017 call_analyzer->is_arg_returned(k)) {
1018 // The call returns arguments.
1019 if (call_ptn != NULL) { // Is call's result used?
1020 assert(call_ptn->is_LocalVar(), "node should be registered");
1021 assert(arg_ptn != NULL, "node should be registered");
1022 add_edge(call_ptn, arg_ptn);
1023 }
1024 }
1025 if (at->isa_oopptr() != NULL &&
1026 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1027 if (!call_analyzer->is_arg_stack(k)) {
1028 // The argument global escapes
1029 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1030 } else {
1031 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1032 if (!call_analyzer->is_arg_local(k)) {
1033 // The argument itself doesn't escape, but any fields might
1034 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1035 }
1036 }
1037 }
1038 }
1039 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1040 // The call returns arguments.
1041 assert(call_ptn->edge_count() > 0, "sanity");
1042 if (!call_analyzer->is_return_local()) {
1043 // Returns also unknown object.
1044 add_edge(call_ptn, phantom_obj);
1045 }
1046 }
1047 break;
1048 }
1049 }
1050 default: {
1051 // Fall-through here if not a Java method or no analyzer information
1052 // or some other type of call, assume the worst case: all arguments
1053 // globally escape.
1054 const TypeTuple* d = call->tf()->domain();
1055 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1056 const Type* at = d->field_at(i);
1057 if (at->isa_oopptr() != NULL) {
1058 Node* arg = call->in(i);
1059 if (arg->is_AddP()) {
1060 arg = get_addp_base(arg);
1061 }
1062 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1063 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1064 }
1065 }
1066 }
1067 }
1068 }
1071 // Finish Graph construction.
1072 bool ConnectionGraph::complete_connection_graph(
1073 GrowableArray<PointsToNode*>& ptnodes_worklist,
1074 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1075 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1076 GrowableArray<FieldNode*>& oop_fields_worklist) {
1077 // Normally only 1-3 passes needed to build Connection Graph depending
1078 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1079 // Set limit to 20 to catch situation when something did go wrong and
1080 // bailout Escape Analysis.
1081 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1082 #define CG_BUILD_ITER_LIMIT 20
1084 // Propagate GlobalEscape and ArgEscape escape states and check that
1085 // we still have non-escaping objects. The method pushs on _worklist
1086 // Field nodes which reference phantom_object.
1087 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1088 return false; // Nothing to do.
1089 }
1090 // Now propagate references to all JavaObject nodes.
1091 int java_objects_length = java_objects_worklist.length();
1092 elapsedTimer time;
1093 bool timeout = false;
1094 int new_edges = 1;
1095 int iterations = 0;
1096 do {
1097 while ((new_edges > 0) &&
1098 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1099 double start_time = time.seconds();
1100 time.start();
1101 new_edges = 0;
1102 // Propagate references to phantom_object for nodes pushed on _worklist
1103 // by find_non_escaped_objects() and find_field_value().
1104 new_edges += add_java_object_edges(phantom_obj, false);
1105 for (int next = 0; next < java_objects_length; ++next) {
1106 JavaObjectNode* ptn = java_objects_worklist.at(next);
1107 new_edges += add_java_object_edges(ptn, true);
1109 #define SAMPLE_SIZE 4
1110 if ((next % SAMPLE_SIZE) == 0) {
1111 // Each 4 iterations calculate how much time it will take
1112 // to complete graph construction.
1113 time.stop();
1114 // Poll for requests from shutdown mechanism to quiesce compiler
1115 // because Connection graph construction may take long time.
1116 CompileBroker::maybe_block();
1117 double stop_time = time.seconds();
1118 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1119 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1120 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1121 timeout = true;
1122 break; // Timeout
1123 }
1124 start_time = stop_time;
1125 time.start();
1126 }
1127 #undef SAMPLE_SIZE
1129 }
1130 if (timeout) break;
1131 if (new_edges > 0) {
1132 // Update escape states on each iteration if graph was updated.
1133 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1134 return false; // Nothing to do.
1135 }
1136 }
1137 time.stop();
1138 if (time.seconds() >= EscapeAnalysisTimeout) {
1139 timeout = true;
1140 break;
1141 }
1142 }
1143 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1144 time.start();
1145 // Find fields which have unknown value.
1146 int fields_length = oop_fields_worklist.length();
1147 for (int next = 0; next < fields_length; next++) {
1148 FieldNode* field = oop_fields_worklist.at(next);
1149 if (field->edge_count() == 0) {
1150 new_edges += find_field_value(field);
1151 // This code may added new edges to phantom_object.
1152 // Need an other cycle to propagate references to phantom_object.
1153 }
1154 }
1155 time.stop();
1156 if (time.seconds() >= EscapeAnalysisTimeout) {
1157 timeout = true;
1158 break;
1159 }
1160 } else {
1161 new_edges = 0; // Bailout
1162 }
1163 } while (new_edges > 0);
1165 // Bailout if passed limits.
1166 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1167 Compile* C = _compile;
1168 if (C->log() != NULL) {
1169 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1170 C->log()->text("%s", timeout ? "time" : "iterations");
1171 C->log()->end_elem(" limit'");
1172 }
1173 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1174 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1175 // Possible infinite build_connection_graph loop,
1176 // bailout (no changes to ideal graph were made).
1177 return false;
1178 }
1179 #ifdef ASSERT
1180 if (Verbose && PrintEscapeAnalysis) {
1181 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1182 iterations, nodes_size(), ptnodes_worklist.length());
1183 }
1184 #endif
1186 #undef CG_BUILD_ITER_LIMIT
1188 // Find fields initialized by NULL for non-escaping Allocations.
1189 int non_escaped_length = non_escaped_worklist.length();
1190 for (int next = 0; next < non_escaped_length; next++) {
1191 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1192 PointsToNode::EscapeState es = ptn->escape_state();
1193 assert(es <= PointsToNode::ArgEscape, "sanity");
1194 if (es == PointsToNode::NoEscape) {
1195 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1196 // Adding references to NULL object does not change escape states
1197 // since it does not escape. Also no fields are added to NULL object.
1198 add_java_object_edges(null_obj, false);
1199 }
1200 }
1201 Node* n = ptn->ideal_node();
1202 if (n->is_Allocate()) {
1203 // The object allocated by this Allocate node will never be
1204 // seen by an other thread. Mark it so that when it is
1205 // expanded no MemBarStoreStore is added.
1206 InitializeNode* ini = n->as_Allocate()->initialization();
1207 if (ini != NULL)
1208 ini->set_does_not_escape();
1209 }
1210 }
1211 return true; // Finished graph construction.
1212 }
1214 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1215 // and check that we still have non-escaping java objects.
1216 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1217 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1218 GrowableArray<PointsToNode*> escape_worklist;
1219 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1220 int ptnodes_length = ptnodes_worklist.length();
1221 for (int next = 0; next < ptnodes_length; ++next) {
1222 PointsToNode* ptn = ptnodes_worklist.at(next);
1223 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1224 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1225 escape_worklist.push(ptn);
1226 }
1227 }
1228 // Set escape states to referenced nodes (edges list).
1229 while (escape_worklist.length() > 0) {
1230 PointsToNode* ptn = escape_worklist.pop();
1231 PointsToNode::EscapeState es = ptn->escape_state();
1232 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1233 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1234 es >= PointsToNode::ArgEscape) {
1235 // GlobalEscape or ArgEscape state of field means it has unknown value.
1236 if (add_edge(ptn, phantom_obj)) {
1237 // New edge was added
1238 add_field_uses_to_worklist(ptn->as_Field());
1239 }
1240 }
1241 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1242 PointsToNode* e = i.get();
1243 if (e->is_Arraycopy()) {
1244 assert(ptn->arraycopy_dst(), "sanity");
1245 // Propagate only fields escape state through arraycopy edge.
1246 if (e->fields_escape_state() < field_es) {
1247 set_fields_escape_state(e, field_es);
1248 escape_worklist.push(e);
1249 }
1250 } else if (es >= field_es) {
1251 // fields_escape_state is also set to 'es' if it is less than 'es'.
1252 if (e->escape_state() < es) {
1253 set_escape_state(e, es);
1254 escape_worklist.push(e);
1255 }
1256 } else {
1257 // Propagate field escape state.
1258 bool es_changed = false;
1259 if (e->fields_escape_state() < field_es) {
1260 set_fields_escape_state(e, field_es);
1261 es_changed = true;
1262 }
1263 if ((e->escape_state() < field_es) &&
1264 e->is_Field() && ptn->is_JavaObject() &&
1265 e->as_Field()->is_oop()) {
1266 // Change escape state of referenced fileds.
1267 set_escape_state(e, field_es);
1268 es_changed = true;;
1269 } else if (e->escape_state() < es) {
1270 set_escape_state(e, es);
1271 es_changed = true;;
1272 }
1273 if (es_changed) {
1274 escape_worklist.push(e);
1275 }
1276 }
1277 }
1278 }
1279 // Remove escaped objects from non_escaped list.
1280 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1281 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1282 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1283 non_escaped_worklist.delete_at(next);
1284 }
1285 if (ptn->escape_state() == PointsToNode::NoEscape) {
1286 // Find fields in non-escaped allocations which have unknown value.
1287 find_init_values(ptn, phantom_obj, NULL);
1288 }
1289 }
1290 return (non_escaped_worklist.length() > 0);
1291 }
1293 // Add all references to JavaObject node by walking over all uses.
1294 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1295 int new_edges = 0;
1296 if (populate_worklist) {
1297 // Populate _worklist by uses of jobj's uses.
1298 for (UseIterator i(jobj); i.has_next(); i.next()) {
1299 PointsToNode* use = i.get();
1300 if (use->is_Arraycopy())
1301 continue;
1302 add_uses_to_worklist(use);
1303 if (use->is_Field() && use->as_Field()->is_oop()) {
1304 // Put on worklist all field's uses (loads) and
1305 // related field nodes (same base and offset).
1306 add_field_uses_to_worklist(use->as_Field());
1307 }
1308 }
1309 }
1310 for (int l = 0; l < _worklist.length(); l++) {
1311 PointsToNode* use = _worklist.at(l);
1312 if (PointsToNode::is_base_use(use)) {
1313 // Add reference from jobj to field and from field to jobj (field's base).
