Thu, 13 Jun 2013 22:02:40 -0700
8014431: cleanup warnings indicated by the -Wunused-value compiler option on linux
Reviewed-by: dholmes, coleenp
Contributed-by: jeremymanson@google.com, calvin.cheung@oracle.com
1 /*
2 * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.hpp"
30 #include "opto/c2compiler.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
38 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
39 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
40 _collecting(true),
41 _verify(false),
42 _compile(C),
43 _igvn(igvn),
44 _node_map(C->comp_arena()) {
45 // Add unknown java object.
46 add_java_object(C->top(), PointsToNode::GlobalEscape);
47 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
48 // Add ConP(#NULL) and ConN(#NULL) nodes.
49 Node* oop_null = igvn->zerocon(T_OBJECT);
50 assert(oop_null->_idx < nodes_size(), "should be created already");
51 add_java_object(oop_null, PointsToNode::NoEscape);
52 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
53 if (UseCompressedOops) {
54 Node* noop_null = igvn->zerocon(T_NARROWOOP);
55 assert(noop_null->_idx < nodes_size(), "should be created already");
56 map_ideal_node(noop_null, null_obj);
57 }
58 _pcmp_neq = NULL; // Should be initialized
59 _pcmp_eq = NULL;
60 }
62 bool ConnectionGraph::has_candidates(Compile *C) {
63 // EA brings benefits only when the code has allocations and/or locks which
64 // are represented by ideal Macro nodes.
65 int cnt = C->macro_count();
66 for (int i = 0; i < cnt; i++) {
67 Node *n = C->macro_node(i);
68 if (n->is_Allocate())
69 return true;
70 if (n->is_Lock()) {
71 Node* obj = n->as_Lock()->obj_node()->uncast();
72 if (!(obj->is_Parm() || obj->is_Con()))
73 return true;
74 }
75 if (n->is_CallStaticJava() &&
76 n->as_CallStaticJava()->is_boxing_method()) {
77 return true;
78 }
79 }
80 return false;
81 }
83 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
84 Compile::TracePhase t2("escapeAnalysis", &Phase::_t_escapeAnalysis, true);
85 ResourceMark rm;
87 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
88 // to create space for them in ConnectionGraph::_nodes[].
89 Node* oop_null = igvn->zerocon(T_OBJECT);
90 Node* noop_null = igvn->zerocon(T_NARROWOOP);
91 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
92 // Perform escape analysis
93 if (congraph->compute_escape()) {
94 // There are non escaping objects.
95 C->set_congraph(congraph);
96 }
97 // Cleanup.
98 if (oop_null->outcnt() == 0)
99 igvn->hash_delete(oop_null);
100 if (noop_null->outcnt() == 0)
101 igvn->hash_delete(noop_null);
102 }
104 bool ConnectionGraph::compute_escape() {
105 Compile* C = _compile;
106 PhaseGVN* igvn = _igvn;
108 // Worklists used by EA.
109 Unique_Node_List delayed_worklist;
110 GrowableArray<Node*> alloc_worklist;
111 GrowableArray<Node*> ptr_cmp_worklist;
112 GrowableArray<Node*> storestore_worklist;
113 GrowableArray<PointsToNode*> ptnodes_worklist;
114 GrowableArray<JavaObjectNode*> java_objects_worklist;
115 GrowableArray<JavaObjectNode*> non_escaped_worklist;
116 GrowableArray<FieldNode*> oop_fields_worklist;
117 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
119 { Compile::TracePhase t3("connectionGraph", &Phase::_t_connectionGraph, true);
121 // 1. Populate Connection Graph (CG) with PointsTo nodes.
122 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
123 // Initialize worklist
124 if (C->root() != NULL) {
125 ideal_nodes.push(C->root());
126 }
127 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
128 Node* n = ideal_nodes.at(next);
129 // Create PointsTo nodes and add them to Connection Graph. Called
130 // only once per ideal node since ideal_nodes is Unique_Node list.
131 add_node_to_connection_graph(n, &delayed_worklist);
132 PointsToNode* ptn = ptnode_adr(n->_idx);
133 if (ptn != NULL) {
134 ptnodes_worklist.append(ptn);
135 if (ptn->is_JavaObject()) {
136 java_objects_worklist.append(ptn->as_JavaObject());
137 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
138 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
139 // Only allocations and java static calls results are interesting.
140 non_escaped_worklist.append(ptn->as_JavaObject());
141 }
142 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
143 oop_fields_worklist.append(ptn->as_Field());
144 }
145 }
146 if (n->is_MergeMem()) {
147 // Collect all MergeMem nodes to add memory slices for
148 // scalar replaceable objects in split_unique_types().
149 _mergemem_worklist.append(n->as_MergeMem());
150 } else if (OptimizePtrCompare && n->is_Cmp() &&
151 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
152 // Collect compare pointers nodes.
153 ptr_cmp_worklist.append(n);
154 } else if (n->is_MemBarStoreStore()) {
155 // Collect all MemBarStoreStore nodes so that depending on the
156 // escape status of the associated Allocate node some of them
157 // may be eliminated.
158 storestore_worklist.append(n);
159 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
160 (n->req() > MemBarNode::Precedent)) {
161 record_for_optimizer(n);
162 #ifdef ASSERT
163 } else if (n->is_AddP()) {
164 // Collect address nodes for graph verification.
165 addp_worklist.append(n);
166 #endif
167 }
168 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
169 Node* m = n->fast_out(i); // Get user
170 ideal_nodes.push(m);
171 }
172 }
173 if (non_escaped_worklist.length() == 0) {
174 _collecting = false;
175 return false; // Nothing to do.
176 }
177 // Add final simple edges to graph.
178 while(delayed_worklist.size() > 0) {
179 Node* n = delayed_worklist.pop();
180 add_final_edges(n);
181 }
182 int ptnodes_length = ptnodes_worklist.length();
184 #ifdef ASSERT
185 if (VerifyConnectionGraph) {
186 // Verify that no new simple edges could be created and all
187 // local vars has edges.
188 _verify = true;
189 for (int next = 0; next < ptnodes_length; ++next) {
190 PointsToNode* ptn = ptnodes_worklist.at(next);
191 add_final_edges(ptn->ideal_node());
192 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
193 ptn->dump();
194 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
195 }
196 }
197 _verify = false;
198 }
199 #endif
201 // 2. Finish Graph construction by propagating references to all
202 // java objects through graph.
203 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
204 java_objects_worklist, oop_fields_worklist)) {
205 // All objects escaped or hit time or iterations limits.
206 _collecting = false;
207 return false;
208 }
210 // 3. Adjust scalar_replaceable state of nonescaping objects and push
211 // scalar replaceable allocations on alloc_worklist for processing
212 // in split_unique_types().
213 int non_escaped_length = non_escaped_worklist.length();
214 for (int next = 0; next < non_escaped_length; next++) {
215 JavaObjectNode* ptn = non_escaped_worklist.at(next);
216 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
217 Node* n = ptn->ideal_node();
218 if (n->is_Allocate()) {
219 n->as_Allocate()->_is_non_escaping = noescape;
220 }
221 if (n->is_CallStaticJava()) {
222 n->as_CallStaticJava()->_is_non_escaping = noescape;
223 }
224 if (noescape && ptn->scalar_replaceable()) {
225 adjust_scalar_replaceable_state(ptn);
226 if (ptn->scalar_replaceable()) {
227 alloc_worklist.append(ptn->ideal_node());
228 }
229 }
230 }
232 #ifdef ASSERT
233 if (VerifyConnectionGraph) {
234 // Verify that graph is complete - no new edges could be added or needed.
235 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
236 java_objects_worklist, addp_worklist);
237 }
238 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
239 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
240 null_obj->edge_count() == 0 &&
241 !null_obj->arraycopy_src() &&
242 !null_obj->arraycopy_dst(), "sanity");
243 #endif
245 _collecting = false;
247 } // TracePhase t3("connectionGraph")
249 // 4. Optimize ideal graph based on EA information.
250 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
251 if (has_non_escaping_obj) {
252 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
253 }
255 #ifndef PRODUCT
256 if (PrintEscapeAnalysis) {
257 dump(ptnodes_worklist); // Dump ConnectionGraph
258 }
259 #endif
261 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
262 #ifdef ASSERT
263 if (VerifyConnectionGraph) {
264 int alloc_length = alloc_worklist.length();
265 for (int next = 0; next < alloc_length; ++next) {
266 Node* n = alloc_worklist.at(next);
267 PointsToNode* ptn = ptnode_adr(n->_idx);
268 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
269 }
270 }
271 #endif
273 // 5. Separate memory graph for scalar replaceable allcations.
274 if (has_scalar_replaceable_candidates &&
275 C->AliasLevel() >= 3 && EliminateAllocations) {
276 // Now use the escape information to create unique types for
277 // scalar replaceable objects.
278 split_unique_types(alloc_worklist);
279 if (C->failing()) return false;
280 C->print_method(PHASE_AFTER_EA, 2);
282 #ifdef ASSERT
283 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
284 tty->print("=== No allocations eliminated for ");
285 C->method()->print_short_name();
286 if(!EliminateAllocations) {
287 tty->print(" since EliminateAllocations is off ===");
288 } else if(!has_scalar_replaceable_candidates) {
289 tty->print(" since there are no scalar replaceable candidates ===");
290 } else if(C->AliasLevel() < 3) {
291 tty->print(" since AliasLevel < 3 ===");
292 }
293 tty->cr();
294 #endif
295 }
296 return has_non_escaping_obj;
297 }
299 // Utility function for nodes that load an object
300 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
301 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
302 // ThreadLocal has RawPtr type.
303 const Type* t = _igvn->type(n);
304 if (t->make_ptr() != NULL) {
305 Node* adr = n->in(MemNode::Address);
306 #ifdef ASSERT
307 if (!adr->is_AddP()) {
308 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
309 } else {
310 assert((ptnode_adr(adr->_idx) == NULL ||
311 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
312 }
313 #endif
314 add_local_var_and_edge(n, PointsToNode::NoEscape,
315 adr, delayed_worklist);
316 }
317 }
319 // Populate Connection Graph with PointsTo nodes and create simple
320 // connection graph edges.
321 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
322 assert(!_verify, "this method sould not be called for verification");
323 PhaseGVN* igvn = _igvn;
324 uint n_idx = n->_idx;
325 PointsToNode* n_ptn = ptnode_adr(n_idx);
326 if (n_ptn != NULL)
327 return; // No need to redefine PointsTo node during first iteration.
329 if (n->is_Call()) {
330 // Arguments to allocation and locking don't escape.
331 if (n->is_AbstractLock()) {
332 // Put Lock and Unlock nodes on IGVN worklist to process them during
333 // first IGVN optimization when escape information is still available.
334 record_for_optimizer(n);
335 } else if (n->is_Allocate()) {
336 add_call_node(n->as_Call());
337 record_for_optimizer(n);
338 } else {
339 if (n->is_CallStaticJava()) {
340 const char* name = n->as_CallStaticJava()->_name;
341 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
342 return; // Skip uncommon traps
343 }
344 // Don't mark as processed since call's arguments have to be processed.
345 delayed_worklist->push(n);
346 // Check if a call returns an object.
347 if ((n->as_Call()->returns_pointer() &&
348 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) ||
349 (n->is_CallStaticJava() &&
350 n->as_CallStaticJava()->is_boxing_method())) {
351 add_call_node(n->as_Call());
352 }
353 }
354 return;
355 }
356 // Put this check here to process call arguments since some call nodes
357 // point to phantom_obj.
358 if (n_ptn == phantom_obj || n_ptn == null_obj)
359 return; // Skip predefined nodes.
361 int opcode = n->Opcode();
362 switch (opcode) {
363 case Op_AddP: {
364 Node* base = get_addp_base(n);
365 PointsToNode* ptn_base = ptnode_adr(base->_idx);
366 // Field nodes are created for all field types. They are used in
367 // adjust_scalar_replaceable_state() and split_unique_types().
368 // Note, non-oop fields will have only base edges in Connection
369 // Graph because such fields are not used for oop loads and stores.
370 int offset = address_offset(n, igvn);
371 add_field(n, PointsToNode::NoEscape, offset);
372 if (ptn_base == NULL) {
373 delayed_worklist->push(n); // Process it later.
374 } else {
375 n_ptn = ptnode_adr(n_idx);
376 add_base(n_ptn->as_Field(), ptn_base);
377 }
378 break;
379 }
380 case Op_CastX2P: {
381 map_ideal_node(n, phantom_obj);
382 break;
383 }
384 case Op_CastPP:
385 case Op_CheckCastPP:
386 case Op_EncodeP:
387 case Op_DecodeN:
388 case Op_EncodePKlass:
389 case Op_DecodeNKlass: {
390 add_local_var_and_edge(n, PointsToNode::NoEscape,
391 n->in(1), delayed_worklist);
392 break;
393 }
394 case Op_CMoveP: {
395 add_local_var(n, PointsToNode::NoEscape);
396 // Do not add edges during first iteration because some could be
397 // not defined yet.