1314 use = PointsToNode::get_use_node(use)->as_Field();
1315 if (add_base(use->as_Field(), jobj)) {
1316 new_edges++;
1317 }
1318 continue;
1319 }
1320 assert(!use->is_JavaObject(), "sanity");
1321 if (use->is_Arraycopy()) {
1322 if (jobj == null_obj) // NULL object does not have field edges
1323 continue;
1324 // Added edge from Arraycopy node to arraycopy's source java object
1325 if (add_edge(use, jobj)) {
1326 jobj->set_arraycopy_src();
1327 new_edges++;
1328 }
1329 // and stop here.
1330 continue;
1331 }
1332 if (!add_edge(use, jobj))
1333 continue; // No new edge added, there was such edge already.
1334 new_edges++;
1335 if (use->is_LocalVar()) {
1336 add_uses_to_worklist(use);
1337 if (use->arraycopy_dst()) {
1338 for (EdgeIterator i(use); i.has_next(); i.next()) {
1339 PointsToNode* e = i.get();
1340 if (e->is_Arraycopy()) {
1341 if (jobj == null_obj) // NULL object does not have field edges
1342 continue;
1343 // Add edge from arraycopy's destination java object to Arraycopy node.
1344 if (add_edge(jobj, e)) {
1345 new_edges++;
1346 jobj->set_arraycopy_dst();
1347 }
1348 }
1349 }
1350 }
1351 } else {
1352 // Added new edge to stored in field values.
1353 // Put on worklist all field's uses (loads) and
1354 // related field nodes (same base and offset).
1355 add_field_uses_to_worklist(use->as_Field());
1356 }
1357 }
1358 _worklist.clear();
1359 _in_worklist.Reset();
1360 return new_edges;
1361 }
1363 // Put on worklist all related field nodes.
1364 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1365 assert(field->is_oop(), "sanity");
1366 int offset = field->offset();
1367 add_uses_to_worklist(field);
1368 // Loop over all bases of this field and push on worklist Field nodes
1369 // with the same offset and base (since they may reference the same field).
1370 for (BaseIterator i(field); i.has_next(); i.next()) {
1371 PointsToNode* base = i.get();
1372 add_fields_to_worklist(field, base);
1373 // Check if the base was source object of arraycopy and go over arraycopy's
1374 // destination objects since values stored to a field of source object are
1375 // accessable by uses (loads) of fields of destination objects.
1376 if (base->arraycopy_src()) {
1377 for (UseIterator j(base); j.has_next(); j.next()) {
1378 PointsToNode* arycp = j.get();
1379 if (arycp->is_Arraycopy()) {
1380 for (UseIterator k(arycp); k.has_next(); k.next()) {
1381 PointsToNode* abase = k.get();
1382 if (abase->arraycopy_dst() && abase != base) {
1383 // Look for the same arracopy reference.
1384 add_fields_to_worklist(field, abase);
1385 }
1386 }
1387 }
1388 }
1389 }
1390 }
1391 }
1393 // Put on worklist all related field nodes.
1394 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1395 int offset = field->offset();
1396 if (base->is_LocalVar()) {
1397 for (UseIterator j(base); j.has_next(); j.next()) {
1398 PointsToNode* f = j.get();
1399 if (PointsToNode::is_base_use(f)) { // Field
1400 f = PointsToNode::get_use_node(f);
1401 if (f == field || !f->as_Field()->is_oop())
1402 continue;
1403 int offs = f->as_Field()->offset();
1404 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1405 add_to_worklist(f);
1406 }
1407 }
1408 }
1409 } else {
1410 assert(base->is_JavaObject(), "sanity");
1411 if (// Skip phantom_object since it is only used to indicate that
1412 // this field's content globally escapes.
1413 (base != phantom_obj) &&
1414 // NULL object node does not have fields.
1415 (base != null_obj)) {
1416 for (EdgeIterator i(base); i.has_next(); i.next()) {
1417 PointsToNode* f = i.get();
1418 // Skip arraycopy edge since store to destination object field
1419 // does not update value in source object field.
1420 if (f->is_Arraycopy()) {
1421 assert(base->arraycopy_dst(), "sanity");
1422 continue;
1423 }
1424 if (f == field || !f->as_Field()->is_oop())
1425 continue;
1426 int offs = f->as_Field()->offset();
1427 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1428 add_to_worklist(f);
1429 }
1430 }
1431 }
1432 }
1433 }
1435 // Find fields which have unknown value.
1436 int ConnectionGraph::find_field_value(FieldNode* field) {
1437 // Escaped fields should have init value already.
1438 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1439 int new_edges = 0;
1440 for (BaseIterator i(field); i.has_next(); i.next()) {
1441 PointsToNode* base = i.get();
1442 if (base->is_JavaObject()) {
1443 // Skip Allocate's fields which will be processed later.
1444 if (base->ideal_node()->is_Allocate())
1445 return 0;
1446 assert(base == null_obj, "only NULL ptr base expected here");
1447 }
1448 }
1449 if (add_edge(field, phantom_obj)) {
1450 // New edge was added
1451 new_edges++;
1452 add_field_uses_to_worklist(field);
1453 }
1454 return new_edges;
1455 }
1457 // Find fields initializing values for allocations.
1458 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1459 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1460 int new_edges = 0;
1461 Node* alloc = pta->ideal_node();
1462 if (init_val == phantom_obj) {
1463 // Do nothing for Allocate nodes since its fields values are "known".
1464 if (alloc->is_Allocate())
1465 return 0;
1466 assert(alloc->as_CallStaticJava(), "sanity");
1467 #ifdef ASSERT
1468 if (alloc->as_CallStaticJava()->method() == NULL) {
1469 const char* name = alloc->as_CallStaticJava()->_name;
1470 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1471 }
1472 #endif
1473 // Non-escaped allocation returned from Java or runtime call have
1474 // unknown values in fields.
1475 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1476 PointsToNode* field = i.get();
1477 if (field->is_Field() && field->as_Field()->is_oop()) {
1478 if (add_edge(field, phantom_obj)) {
1479 // New edge was added
1480 new_edges++;
1481 add_field_uses_to_worklist(field->as_Field());
1482 }
1483 }
1484 }
1485 return new_edges;
1486 }
1487 assert(init_val == null_obj, "sanity");
1488 // Do nothing for Call nodes since its fields values are unknown.
1489 if (!alloc->is_Allocate())
1490 return 0;
1492 InitializeNode* ini = alloc->as_Allocate()->initialization();
1493 Compile* C = _compile;
1494 bool visited_bottom_offset = false;
1495 GrowableArray<int> offsets_worklist;
1497 // Check if an oop field's initializing value is recorded and add
1498 // a corresponding NULL if field's value if it is not recorded.
1499 // Connection Graph does not record a default initialization by NULL
1500 // captured by Initialize node.
1501 //
1502 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1503 PointsToNode* field = i.get(); // Field (AddP)
1504 if (!field->is_Field() || !field->as_Field()->is_oop())
1505 continue; // Not oop field
1506 int offset = field->as_Field()->offset();
1507 if (offset == Type::OffsetBot) {
1508 if (!visited_bottom_offset) {
1509 // OffsetBot is used to reference array's element,
1510 // always add reference to NULL to all Field nodes since we don't
1511 // known which element is referenced.
1512 if (add_edge(field, null_obj)) {
1513 // New edge was added
1514 new_edges++;
1515 add_field_uses_to_worklist(field->as_Field());
1516 visited_bottom_offset = true;
1517 }
1518 }
1519 } else {
1520 // Check only oop fields.
1521 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1522 if (adr_type->isa_rawptr()) {
1523 #ifdef ASSERT
1524 // Raw pointers are used for initializing stores so skip it
1525 // since it should be recorded already
1526 Node* base = get_addp_base(field->ideal_node());
1527 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1528 (base->in(0) == alloc),"unexpected pointer type");
1529 #endif
1530 continue;
1531 }
1532 if (!offsets_worklist.contains(offset)) {
1533 offsets_worklist.append(offset);
1534 Node* value = NULL;
1535 if (ini != NULL) {
1536 // StoreP::memory_type() == T_ADDRESS
1537 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1538 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1539 // Make sure initializing store has the same type as this AddP.
1540 // This AddP may reference non existing field because it is on a
1541 // dead branch of bimorphic call which is not eliminated yet.
1542 if (store != NULL && store->is_Store() &&
1543 store->as_Store()->memory_type() == ft) {
1544 value = store->in(MemNode::ValueIn);
1545 #ifdef ASSERT
1546 if (VerifyConnectionGraph) {
1547 // Verify that AddP already points to all objects the value points to.
1548 PointsToNode* val = ptnode_adr(value->_idx);
1549 assert((val != NULL), "should be processed already");
1550 PointsToNode* missed_obj = NULL;
1551 if (val->is_JavaObject()) {
1552 if (!field->points_to(val->as_JavaObject())) {
1553 missed_obj = val;
1554 }
1555 } else {
1556 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1557 tty->print_cr("----------init store has invalid value -----");
1558 store->dump();
1559 val->dump();
1560 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1561 }
1562 for (EdgeIterator j(val); j.has_next(); j.next()) {
1563 PointsToNode* obj = j.get();
1564 if (obj->is_JavaObject()) {
1565 if (!field->points_to(obj->as_JavaObject())) {
1566 missed_obj = obj;
1567 break;
1568 }
1569 }
1570 }
1571 }
1572 if (missed_obj != NULL) {
1573 tty->print_cr("----------field---------------------------------");
1574 field->dump();
1575 tty->print_cr("----------missed referernce to object-----------");
1576 missed_obj->dump();
1577 tty->print_cr("----------object referernced by init store -----");
1578 store->dump();
1579 val->dump();
1580 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1581 }
1582 }
1583 #endif
1584 } else {
1585 // There could be initializing stores which follow allocation.
1586 // For example, a volatile field store is not collected
1587 // by Initialize node.
1588 //
1589 // Need to check for dependent loads to separate such stores from
1590 // stores which follow loads. For now, add initial value NULL so
1591 // that compare pointers optimization works correctly.
1592 }
1593 }
1594 if (value == NULL) {
1595 // A field's initializing value was not recorded. Add NULL.
1596 if (add_edge(field, null_obj)) {
1597 // New edge was added
1598 new_edges++;
1599 add_field_uses_to_worklist(field->as_Field());
1600 }
1601 }
1602 }
1603 }
1604 }
1605 return new_edges;
1606 }
1608 // Adjust scalar_replaceable state after Connection Graph is built.
1609 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1610 // Search for non-escaping objects which are not scalar replaceable
1611 // and mark them to propagate the state to referenced objects.
1613 // 1. An object is not scalar replaceable if the field into which it is
1614 // stored has unknown offset (stored into unknown element of an array).
1615 //
1616 for (UseIterator i(jobj); i.has_next(); i.next()) {
1617 PointsToNode* use = i.get();
1618 assert(!use->is_Arraycopy(), "sanity");
1619 if (use->is_Field()) {
1620 FieldNode* field = use->as_Field();
1621 assert(field->is_oop() && field->scalar_replaceable() &&
1622 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1623 if (field->offset() == Type::OffsetBot) {
1624 jobj->set_scalar_replaceable(false);
1625 return;
1626 }
1627 // 2. An object is not scalar replaceable if the field into which it is
1628 // stored has multiple bases one of which is null.
1629 if (field->base_count() > 1) {
1630 for (BaseIterator i(field); i.has_next(); i.next()) {
1631 PointsToNode* base = i.get();
1632 if (base == null_obj) {
1633 jobj->set_scalar_replaceable(false);
1634 return;
1635 }
1636 }
1637 }
1638 }
1639 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1640 // 3. An object is not scalar replaceable if it is merged with other objects.
1641 for (EdgeIterator j(use); j.has_next(); j.next()) {
1642 PointsToNode* ptn = j.get();
1643 if (ptn->is_JavaObject() && ptn != jobj) {
1644 // Mark all objects.
1645 jobj->set_scalar_replaceable(false);
1646 ptn->set_scalar_replaceable(false);
1647 }
1648 }
1649 if (!jobj->scalar_replaceable()) {
1650 return;
1651 }
1652 }
1654 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1655 // Non-escaping object node should point only to field nodes.
1656 FieldNode* field = j.get()->as_Field();
1657 int offset = field->as_Field()->offset();
1659 // 4. An object is not scalar replaceable if it has a field with unknown
1660 // offset (array's element is accessed in loop).
1661 if (offset == Type::OffsetBot) {
1662 jobj->set_scalar_replaceable(false);
1663 return;
1664 }
1665 // 5. Currently an object is not scalar replaceable if a LoadStore node
1666 // access its field since the field value is unknown after it.
1667 //
1668 Node* n = field->ideal_node();
1669 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1670 if (n->fast_out(i)->is_LoadStore()) {
1671 jobj->set_scalar_replaceable(false);
1672 return;
1673 }
1674 }
1676 // 6. Or the address may point to more then one object. This may produce
1677 // the false positive result (set not scalar replaceable)
1678 // since the flow-insensitive escape analysis can't separate
1679 // the case when stores overwrite the field's value from the case
1680 // when stores happened on different control branches.
1681 //
1682 // Note: it will disable scalar replacement in some cases:
1683 //
1684 // Point p[] = new Point[1];
1685 // p[0] = new Point(); // Will be not scalar replaced
1686 //
1687 // but it will save us from incorrect optimizations in next cases:
1688 //
1689 // Point p[] = new Point[1];
1690 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1691 //
1692 if (field->base_count() > 1) {
1693 for (BaseIterator i(field); i.has_next(); i.next()) {
1694 PointsToNode* base = i.get();
1695 // Don't take into account LocalVar nodes which
1696 // may point to only one object which should be also
1697 // this field's base by now.
1698 if (base->is_JavaObject() && base != jobj) {
1699 // Mark all bases.
1700 jobj->set_scalar_replaceable(false);
1701 base->set_scalar_replaceable(false);
1702 }
1703 }
1704 }
1705 }
1706 }
1708 #ifdef ASSERT
1709 void ConnectionGraph::verify_connection_graph(
1710 GrowableArray<PointsToNode*>& ptnodes_worklist,
1711 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1712 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1713 GrowableArray<Node*>& addp_worklist) {
1714 // Verify that graph is complete - no new edges could be added.
1715 int java_objects_length = java_objects_worklist.length();
1716 int non_escaped_length = non_escaped_worklist.length();
1717 int new_edges = 0;
1718 for (int next = 0; next < java_objects_length; ++next) {
1719 JavaObjectNode* ptn = java_objects_worklist.at(next);
1720 new_edges += add_java_object_edges(ptn, true);
1721 }
1722 assert(new_edges == 0, "graph was not complete");
1723 // Verify that escape state is final.
1724 int length = non_escaped_worklist.length();
1725 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1726 assert((non_escaped_length == non_escaped_worklist.length()) &&
1727 (non_escaped_length == length) &&
1728 (_worklist.length() == 0), "escape state was not final");
1730 // Verify fields information.
1731 int addp_length = addp_worklist.length();
1732 for (int next = 0; next < addp_length; ++next ) {
1733 Node* n = addp_worklist.at(next);
1734 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1735 if (field->is_oop()) {
1736 // Verify that field has all bases
1737 Node* base = get_addp_base(n);
1738 PointsToNode* ptn = ptnode_adr(base->_idx);
1739 if (ptn->is_JavaObject()) {
1740 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1741 } else {
1742 assert(ptn->is_LocalVar(), "sanity");
1743 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1744 PointsToNode* e = i.get();
1745 if (e->is_JavaObject()) {
1746 assert(field->has_base(e->as_JavaObject()), "sanity");
1747 }
1748 }
1749 }
1750 // Verify that all fields have initializing values.
1751 if (field->edge_count() == 0) {
1752 tty->print_cr("----------field does not have references----------");
1753 field->dump();
1754 for (BaseIterator i(field); i.has_next(); i.next()) {
1755 PointsToNode* base = i.get();
1756 tty->print_cr("----------field has next base---------------------");
1757 base->dump();
1758 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1759 tty->print_cr("----------base has fields-------------------------");
1760 for (EdgeIterator j(base); j.has_next(); j.next()) {
1761 j.get()->dump();
1762 }
1763 tty->print_cr("----------base has references---------------------");
1764 for (UseIterator j(base); j.has_next(); j.next()) {
1765 j.get()->dump();
1766 }
1767 }
1768 }
1769 for (UseIterator i(field); i.has_next(); i.next()) {
1770 i.get()->dump();
1771 }
1772 assert(field->edge_count() > 0, "sanity");
1773 }
1774 }
1775 }
1776 }
1777 #endif
1779 // Optimize ideal graph.
1780 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1781 GrowableArray<Node*>& storestore_worklist) {
1782 Compile* C = _compile;
1783 PhaseIterGVN* igvn = _igvn;
1784 if (EliminateLocks) {
1785 // Mark locks before changing ideal graph.
1786 int cnt = C->macro_count();
1787 for( int i=0; i < cnt; i++ ) {
1788 Node *n = C->macro_node(i);
1789 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1790 AbstractLockNode* alock = n->as_AbstractLock();
1791 if (!alock->is_non_esc_obj()) {
1792 if (not_global_escape(alock->obj_node())) {
1793 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1794 // The lock could be marked eliminated by lock coarsening
1795 // code during first IGVN before EA. Replace coarsened flag
1796 // to eliminate all associated locks/unlocks.
1797 #ifdef ASSERT
1798 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1799 #endif
1800 alock->set_non_esc_obj();
1801 }
1802 }
1803 }
1804 }
1805 }
1807 if (OptimizePtrCompare) {
1808 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1809 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1810 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1811 // Optimize objects compare.
1812 while (ptr_cmp_worklist.length() != 0) {
1813 Node *n = ptr_cmp_worklist.pop();
1814 Node *res = optimize_ptr_compare(n);
1815 if (res != NULL) {
1816 #ifndef PRODUCT
1817 if (PrintOptimizePtrCompare) {
1818 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"));
1819 if (Verbose) {
1820 n->dump(1);
1821 }
1822 }
1823 #endif
1824 igvn->replace_node(n, res);
1825 }
1826 }
1827 // cleanup
1828 if (_pcmp_neq->outcnt() == 0)
1829 igvn->hash_delete(_pcmp_neq);
1830 if (_pcmp_eq->outcnt() == 0)
1831 igvn->hash_delete(_pcmp_eq);
1832 }
1834 // For MemBarStoreStore nodes added in library_call.cpp, check
1835 // escape status of associated AllocateNode and optimize out
1836 // MemBarStoreStore node if the allocated object never escapes.
1837 while (storestore_worklist.length() != 0) {
1838 Node *n = storestore_worklist.pop();
1839 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1840 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1841 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1842 if (not_global_escape(alloc)) {
1843 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1844 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1845 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1846 igvn->register_new_node_with_optimizer(mb);
1847 igvn->replace_node(storestore, mb);
1848 }
1849 }
1850 }
1852 // Optimize objects compare.
1853 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1854 assert(OptimizePtrCompare, "sanity");
1855 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1856 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1857 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1858 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1859 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1860 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1862 // Check simple cases first.
1863 if (jobj1 != NULL) {
1864 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1865 if (jobj1 == jobj2) {
1866 // Comparing the same not escaping object.
1867 return _pcmp_eq;
1868 }
1869 Node* obj = jobj1->ideal_node();
1870 // Comparing not escaping allocation.
1871 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1872 !ptn2->points_to(jobj1)) {
1873 return _pcmp_neq; // This includes nullness check.
1874 }
1875 }
1876 }
1877 if (jobj2 != NULL) {
1878 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1879 Node* obj = jobj2->ideal_node();
1880 // Comparing not escaping allocation.
1881 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1882 !ptn1->points_to(jobj2)) {
1883 return _pcmp_neq; // This includes nullness check.
1884 }
1885 }
1886 }
1887 if (jobj1 != NULL && jobj1 != phantom_obj &&
1888 jobj2 != NULL && jobj2 != phantom_obj &&
1889 jobj1->ideal_node()->is_Con() &&
1890 jobj2->ideal_node()->is_Con()) {
1891 // Klass or String constants compare. Need to be careful with
1892 // compressed pointers - compare types of ConN and ConP instead of nodes.
1893 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1894 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1895 if (t1->make_ptr() == t2->make_ptr()) {
1896 return _pcmp_eq;
1897 } else {
1898 return _pcmp_neq;
1899 }
1900 }
1901 if (ptn1->meet(ptn2)) {
1902 return NULL; // Sets are not disjoint
1903 }
1905 // Sets are disjoint.