398 delayed_worklist->push(n);
399 break;
400 }
401 case Op_ConP:
402 case Op_ConN:
403 case Op_ConNKlass: {
404 // assume all oop constants globally escape except for null
405 PointsToNode::EscapeState es;
406 const Type* t = igvn->type(n);
407 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
408 es = PointsToNode::NoEscape;
409 } else {
410 es = PointsToNode::GlobalEscape;
411 }
412 add_java_object(n, es);
413 break;
414 }
415 case Op_CreateEx: {
416 // assume that all exception objects globally escape
417 add_java_object(n, PointsToNode::GlobalEscape);
418 break;
419 }
420 case Op_LoadKlass:
421 case Op_LoadNKlass: {
422 // Unknown class is loaded
423 map_ideal_node(n, phantom_obj);
424 break;
425 }
426 case Op_LoadP:
427 case Op_LoadN:
428 case Op_LoadPLocked: {
429 add_objload_to_connection_graph(n, delayed_worklist);
430 break;
431 }
432 case Op_Parm: {
433 map_ideal_node(n, phantom_obj);
434 break;
435 }
436 case Op_PartialSubtypeCheck: {
437 // Produces Null or notNull and is used in only in CmpP so
438 // phantom_obj could be used.
439 map_ideal_node(n, phantom_obj); // Result is unknown
440 break;
441 }
442 case Op_Phi: {
443 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
444 // ThreadLocal has RawPtr type.
445 const Type* t = n->as_Phi()->type();
446 if (t->make_ptr() != NULL) {
447 add_local_var(n, PointsToNode::NoEscape);
448 // Do not add edges during first iteration because some could be
449 // not defined yet.
450 delayed_worklist->push(n);
451 }
452 break;
453 }
454 case Op_Proj: {
455 // we are only interested in the oop result projection from a call
456 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
457 n->in(0)->as_Call()->returns_pointer()) {
458 add_local_var_and_edge(n, PointsToNode::NoEscape,
459 n->in(0), delayed_worklist);
460 }
461 break;
462 }
463 case Op_Rethrow: // Exception object escapes
464 case Op_Return: {
465 if (n->req() > TypeFunc::Parms &&
466 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
467 // Treat Return value as LocalVar with GlobalEscape escape state.
468 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
469 n->in(TypeFunc::Parms), delayed_worklist);
470 }
471 break;
472 }
473 case Op_GetAndSetP:
474 case Op_GetAndSetN: {
475 add_objload_to_connection_graph(n, delayed_worklist);
476 // fallthrough
477 }
478 case Op_StoreP:
479 case Op_StoreN:
480 case Op_StoreNKlass:
481 case Op_StorePConditional:
482 case Op_CompareAndSwapP:
483 case Op_CompareAndSwapN: {
484 Node* adr = n->in(MemNode::Address);
485 const Type *adr_type = igvn->type(adr);
486 adr_type = adr_type->make_ptr();
487 if (adr_type == NULL) {
488 break; // skip dead nodes
489 }
490 if (adr_type->isa_oopptr() ||
491 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
492 (adr_type == TypeRawPtr::NOTNULL &&
493 adr->in(AddPNode::Address)->is_Proj() &&
494 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
495 delayed_worklist->push(n); // Process it later.
496 #ifdef ASSERT
497 assert(adr->is_AddP(), "expecting an AddP");
498 if (adr_type == TypeRawPtr::NOTNULL) {
499 // Verify a raw address for a store captured by Initialize node.
500 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
501 assert(offs != Type::OffsetBot, "offset must be a constant");
502 }
503 #endif
504 } else {
505 // Ignore copy the displaced header to the BoxNode (OSR compilation).
506 if (adr->is_BoxLock())
507 break;
508 // Stored value escapes in unsafe access.
509 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
510 // Pointer stores in G1 barriers looks like unsafe access.
511 // Ignore such stores to be able scalar replace non-escaping
512 // allocations.
513 if (UseG1GC && adr->is_AddP()) {
514 Node* base = get_addp_base(adr);
515 if (base->Opcode() == Op_LoadP &&
516 base->in(MemNode::Address)->is_AddP()) {
517 adr = base->in(MemNode::Address);
518 Node* tls = get_addp_base(adr);
519 if (tls->Opcode() == Op_ThreadLocal) {
520 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
521 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() +
522 PtrQueue::byte_offset_of_buf())) {
523 break; // G1 pre barier previous oop value store.
524 }
525 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() +
526 PtrQueue::byte_offset_of_buf())) {
527 break; // G1 post barier card address store.
528 }
529 }
530 }
531 }
532 delayed_worklist->push(n); // Process unsafe access later.
533 break;
534 }
535 #ifdef ASSERT
536 n->dump(1);
537 assert(false, "not unsafe or G1 barrier raw StoreP");
538 #endif
539 }
540 break;
541 }
542 case Op_AryEq:
543 case Op_StrComp:
544 case Op_StrEquals:
545 case Op_StrIndexOf:
546 case Op_EncodeISOArray: {
547 add_local_var(n, PointsToNode::ArgEscape);
548 delayed_worklist->push(n); // Process it later.
549 break;
550 }
551 case Op_ThreadLocal: {
552 add_java_object(n, PointsToNode::ArgEscape);
553 break;
554 }
555 default:
556 ; // Do nothing for nodes not related to EA.
557 }
558 return;
559 }
561 #ifdef ASSERT
562 #define ELSE_FAIL(name) \
563 /* Should not be called for not pointer type. */ \
564 n->dump(1); \
565 assert(false, name); \
566 break;
567 #else
568 #define ELSE_FAIL(name) \
569 break;
570 #endif
572 // Add final simple edges to graph.
573 void ConnectionGraph::add_final_edges(Node *n) {
574 PointsToNode* n_ptn = ptnode_adr(n->_idx);
575 #ifdef ASSERT
576 if (_verify && n_ptn->is_JavaObject())
577 return; // This method does not change graph for JavaObject.
578 #endif
580 if (n->is_Call()) {
581 process_call_arguments(n->as_Call());
582 return;
583 }
584 assert(n->is_Store() || n->is_LoadStore() ||
585 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
586 "node should be registered already");
587 int opcode = n->Opcode();
588 switch (opcode) {
589 case Op_AddP: {
590 Node* base = get_addp_base(n);
591 PointsToNode* ptn_base = ptnode_adr(base->_idx);
592 assert(ptn_base != NULL, "field's base should be registered");
593 add_base(n_ptn->as_Field(), ptn_base);
594 break;
595 }
596 case Op_CastPP:
597 case Op_CheckCastPP:
598 case Op_EncodeP:
599 case Op_DecodeN:
600 case Op_EncodePKlass:
601 case Op_DecodeNKlass: {
602 add_local_var_and_edge(n, PointsToNode::NoEscape,
603 n->in(1), NULL);
604 break;
605 }
606 case Op_CMoveP: {
607 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
608 Node* in = n->in(i);
609 if (in == NULL)
610 continue; // ignore NULL
611 Node* uncast_in = in->uncast();
612 if (uncast_in->is_top() || uncast_in == n)
613 continue; // ignore top or inputs which go back this node
614 PointsToNode* ptn = ptnode_adr(in->_idx);
615 assert(ptn != NULL, "node should be registered");
616 add_edge(n_ptn, ptn);
617 }
618 break;
619 }
620 case Op_LoadP:
621 case Op_LoadN:
622 case Op_LoadPLocked: {
623 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
624 // ThreadLocal has RawPtr type.
625 const Type* t = _igvn->type(n);
626 if (t->make_ptr() != NULL) {
627 Node* adr = n->in(MemNode::Address);
628 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
629 break;
630 }
631 ELSE_FAIL("Op_LoadP");
632 }
633 case Op_Phi: {
634 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
635 // ThreadLocal has RawPtr type.
636 const Type* t = n->as_Phi()->type();
637 if (t->make_ptr() != NULL) {
638 for (uint i = 1; i < n->req(); i++) {
639 Node* in = n->in(i);
640 if (in == NULL)
641 continue; // ignore NULL
642 Node* uncast_in = in->uncast();
643 if (uncast_in->is_top() || uncast_in == n)
644 continue; // ignore top or inputs which go back this node
645 PointsToNode* ptn = ptnode_adr(in->_idx);
646 assert(ptn != NULL, "node should be registered");
647 add_edge(n_ptn, ptn);
648 }
649 break;
650 }
651 ELSE_FAIL("Op_Phi");
652 }
653 case Op_Proj: {
654 // we are only interested in the oop result projection from a call
655 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
656 n->in(0)->as_Call()->returns_pointer()) {
657 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
658 break;
659 }
660 ELSE_FAIL("Op_Proj");
661 }
662 case Op_Rethrow: // Exception object escapes
663 case Op_Return: {
664 if (n->req() > TypeFunc::Parms &&
665 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
666 // Treat Return value as LocalVar with GlobalEscape escape state.
667 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
668 n->in(TypeFunc::Parms), NULL);
669 break;
670 }
671 ELSE_FAIL("Op_Return");
672 }
673 case Op_StoreP:
674 case Op_StoreN:
675 case Op_StoreNKlass:
676 case Op_StorePConditional:
677 case Op_CompareAndSwapP:
678 case Op_CompareAndSwapN:
679 case Op_GetAndSetP:
680 case Op_GetAndSetN: {
681 Node* adr = n->in(MemNode::Address);
682 const Type *adr_type = _igvn->type(adr);
683 adr_type = adr_type->make_ptr();
684 #ifdef ASSERT
685 if (adr_type == NULL) {
686 n->dump(1);
687 assert(adr_type != NULL, "dead node should not be on list");
688 break;
689 }
690 #endif
691 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) {
692 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
693 }
694 if (adr_type->isa_oopptr() ||
695 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
696 (adr_type == TypeRawPtr::NOTNULL &&
697 adr->in(AddPNode::Address)->is_Proj() &&
698 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
699 // Point Address to Value
700 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
701 assert(adr_ptn != NULL &&
702 adr_ptn->as_Field()->is_oop(), "node should be registered");
703 Node *val = n->in(MemNode::ValueIn);
704 PointsToNode* ptn = ptnode_adr(val->_idx);
705 assert(ptn != NULL, "node should be registered");
706 add_edge(adr_ptn, ptn);
707 break;
708 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
709 // Stored value escapes in unsafe access.
710 Node *val = n->in(MemNode::ValueIn);
711 PointsToNode* ptn = ptnode_adr(val->_idx);
712 assert(ptn != NULL, "node should be registered");
713 ptn->set_escape_state(PointsToNode::GlobalEscape);
714 // Add edge to object for unsafe access with offset.
715 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
716 assert(adr_ptn != NULL, "node should be registered");
717 if (adr_ptn->is_Field()) {
718 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
719 add_edge(adr_ptn, ptn);
720 }
721 break;
722 }
723 ELSE_FAIL("Op_StoreP");
724 }
725 case Op_AryEq:
726 case Op_StrComp:
727 case Op_StrEquals:
728 case Op_StrIndexOf:
729 case Op_EncodeISOArray: {
730 // char[] arrays passed to string intrinsic do not escape but
731 // they are not scalar replaceable. Adjust escape state for them.
732 // Start from in(2) edge since in(1) is memory edge.
733 for (uint i = 2; i < n->req(); i++) {
734 Node* adr = n->in(i);
735 const Type* at = _igvn->type(adr);
736 if (!adr->is_top() && at->isa_ptr()) {
737 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
738 at->isa_ptr() != NULL, "expecting a pointer");
739 if (adr->is_AddP()) {
740 adr = get_addp_base(adr);
741 }
742 PointsToNode* ptn = ptnode_adr(adr->_idx);
743 assert(ptn != NULL, "node should be registered");
744 add_edge(n_ptn, ptn);
745 }
746 }
747 break;
748 }
749 default: {
750 // This method should be called only for EA specific nodes which may
751 // miss some edges when they were created.
752 #ifdef ASSERT
753 n->dump(1);
754 #endif
755 guarantee(false, "unknown node");
756 }
757 }
758 return;
759 }
761 void ConnectionGraph::add_call_node(CallNode* call) {
762 assert(call->returns_pointer(), "only for call which returns pointer");
763 uint call_idx = call->_idx;
764 if (call->is_Allocate()) {
765 Node* k = call->in(AllocateNode::KlassNode);
766 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
767 assert(kt != NULL, "TypeKlassPtr required.");
768 ciKlass* cik = kt->klass();
769 PointsToNode::EscapeState es = PointsToNode::NoEscape;
770 bool scalar_replaceable = true;
771 if (call->is_AllocateArray()) {
772 if (!cik->is_array_klass()) { // StressReflectiveCode
773 es = PointsToNode::GlobalEscape;
774 } else {
775 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
776 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
777 // Not scalar replaceable if the length is not constant or too big.