1906 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1907 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1908 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1909 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1910 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1911 set2_has_unknown_ptr && set1_has_null_ptr) {
1912 // Check nullness of unknown object.
1913 return NULL;
1914 }
1916 // Disjointness by itself is not sufficient since
1917 // alias analysis is not complete for escaped objects.
1918 // Disjoint sets are definitely unrelated only when
1919 // at least one set has only not escaping allocations.
1920 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1921 if (ptn1->non_escaping_allocation()) {
1922 return _pcmp_neq;
1923 }
1924 }
1925 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1926 if (ptn2->non_escaping_allocation()) {
1927 return _pcmp_neq;
1928 }
1929 }
1930 return NULL;
1931 }
1933 // Connection Graph constuction functions.
1935 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1936 PointsToNode* ptadr = _nodes.at(n->_idx);
1937 if (ptadr != NULL) {
1938 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1939 return;
1940 }
1941 Compile* C = _compile;
1942 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
1943 _nodes.at_put(n->_idx, ptadr);
1944 }
1946 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1947 PointsToNode* ptadr = _nodes.at(n->_idx);
1948 if (ptadr != NULL) {
1949 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1950 return;
1951 }
1952 Compile* C = _compile;
1953 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
1954 _nodes.at_put(n->_idx, ptadr);
1955 }
1957 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1958 PointsToNode* ptadr = _nodes.at(n->_idx);
1959 if (ptadr != NULL) {
1960 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1961 return;
1962 }
1963 bool unsafe = false;
1964 bool is_oop = is_oop_field(n, offset, &unsafe);
1965 if (unsafe) {
1966 es = PointsToNode::GlobalEscape;
1967 }
1968 Compile* C = _compile;
1969 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
1970 _nodes.at_put(n->_idx, field);
1971 }
1973 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1974 PointsToNode* src, PointsToNode* dst) {
1975 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1976 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1977 PointsToNode* ptadr = _nodes.at(n->_idx);
1978 if (ptadr != NULL) {
1979 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1980 return;
1981 }
1982 Compile* C = _compile;
1983 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
1984 _nodes.at_put(n->_idx, ptadr);
1985 // Add edge from arraycopy node to source object.
1986 (void)add_edge(ptadr, src);
1987 src->set_arraycopy_src();
1988 // Add edge from destination object to arraycopy node.
1989 (void)add_edge(dst, ptadr);
1990 dst->set_arraycopy_dst();
1991 }
1993 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1994 const Type* adr_type = n->as_AddP()->bottom_type();
1995 BasicType bt = T_INT;
1996 if (offset == Type::OffsetBot) {
1997 // Check only oop fields.
1998 if (!adr_type->isa_aryptr() ||
1999 (adr_type->isa_aryptr()->klass() == NULL) ||
2000 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2001 // OffsetBot is used to reference array's element. Ignore first AddP.
2002 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2003 bt = T_OBJECT;
2004 }
2005 }
2006 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2007 if (adr_type->isa_instptr()) {
2008 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
2009 if (field != NULL) {
2010 bt = field->layout_type();
2011 } else {
2012 // Check for unsafe oop field access
2013 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2014 int opcode = n->fast_out(i)->Opcode();
2015 if (opcode == Op_StoreP || opcode == Op_LoadP ||
2016 opcode == Op_StoreN || opcode == Op_LoadN) {
2017 bt = T_OBJECT;
2018 (*unsafe) = true;
2019 break;
2020 }
2021 }
2022 }
2023 } else if (adr_type->isa_aryptr()) {
2024 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2025 // Ignore array length load.
2026 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2027 // Ignore first AddP.
2028 } else {
2029 const Type* elemtype = adr_type->isa_aryptr()->elem();
2030 bt = elemtype->array_element_basic_type();
2031 }
2032 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2033 // Allocation initialization, ThreadLocal field access, unsafe access
2034 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2035 int opcode = n->fast_out(i)->Opcode();
2036 if (opcode == Op_StoreP || opcode == Op_LoadP ||
2037 opcode == Op_StoreN || opcode == Op_LoadN) {
2038 bt = T_OBJECT;
2039 break;
2040 }
2041 }
2042 }
2043 }
2044 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2045 }
2047 // Returns unique pointed java object or NULL.
2048 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2049 assert(!_collecting, "should not call when contructed graph");
2050 // If the node was created after the escape computation we can't answer.
2051 uint idx = n->_idx;
2052 if (idx >= nodes_size()) {
2053 return NULL;
2054 }
2055 PointsToNode* ptn = ptnode_adr(idx);
2056 if (ptn->is_JavaObject()) {
2057 return ptn->as_JavaObject();
2058 }
2059 assert(ptn->is_LocalVar(), "sanity");
2060 // Check all java objects it points to.
2061 JavaObjectNode* jobj = NULL;
2062 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2063 PointsToNode* e = i.get();
2064 if (e->is_JavaObject()) {
2065 if (jobj == NULL) {
2066 jobj = e->as_JavaObject();
2067 } else if (jobj != e) {
2068 return NULL;
2069 }
2070 }
2071 }
2072 return jobj;
2073 }
2075 // Return true if this node points only to non-escaping allocations.
2076 bool PointsToNode::non_escaping_allocation() {
2077 if (is_JavaObject()) {
2078 Node* n = ideal_node();
2079 if (n->is_Allocate() || n->is_CallStaticJava()) {
2080 return (escape_state() == PointsToNode::NoEscape);
2081 } else {
2082 return false;
2083 }
2084 }
2085 assert(is_LocalVar(), "sanity");
2086 // Check all java objects it points to.
2087 for (EdgeIterator i(this); i.has_next(); i.next()) {
2088 PointsToNode* e = i.get();
2089 if (e->is_JavaObject()) {
2090 Node* n = e->ideal_node();
2091 if ((e->escape_state() != PointsToNode::NoEscape) ||
2092 !(n->is_Allocate() || n->is_CallStaticJava())) {
2093 return false;
2094 }
2095 }
2096 }
2097 return true;
2098 }
2100 // Return true if we know the node does not escape globally.
2101 bool ConnectionGraph::not_global_escape(Node *n) {
2102 assert(!_collecting, "should not call during graph construction");
2103 // If the node was created after the escape computation we can't answer.
2104 uint idx = n->_idx;
2105 if (idx >= nodes_size()) {
2106 return false;
2107 }
2108 PointsToNode* ptn = ptnode_adr(idx);
2109 PointsToNode::EscapeState es = ptn->escape_state();
2110 // If we have already computed a value, return it.
2111 if (es >= PointsToNode::GlobalEscape)
2112 return false;
2113 if (ptn->is_JavaObject()) {
2114 return true; // (es < PointsToNode::GlobalEscape);
2115 }
2116 assert(ptn->is_LocalVar(), "sanity");
2117 // Check all java objects it points to.
2118 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2119 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2120 return false;
2121 }
2122 return true;
2123 }
2126 // Helper functions
2128 // Return true if this node points to specified node or nodes it points to.
2129 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2130 if (is_JavaObject()) {
2131 return (this == ptn);
2132 }
2133 assert(is_LocalVar() || is_Field(), "sanity");
2134 for (EdgeIterator i(this); i.has_next(); i.next()) {
2135 if (i.get() == ptn)
2136 return true;
2137 }
2138 return false;
2139 }
2141 // Return true if one node points to an other.
2142 bool PointsToNode::meet(PointsToNode* ptn) {
2143 if (this == ptn) {
2144 return true;
2145 } else if (ptn->is_JavaObject()) {
2146 return this->points_to(ptn->as_JavaObject());
2147 } else if (this->is_JavaObject()) {
2148 return ptn->points_to(this->as_JavaObject());
2149 }
2150 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2151 int ptn_count = ptn->edge_count();
2152 for (EdgeIterator i(this); i.has_next(); i.next()) {
2153 PointsToNode* this_e = i.get();
2154 for (int j = 0; j < ptn_count; j++) {
2155 if (this_e == ptn->edge(j))
2156 return true;
2157 }
2158 }
2159 return false;
2160 }
2162 #ifdef ASSERT
2163 // Return true if bases point to this java object.
2164 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2165 for (BaseIterator i(this); i.has_next(); i.next()) {
2166 if (i.get() == jobj)
2167 return true;
2168 }
2169 return false;
2170 }
2171 #endif
2173 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2174 const Type *adr_type = phase->type(adr);
2175 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2176 adr->in(AddPNode::Address)->is_Proj() &&
2177 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2178 // We are computing a raw address for a store captured by an Initialize
2179 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2180 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2181 assert(offs != Type::OffsetBot ||
2182 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2183 "offset must be a constant or it is initialization of array");
2184 return offs;
2185 }
2186 const TypePtr *t_ptr = adr_type->isa_ptr();
2187 assert(t_ptr != NULL, "must be a pointer type");
2188 return t_ptr->offset();
2189 }
2191 Node* ConnectionGraph::get_addp_base(Node *addp) {
2192 assert(addp->is_AddP(), "must be AddP");
2193 //
2194 // AddP cases for Base and Address inputs:
2195 // case #1. Direct object's field reference:
2196 // Allocate
2197 // |
2198 // Proj #5 ( oop result )
2199 // |
2200 // CheckCastPP (cast to instance type)
2201 // | |
2202 // AddP ( base == address )
2203 //
2204 // case #2. Indirect object's field reference:
2205 // Phi
2206 // |
2207 // CastPP (cast to instance type)
2208 // | |
2209 // AddP ( base == address )
2210 //
2211 // case #3. Raw object's field reference for Initialize node:
2212 // Allocate
2213 // |
2214 // Proj #5 ( oop result )
2215 // top |
2216 // \ |
2217 // AddP ( base == top )
2218 //
2219 // case #4. Array's element reference:
2220 // {CheckCastPP | CastPP}
2221 // | | |
2222 // | AddP ( array's element offset )
2223 // | |
2224 // AddP ( array's offset )
2225 //
2226 // case #5. Raw object's field reference for arraycopy stub call:
2227 // The inline_native_clone() case when the arraycopy stub is called
2228 // after the allocation before Initialize and CheckCastPP nodes.
2229 // Allocate
2230 // |
2231 // Proj #5 ( oop result )
2232 // | |
2233 // AddP ( base == address )
2234 //
2235 // case #6. Constant Pool, ThreadLocal, CastX2P or
2236 // Raw object's field reference:
2237 // {ConP, ThreadLocal, CastX2P, raw Load}
2238 // top |
2239 // \ |
2240 // AddP ( base == top )
2241 //
2242 // case #7. Klass's field reference.
2243 // LoadKlass
2244 // | |
2245 // AddP ( base == address )
2246 //
2247 // case #8. narrow Klass's field reference.
2248 // LoadNKlass
2249 // |
2250 // DecodeN
2251 // | |
2252 // AddP ( base == address )
2253 //
2254 Node *base = addp->in(AddPNode::Base);
2255 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2256 base = addp->in(AddPNode::Address);
2257 while (base->is_AddP()) {
2258 // Case #6 (unsafe access) may have several chained AddP nodes.
2259 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2260 base = base->in(AddPNode::Address);
2261 }
2262 Node* uncast_base = base->uncast();
2263 int opcode = uncast_base->Opcode();
2264 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2265 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2266 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2267 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2268 }
2269 return base;
2270 }
2272 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2273 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2274 Node* addp2 = addp->raw_out(0);
2275 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2276 addp2->in(AddPNode::Base) == n &&
2277 addp2->in(AddPNode::Address) == addp) {
2278 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2279 //
2280 // Find array's offset to push it on worklist first and
2281 // as result process an array's element offset first (pushed second)
2282 // to avoid CastPP for the array's offset.
2283 // Otherwise the inserted CastPP (LocalVar) will point to what
2284 // the AddP (Field) points to. Which would be wrong since
2285 // the algorithm expects the CastPP has the same point as
2286 // as AddP's base CheckCastPP (LocalVar).
2287 //
2288 // ArrayAllocation
2289 // |
2290 // CheckCastPP
2291 // |
2292 // memProj (from ArrayAllocation CheckCastPP)
2293 // | ||
2294 // | || Int (element index)
2295 // | || | ConI (log(element size))
2296 // | || | /
2297 // | || LShift
2298 // | || /
2299 // | AddP (array's element offset)
2300 // | |
2301 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2302 // | / /
2303 // AddP (array's offset)
2304 // |
2305 // Load/Store (memory operation on array's element)
2306 //
2307 return addp2;
2308 }
2309 return NULL;
2310 }
2312 //
2313 // Adjust the type and inputs of an AddP which computes the
2314 // address of a field of an instance
2315 //
2316 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2317 PhaseGVN* igvn = _igvn;
2318 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2319 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2320 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2321 if (t == NULL) {
2322 // We are computing a raw address for a store captured by an Initialize
2323 // compute an appropriate address type (cases #3 and #5).
2324 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2325 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2326 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2327 assert(offs != Type::OffsetBot, "offset must be a constant");
2328 t = base_t->add_offset(offs)->is_oopptr();
2329 }
2330 int inst_id = base_t->instance_id();
2331 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2332 "old type must be non-instance or match new type");
2334 // The type 't' could be subclass of 'base_t'.
2335 // As result t->offset() could be large then base_t's size and it will
2336 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2337 // constructor verifies correctness of the offset.
2338 //
2339 // It could happened on subclass's branch (from the type profiling
2340 // inlining) which was not eliminated during parsing since the exactness
2341 // of the allocation type was not propagated to the subclass type check.
2342 //
2343 // Or the type 't' could be not related to 'base_t' at all.
2344 // It could happened when CHA type is different from MDO type on a dead path
2345 // (for example, from instanceof check) which is not collapsed during parsing.
2346 //
2347 // Do nothing for such AddP node and don't process its users since
2348 // this code branch will go away.
2349 //
2350 if (!t->is_known_instance() &&
2351 !base_t->klass()->is_subtype_of(t->klass())) {
2352 return false; // bail out
2353 }
2354 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2355 // Do NOT remove the next line: ensure a new alias index is allocated
2356 // for the instance type. Note: C++ will not remove it since the call
2357 // has side effect.
2358 int alias_idx = _compile->get_alias_index(tinst);
2359 igvn->set_type(addp, tinst);
2360 // record the allocation in the node map
2361 set_map(addp, get_map(base->_idx));
2362 // Set addp's Base and Address to 'base'.
2363 Node *abase = addp->in(AddPNode::Base);
2364 Node *adr = addp->in(AddPNode::Address);
2365 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2366 adr->in(0)->_idx == (uint)inst_id) {
2367 // Skip AddP cases #3 and #5.
2368 } else {
2369 assert(!abase->is_top(), "sanity"); // AddP case #3
2370 if (abase != base) {
2371 igvn->hash_delete(addp);
2372 addp->set_req(AddPNode::Base, base);
2373 if (abase == adr) {
2374 addp->set_req(AddPNode::Address, base);
2375 } else {
2376 // AddP case #4 (adr is array's element offset AddP node)
2377 #ifdef ASSERT
2378 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2379 assert(adr->is_AddP() && atype != NULL &&
2380 atype->instance_id() == inst_id, "array's element offset should be processed first");
2381 #endif
2382 }
2383 igvn->hash_insert(addp);
2384 }
2385 }
2386 // Put on IGVN worklist since at least addp's type was changed above.
2387 record_for_optimizer(addp);
2388 return true;
2389 }
2391 //
2392 // Create a new version of orig_phi if necessary. Returns either the newly
2393 // created phi or an existing phi. Sets create_new to indicate whether a new
2394 // phi was created. Cache the last newly created phi in the node map.
2395 //
2396 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2397 Compile *C = _compile;
2398 PhaseGVN* igvn = _igvn;
2399 new_created = false;
2400 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2401 // nothing to do if orig_phi is bottom memory or matches alias_idx
2402 if (phi_alias_idx == alias_idx) {
2403 return orig_phi;
2404 }
2405 // Have we recently created a Phi for this alias index?
2406 PhiNode *result = get_map_phi(orig_phi->_idx);
2407 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2408 return result;
2409 }
2410 // Previous check may fail when the same wide memory Phi was split into Phis
2411 // for different memory slices. Search all Phis for this region.
2412 if (result != NULL) {
2413 Node* region = orig_phi->in(0);
2414 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2415 Node* phi = region->fast_out(i);
2416 if (phi->is_Phi() &&
2417 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2418 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2419 return phi->as_Phi();
2420 }
2421 }
2422 }
2423 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2424 if (C->do_escape_analysis() == true && !C->failing()) {
2425 // Retry compilation without escape analysis.
2426 // If this is the first failure, the sentinel string will "stick"
2427 // to the Compile object, and the C2Compiler will see it and retry.
2428 C->record_failure(C2Compiler::retry_no_escape_analysis());
2429 }
2430 return NULL;
2431 }
2432 orig_phi_worklist.append_if_missing(orig_phi);
2433 const TypePtr *atype = C->get_adr_type(alias_idx);
2434 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2435 C->copy_node_notes_to(result, orig_phi);
2436 igvn->set_type(result, result->bottom_type());
2437 record_for_optimizer(result);
2438 set_map(orig_phi, result);
2439 new_created = true;
2440 return result;
2441 }
2443 //
2444 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2445 // specified alias index.
2446 //
2447 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2448 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2449 Compile *C = _compile;
2450 PhaseGVN* igvn = _igvn;
2451 bool new_phi_created;
2452 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2453 if (!new_phi_created) {
2454 return result;
2455 }
2456 GrowableArray<PhiNode *> phi_list;
2457 GrowableArray<uint> cur_input;
2458 PhiNode *phi = orig_phi;
2459 uint idx = 1;
2460 bool finished = false;
2461 while(!finished) {
2462 while (idx < phi->req()) {
2463 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2464 if (mem != NULL && mem->is_Phi()) {
2465 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2466 if (new_phi_created) {
2467 // found an phi for which we created a new split, push current one on worklist and begin
2468 // processing new one
2469 phi_list.push(phi);
2470 cur_input.push(idx);
2471 phi = mem->as_Phi();
2472 result = newphi;
2473 idx = 1;
2474 continue;
2475 } else {
2476 mem = newphi;
2477 }
2478 }
2479 if (C->failing()) {
2480 return NULL;
2481 }
2482 result->set_req(idx++, mem);
2483 }
2484 #ifdef ASSERT
2485 // verify that the new Phi has an input for each input of the original
2486 assert( phi->req() == result->req(), "must have same number of inputs.");
2487 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2488 #endif
2489 // Check if all new phi's inputs have specified alias index.
2490 // Otherwise use old phi.
2491 for (uint i = 1; i < phi->req(); i++) {
2492 Node* in = result->in(i);
2493 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2494 }
2495 // we have finished processing a Phi, see if there are any more to do
2496 finished = (phi_list.length() == 0 );
2497 if (!finished) {
2498 phi = phi_list.pop();
2499 idx = cur_input.pop();
2500 PhiNode *prev_result = get_map_phi(phi->_idx);
2501 prev_result->set_req(idx++, result);
2502 result = prev_result;
2503 }
2504 }
2505 return result;
2506 }
2508 //
2509 // The next methods are derived from methods in MemNode.
2510 //
2511 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2512 Node *mem = mmem;
2513 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2514 // means an array I have not precisely typed yet. Do not do any
2515 // alias stuff with it any time soon.
2516 if (toop->base() != Type::AnyPtr &&
2517 !(toop->klass() != NULL &&
2518 toop->klass()->is_java_lang_Object() &&
2519 toop->offset() == Type::OffsetBot)) {
2520 mem = mmem->memory_at(alias_idx);
2521 // Update input if it is progress over what we have now
2522 }
2523 return mem;
2524 }
2526 //
2527 // Move memory users to their memory slices.
2528 //
2529 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2530 Compile* C = _compile;
2531 PhaseGVN* igvn = _igvn;
2532 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2533 assert(tp != NULL, "ptr type");
2534 int alias_idx = C->get_alias_index(tp);
2535 int general_idx = C->get_general_index(alias_idx);
2537 // Move users first
2538 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2539 Node* use = n->fast_out(i);
2540 if (use->is_MergeMem()) {
2541 MergeMemNode* mmem = use->as_MergeMem();
2542 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2543 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2544 continue; // Nothing to do
2545 }
2546 // Replace previous general reference to mem node.