778 scalar_replaceable = false;
779 }
780 }
781 } else { // Allocate instance
782 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
783 !cik->is_instance_klass() || // StressReflectiveCode
784 cik->as_instance_klass()->has_finalizer()) {
785 es = PointsToNode::GlobalEscape;
786 }
787 }
788 add_java_object(call, es);
789 PointsToNode* ptn = ptnode_adr(call_idx);
790 if (!scalar_replaceable && ptn->scalar_replaceable()) {
791 ptn->set_scalar_replaceable(false);
792 }
793 } else if (call->is_CallStaticJava()) {
794 // Call nodes could be different types:
795 //
796 // 1. CallDynamicJavaNode (what happened during call is unknown):
797 //
798 // - mapped to GlobalEscape JavaObject node if oop is returned;
799 //
800 // - all oop arguments are escaping globally;
801 //
802 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
803 //
804 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
805 //
806 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
807 // - mapped to NoEscape JavaObject node if non-escaping object allocated
808 // during call is returned;
809 // - mapped to ArgEscape LocalVar node pointed to object arguments
810 // which are returned and does not escape during call;
811 //
812 // - oop arguments escaping status is defined by bytecode analysis;
813 //
814 // For a static call, we know exactly what method is being called.
815 // Use bytecode estimator to record whether the call's return value escapes.
816 ciMethod* meth = call->as_CallJava()->method();
817 if (meth == NULL) {
818 const char* name = call->as_CallStaticJava()->_name;
819 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
820 // Returns a newly allocated unescaped object.
821 add_java_object(call, PointsToNode::NoEscape);
822 ptnode_adr(call_idx)->set_scalar_replaceable(false);
823 } else if (meth->is_boxing_method()) {
824 // Returns boxing object
825 PointsToNode::EscapeState es;
826 vmIntrinsics::ID intr = meth->intrinsic_id();
827 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
828 // It does not escape if object is always allocated.
829 es = PointsToNode::NoEscape;
830 } else {
831 // It escapes globally if object could be loaded from cache.
832 es = PointsToNode::GlobalEscape;
833 }
834 add_java_object(call, es);
835 } else {
836 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
837 call_analyzer->copy_dependencies(_compile->dependencies());
838 if (call_analyzer->is_return_allocated()) {
839 // Returns a newly allocated unescaped object, simply
840 // update dependency information.
841 // Mark it as NoEscape so that objects referenced by
842 // it's fields will be marked as NoEscape at least.
843 add_java_object(call, PointsToNode::NoEscape);
844 ptnode_adr(call_idx)->set_scalar_replaceable(false);
845 } else {
846 // Determine whether any arguments are returned.
847 const TypeTuple* d = call->tf()->domain();
848 bool ret_arg = false;
849 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
850 if (d->field_at(i)->isa_ptr() != NULL &&
851 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
852 ret_arg = true;
853 break;
854 }
855 }
856 if (ret_arg) {
857 add_local_var(call, PointsToNode::ArgEscape);
858 } else {
859 // Returns unknown object.
860 map_ideal_node(call, phantom_obj);
861 }
862 }
863 }
864 } else {
865 // An other type of call, assume the worst case:
866 // returned value is unknown and globally escapes.
867 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
868 map_ideal_node(call, phantom_obj);
869 }
870 }
872 void ConnectionGraph::process_call_arguments(CallNode *call) {
873 bool is_arraycopy = false;
874 switch (call->Opcode()) {
875 #ifdef ASSERT
876 case Op_Allocate:
877 case Op_AllocateArray:
878 case Op_Lock:
879 case Op_Unlock:
880 assert(false, "should be done already");
881 break;
882 #endif
883 case Op_CallLeafNoFP:
884 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
885 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
886 // fall through
887 case Op_CallLeaf: {
888 // Stub calls, objects do not escape but they are not scale replaceable.
889 // Adjust escape state for outgoing arguments.
890 const TypeTuple * d = call->tf()->domain();
891 bool src_has_oops = false;
892 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
893 const Type* at = d->field_at(i);
894 Node *arg = call->in(i);
895 const Type *aat = _igvn->type(arg);
896 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
897 continue;
898 if (arg->is_AddP()) {
899 //
900 // The inline_native_clone() case when the arraycopy stub is called
901 // after the allocation before Initialize and CheckCastPP nodes.
902 // Or normal arraycopy for object arrays case.
903 //
904 // Set AddP's base (Allocate) as not scalar replaceable since
905 // pointer to the base (with offset) is passed as argument.
906 //
907 arg = get_addp_base(arg);
908 }
909 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
910 assert(arg_ptn != NULL, "should be registered");
911 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
912 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
913 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
914 aat->isa_ptr() != NULL, "expecting an Ptr");
915 bool arg_has_oops = aat->isa_oopptr() &&
916 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
917 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
918 if (i == TypeFunc::Parms) {
919 src_has_oops = arg_has_oops;
920 }
921 //
922 // src or dst could be j.l.Object when other is basic type array:
923 //
924 // arraycopy(char[],0,Object*,0,size);
925 // arraycopy(Object*,0,char[],0,size);
926 //
927 // Don't add edges in such cases.
928 //
929 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
930 arg_has_oops && (i > TypeFunc::Parms);
931 #ifdef ASSERT
932 if (!(is_arraycopy ||
933 (call->as_CallLeaf()->_name != NULL &&
934 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
935 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
936 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
937 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
938 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
939 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0)
940 ))) {
941 call->dump();
942 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
943 }
944 #endif
945 // Always process arraycopy's destination object since
946 // we need to add all possible edges to references in
947 // source object.
948 if (arg_esc >= PointsToNode::ArgEscape &&
949 !arg_is_arraycopy_dest) {
950 continue;
951 }
952 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
953 if (arg_is_arraycopy_dest) {
954 Node* src = call->in(TypeFunc::Parms);
955 if (src->is_AddP()) {
956 src = get_addp_base(src);
957 }
958 PointsToNode* src_ptn = ptnode_adr(src->_idx);
959 assert(src_ptn != NULL, "should be registered");
960 if (arg_ptn != src_ptn) {
961 // Special arraycopy edge:
962 // A destination object's field can't have the source object
963 // as base since objects escape states are not related.
964 // Only escape state of destination object's fields affects
965 // escape state of fields in source object.
966 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
967 }
968 }
969 }
970 }
971 break;
972 }
973 case Op_CallStaticJava: {
974 // For a static call, we know exactly what method is being called.
975 // Use bytecode estimator to record the call's escape affects
976 #ifdef ASSERT
977 const char* name = call->as_CallStaticJava()->_name;
978 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
979 #endif
980 ciMethod* meth = call->as_CallJava()->method();
981 if ((meth != NULL) && meth->is_boxing_method()) {
982 break; // Boxing methods do not modify any oops.
983 }
984 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
985 // fall-through if not a Java method or no analyzer information
986 if (call_analyzer != NULL) {
987 PointsToNode* call_ptn = ptnode_adr(call->_idx);
988 const TypeTuple* d = call->tf()->domain();
989 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
990 const Type* at = d->field_at(i);
991 int k = i - TypeFunc::Parms;
992 Node* arg = call->in(i);
993 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
994 if (at->isa_ptr() != NULL &&
995 call_analyzer->is_arg_returned(k)) {
996 // The call returns arguments.
997 if (call_ptn != NULL) { // Is call's result used?
998 assert(call_ptn->is_LocalVar(), "node should be registered");
999 assert(arg_ptn != NULL, "node should be registered");
1000 add_edge(call_ptn, arg_ptn);
1001 }
1002 }
1003 if (at->isa_oopptr() != NULL &&
1004 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1005 if (!call_analyzer->is_arg_stack(k)) {
1006 // The argument global escapes
1007 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1008 } else {
1009 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1010 if (!call_analyzer->is_arg_local(k)) {
1011 // The argument itself doesn't escape, but any fields might
1012 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1013 }
1014 }
1015 }
1016 }
1017 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1018 // The call returns arguments.
1019 assert(call_ptn->edge_count() > 0, "sanity");
1020 if (!call_analyzer->is_return_local()) {
1021 // Returns also unknown object.
1022 add_edge(call_ptn, phantom_obj);
1023 }
1024 }
1025 break;
1026 }
1027 }
1028 default: {
1029 // Fall-through here if not a Java method or no analyzer information
1030 // or some other type of call, assume the worst case: all arguments
1031 // globally escape.
1032 const TypeTuple* d = call->tf()->domain();
1033 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1034 const Type* at = d->field_at(i);
1035 if (at->isa_oopptr() != NULL) {
1036 Node* arg = call->in(i);
1037 if (arg->is_AddP()) {
1038 arg = get_addp_base(arg);
1039 }
1040 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1041 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1042 }
1043 }
1044 }
1045 }
1046 }
1049 // Finish Graph construction.
1050 bool ConnectionGraph::complete_connection_graph(
1051 GrowableArray<PointsToNode*>& ptnodes_worklist,
1052 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1053 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1054 GrowableArray<FieldNode*>& oop_fields_worklist) {
1055 // Normally only 1-3 passes needed to build Connection Graph depending
1056 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1057 // Set limit to 20 to catch situation when something did go wrong and
1058 // bailout Escape Analysis.
1059 // Also limit build time to 30 sec (60 in debug VM).
1060 #define CG_BUILD_ITER_LIMIT 20
1061 #ifdef ASSERT
1062 #define CG_BUILD_TIME_LIMIT 60.0
1063 #else
1064 #define CG_BUILD_TIME_LIMIT 30.0
1065 #endif
1067 // Propagate GlobalEscape and ArgEscape escape states and check that
1068 // we still have non-escaping objects. The method pushs on _worklist
1069 // Field nodes which reference phantom_object.
1070 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1071 return false; // Nothing to do.
1072 }
1073 // Now propagate references to all JavaObject nodes.
1074 int java_objects_length = java_objects_worklist.length();
1075 elapsedTimer time;
1076 int new_edges = 1;
1077 int iterations = 0;
1078 do {
1079 while ((new_edges > 0) &&
1080 (iterations++ < CG_BUILD_ITER_LIMIT) &&
1081 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1082 time.start();
1083 new_edges = 0;
1084 // Propagate references to phantom_object for nodes pushed on _worklist
1085 // by find_non_escaped_objects() and find_field_value().
1086 new_edges += add_java_object_edges(phantom_obj, false);
1087 for (int next = 0; next < java_objects_length; ++next) {
1088 JavaObjectNode* ptn = java_objects_worklist.at(next);
1089 new_edges += add_java_object_edges(ptn, true);
1090 }
1091 if (new_edges > 0) {
1092 // Update escape states on each iteration if graph was updated.
1093 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1094 return false; // Nothing to do.
1095 }
1096 }
1097 time.stop();
1098 }
1099 if ((iterations < CG_BUILD_ITER_LIMIT) &&
1100 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1101 time.start();
1102 // Find fields which have unknown value.
1103 int fields_length = oop_fields_worklist.length();
1104 for (int next = 0; next < fields_length; next++) {
1105 FieldNode* field = oop_fields_worklist.at(next);
1106 if (field->edge_count() == 0) {
1107 new_edges += find_field_value(field);
1108 // This code may added new edges to phantom_object.
1109 // Need an other cycle to propagate references to phantom_object.
1110 }
1111 }
1112 time.stop();
1113 } else {
1114 new_edges = 0; // Bailout
1115 }
1116 } while (new_edges > 0);
1118 // Bailout if passed limits.
1119 if ((iterations >= CG_BUILD_ITER_LIMIT) ||
1120 (time.seconds() >= CG_BUILD_TIME_LIMIT)) {
1121 Compile* C = _compile;
1122 if (C->log() != NULL) {
1123 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1124 C->log()->text("%s", (iterations >= CG_BUILD_ITER_LIMIT) ? "iterations" : "time");
1125 C->log()->end_elem(" limit'");
1126 }
1127 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1128 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1129 // Possible infinite build_connection_graph loop,
1130 // bailout (no changes to ideal graph were made).
1131 return false;
1132 }
1133 #ifdef ASSERT
1134 if (Verbose && PrintEscapeAnalysis) {
1135 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1136 iterations, nodes_size(), ptnodes_worklist.length());
1137 }
1138 #endif
1140 #undef CG_BUILD_ITER_LIMIT
1141 #undef CG_BUILD_TIME_LIMIT
1143 // Find fields initialized by NULL for non-escaping Allocations.
1144 int non_escaped_length = non_escaped_worklist.length();
1145 for (int next = 0; next < non_escaped_length; next++) {
1146 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1147 PointsToNode::EscapeState es = ptn->escape_state();
1148 assert(es <= PointsToNode::ArgEscape, "sanity");
1149 if (es == PointsToNode::NoEscape) {
1150 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1151 // Adding references to NULL object does not change escape states
1152 // since it does not escape. Also no fields are added to NULL object.