2547 uint orig_uniq = C->unique();
2548 Node* m = find_inst_mem(n, general_idx, orig_phis);
2549 assert(orig_uniq == C->unique(), "no new nodes");
2550 mmem->set_memory_at(general_idx, m);
2551 --imax;
2552 --i;
2553 } else if (use->is_MemBar()) {
2554 assert(!use->is_Initialize(), "initializing stores should not be moved");
2555 if (use->req() > MemBarNode::Precedent &&
2556 use->in(MemBarNode::Precedent) == n) {
2557 // Don't move related membars.
2558 record_for_optimizer(use);
2559 continue;
2560 }
2561 tp = use->as_MemBar()->adr_type()->isa_ptr();
2562 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2563 alias_idx == general_idx) {
2564 continue; // Nothing to do
2565 }
2566 // Move to general memory slice.
2567 uint orig_uniq = C->unique();
2568 Node* m = find_inst_mem(n, general_idx, orig_phis);
2569 assert(orig_uniq == C->unique(), "no new nodes");
2570 igvn->hash_delete(use);
2571 imax -= use->replace_edge(n, m);
2572 igvn->hash_insert(use);
2573 record_for_optimizer(use);
2574 --i;
2575 #ifdef ASSERT
2576 } else if (use->is_Mem()) {
2577 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2578 // Don't move related cardmark.
2579 continue;
2580 }
2581 // Memory nodes should have new memory input.
2582 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2583 assert(tp != NULL, "ptr type");
2584 int idx = C->get_alias_index(tp);
2585 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2586 "Following memory nodes should have new memory input or be on the same memory slice");
2587 } else if (use->is_Phi()) {
2588 // Phi nodes should be split and moved already.
2589 tp = use->as_Phi()->adr_type()->isa_ptr();
2590 assert(tp != NULL, "ptr type");
2591 int idx = C->get_alias_index(tp);
2592 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2593 } else {
2594 use->dump();
2595 assert(false, "should not be here");
2596 #endif
2597 }
2598 }
2599 }
2601 //
2602 // Search memory chain of "mem" to find a MemNode whose address
2603 // is the specified alias index.
2604 //
2605 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2606 if (orig_mem == NULL)
2607 return orig_mem;
2608 Compile* C = _compile;
2609 PhaseGVN* igvn = _igvn;
2610 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2611 bool is_instance = (toop != NULL) && toop->is_known_instance();
2612 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2613 Node *prev = NULL;
2614 Node *result = orig_mem;
2615 while (prev != result) {
2616 prev = result;
2617 if (result == start_mem)
2618 break; // hit one of our sentinels
2619 if (result->is_Mem()) {
2620 const Type *at = igvn->type(result->in(MemNode::Address));
2621 if (at == Type::TOP)
2622 break; // Dead
2623 assert (at->isa_ptr() != NULL, "pointer type required.");
2624 int idx = C->get_alias_index(at->is_ptr());
2625 if (idx == alias_idx)
2626 break; // Found
2627 if (!is_instance && (at->isa_oopptr() == NULL ||
2628 !at->is_oopptr()->is_known_instance())) {
2629 break; // Do not skip store to general memory slice.
2630 }
2631 result = result->in(MemNode::Memory);
2632 }
2633 if (!is_instance)
2634 continue; // don't search further for non-instance types
2635 // skip over a call which does not affect this memory slice
2636 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2637 Node *proj_in = result->in(0);
2638 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2639 break; // hit one of our sentinels
2640 } else if (proj_in->is_Call()) {
2641 CallNode *call = proj_in->as_Call();
2642 if (!call->may_modify(toop, igvn)) {
2643 result = call->in(TypeFunc::Memory);
2644 }
2645 } else if (proj_in->is_Initialize()) {
2646 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2647 // Stop if this is the initialization for the object instance which
2648 // which contains this memory slice, otherwise skip over it.
2649 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2650 result = proj_in->in(TypeFunc::Memory);
2651 }
2652 } else if (proj_in->is_MemBar()) {
2653 result = proj_in->in(TypeFunc::Memory);
2654 }
2655 } else if (result->is_MergeMem()) {
2656 MergeMemNode *mmem = result->as_MergeMem();
2657 result = step_through_mergemem(mmem, alias_idx, toop);
2658 if (result == mmem->base_memory()) {
2659 // Didn't find instance memory, search through general slice recursively.
2660 result = mmem->memory_at(C->get_general_index(alias_idx));
2661 result = find_inst_mem(result, alias_idx, orig_phis);
2662 if (C->failing()) {
2663 return NULL;
2664 }
2665 mmem->set_memory_at(alias_idx, result);
2666 }
2667 } else if (result->is_Phi() &&
2668 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2669 Node *un = result->as_Phi()->unique_input(igvn);
2670 if (un != NULL) {
2671 orig_phis.append_if_missing(result->as_Phi());
2672 result = un;
2673 } else {
2674 break;
2675 }
2676 } else if (result->is_ClearArray()) {
2677 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2678 // Can not bypass initialization of the instance
2679 // we are looking for.
2680 break;
2681 }
2682 // Otherwise skip it (the call updated 'result' value).
2683 } else if (result->Opcode() == Op_SCMemProj) {
2684 Node* mem = result->in(0);
2685 Node* adr = NULL;
2686 if (mem->is_LoadStore()) {
2687 adr = mem->in(MemNode::Address);
2688 } else {
2689 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2690 adr = mem->in(3); // Memory edge corresponds to destination array
2691 }
2692 const Type *at = igvn->type(adr);
2693 if (at != Type::TOP) {
2694 assert (at->isa_ptr() != NULL, "pointer type required.");
2695 int idx = C->get_alias_index(at->is_ptr());
2696 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2697 break;
2698 }
2699 result = mem->in(MemNode::Memory);
2700 }
2701 }
2702 if (result->is_Phi()) {
2703 PhiNode *mphi = result->as_Phi();
2704 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2705 const TypePtr *t = mphi->adr_type();
2706 if (!is_instance) {
2707 // Push all non-instance Phis on the orig_phis worklist to update inputs
2708 // during Phase 4 if needed.
2709 orig_phis.append_if_missing(mphi);
2710 } else if (C->get_alias_index(t) != alias_idx) {
2711 // Create a new Phi with the specified alias index type.
2712 result = split_memory_phi(mphi, alias_idx, orig_phis);
2713 }
2714 }
2715 // the result is either MemNode, PhiNode, InitializeNode.
2716 return result;
2717 }
2719 //
2720 // Convert the types of unescaped object to instance types where possible,
2721 // propagate the new type information through the graph, and update memory
2722 // edges and MergeMem inputs to reflect the new type.
2723 //
2724 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2725 // The processing is done in 4 phases:
2726 //
2727 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2728 // types for the CheckCastPP for allocations where possible.
2729 // Propagate the the new types through users as follows:
2730 // casts and Phi: push users on alloc_worklist
2731 // AddP: cast Base and Address inputs to the instance type
2732 // push any AddP users on alloc_worklist and push any memnode
2733 // users onto memnode_worklist.
2734 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2735 // search the Memory chain for a store with the appropriate type
2736 // address type. If a Phi is found, create a new version with
2737 // the appropriate memory slices from each of the Phi inputs.
2738 // For stores, process the users as follows:
2739 // MemNode: push on memnode_worklist
2740 // MergeMem: push on mergemem_worklist
2741 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2742 // moving the first node encountered of each instance type to the
2743 // the input corresponding to its alias index.
2744 // appropriate memory slice.
2745 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2746 //
2747 // In the following example, the CheckCastPP nodes are the cast of allocation
2748 // results and the allocation of node 29 is unescaped and eligible to be an
2749 // instance type.
2750 //
2751 // We start with:
2752 //
2753 // 7 Parm #memory
2754 // 10 ConI "12"
2755 // 19 CheckCastPP "Foo"
2756 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2757 // 29 CheckCastPP "Foo"
2758 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2759 //
2760 // 40 StoreP 25 7 20 ... alias_index=4
2761 // 50 StoreP 35 40 30 ... alias_index=4
2762 // 60 StoreP 45 50 20 ... alias_index=4
2763 // 70 LoadP _ 60 30 ... alias_index=4
2764 // 80 Phi 75 50 60 Memory alias_index=4
2765 // 90 LoadP _ 80 30 ... alias_index=4
2766 // 100 LoadP _ 80 20 ... alias_index=4
2767 //
2768 //
2769 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2770 // and creating a new alias index for node 30. This gives:
2771 //
2772 // 7 Parm #memory
2773 // 10 ConI "12"
2774 // 19 CheckCastPP "Foo"
2775 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2776 // 29 CheckCastPP "Foo" iid=24
2777 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2778 //
2779 // 40 StoreP 25 7 20 ... alias_index=4
2780 // 50 StoreP 35 40 30 ... alias_index=6
2781 // 60 StoreP 45 50 20 ... alias_index=4
2782 // 70 LoadP _ 60 30 ... alias_index=6
2783 // 80 Phi 75 50 60 Memory alias_index=4
2784 // 90 LoadP _ 80 30 ... alias_index=6
2785 // 100 LoadP _ 80 20 ... alias_index=4
2786 //
2787 // In phase 2, new memory inputs are computed for the loads and stores,
2788 // And a new version of the phi is created. In phase 4, the inputs to
2789 // node 80 are updated and then the memory nodes are updated with the
2790 // values computed in phase 2. This results in:
2791 //
2792 // 7 Parm #memory
2793 // 10 ConI "12"
2794 // 19 CheckCastPP "Foo"
2795 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2796 // 29 CheckCastPP "Foo" iid=24
2797 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2798 //
2799 // 40 StoreP 25 7 20 ... alias_index=4
2800 // 50 StoreP 35 7 30 ... alias_index=6
2801 // 60 StoreP 45 40 20 ... alias_index=4
2802 // 70 LoadP _ 50 30 ... alias_index=6
2803 // 80 Phi 75 40 60 Memory alias_index=4
2804 // 120 Phi 75 50 50 Memory alias_index=6
2805 // 90 LoadP _ 120 30 ... alias_index=6
2806 // 100 LoadP _ 80 20 ... alias_index=4
2807 //
2808 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2809 GrowableArray<Node *> memnode_worklist;
2810 GrowableArray<PhiNode *> orig_phis;
2811 PhaseIterGVN *igvn = _igvn;
2812 uint new_index_start = (uint) _compile->num_alias_types();
2813 Arena* arena = Thread::current()->resource_area();
2814 VectorSet visited(arena);
2815 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2816 uint unique_old = _compile->unique();
2818 // Phase 1: Process possible allocations from alloc_worklist.