1153 add_java_object_edges(null_obj, false);
1154 }
1155 }
1156 Node* n = ptn->ideal_node();
1157 if (n->is_Allocate()) {
1158 // The object allocated by this Allocate node will never be
1159 // seen by an other thread. Mark it so that when it is
1160 // expanded no MemBarStoreStore is added.
1161 InitializeNode* ini = n->as_Allocate()->initialization();
1162 if (ini != NULL)
1163 ini->set_does_not_escape();
1164 }
1165 }
1166 return true; // Finished graph construction.
1167 }
1169 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1170 // and check that we still have non-escaping java objects.
1171 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1172 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1173 GrowableArray<PointsToNode*> escape_worklist;
1174 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1175 int ptnodes_length = ptnodes_worklist.length();
1176 for (int next = 0; next < ptnodes_length; ++next) {
1177 PointsToNode* ptn = ptnodes_worklist.at(next);
1178 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1179 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1180 escape_worklist.push(ptn);
1181 }
1182 }
1183 // Set escape states to referenced nodes (edges list).
1184 while (escape_worklist.length() > 0) {
1185 PointsToNode* ptn = escape_worklist.pop();
1186 PointsToNode::EscapeState es = ptn->escape_state();
1187 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1188 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1189 es >= PointsToNode::ArgEscape) {
1190 // GlobalEscape or ArgEscape state of field means it has unknown value.
1191 if (add_edge(ptn, phantom_obj)) {
1192 // New edge was added
1193 add_field_uses_to_worklist(ptn->as_Field());
1194 }
1195 }
1196 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1197 PointsToNode* e = i.get();
1198 if (e->is_Arraycopy()) {
1199 assert(ptn->arraycopy_dst(), "sanity");
1200 // Propagate only fields escape state through arraycopy edge.
1201 if (e->fields_escape_state() < field_es) {
1202 set_fields_escape_state(e, field_es);
1203 escape_worklist.push(e);
1204 }
1205 } else if (es >= field_es) {
1206 // fields_escape_state is also set to 'es' if it is less than 'es'.
1207 if (e->escape_state() < es) {
1208 set_escape_state(e, es);
1209 escape_worklist.push(e);
1210 }
1211 } else {
1212 // Propagate field escape state.
1213 bool es_changed = false;
1214 if (e->fields_escape_state() < field_es) {
1215 set_fields_escape_state(e, field_es);
1216 es_changed = true;
1217 }
1218 if ((e->escape_state() < field_es) &&
1219 e->is_Field() && ptn->is_JavaObject() &&
1220 e->as_Field()->is_oop()) {
1221 // Change escape state of referenced fileds.
1222 set_escape_state(e, field_es);
1223 es_changed = true;;
1224 } else if (e->escape_state() < es) {
1225 set_escape_state(e, es);
1226 es_changed = true;;
1227 }
1228 if (es_changed) {
1229 escape_worklist.push(e);
1230 }
1231 }
1232 }
1233 }
1234 // Remove escaped objects from non_escaped list.
1235 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1236 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1237 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1238 non_escaped_worklist.delete_at(next);
1239 }
1240 if (ptn->escape_state() == PointsToNode::NoEscape) {
1241 // Find fields in non-escaped allocations which have unknown value.
1242 find_init_values(ptn, phantom_obj, NULL);
1243 }
1244 }
1245 return (non_escaped_worklist.length() > 0);
1246 }
1248 // Add all references to JavaObject node by walking over all uses.
1249 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1250 int new_edges = 0;
1251 if (populate_worklist) {
1252 // Populate _worklist by uses of jobj's uses.
1253 for (UseIterator i(jobj); i.has_next(); i.next()) {
1254 PointsToNode* use = i.get();
1255 if (use->is_Arraycopy())
1256 continue;
1257 add_uses_to_worklist(use);
1258 if (use->is_Field() && use->as_Field()->is_oop()) {
1259 // Put on worklist all field's uses (loads) and
1260 // related field nodes (same base and offset).
1261 add_field_uses_to_worklist(use->as_Field());
1262 }
1263 }
1264 }
1265 while(_worklist.length() > 0) {
1266 PointsToNode* use = _worklist.pop();
1267 if (PointsToNode::is_base_use(use)) {
1268 // Add reference from jobj to field and from field to jobj (field's base).
1269 use = PointsToNode::get_use_node(use)->as_Field();
1270 if (add_base(use->as_Field(), jobj)) {
1271 new_edges++;
1272 }
1273 continue;
1274 }
1275 assert(!use->is_JavaObject(), "sanity");
1276 if (use->is_Arraycopy()) {
1277 if (jobj == null_obj) // NULL object does not have field edges
1278 continue;
1279 // Added edge from Arraycopy node to arraycopy's source java object
1280 if (add_edge(use, jobj)) {
1281 jobj->set_arraycopy_src();
1282 new_edges++;
1283 }
1284 // and stop here.
1285 continue;
1286 }
1287 if (!add_edge(use, jobj))
1288 continue; // No new edge added, there was such edge already.
1289 new_edges++;
1290 if (use->is_LocalVar()) {
1291 add_uses_to_worklist(use);
1292 if (use->arraycopy_dst()) {
1293 for (EdgeIterator i(use); i.has_next(); i.next()) {
1294 PointsToNode* e = i.get();
1295 if (e->is_Arraycopy()) {
1296 if (jobj == null_obj) // NULL object does not have field edges
1297 continue;
1298 // Add edge from arraycopy's destination java object to Arraycopy node.
1299 if (add_edge(jobj, e)) {
1300 new_edges++;
1301 jobj->set_arraycopy_dst();
1302 }
1303 }
1304 }
1305 }
1306 } else {
1307 // Added new edge to stored in field values.
1308 // Put on worklist all field's uses (loads) and
1309 // related field nodes (same base and offset).
1310 add_field_uses_to_worklist(use->as_Field());
1311 }
1312 }
1313 return new_edges;
1314 }
1316 // Put on worklist all related field nodes.
1317 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1318 assert(field->is_oop(), "sanity");
1319 int offset = field->offset();
1320 add_uses_to_worklist(field);
1321 // Loop over all bases of this field and push on worklist Field nodes
1322 // with the same offset and base (since they may reference the same field).
1323 for (BaseIterator i(field); i.has_next(); i.next()) {
1324 PointsToNode* base = i.get();
1325 add_fields_to_worklist(field, base);
1326 // Check if the base was source object of arraycopy and go over arraycopy's
1327 // destination objects since values stored to a field of source object are
1328 // accessable by uses (loads) of fields of destination objects.
1329 if (base->arraycopy_src()) {
1330 for (UseIterator j(base); j.has_next(); j.next()) {
1331 PointsToNode* arycp = j.get();
1332 if (arycp->is_Arraycopy()) {
1333 for (UseIterator k(arycp); k.has_next(); k.next()) {
1334 PointsToNode* abase = k.get();
1335 if (abase->arraycopy_dst() && abase != base) {
1336 // Look for the same arracopy reference.
1337 add_fields_to_worklist(field, abase);
1338 }
1339 }
1340 }
1341 }
1342 }
1343 }
1344 }
1346 // Put on worklist all related field nodes.
1347 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1348 int offset = field->offset();
1349 if (base->is_LocalVar()) {
1350 for (UseIterator j(base); j.has_next(); j.next()) {
1351 PointsToNode* f = j.get();
1352 if (PointsToNode::is_base_use(f)) { // Field
1353 f = PointsToNode::get_use_node(f);
1354 if (f == field || !f->as_Field()->is_oop())
1355 continue;
1356 int offs = f->as_Field()->offset();
1357 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1358 add_to_worklist(f);
1359 }
1360 }
1361 }
1362 } else {
1363 assert(base->is_JavaObject(), "sanity");
1364 if (// Skip phantom_object since it is only used to indicate that
1365 // this field's content globally escapes.
1366 (base != phantom_obj) &&
1367 // NULL object node does not have fields.
1368 (base != null_obj)) {
1369 for (EdgeIterator i(base); i.has_next(); i.next()) {
1370 PointsToNode* f = i.get();
1371 // Skip arraycopy edge since store to destination object field
1372 // does not update value in source object field.
1373 if (f->is_Arraycopy()) {
1374 assert(base->arraycopy_dst(), "sanity");
1375 continue;
1376 }
1377 if (f == field || !f->as_Field()->is_oop())
1378 continue;
1379 int offs = f->as_Field()->offset();
1380 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1381 add_to_worklist(f);
1382 }
1383 }
1384 }
1385 }
1386 }
1388 // Find fields which have unknown value.
1389 int ConnectionGraph::find_field_value(FieldNode* field) {
1390 // Escaped fields should have init value already.
1391 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1392 int new_edges = 0;
1393 for (BaseIterator i(field); i.has_next(); i.next()) {
1394 PointsToNode* base = i.get();
1395 if (base->is_JavaObject()) {
1396 // Skip Allocate's fields which will be processed later.
1397 if (base->ideal_node()->is_Allocate())
1398 return 0;
1399 assert(base == null_obj, "only NULL ptr base expected here");
1400 }
1401 }
1402 if (add_edge(field, phantom_obj)) {
1403 // New edge was added
1404 new_edges++;
1405 add_field_uses_to_worklist(field);
1406 }
1407 return new_edges;
1408 }
1410 // Find fields initializing values for allocations.
1411 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1412 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1413 int new_edges = 0;
1414 Node* alloc = pta->ideal_node();
1415 if (init_val == phantom_obj) {
1416 // Do nothing for Allocate nodes since its fields values are "known".
1417 if (alloc->is_Allocate())
1418 return 0;
1419 assert(alloc->as_CallStaticJava(), "sanity");
1420 #ifdef ASSERT
1421 if (alloc->as_CallStaticJava()->method() == NULL) {
1422 const char* name = alloc->as_CallStaticJava()->_name;
1423 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1424 }
1425 #endif
1426 // Non-escaped allocation returned from Java or runtime call have
1427 // unknown values in fields.
1428 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1429 PointsToNode* field = i.get();
1430 if (field->is_Field() && field->as_Field()->is_oop()) {
1431 if (add_edge(field, phantom_obj)) {
1432 // New edge was added
1433 new_edges++;
1434 add_field_uses_to_worklist(field->as_Field());
1435 }
1436 }
1437 }
1438 return new_edges;
1439 }
1440 assert(init_val == null_obj, "sanity");
1441 // Do nothing for Call nodes since its fields values are unknown.
1442 if (!alloc->is_Allocate())
1443 return 0;
1445 InitializeNode* ini = alloc->as_Allocate()->initialization();
1446 Compile* C = _compile;
1447 bool visited_bottom_offset = false;
1448 GrowableArray<int> offsets_worklist;
1450 // Check if an oop field's initializing value is recorded and add
1451 // a corresponding NULL if field's value if it is not recorded.
1452 // Connection Graph does not record a default initialization by NULL
1453 // captured by Initialize node.
1454 //
1455 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1456 PointsToNode* field = i.get(); // Field (AddP)
1457 if (!field->is_Field() || !field->as_Field()->is_oop())
1458 continue; // Not oop field
1459 int offset = field->as_Field()->offset();
1460 if (offset == Type::OffsetBot) {
1461 if (!visited_bottom_offset) {
1462 // OffsetBot is used to reference array's element,
1463 // always add reference to NULL to all Field nodes since we don't
1464 // known which element is referenced.
1465 if (add_edge(field, null_obj)) {
1466 // New edge was added
1467 new_edges++;
1468 add_field_uses_to_worklist(field->as_Field());
1469 visited_bottom_offset = true;
1470 }
1471 }
1472 } else {
1473 // Check only oop fields.
1474 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1475 if (adr_type->isa_rawptr()) {
1476 #ifdef ASSERT
1477 // Raw pointers are used for initializing stores so skip it
1478 // since it should be recorded already
1479 Node* base = get_addp_base(field->ideal_node());
1480 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1481 (base->in(0) == alloc),"unexpected pointer type");
1482 #endif
1483 continue;
1484 }
1485 if (!offsets_worklist.contains(offset)) {
1486 offsets_worklist.append(offset);
1487 Node* value = NULL;
1488 if (ini != NULL) {
1489 // StoreP::memory_type() == T_ADDRESS
1490 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1491 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1492 // Make sure initializing store has the same type as this AddP.
1493 // This AddP may reference non existing field because it is on a
1494 // dead branch of bimorphic call which is not eliminated yet.
1495 if (store != NULL && store->is_Store() &&
1496 store->as_Store()->memory_type() == ft) {
1497 value = store->in(MemNode::ValueIn);
1498 #ifdef ASSERT
1499 if (VerifyConnectionGraph) {
1500 // Verify that AddP already points to all objects the value points to.
1501 PointsToNode* val = ptnode_adr(value->_idx);
1502 assert((val != NULL), "should be processed already");
1503 PointsToNode* missed_obj = NULL;
1504 if (val->is_JavaObject()) {
1505 if (!field->points_to(val->as_JavaObject())) {
1506 missed_obj = val;
1507 }
1508 } else {
1509 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1510 tty->print_cr("----------init store has invalid value -----");
1511 store->dump();
1512 val->dump();
1513 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1514 }
1515 for (EdgeIterator j(val); j.has_next(); j.next()) {
1516 PointsToNode* obj = j.get();
1517 if (obj->is_JavaObject()) {
1518 if (!field->points_to(obj->as_JavaObject())) {
1519 missed_obj = obj;
1520 break;
1521 }
1522 }
1523 }
1524 }
1525 if (missed_obj != NULL) {
1526 tty->print_cr("----------field---------------------------------");
1527 field->dump();
1528 tty->print_cr("----------missed referernce to object-----------");
1529 missed_obj->dump();
1530 tty->print_cr("----------object referernced by init store -----");
1531 store->dump();
1532 val->dump();
1533 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1534 }
1535 }
1536 #endif
1537 } else {
1538 // There could be initializing stores which follow allocation.
1539 // For example, a volatile field store is not collected
1540 // by Initialize node.
1541 //
1542 // Need to check for dependent loads to separate such stores from
1543 // stores which follow loads. For now, add initial value NULL so
1544 // that compare pointers optimization works correctly.
1545 }
1546 }
1547 if (value == NULL) {
1548 // A field's initializing value was not recorded. Add NULL.
1549 if (add_edge(field, null_obj)) {
1550 // New edge was added
1551 new_edges++;
1552 add_field_uses_to_worklist(field->as_Field());
1553 }
1554 }
1555 }
1556 }
1557 }
1558 return new_edges;
1559 }
1561 // Adjust scalar_replaceable state after Connection Graph is built.
1562 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1563 // Search for non-escaping objects which are not scalar replaceable
1564 // and mark them to propagate the state to referenced objects.
1566 // 1. An object is not scalar replaceable if the field into which it is
1567 // stored has unknown offset (stored into unknown element of an array).
1568 //
1569 for (UseIterator i(jobj); i.has_next(); i.next()) {
1570 PointsToNode* use = i.get();
1571 assert(!use->is_Arraycopy(), "sanity");
1572 if (use->is_Field()) {
1573 FieldNode* field = use->as_Field();
1574 assert(field->is_oop() && field->scalar_replaceable() &&
1575 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1576 if (field->offset() == Type::OffsetBot) {
1577 jobj->set_scalar_replaceable(false);
1578 return;
1579 }
1580 }
1581 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1582 // 2. An object is not scalar replaceable if it is merged with other objects.
1583 for (EdgeIterator j(use); j.has_next(); j.next()) {
1584 PointsToNode* ptn = j.get();
1585 if (ptn->is_JavaObject() && ptn != jobj) {
1586 // Mark all objects.
1587 jobj->set_scalar_replaceable(false);
1588 ptn->set_scalar_replaceable(false);
1589 }
1590 }
1591 if (!jobj->scalar_replaceable()) {
1592 return;
1593 }
1594 }
1596 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1597 // Non-escaping object node should point only to field nodes.
1598 FieldNode* field = j.get()->as_Field();
1599 int offset = field->as_Field()->offset();
1601 // 3. An object is not scalar replaceable if it has a field with unknown
1602 // offset (array's element is accessed in loop).
1603 if (offset == Type::OffsetBot) {
1604 jobj->set_scalar_replaceable(false);
1605 return;
1606 }
1607 // 4. Currently an object is not scalar replaceable if a LoadStore node
1608 // access its field since the field value is unknown after it.
1609 //
1610 Node* n = field->ideal_node();
1611 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1612 if (n->fast_out(i)->is_LoadStore()) {
1613 jobj->set_scalar_replaceable(false);
1614 return;
1615 }
1616 }
1618 // 5. Or the address may point to more then one object. This may produce
1619 // the false positive result (set not scalar replaceable)
1620 // since the flow-insensitive escape analysis can't separate
1621 // the case when stores overwrite the field's value from the case
1622 // when stores happened on different control branches.
1623 //
1624 // Note: it will disable scalar replacement in some cases:
1625 //
1626 // Point p[] = new Point[1];
1627 // p[0] = new Point(); // Will be not scalar replaced
1628 //
1629 // but it will save us from incorrect optimizations in next cases:
1630 //
1631 // Point p[] = new Point[1];
1632 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1633 //
1634 if (field->base_count() > 1) {
1635 for (BaseIterator i(field); i.has_next(); i.next()) {
1636 PointsToNode* base = i.get();
1637 // Don't take into account LocalVar nodes which
1638 // may point to only one object which should be also
1639 // this field's base by now.
1640 if (base->is_JavaObject() && base != jobj) {
1641 // Mark all bases.
1642 jobj->set_scalar_replaceable(false);
1643 base->set_scalar_replaceable(false);
1644 }
1645 }
1646 }
1647 }
1648 }
1650 #ifdef ASSERT
1651 void ConnectionGraph::verify_connection_graph(
1652 GrowableArray<PointsToNode*>& ptnodes_worklist,
1653 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1654 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1655 GrowableArray<Node*>& addp_worklist) {
1656 // Verify that graph is complete - no new edges could be added.
1657 int java_objects_length = java_objects_worklist.length();
1658 int non_escaped_length = non_escaped_worklist.length();
1659 int new_edges = 0;
1660 for (int next = 0; next < java_objects_length; ++next) {
1661 JavaObjectNode* ptn = java_objects_worklist.at(next);
1662 new_edges += add_java_object_edges(ptn, true);
1663 }
1664 assert(new_edges == 0, "graph was not complete");
1665 // Verify that escape state is final.
1666 int length = non_escaped_worklist.length();
1667 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1668 assert((non_escaped_length == non_escaped_worklist.length()) &&
1669 (non_escaped_length == length) &&
1670 (_worklist.length() == 0), "escape state was not final");
1672 // Verify fields information.
1673 int addp_length = addp_worklist.length();
1674 for (int next = 0; next < addp_length; ++next ) {
1675 Node* n = addp_worklist.at(next);
1676 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1677 if (field->is_oop()) {
1678 // Verify that field has all bases
1679 Node* base = get_addp_base(n);
1680 PointsToNode* ptn = ptnode_adr(base->_idx);
1681 if (ptn->is_JavaObject()) {
1682 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1683 } else {
1684 assert(ptn->is_LocalVar(), "sanity");
1685 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1686 PointsToNode* e = i.get();
1687 if (e->is_JavaObject()) {
1688 assert(field->has_base(e->as_JavaObject()), "sanity");
1689 }
1690 }
1691 }
1692 // Verify that all fields have initializing values.
1693 if (field->edge_count() == 0) {
1694 tty->print_cr("----------field does not have references----------");
1695 field->dump();
1696 for (BaseIterator i(field); i.has_next(); i.next()) {
1697 PointsToNode* base = i.get();
1698 tty->print_cr("----------field has next base---------------------");
1699 base->dump();
1700 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1701 tty->print_cr("----------base has fields-------------------------");
1702 for (EdgeIterator j(base); j.has_next(); j.next()) {
1703 j.get()->dump();
1704 }
1705 tty->print_cr("----------base has references---------------------");
1706 for (UseIterator j(base); j.has_next(); j.next()) {
1707 j.get()->dump();
1708 }
1709 }
1710 }
1711 for (UseIterator i(field); i.has_next(); i.next()) {
1712 i.get()->dump();
1713 }
1714 assert(field->edge_count() > 0, "sanity");
1715 }
1716 }
1717 }
1718 }
1719 #endif
1721 // Optimize ideal graph.
1722 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1723 GrowableArray<Node*>& storestore_worklist) {
1724 Compile* C = _compile;
1725 PhaseIterGVN* igvn = _igvn;
1726 if (EliminateLocks) {
1727 // Mark locks before changing ideal graph.
1728 int cnt = C->macro_count();
1729 for( int i=0; i < cnt; i++ ) {
1730 Node *n = C->macro_node(i);
1731 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1732 AbstractLockNode* alock = n->as_AbstractLock();
1733 if (!alock->is_non_esc_obj()) {
1734 if (not_global_escape(alock->obj_node())) {
1735 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1736 // The lock could be marked eliminated by lock coarsening
1737 // code during first IGVN before EA. Replace coarsened flag
1738 // to eliminate all associated locks/unlocks.
1739 alock->set_non_esc_obj();
1740 }
1741 }
1742 }
1743 }
1744 }
1746 if (OptimizePtrCompare) {
1747 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1748 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1749 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1750 // Optimize objects compare.
1751 while (ptr_cmp_worklist.length() != 0) {
1752 Node *n = ptr_cmp_worklist.pop();
1753 Node *res = optimize_ptr_compare(n);
1754 if (res != NULL) {
1755 #ifndef PRODUCT
1756 if (PrintOptimizePtrCompare) {
1757 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"));
1758 if (Verbose) {
1759 n->dump(1);
1760 }
1761 }
1762 #endif
1763 igvn->replace_node(n, res);
1764 }
1765 }
1766 // cleanup
1767 if (_pcmp_neq->outcnt() == 0)
1768 igvn->hash_delete(_pcmp_neq);
1769 if (_pcmp_eq->outcnt() == 0)
1770 igvn->hash_delete(_pcmp_eq);
1771 }
1773 // For MemBarStoreStore nodes added in library_call.cpp, check
1774 // escape status of associated AllocateNode and optimize out
1775 // MemBarStoreStore node if the allocated object never escapes.
1776 while (storestore_worklist.length() != 0) {
1777 Node *n = storestore_worklist.pop();
1778 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1779 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1780 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1781 if (not_global_escape(alloc)) {
1782 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1783 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1784 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1785 igvn->register_new_node_with_optimizer(mb);
1786 igvn->replace_node(storestore, mb);
1787 }
1788 }
1789 }
1791 // Optimize objects compare.
1792 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1793 assert(OptimizePtrCompare, "sanity");
1794 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1795 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1796 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1797 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1798 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1799 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1801 // Check simple cases first.
1802 if (jobj1 != NULL) {
1803 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1804 if (jobj1 == jobj2) {
1805 // Comparing the same not escaping object.
1806 return _pcmp_eq;
1807 }
1808 Node* obj = jobj1->ideal_node();
1809 // Comparing not escaping allocation.
1810 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1811 !ptn2->points_to(jobj1)) {
1812 return _pcmp_neq; // This includes nullness check.
1813 }
1814 }
1815 }
1816 if (jobj2 != NULL) {
1817 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1818 Node* obj = jobj2->ideal_node();
1819 // Comparing not escaping allocation.
1820 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1821 !ptn1->points_to(jobj2)) {
1822 return _pcmp_neq; // This includes nullness check.
1823 }
1824 }
1825 }
1826 if (jobj1 != NULL && jobj1 != phantom_obj &&
1827 jobj2 != NULL && jobj2 != phantom_obj &&
1828 jobj1->ideal_node()->is_Con() &&
1829 jobj2->ideal_node()->is_Con()) {
1830 // Klass or String constants compare. Need to be careful with
1831 // compressed pointers - compare types of ConN and ConP instead of nodes.
1832 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1833 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1834 if (t1->make_ptr() == t2->make_ptr()) {
1835 return _pcmp_eq;
1836 } else {
1837 return _pcmp_neq;
1838 }
1839 }
1840 if (ptn1->meet(ptn2)) {
1841 return NULL; // Sets are not disjoint
1842 }
1844 // Sets are disjoint.
1845 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1846 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1847 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1848 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1849 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1850 set2_has_unknown_ptr && set1_has_null_ptr) {
1851 // Check nullness of unknown object.
1852 return NULL;
1853 }
1855 // Disjointness by itself is not sufficient since
1856 // alias analysis is not complete for escaped objects.
1857 // Disjoint sets are definitely unrelated only when
1858 // at least one set has only not escaping allocations.
1859 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1860 if (ptn1->non_escaping_allocation()) {
1861 return _pcmp_neq;
1862 }
1863 }
1864 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1865 if (ptn2->non_escaping_allocation()) {
1866 return _pcmp_neq;
1867 }
1868 }
1869 return NULL;
1870 }
1872 // Connection Graph constuction functions.
1874 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1875 PointsToNode* ptadr = _nodes.at(n->_idx);
1876 if (ptadr != NULL) {
1877 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1878 return;
1879 }
1880 Compile* C = _compile;
1881 ptadr = new (C->comp_arena()) LocalVarNode(C, n, es);
1882 _nodes.at_put(n->_idx, ptadr);
1883 }
1885 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1886 PointsToNode* ptadr = _nodes.at(n->_idx);
1887 if (ptadr != NULL) {
1888 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1889 return;
1890 }
1891 Compile* C = _compile;
1892 ptadr = new (C->comp_arena()) JavaObjectNode(C, n, es);
1893 _nodes.at_put(n->_idx, ptadr);
1894 }
1896 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1897 PointsToNode* ptadr = _nodes.at(n->_idx);
1898 if (ptadr != NULL) {
1899 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1900 return;
1901 }
1902 bool unsafe = false;
1903 bool is_oop = is_oop_field(n, offset, &unsafe);
1904 if (unsafe) {
1905 es = PointsToNode::GlobalEscape;
1906 }
1907 Compile* C = _compile;
1908 FieldNode* field = new (C->comp_arena()) FieldNode(C, n, es, offset, is_oop);
1909 _nodes.at_put(n->_idx, field);
1910 }
1912 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1913 PointsToNode* src, PointsToNode* dst) {
1914 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1915 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1916 PointsToNode* ptadr = _nodes.at(n->_idx);
1917 if (ptadr != NULL) {
1918 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1919 return;
1920 }
1921 Compile* C = _compile;
1922 ptadr = new (C->comp_arena()) ArraycopyNode(C, n, es);
1923 _nodes.at_put(n->_idx, ptadr);
1924 // Add edge from arraycopy node to source object.