2819 // Create instance types for the CheckCastPP for allocations where possible.
2820 //
2821 // (Note: don't forget to change the order of the second AddP node on
2822 // the alloc_worklist if the order of the worklist processing is changed,
2823 // see the comment in find_second_addp().)
2824 //
2825 while (alloc_worklist.length() != 0) {
2826 Node *n = alloc_worklist.pop();
2827 uint ni = n->_idx;
2828 if (n->is_Call()) {
2829 CallNode *alloc = n->as_Call();
2830 // copy escape information to call node
2831 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2832 PointsToNode::EscapeState es = ptn->escape_state();
2833 // We have an allocation or call which returns a Java object,
2834 // see if it is unescaped.
2835 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2836 continue;
2837 // Find CheckCastPP for the allocate or for the return value of a call
2838 n = alloc->result_cast();
2839 if (n == NULL) { // No uses except Initialize node
2840 if (alloc->is_Allocate()) {
2841 // Set the scalar_replaceable flag for allocation
2842 // so it could be eliminated if it has no uses.
2843 alloc->as_Allocate()->_is_scalar_replaceable = true;
2844 }
2845 if (alloc->is_CallStaticJava()) {
2846 // Set the scalar_replaceable flag for boxing method
2847 // so it could be eliminated if it has no uses.
2848 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2849 }
2850 continue;
2851 }
2852 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2853 assert(!alloc->is_Allocate(), "allocation should have unique type");
2854 continue;
2855 }
2857 // The inline code for Object.clone() casts the allocation result to
2858 // java.lang.Object and then to the actual type of the allocated
2859 // object. Detect this case and use the second cast.
2860 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2861 // the allocation result is cast to java.lang.Object and then
2862 // to the actual Array type.
2863 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2864 && (alloc->is_AllocateArray() ||
2865 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2866 Node *cast2 = NULL;
2867 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2868 Node *use = n->fast_out(i);
2869 if (use->is_CheckCastPP()) {
2870 cast2 = use;
2871 break;
2872 }
2873 }
2874 if (cast2 != NULL) {
2875 n = cast2;
2876 } else {
2877 // Non-scalar replaceable if the allocation type is unknown statically
2878 // (reflection allocation), the object can't be restored during
2879 // deoptimization without precise type.
2880 continue;
2881 }
2882 }
2884 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2885 if (t == NULL)
2886 continue; // not a TypeOopPtr
2887 if (!t->klass_is_exact())
2888 continue; // not an unique type
2890 if (alloc->is_Allocate()) {
2891 // Set the scalar_replaceable flag for allocation
2892 // so it could be eliminated.
2893 alloc->as_Allocate()->_is_scalar_replaceable = true;
2894 }
2895 if (alloc->is_CallStaticJava()) {
2896 // Set the scalar_replaceable flag for boxing method
2897 // so it could be eliminated.
2898 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2899 }
2900 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2901 // in order for an object to be scalar-replaceable, it must be:
2902 // - a direct allocation (not a call returning an object)
2903 // - non-escaping
2904 // - eligible to be a unique type
2905 // - not determined to be ineligible by escape analysis
2906 set_map(alloc, n);
2907 set_map(n, alloc);
2908 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
2909 igvn->hash_delete(n);
2910 igvn->set_type(n, tinst);
2911 n->raise_bottom_type(tinst);
2912 igvn->hash_insert(n);
2913 record_for_optimizer(n);
2914 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2916 // First, put on the worklist all Field edges from Connection Graph
2917 // which is more accurate then putting immediate users from Ideal Graph.
2918 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2919 PointsToNode* tgt = e.get();
2920 Node* use = tgt->ideal_node();
2921 assert(tgt->is_Field() && use->is_AddP(),
2922 "only AddP nodes are Field edges in CG");
2923 if (use->outcnt() > 0) { // Don't process dead nodes
2924 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2925 if (addp2 != NULL) {
2926 assert(alloc->is_AllocateArray(),"array allocation was expected");
2927 alloc_worklist.append_if_missing(addp2);
2928 }
2929 alloc_worklist.append_if_missing(use);
2930 }
2931 }
2933 // An allocation may have an Initialize which has raw stores. Scan
2934 // the users of the raw allocation result and push AddP users
2935 // on alloc_worklist.
2936 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2937 assert (raw_result != NULL, "must have an allocation result");
2938 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2939 Node *use = raw_result->fast_out(i);
2940 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2941 Node* addp2 = find_second_addp(use, raw_result);
2942 if (addp2 != NULL) {
2943 assert(alloc->is_AllocateArray(),"array allocation was expected");
2944 alloc_worklist.append_if_missing(addp2);
2945 }
2946 alloc_worklist.append_if_missing(use);
2947 } else if (use->is_MemBar()) {
2948 memnode_worklist.append_if_missing(use);
2949 }
2950 }
2951 }
2952 } else if (n->is_AddP()) {
2953 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2954 if (jobj == NULL || jobj == phantom_obj) {
2955 #ifdef ASSERT
2956 ptnode_adr(get_addp_base(n)->_idx)->dump();
2957 ptnode_adr(n->_idx)->dump();
2958 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2959 #endif
2960 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2961 return;
2962 }
2963 Node *base = get_map(jobj->idx()); // CheckCastPP node
2964 if (!split_AddP(n, base)) continue; // wrong type from dead path
2965 } else if (n->is_Phi() ||
2966 n->is_CheckCastPP() ||
2967 n->is_EncodeP() ||
2968 n->is_DecodeN() ||
2969 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2970 if (visited.test_set(n->_idx)) {
2971 assert(n->is_Phi(), "loops only through Phi's");
2972 continue; // already processed
2973 }
2974 JavaObjectNode* jobj = unique_java_object(n);
2975 if (jobj == NULL || jobj == phantom_obj) {
2976 #ifdef ASSERT
2977 ptnode_adr(n->_idx)->dump();
2978 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2979 #endif
2980 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2981 return;
2982 } else {
2983 Node *val = get_map(jobj->idx()); // CheckCastPP node
2984 TypeNode *tn = n->as_Type();
2985 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2986 assert(tinst != NULL && tinst->is_known_instance() &&
2987 tinst->instance_id() == jobj->idx() , "instance type expected.");
2989 const Type *tn_type = igvn->type(tn);
2990 const TypeOopPtr *tn_t;
2991 if (tn_type->isa_narrowoop()) {
2992 tn_t = tn_type->make_ptr()->isa_oopptr();
2993 } else {
2994 tn_t = tn_type->isa_oopptr();
2995 }
2996 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2997 if (tn_type->isa_narrowoop()) {
2998 tn_type = tinst->make_narrowoop();
2999 } else {
3000 tn_type = tinst;
3001 }
3002 igvn->hash_delete(tn);
3003 igvn->set_type(tn, tn_type);
3004 tn->set_type(tn_type);
3005 igvn->hash_insert(tn);
3006 record_for_optimizer(n);
3007 } else {
3008 assert(tn_type == TypePtr::NULL_PTR ||
3009 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3010 "unexpected type");
3011 continue; // Skip dead path with different type
3012 }
3013 }
3014 } else {
3015 debug_only(n->dump();)
3016 assert(false, "EA: unexpected node");
3017 continue;
3018 }
3019 // push allocation's users on appropriate worklist
3020 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3021 Node *use = n->fast_out(i);
3022 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3023 // Load/store to instance's field
3024 memnode_worklist.append_if_missing(use);
3025 } else if (use->is_MemBar()) {
3026 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3027 memnode_worklist.append_if_missing(use);
3028 }
3029 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3030 Node* addp2 = find_second_addp(use, n);
3031 if (addp2 != NULL) {
3032 alloc_worklist.append_if_missing(addp2);
3033 }
3034 alloc_worklist.append_if_missing(use);
3035 } else if (use->is_Phi() ||
3036 use->is_CheckCastPP() ||
3037 use->is_EncodeNarrowPtr() ||
3038 use->is_DecodeNarrowPtr() ||
3039 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3040 alloc_worklist.append_if_missing(use);
3041 #ifdef ASSERT
3042 } else if (use->is_Mem()) {
3043 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3044 } else if (use->is_MergeMem()) {
3045 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3046 } else if (use->is_SafePoint()) {
3047 // Look for MergeMem nodes for calls which reference unique allocation
3048 // (through CheckCastPP nodes) even for debug info.
3049 Node* m = use->in(TypeFunc::Memory);
3050 if (m->is_MergeMem()) {
3051 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3052 }
3053 } else if (use->Opcode() == Op_EncodeISOArray) {
3054 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3055 // EncodeISOArray overwrites destination array
3056 memnode_worklist.append_if_missing(use);
3057 }
3058 } else {
3059 uint op = use->Opcode();
3060 if (!(op == Op_CmpP || op == Op_Conv2B ||
3061 op == Op_CastP2X || op == Op_StoreCM ||
3062 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
3063 op == Op_StrEquals || op == Op_StrIndexOf)) {
3064 n->dump();
3065 use->dump();
3066 assert(false, "EA: missing allocation reference path");
3067 }
3068 #endif
3069 }
3070 }
3072 }
3073 // New alias types were created in split_AddP().
3074 uint new_index_end = (uint) _compile->num_alias_types();
3075 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3077 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3078 // compute new values for Memory inputs (the Memory inputs are not
3079 // actually updated until phase 4.)
3080 if (memnode_worklist.length() == 0)
3081 return; // nothing to do
3082 while (memnode_worklist.length() != 0) {
3083 Node *n = memnode_worklist.pop();
3084 if (visited.test_set(n->_idx))
3085 continue;
3086 if (n->is_Phi() || n->is_ClearArray()) {
3087 // we don't need to do anything, but the users must be pushed
3088 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3089 // we don't need to do anything, but the users must be pushed
3090 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3091 if (n == NULL)
3092 continue;
3093 } else if (n->Opcode() == Op_EncodeISOArray) {
3094 // get the memory projection
3095 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3096 Node *use = n->fast_out(i);
3097 if (use->Opcode() == Op_SCMemProj) {
3098 n = use;
3099 break;
3100 }
3101 }
3102 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3103 } else {
3104 assert(n->is_Mem(), "memory node required.");
3105 Node *addr = n->in(MemNode::Address);
3106 const Type *addr_t = igvn->type(addr);
3107 if (addr_t == Type::TOP)
3108 continue;
3109 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3110 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3111 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3112 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3113 if (_compile->failing()) {
3114 return;
3115 }
3116 if (mem != n->in(MemNode::Memory)) {
3117 // We delay the memory edge update since we need old one in
3118 // MergeMem code below when instances memory slices are separated.