1925 (void)add_edge(ptadr, src);
1926 src->set_arraycopy_src();
1927 // Add edge from destination object to arraycopy node.
1928 (void)add_edge(dst, ptadr);
1929 dst->set_arraycopy_dst();
1930 }
1932 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1933 const Type* adr_type = n->as_AddP()->bottom_type();
1934 BasicType bt = T_INT;
1935 if (offset == Type::OffsetBot) {
1936 // Check only oop fields.
1937 if (!adr_type->isa_aryptr() ||
1938 (adr_type->isa_aryptr()->klass() == NULL) ||
1939 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1940 // OffsetBot is used to reference array's element. Ignore first AddP.
1941 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
1942 bt = T_OBJECT;
1943 }
1944 }
1945 } else if (offset != oopDesc::klass_offset_in_bytes()) {
1946 if (adr_type->isa_instptr()) {
1947 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
1948 if (field != NULL) {
1949 bt = field->layout_type();
1950 } else {
1951 // Check for unsafe oop field access
1952 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1953 int opcode = n->fast_out(i)->Opcode();
1954 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1955 opcode == Op_StoreN || opcode == Op_LoadN) {
1956 bt = T_OBJECT;
1957 (*unsafe) = true;
1958 break;
1959 }
1960 }
1961 }
1962 } else if (adr_type->isa_aryptr()) {
1963 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1964 // Ignore array length load.
1965 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
1966 // Ignore first AddP.
1967 } else {
1968 const Type* elemtype = adr_type->isa_aryptr()->elem();
1969 bt = elemtype->array_element_basic_type();
1970 }
1971 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
1972 // Allocation initialization, ThreadLocal field access, unsafe access
1973 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1974 int opcode = n->fast_out(i)->Opcode();
1975 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1976 opcode == Op_StoreN || opcode == Op_LoadN) {
1977 bt = T_OBJECT;
1978 break;
1979 }
1980 }
1981 }
1982 }
1983 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
1984 }
1986 // Returns unique pointed java object or NULL.
1987 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
1988 assert(!_collecting, "should not call when contructed graph");
1989 // If the node was created after the escape computation we can't answer.
1990 uint idx = n->_idx;
1991 if (idx >= nodes_size()) {
1992 return NULL;
1993 }
1994 PointsToNode* ptn = ptnode_adr(idx);
1995 if (ptn->is_JavaObject()) {
1996 return ptn->as_JavaObject();
1997 }
1998 assert(ptn->is_LocalVar(), "sanity");
1999 // Check all java objects it points to.
2000 JavaObjectNode* jobj = NULL;
2001 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2002 PointsToNode* e = i.get();
2003 if (e->is_JavaObject()) {
2004 if (jobj == NULL) {
2005 jobj = e->as_JavaObject();
2006 } else if (jobj != e) {
2007 return NULL;
2008 }
2009 }
2010 }
2011 return jobj;
2012 }
2014 // Return true if this node points only to non-escaping allocations.
2015 bool PointsToNode::non_escaping_allocation() {
2016 if (is_JavaObject()) {
2017 Node* n = ideal_node();
2018 if (n->is_Allocate() || n->is_CallStaticJava()) {
2019 return (escape_state() == PointsToNode::NoEscape);
2020 } else {
2021 return false;
2022 }
2023 }
2024 assert(is_LocalVar(), "sanity");
2025 // Check all java objects it points to.
2026 for (EdgeIterator i(this); i.has_next(); i.next()) {
2027 PointsToNode* e = i.get();
2028 if (e->is_JavaObject()) {
2029 Node* n = e->ideal_node();
2030 if ((e->escape_state() != PointsToNode::NoEscape) ||
2031 !(n->is_Allocate() || n->is_CallStaticJava())) {
2032 return false;
2033 }
2034 }
2035 }
2036 return true;
2037 }
2039 // Return true if we know the node does not escape globally.
2040 bool ConnectionGraph::not_global_escape(Node *n) {
2041 assert(!_collecting, "should not call during graph construction");
2042 // If the node was created after the escape computation we can't answer.
2043 uint idx = n->_idx;
2044 if (idx >= nodes_size()) {
2045 return false;
2046 }
2047 PointsToNode* ptn = ptnode_adr(idx);
2048 PointsToNode::EscapeState es = ptn->escape_state();
2049 // If we have already computed a value, return it.
2050 if (es >= PointsToNode::GlobalEscape)
2051 return false;
2052 if (ptn->is_JavaObject()) {
2053 return true; // (es < PointsToNode::GlobalEscape);
2054 }
2055 assert(ptn->is_LocalVar(), "sanity");
2056 // Check all java objects it points to.
2057 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2058 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2059 return false;
2060 }
2061 return true;
2062 }
2065 // Helper functions
2067 // Return true if this node points to specified node or nodes it points to.
2068 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2069 if (is_JavaObject()) {
2070 return (this == ptn);
2071 }
2072 assert(is_LocalVar() || is_Field(), "sanity");
2073 for (EdgeIterator i(this); i.has_next(); i.next()) {
2074 if (i.get() == ptn)
2075 return true;
2076 }
2077 return false;
2078 }
2080 // Return true if one node points to an other.
2081 bool PointsToNode::meet(PointsToNode* ptn) {
2082 if (this == ptn) {
2083 return true;
2084 } else if (ptn->is_JavaObject()) {
2085 return this->points_to(ptn->as_JavaObject());
2086 } else if (this->is_JavaObject()) {
2087 return ptn->points_to(this->as_JavaObject());
2088 }
2089 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2090 int ptn_count = ptn->edge_count();
2091 for (EdgeIterator i(this); i.has_next(); i.next()) {
2092 PointsToNode* this_e = i.get();
2093 for (int j = 0; j < ptn_count; j++) {
2094 if (this_e == ptn->edge(j))
2095 return true;
2096 }
2097 }
2098 return false;
2099 }
2101 #ifdef ASSERT
2102 // Return true if bases point to this java object.
2103 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2104 for (BaseIterator i(this); i.has_next(); i.next()) {
2105 if (i.get() == jobj)
2106 return true;
2107 }
2108 return false;
2109 }
2110 #endif
2112 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2113 const Type *adr_type = phase->type(adr);
2114 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2115 adr->in(AddPNode::Address)->is_Proj() &&
2116 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2117 // We are computing a raw address for a store captured by an Initialize
2118 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2119 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2120 assert(offs != Type::OffsetBot ||
2121 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2122 "offset must be a constant or it is initialization of array");
2123 return offs;
2124 }
2125 const TypePtr *t_ptr = adr_type->isa_ptr();
2126 assert(t_ptr != NULL, "must be a pointer type");
2127 return t_ptr->offset();
2128 }
2130 Node* ConnectionGraph::get_addp_base(Node *addp) {
2131 assert(addp->is_AddP(), "must be AddP");
2132 //
2133 // AddP cases for Base and Address inputs:
2134 // case #1. Direct object's field reference:
2135 // Allocate
2136 // |
2137 // Proj #5 ( oop result )
2138 // |
2139 // CheckCastPP (cast to instance type)
2140 // | |
2141 // AddP ( base == address )
2142 //
2143 // case #2. Indirect object's field reference:
2144 // Phi
2145 // |
2146 // CastPP (cast to instance type)
2147 // | |
2148 // AddP ( base == address )
2149 //
2150 // case #3. Raw object's field reference for Initialize node:
2151 // Allocate
2152 // |
2153 // Proj #5 ( oop result )
2154 // top |
2155 // \ |
2156 // AddP ( base == top )
2157 //
2158 // case #4. Array's element reference:
2159 // {CheckCastPP | CastPP}
2160 // | | |
2161 // | AddP ( array's element offset )
2162 // | |
2163 // AddP ( array's offset )
2164 //
2165 // case #5. Raw object's field reference for arraycopy stub call:
2166 // The inline_native_clone() case when the arraycopy stub is called
2167 // after the allocation before Initialize and CheckCastPP nodes.
2168 // Allocate
2169 // |
2170 // Proj #5 ( oop result )
2171 // | |
2172 // AddP ( base == address )
2173 //
2174 // case #6. Constant Pool, ThreadLocal, CastX2P or
2175 // Raw object's field reference:
2176 // {ConP, ThreadLocal, CastX2P, raw Load}
2177 // top |
2178 // \ |
2179 // AddP ( base == top )
2180 //
2181 // case #7. Klass's field reference.
2182 // LoadKlass
2183 // | |
2184 // AddP ( base == address )
2185 //
2186 // case #8. narrow Klass's field reference.
2187 // LoadNKlass
2188 // |
2189 // DecodeN
2190 // | |
2191 // AddP ( base == address )
2192 //
2193 Node *base = addp->in(AddPNode::Base);
2194 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2195 base = addp->in(AddPNode::Address);
2196 while (base->is_AddP()) {
2197 // Case #6 (unsafe access) may have several chained AddP nodes.
2198 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2199 base = base->in(AddPNode::Address);
2200 }
2201 Node* uncast_base = base->uncast();
2202 int opcode = uncast_base->Opcode();
2203 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2204 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2205 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2206 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2207 }
2208 return base;
2209 }
2211 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2212 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2213 Node* addp2 = addp->raw_out(0);
2214 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2215 addp2->in(AddPNode::Base) == n &&
2216 addp2->in(AddPNode::Address) == addp) {
2217 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2218 //
2219 // Find array's offset to push it on worklist first and
2220 // as result process an array's element offset first (pushed second)
2221 // to avoid CastPP for the array's offset.
2222 // Otherwise the inserted CastPP (LocalVar) will point to what
2223 // the AddP (Field) points to. Which would be wrong since
2224 // the algorithm expects the CastPP has the same point as
2225 // as AddP's base CheckCastPP (LocalVar).
2226 //
2227 // ArrayAllocation
2228 // |
2229 // CheckCastPP
2230 // |
2231 // memProj (from ArrayAllocation CheckCastPP)
2232 // | ||
2233 // | || Int (element index)
2234 // | || | ConI (log(element size))
2235 // | || | /
2236 // | || LShift
2237 // | || /
2238 // | AddP (array's element offset)
2239 // | |
2240 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2241 // | / /
2242 // AddP (array's offset)
2243 // |
2244 // Load/Store (memory operation on array's element)
2245 //
2246 return addp2;
2247 }
2248 return NULL;
2249 }
2251 //
2252 // Adjust the type and inputs of an AddP which computes the
2253 // address of a field of an instance
2254 //
2255 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2256 PhaseGVN* igvn = _igvn;
2257 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2258 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2259 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2260 if (t == NULL) {
2261 // We are computing a raw address for a store captured by an Initialize
2262 // compute an appropriate address type (cases #3 and #5).
2263 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2264 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2265 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2266 assert(offs != Type::OffsetBot, "offset must be a constant");
2267 t = base_t->add_offset(offs)->is_oopptr();
2268 }
2269 int inst_id = base_t->instance_id();
2270 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2271 "old type must be non-instance or match new type");
2273 // The type 't' could be subclass of 'base_t'.
2274 // As result t->offset() could be large then base_t's size and it will
2275 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2276 // constructor verifies correctness of the offset.
2277 //
2278 // It could happened on subclass's branch (from the type profiling
2279 // inlining) which was not eliminated during parsing since the exactness
2280 // of the allocation type was not propagated to the subclass type check.
2281 //
2282 // Or the type 't' could be not related to 'base_t' at all.
2283 // It could happened when CHA type is different from MDO type on a dead path
2284 // (for example, from instanceof check) which is not collapsed during parsing.
2285 //
2286 // Do nothing for such AddP node and don't process its users since
2287 // this code branch will go away.
2288 //
2289 if (!t->is_known_instance() &&
2290 !base_t->klass()->is_subtype_of(t->klass())) {
2291 return false; // bail out
2292 }
2293 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2294 // Do NOT remove the next line: ensure a new alias index is allocated
2295 // for the instance type. Note: C++ will not remove it since the call
2296 // has side effect.
2297 int alias_idx = _compile->get_alias_index(tinst);
2298 igvn->set_type(addp, tinst);
2299 // record the allocation in the node map
2300 set_map(addp, get_map(base->_idx));
2301 // Set addp's Base and Address to 'base'.