3119 set_map(n, mem);
3120 }
3121 if (n->is_Load()) {
3122 continue; // don't push users
3123 } else if (n->is_LoadStore()) {
3124 // get the memory projection
3125 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3126 Node *use = n->fast_out(i);
3127 if (use->Opcode() == Op_SCMemProj) {
3128 n = use;
3129 break;
3130 }
3131 }
3132 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3133 }
3134 }
3135 // push user on appropriate worklist
3136 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3137 Node *use = n->fast_out(i);
3138 if (use->is_Phi() || use->is_ClearArray()) {
3139 memnode_worklist.append_if_missing(use);
3140 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3141 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3142 continue;
3143 memnode_worklist.append_if_missing(use);
3144 } else if (use->is_MemBar()) {
3145 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3146 memnode_worklist.append_if_missing(use);
3147 }
3148 #ifdef ASSERT
3149 } else if(use->is_Mem()) {
3150 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3151 } else if (use->is_MergeMem()) {
3152 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3153 } else if (use->Opcode() == Op_EncodeISOArray) {
3154 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3155 // EncodeISOArray overwrites destination array
3156 memnode_worklist.append_if_missing(use);
3157 }
3158 } else {
3159 uint op = use->Opcode();
3160 if (!(op == Op_StoreCM ||
3161 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3162 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3163 op == Op_AryEq || op == Op_StrComp ||
3164 op == Op_StrEquals || op == Op_StrIndexOf)) {
3165 n->dump();
3166 use->dump();
3167 assert(false, "EA: missing memory path");
3168 }
3169 #endif
3170 }
3171 }
3172 }
3174 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3175 // Walk each memory slice moving the first node encountered of each
3176 // instance type to the the input corresponding to its alias index.
3177 uint length = _mergemem_worklist.length();
3178 for( uint next = 0; next < length; ++next ) {
3179 MergeMemNode* nmm = _mergemem_worklist.at(next);
3180 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3181 // Note: we don't want to use MergeMemStream here because we only want to
3182 // scan inputs which exist at the start, not ones we add during processing.
3183 // Note 2: MergeMem may already contains instance memory slices added
3184 // during find_inst_mem() call when memory nodes were processed above.
3185 igvn->hash_delete(nmm);
3186 uint nslices = MIN2(nmm->req(), new_index_start);
3187 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3188 Node* mem = nmm->in(i);
3189 Node* cur = NULL;
3190 if (mem == NULL || mem->is_top())
3191 continue;
3192 // First, update mergemem by moving memory nodes to corresponding slices
3193 // if their type became more precise since this mergemem was created.
3194 while (mem->is_Mem()) {
3195 const Type *at = igvn->type(mem->in(MemNode::Address));
3196 if (at != Type::TOP) {
3197 assert (at->isa_ptr() != NULL, "pointer type required.");
3198 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3199 if (idx == i) {
3200 if (cur == NULL)
3201 cur = mem;
3202 } else {
3203 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3204 nmm->set_memory_at(idx, mem);
3205 }
3206 }
3207 }
3208 mem = mem->in(MemNode::Memory);
3209 }
3210 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3211 // Find any instance of the current type if we haven't encountered
3212 // already a memory slice of the instance along the memory chain.
3213 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3214 if((uint)_compile->get_general_index(ni) == i) {
3215 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3216 if (nmm->is_empty_memory(m)) {
3217 Node* result = find_inst_mem(mem, ni, orig_phis);
3218 if (_compile->failing()) {
3219 return;
3220 }
3221 nmm->set_memory_at(ni, result);
3222 }
3223 }
3224 }
3225 }
3226 // Find the rest of instances values
3227 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3228 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3229 Node* result = step_through_mergemem(nmm, ni, tinst);
3230 if (result == nmm->base_memory()) {
3231 // Didn't find instance memory, search through general slice recursively.
3232 result = nmm->memory_at(_compile->get_general_index(ni));
3233 result = find_inst_mem(result, ni, orig_phis);
3234 if (_compile->failing()) {
3235 return;
3236 }
3237 nmm->set_memory_at(ni, result);
3238 }
3239 }
3240 igvn->hash_insert(nmm);
3241 record_for_optimizer(nmm);
3242 }
3244 // Phase 4: Update the inputs of non-instance memory Phis and
3245 // the Memory input of memnodes
3246 // First update the inputs of any non-instance Phi's from
3247 // which we split out an instance Phi. Note we don't have
3248 // to recursively process Phi's encounted on the input memory
3249 // chains as is done in split_memory_phi() since they will
3250 // also be processed here.
3251 for (int j = 0; j < orig_phis.length(); j++) {
3252 PhiNode *phi = orig_phis.at(j);
3253 int alias_idx = _compile->get_alias_index(phi->adr_type());
3254 igvn->hash_delete(phi);
3255 for (uint i = 1; i < phi->req(); i++) {
3256 Node *mem = phi->in(i);
3257 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3258 if (_compile->failing()) {
3259 return;
3260 }
3261 if (mem != new_mem) {
3262 phi->set_req(i, new_mem);
3263 }
3264 }
3265 igvn->hash_insert(phi);
3266 record_for_optimizer(phi);
3267 }
3269 // Update the memory inputs of MemNodes with the value we computed
3270 // in Phase 2 and move stores memory users to corresponding memory slices.
3271 // Disable memory split verification code until the fix for 6984348.
3272 // Currently it produces false negative results since it does not cover all cases.
3273 #if 0 // ifdef ASSERT
3274 visited.Reset();
3275 Node_Stack old_mems(arena, _compile->unique() >> 2);
3276 #endif
3277 for (uint i = 0; i < ideal_nodes.size(); i++) {
3278 Node* n = ideal_nodes.at(i);
3279 Node* nmem = get_map(n->_idx);
3280 assert(nmem != NULL, "sanity");
3281 if (n->is_Mem()) {
3282 #if 0 // ifdef ASSERT
3283 Node* old_mem = n->in(MemNode::Memory);
3284 if (!visited.test_set(old_mem->_idx)) {
3285 old_mems.push(old_mem, old_mem->outcnt());
3286 }
3287 #endif
3288 assert(n->in(MemNode::Memory) != nmem, "sanity");
3289 if (!n->is_Load()) {
3290 // Move memory users of a store first.
3291 move_inst_mem(n, orig_phis);
3292 }
3293 // Now update memory input
3294 igvn->hash_delete(n);
3295 n->set_req(MemNode::Memory, nmem);
3296 igvn->hash_insert(n);
3297 record_for_optimizer(n);
3298 } else {
3299 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3300 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3301 }
3302 }
3303 #if 0 // ifdef ASSERT
3304 // Verify that memory was split correctly
3305 while (old_mems.is_nonempty()) {
3306 Node* old_mem = old_mems.node();
3307 uint old_cnt = old_mems.index();
3308 old_mems.pop();
3309 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3310 }
3311 #endif
3312 }
3314 #ifndef PRODUCT
3315 static const char *node_type_names[] = {
3316 "UnknownType",
3317 "JavaObject",
3318 "LocalVar",
3319 "Field",
3320 "Arraycopy"
3321 };
3323 static const char *esc_names[] = {
3324 "UnknownEscape",
3325 "NoEscape",
3326 "ArgEscape",
3327 "GlobalEscape"
3328 };
3330 void PointsToNode::dump(bool print_state) const {
3331 NodeType nt = node_type();
3332 tty->print("%s ", node_type_names[(int) nt]);
3333 if (print_state) {
3334 EscapeState es = escape_state();
3335 EscapeState fields_es = fields_escape_state();
3336 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3337 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3338 tty->print("NSR ");
3339 }
3340 if (is_Field()) {
3341 FieldNode* f = (FieldNode*)this;
3342 if (f->is_oop())
3343 tty->print("oop ");
3344 if (f->offset() > 0)
3345 tty->print("+%d ", f->offset());
3346 tty->print("(");
3347 for (BaseIterator i(f); i.has_next(); i.next()) {
3348 PointsToNode* b = i.get();
3349 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3350 }
3351 tty->print(" )");
3352 }
3353 tty->print("[");
3354 for (EdgeIterator i(this); i.has_next(); i.next()) {
3355 PointsToNode* e = i.get();
3356 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3357 }
3358 tty->print(" [");
3359 for (UseIterator i(this); i.has_next(); i.next()) {
3360 PointsToNode* u = i.get();
3361 bool is_base = false;
3362 if (PointsToNode::is_base_use(u)) {
3363 is_base = true;
3364 u = PointsToNode::get_use_node(u)->as_Field();
3365 }
3366 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3367 }
3368 tty->print(" ]] ");
3369 if (_node == NULL)
3370 tty->print_cr("<null>");
3371 else
3372 _node->dump();
3373 }
3375 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3376 bool first = true;
3377 int ptnodes_length = ptnodes_worklist.length();
3378 for (int i = 0; i < ptnodes_length; i++) {
3379 PointsToNode *ptn = ptnodes_worklist.at(i);
3380 if (ptn == NULL || !ptn->is_JavaObject())
3381 continue;
3382 PointsToNode::EscapeState es = ptn->escape_state();
3383 if ((es != PointsToNode::NoEscape) && !Verbose) {
3384 continue;
3385 }
3386 Node* n = ptn->ideal_node();
3387 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3388 n->as_CallStaticJava()->is_boxing_method())) {
3389 if (first) {
3390 tty->cr();
3391 tty->print("======== Connection graph for ");
3392 _compile->method()->print_short_name();
3393 tty->cr();
3394 first = false;
3395 }
3396 ptn->dump();
3397 // Print all locals and fields which reference this allocation
3398 for (UseIterator j(ptn); j.has_next(); j.next()) {
3399 PointsToNode* use = j.get();
3400 if (use->is_LocalVar()) {
3401 use->dump(Verbose);
3402 } else if (Verbose) {
3403 use->dump();
3404 }
3405 }
3406 tty->cr();
3407 }
3408 }
3409 }
3410 #endif