2302 Node *abase = addp->in(AddPNode::Base);
2303 Node *adr = addp->in(AddPNode::Address);
2304 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2305 adr->in(0)->_idx == (uint)inst_id) {
2306 // Skip AddP cases #3 and #5.
2307 } else {
2308 assert(!abase->is_top(), "sanity"); // AddP case #3
2309 if (abase != base) {
2310 igvn->hash_delete(addp);
2311 addp->set_req(AddPNode::Base, base);
2312 if (abase == adr) {
2313 addp->set_req(AddPNode::Address, base);
2314 } else {
2315 // AddP case #4 (adr is array's element offset AddP node)
2316 #ifdef ASSERT
2317 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2318 assert(adr->is_AddP() && atype != NULL &&
2319 atype->instance_id() == inst_id, "array's element offset should be processed first");
2320 #endif
2321 }
2322 igvn->hash_insert(addp);
2323 }
2324 }
2325 // Put on IGVN worklist since at least addp's type was changed above.
2326 record_for_optimizer(addp);
2327 return true;
2328 }
2330 //
2331 // Create a new version of orig_phi if necessary. Returns either the newly
2332 // created phi or an existing phi. Sets create_new to indicate whether a new
2333 // phi was created. Cache the last newly created phi in the node map.
2334 //
2335 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2336 Compile *C = _compile;
2337 PhaseGVN* igvn = _igvn;
2338 new_created = false;
2339 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2340 // nothing to do if orig_phi is bottom memory or matches alias_idx
2341 if (phi_alias_idx == alias_idx) {
2342 return orig_phi;
2343 }
2344 // Have we recently created a Phi for this alias index?
2345 PhiNode *result = get_map_phi(orig_phi->_idx);
2346 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2347 return result;
2348 }
2349 // Previous check may fail when the same wide memory Phi was split into Phis
2350 // for different memory slices. Search all Phis for this region.
2351 if (result != NULL) {
2352 Node* region = orig_phi->in(0);
2353 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2354 Node* phi = region->fast_out(i);
2355 if (phi->is_Phi() &&
2356 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2357 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2358 return phi->as_Phi();
2359 }
2360 }
2361 }
2362 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) {
2363 if (C->do_escape_analysis() == true && !C->failing()) {
2364 // Retry compilation without escape analysis.
2365 // If this is the first failure, the sentinel string will "stick"
2366 // to the Compile object, and the C2Compiler will see it and retry.
2367 C->record_failure(C2Compiler::retry_no_escape_analysis());
2368 }
2369 return NULL;
2370 }
2371 orig_phi_worklist.append_if_missing(orig_phi);
2372 const TypePtr *atype = C->get_adr_type(alias_idx);
2373 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2374 C->copy_node_notes_to(result, orig_phi);
2375 igvn->set_type(result, result->bottom_type());
2376 record_for_optimizer(result);
2377 set_map(orig_phi, result);
2378 new_created = true;
2379 return result;
2380 }
2382 //
2383 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2384 // specified alias index.
2385 //
2386 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2387 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2388 Compile *C = _compile;
2389 PhaseGVN* igvn = _igvn;
2390 bool new_phi_created;
2391 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2392 if (!new_phi_created) {
2393 return result;
2394 }
2395 GrowableArray<PhiNode *> phi_list;
2396 GrowableArray<uint> cur_input;
2397 PhiNode *phi = orig_phi;
2398 uint idx = 1;
2399 bool finished = false;
2400 while(!finished) {
2401 while (idx < phi->req()) {
2402 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2403 if (mem != NULL && mem->is_Phi()) {
2404 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2405 if (new_phi_created) {
2406 // found an phi for which we created a new split, push current one on worklist and begin
2407 // processing new one
2408 phi_list.push(phi);
2409 cur_input.push(idx);
2410 phi = mem->as_Phi();
2411 result = newphi;
2412 idx = 1;
2413 continue;
2414 } else {
2415 mem = newphi;
2416 }
2417 }
2418 if (C->failing()) {
2419 return NULL;
2420 }
2421 result->set_req(idx++, mem);
2422 }
2423 #ifdef ASSERT
2424 // verify that the new Phi has an input for each input of the original
2425 assert( phi->req() == result->req(), "must have same number of inputs.");
2426 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2427 #endif
2428 // Check if all new phi's inputs have specified alias index.
2429 // Otherwise use old phi.
2430 for (uint i = 1; i < phi->req(); i++) {
2431 Node* in = result->in(i);
2432 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2433 }
2434 // we have finished processing a Phi, see if there are any more to do
2435 finished = (phi_list.length() == 0 );
2436 if (!finished) {
2437 phi = phi_list.pop();
2438 idx = cur_input.pop();
2439 PhiNode *prev_result = get_map_phi(phi->_idx);
2440 prev_result->set_req(idx++, result);
2441 result = prev_result;
2442 }
2443 }
2444 return result;
2445 }
2447 //
2448 // The next methods are derived from methods in MemNode.
2449 //
2450 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2451 Node *mem = mmem;
2452 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2453 // means an array I have not precisely typed yet. Do not do any
2454 // alias stuff with it any time soon.
2455 if (toop->base() != Type::AnyPtr &&
2456 !(toop->klass() != NULL &&
2457 toop->klass()->is_java_lang_Object() &&
2458 toop->offset() == Type::OffsetBot)) {
2459 mem = mmem->memory_at(alias_idx);
2460 // Update input if it is progress over what we have now
2461 }
2462 return mem;
2463 }
2465 //
2466 // Move memory users to their memory slices.
2467 //
2468 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2469 Compile* C = _compile;
2470 PhaseGVN* igvn = _igvn;
2471 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2472 assert(tp != NULL, "ptr type");
2473 int alias_idx = C->get_alias_index(tp);
2474 int general_idx = C->get_general_index(alias_idx);
2476 // Move users first
2477 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2478 Node* use = n->fast_out(i);
2479 if (use->is_MergeMem()) {
2480 MergeMemNode* mmem = use->as_MergeMem();
2481 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2482 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2483 continue; // Nothing to do
2484 }
2485 // Replace previous general reference to mem node.
2486 uint orig_uniq = C->unique();
2487 Node* m = find_inst_mem(n, general_idx, orig_phis);
2488 assert(orig_uniq == C->unique(), "no new nodes");
2489 mmem->set_memory_at(general_idx, m);
2490 --imax;
2491 --i;
2492 } else if (use->is_MemBar()) {
2493 assert(!use->is_Initialize(), "initializing stores should not be moved");
2494 if (use->req() > MemBarNode::Precedent &&
2495 use->in(MemBarNode::Precedent) == n) {
2496 // Don't move related membars.
2497 record_for_optimizer(use);
2498 continue;
2499 }
2500 tp = use->as_MemBar()->adr_type()->isa_ptr();
2501 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2502 alias_idx == general_idx) {
2503 continue; // Nothing to do
2504 }
2505 // Move to general memory slice.
2506 uint orig_uniq = C->unique();
2507 Node* m = find_inst_mem(n, general_idx, orig_phis);
2508 assert(orig_uniq == C->unique(), "no new nodes");
2509 igvn->hash_delete(use);
2510 imax -= use->replace_edge(n, m);
2511 igvn->hash_insert(use);
2512 record_for_optimizer(use);
2513 --i;
2514 #ifdef ASSERT
2515 } else if (use->is_Mem()) {
2516 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2517 // Don't move related cardmark.
2518 continue;
2519 }
2520 // Memory nodes should have new memory input.
2521 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2522 assert(tp != NULL, "ptr type");
2523 int idx = C->get_alias_index(tp);
2524 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2525 "Following memory nodes should have new memory input or be on the same memory slice");
2526 } else if (use->is_Phi()) {
2527 // Phi nodes should be split and moved already.
2528 tp = use->as_Phi()->adr_type()->isa_ptr();
2529 assert(tp != NULL, "ptr type");
2530 int idx = C->get_alias_index(tp);
2531 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2532 } else {
2533 use->dump();
2534 assert(false, "should not be here");
2535 #endif
2536 }
2537 }
2538 }
2540 //
2541 // Search memory chain of "mem" to find a MemNode whose address
2542 // is the specified alias index.
2543 //
2544 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2545 if (orig_mem == NULL)
2546 return orig_mem;
2547 Compile* C = _compile;
2548 PhaseGVN* igvn = _igvn;
2549 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2550 bool is_instance = (toop != NULL) && toop->is_known_instance();
2551 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2552 Node *prev = NULL;
2553 Node *result = orig_mem;
2554 while (prev != result) {
2555 prev = result;
2556 if (result == start_mem)
2557 break; // hit one of our sentinels
2558 if (result->is_Mem()) {
2559 const Type *at = igvn->type(result->in(MemNode::Address));
2560 if (at == Type::TOP)
2561 break; // Dead
2562 assert (at->isa_ptr() != NULL, "pointer type required.");
2563 int idx = C->get_alias_index(at->is_ptr());
2564 if (idx == alias_idx)
2565 break; // Found
2566 if (!is_instance && (at->isa_oopptr() == NULL ||
2567 !at->is_oopptr()->is_known_instance())) {
2568 break; // Do not skip store to general memory slice.
2569 }
2570 result = result->in(MemNode::Memory);
2571 }
2572 if (!is_instance)
2573 continue; // don't search further for non-instance types
2574 // skip over a call which does not affect this memory slice
2575 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2576 Node *proj_in = result->in(0);
2577 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2578 break; // hit one of our sentinels
2579 } else if (proj_in->is_Call()) {
2580 CallNode *call = proj_in->as_Call();
2581 if (!call->may_modify(toop, igvn)) {
2582 result = call->in(TypeFunc::Memory);
2583 }
2584 } else if (proj_in->is_Initialize()) {
2585 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2586 // Stop if this is the initialization for the object instance which
2587 // which contains this memory slice, otherwise skip over it.
2588 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2589 result = proj_in->in(TypeFunc::Memory);
2590 }
2591 } else if (proj_in->is_MemBar()) {
2592 result = proj_in->in(TypeFunc::Memory);
2593 }
2594 } else if (result->is_MergeMem()) {
2595 MergeMemNode *mmem = result->as_MergeMem();
2596 result = step_through_mergemem(mmem, alias_idx, toop);
2597 if (result == mmem->base_memory()) {
2598 // Didn't find instance memory, search through general slice recursively.
2599 result = mmem->memory_at(C->get_general_index(alias_idx));
2600 result = find_inst_mem(result, alias_idx, orig_phis);
2601 if (C->failing()) {
2602 return NULL;
2603 }
2604 mmem->set_memory_at(alias_idx, result);
2605 }
2606 } else if (result->is_Phi() &&
2607 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2608 Node *un = result->as_Phi()->unique_input(igvn);
2609 if (un != NULL) {
2610 orig_phis.append_if_missing(result->as_Phi());
2611 result = un;
2612 } else {
2613 break;
2614 }
2615 } else if (result->is_ClearArray()) {
2616 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2617 // Can not bypass initialization of the instance
2618 // we are looking for.
2619 break;
2620 }
2621 // Otherwise skip it (the call updated 'result' value).
2622 } else if (result->Opcode() == Op_SCMemProj) {
2623 Node* mem = result->in(0);
2624 Node* adr = NULL;
2625 if (mem->is_LoadStore()) {
2626 adr = mem->in(MemNode::Address);
2627 } else {
2628 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2629 adr = mem->in(3); // Memory edge corresponds to destination array
2630 }
2631 const Type *at = igvn->type(adr);
2632 if (at != Type::TOP) {
2633 assert (at->isa_ptr() != NULL, "pointer type required.");
2634 int idx = C->get_alias_index(at->is_ptr());
2635 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2636 break;
2637 }
2638 result = mem->in(MemNode::Memory);
2639 }
2640 }
2641 if (result->is_Phi()) {
2642 PhiNode *mphi = result->as_Phi();
2643 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2644 const TypePtr *t = mphi->adr_type();
2645 if (!is_instance) {
2646 // Push all non-instance Phis on the orig_phis worklist to update inputs
2647 // during Phase 4 if needed.
2648 orig_phis.append_if_missing(mphi);
2649 } else if (C->get_alias_index(t) != alias_idx) {
2650 // Create a new Phi with the specified alias index type.
2651 result = split_memory_phi(mphi, alias_idx, orig_phis);
2652 }
2653 }
2654 // the result is either MemNode, PhiNode, InitializeNode.
2655 return result;
2656 }
2658 //
2659 // Convert the types of unescaped object to instance types where possible,
2660 // propagate the new type information through the graph, and update memory
2661 // edges and MergeMem inputs to reflect the new type.
2662 //
2663 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2664 // The processing is done in 4 phases:
2665 //
2666 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2667 // types for the CheckCastPP for allocations where possible.
2668 // Propagate the the new types through users as follows:
2669 // casts and Phi: push users on alloc_worklist
2670 // AddP: cast Base and Address inputs to the instance type
2671 // push any AddP users on alloc_worklist and push any memnode
2672 // users onto memnode_worklist.
2673 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2674 // search the Memory chain for a store with the appropriate type
2675 // address type. If a Phi is found, create a new version with
2676 // the appropriate memory slices from each of the Phi inputs.
2677 // For stores, process the users as follows:
2678 // MemNode: push on memnode_worklist
2679 // MergeMem: push on mergemem_worklist
2680 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2681 // moving the first node encountered of each instance type to the
2682 // the input corresponding to its alias index.
2683 // appropriate memory slice.
2684 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2685 //
2686 // In the following example, the CheckCastPP nodes are the cast of allocation
2687 // results and the allocation of node 29 is unescaped and eligible to be an
2688 // instance type.
2689 //
2690 // We start with:
2691 //
2692 // 7 Parm #memory
2693 // 10 ConI "12"
2694 // 19 CheckCastPP "Foo"
2695 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2696 // 29 CheckCastPP "Foo"
2697 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2698 //
2699 // 40 StoreP 25 7 20 ... alias_index=4
2700 // 50 StoreP 35 40 30 ... alias_index=4
2701 // 60 StoreP 45 50 20 ... alias_index=4
2702 // 70 LoadP _ 60 30 ... alias_index=4
2703 // 80 Phi 75 50 60 Memory alias_index=4
2704 // 90 LoadP _ 80 30 ... alias_index=4
2705 // 100 LoadP _ 80 20 ... alias_index=4
2706 //
2707 //
2708 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2709 // and creating a new alias index for node 30. This gives:
2710 //
2711 // 7 Parm #memory
2712 // 10 ConI "12"
2713 // 19 CheckCastPP "Foo"
2714 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2715 // 29 CheckCastPP "Foo" iid=24
2716 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2717 //
2718 // 40 StoreP 25 7 20 ... alias_index=4
2719 // 50 StoreP 35 40 30 ... alias_index=6
2720 // 60 StoreP 45 50 20 ... alias_index=4
2721 // 70 LoadP _ 60 30 ... alias_index=6
2722 // 80 Phi 75 50 60 Memory alias_index=4
2723 // 90 LoadP _ 80 30 ... alias_index=6
2724 // 100 LoadP _ 80 20 ... alias_index=4
2725 //
2726 // In phase 2, new memory inputs are computed for the loads and stores,
2727 // And a new version of the phi is created. In phase 4, the inputs to
2728 // node 80 are updated and then the memory nodes are updated with the
2729 // values computed in phase 2. This results in:
2730 //
2731 // 7 Parm #memory
2732 // 10 ConI "12"
2733 // 19 CheckCastPP "Foo"
2734 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2735 // 29 CheckCastPP "Foo" iid=24
2736 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2737 //
2738 // 40 StoreP 25 7 20 ... alias_index=4
2739 // 50 StoreP 35 7 30 ... alias_index=6
2740 // 60 StoreP 45 40 20 ... alias_index=4
2741 // 70 LoadP _ 50 30 ... alias_index=6
2742 // 80 Phi 75 40 60 Memory alias_index=4
2743 // 120 Phi 75 50 50 Memory alias_index=6
2744 // 90 LoadP _ 120 30 ... alias_index=6
2745 // 100 LoadP _ 80 20 ... alias_index=4
2746 //
2747 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2748 GrowableArray<Node *> memnode_worklist;
2749 GrowableArray<PhiNode *> orig_phis;
2750 PhaseIterGVN *igvn = _igvn;
2751 uint new_index_start = (uint) _compile->num_alias_types();
2752 Arena* arena = Thread::current()->resource_area();
2753 VectorSet visited(arena);
2754 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2755 uint unique_old = _compile->unique();
2757 // Phase 1: Process possible allocations from alloc_worklist.
2758 // Create instance types for the CheckCastPP for allocations where possible.
2759 //
2760 // (Note: don't forget to change the order of the second AddP node on
2761 // the alloc_worklist if the order of the worklist processing is changed,
2762 // see the comment in find_second_addp().)
2763 //
2764 while (alloc_worklist.length() != 0) {
2765 Node *n = alloc_worklist.pop();
2766 uint ni = n->_idx;
2767 if (n->is_Call()) {
2768 CallNode *alloc = n->as_Call();
2769 // copy escape information to call node
2770 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2771 PointsToNode::EscapeState es = ptn->escape_state();
2772 // We have an allocation or call which returns a Java object,
2773 // see if it is unescaped.
2774 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2775 continue;
2776 // Find CheckCastPP for the allocate or for the return value of a call
2777 n = alloc->result_cast();
2778 if (n == NULL) { // No uses except Initialize node
2779 if (alloc->is_Allocate()) {
2780 // Set the scalar_replaceable flag for allocation
2781 // so it could be eliminated if it has no uses.
2782 alloc->as_Allocate()->_is_scalar_replaceable = true;
2783 }
2784 if (alloc->is_CallStaticJava()) {
2785 // Set the scalar_replaceable flag for boxing method
2786 // so it could be eliminated if it has no uses.
2787 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2788 }
2789 continue;
2790 }
2791 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2792 assert(!alloc->is_Allocate(), "allocation should have unique type");
2793 continue;
2794 }
2796 // The inline code for Object.clone() casts the allocation result to
2797 // java.lang.Object and then to the actual type of the allocated
2798 // object. Detect this case and use the second cast.
2799 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2800 // the allocation result is cast to java.lang.Object and then
2801 // to the actual Array type.
2802 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2803 && (alloc->is_AllocateArray() ||
2804 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2805 Node *cast2 = NULL;
2806 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2807 Node *use = n->fast_out(i);
2808 if (use->is_CheckCastPP()) {
2809 cast2 = use;
2810 break;
2811 }
2812 }
2813 if (cast2 != NULL) {
2814 n = cast2;
2815 } else {
2816 // Non-scalar replaceable if the allocation type is unknown statically
2817 // (reflection allocation), the object can't be restored during
2818 // deoptimization without precise type.
2819 continue;
2820 }
2821 }
2822 if (alloc->is_Allocate()) {
2823 // Set the scalar_replaceable flag for allocation
2824 // so it could be eliminated.
2825 alloc->as_Allocate()->_is_scalar_replaceable = true;
2826 }
2827 if (alloc->is_CallStaticJava()) {
2828 // Set the scalar_replaceable flag for boxing method
2829 // so it could be eliminated.
2830 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2831 }
2832 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2833 // in order for an object to be scalar-replaceable, it must be:
2834 // - a direct allocation (not a call returning an object)
2835 // - non-escaping
2836 // - eligible to be a unique type
2837 // - not determined to be ineligible by escape analysis
2838 set_map(alloc, n);
2839 set_map(n, alloc);
2840 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2841 if (t == NULL)
2842 continue; // not a TypeOopPtr
2843 const TypeOopPtr* tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
2844 igvn->hash_delete(n);
2845 igvn->set_type(n, tinst);
2846 n->raise_bottom_type(tinst);
2847 igvn->hash_insert(n);
2848 record_for_optimizer(n);
2849 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2851 // First, put on the worklist all Field edges from Connection Graph
2852 // which is more accurate then putting immediate users from Ideal Graph.
2853 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2854 PointsToNode* tgt = e.get();
2855 Node* use = tgt->ideal_node();
2856 assert(tgt->is_Field() && use->is_AddP(),
2857 "only AddP nodes are Field edges in CG");
2858 if (use->outcnt() > 0) { // Don't process dead nodes
2859 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2860 if (addp2 != NULL) {
2861 assert(alloc->is_AllocateArray(),"array allocation was expected");
2862 alloc_worklist.append_if_missing(addp2);
2863 }
2864 alloc_worklist.append_if_missing(use);
2865 }
2866 }
2868 // An allocation may have an Initialize which has raw stores. Scan
2869 // the users of the raw allocation result and push AddP users
2870 // on alloc_worklist.
2871 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2872 assert (raw_result != NULL, "must have an allocation result");
2873 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2874 Node *use = raw_result->fast_out(i);
2875 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2876 Node* addp2 = find_second_addp(use, raw_result);
2877 if (addp2 != NULL) {
2878 assert(alloc->is_AllocateArray(),"array allocation was expected");
2879 alloc_worklist.append_if_missing(addp2);
2880 }
2881 alloc_worklist.append_if_missing(use);
2882 } else if (use->is_MemBar()) {
2883 memnode_worklist.append_if_missing(use);
2884 }
2885 }
2886 }
2887 } else if (n->is_AddP()) {
2888 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2889 if (jobj == NULL || jobj == phantom_obj) {
2890 #ifdef ASSERT
2891 ptnode_adr(get_addp_base(n)->_idx)->dump();
2892 ptnode_adr(n->_idx)->dump();
2893 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2894 #endif
2895 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2896 return;
2897 }
2898 Node *base = get_map(jobj->idx()); // CheckCastPP node
2899 if (!split_AddP(n, base)) continue; // wrong type from dead path
2900 } else if (n->is_Phi() ||
2901 n->is_CheckCastPP() ||
2902 n->is_EncodeP() ||
2903 n->is_DecodeN() ||
2904 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2905 if (visited.test_set(n->_idx)) {
2906 assert(n->is_Phi(), "loops only through Phi's");
2907 continue; // already processed
2908 }
2909 JavaObjectNode* jobj = unique_java_object(n);
2910 if (jobj == NULL || jobj == phantom_obj) {
2911 #ifdef ASSERT
2912 ptnode_adr(n->_idx)->dump();
2913 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2914 #endif
2915 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2916 return;
2917 } else {
2918 Node *val = get_map(jobj->idx()); // CheckCastPP node
2919 TypeNode *tn = n->as_Type();
2920 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2921 assert(tinst != NULL && tinst->is_known_instance() &&
2922 tinst->instance_id() == jobj->idx() , "instance type expected.");
2924 const Type *tn_type = igvn->type(tn);
2925 const TypeOopPtr *tn_t;
2926 if (tn_type->isa_narrowoop()) {
2927 tn_t = tn_type->make_ptr()->isa_oopptr();
2928 } else {
2929 tn_t = tn_type->isa_oopptr();
2930 }
2931 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2932 if (tn_type->isa_narrowoop()) {
2933 tn_type = tinst->make_narrowoop();
2934 } else {
2935 tn_type = tinst;
2936 }
2937 igvn->hash_delete(tn);
2938 igvn->set_type(tn, tn_type);
2939 tn->set_type(tn_type);
2940 igvn->hash_insert(tn);
2941 record_for_optimizer(n);
2942 } else {
2943 assert(tn_type == TypePtr::NULL_PTR ||
2944 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
2945 "unexpected type");
2946 continue; // Skip dead path with different type
2947 }
2948 }
2949 } else {
2950 debug_only(n->dump();)
2951 assert(false, "EA: unexpected node");
2952 continue;
2953 }
2954 // push allocation's users on appropriate worklist
2955 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2956 Node *use = n->fast_out(i);
2957 if(use->is_Mem() && use->in(MemNode::Address) == n) {
2958 // Load/store to instance's field
2959 memnode_worklist.append_if_missing(use);
2960 } else if (use->is_MemBar()) {
2961 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
2962 memnode_worklist.append_if_missing(use);
2963 }
2964 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
2965 Node* addp2 = find_second_addp(use, n);
2966 if (addp2 != NULL) {
2967 alloc_worklist.append_if_missing(addp2);
2968 }
2969 alloc_worklist.append_if_missing(use);
2970 } else if (use->is_Phi() ||
2971 use->is_CheckCastPP() ||
2972 use->is_EncodeNarrowPtr() ||
2973 use->is_DecodeNarrowPtr() ||
2974 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
2975 alloc_worklist.append_if_missing(use);
2976 #ifdef ASSERT
2977 } else if (use->is_Mem()) {
2978 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
2979 } else if (use->is_MergeMem()) {
2980 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
2981 } else if (use->is_SafePoint()) {
2982 // Look for MergeMem nodes for calls which reference unique allocation
2983 // (through CheckCastPP nodes) even for debug info.
2984 Node* m = use->in(TypeFunc::Memory);
2985 if (m->is_MergeMem()) {
2986 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
2987 }
2988 } else if (use->Opcode() == Op_EncodeISOArray) {
2989 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
2990 // EncodeISOArray overwrites destination array
2991 memnode_worklist.append_if_missing(use);
2992 }
2993 } else {
2994 uint op = use->Opcode();
2995 if (!(op == Op_CmpP || op == Op_Conv2B ||
2996 op == Op_CastP2X || op == Op_StoreCM ||
2997 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
2998 op == Op_StrEquals || op == Op_StrIndexOf)) {
2999 n->dump();
3000 use->dump();
3001 assert(false, "EA: missing allocation reference path");
3002 }
3003 #endif
3004 }
3005 }
3007 }
3008 // New alias types were created in split_AddP().
3009 uint new_index_end = (uint) _compile->num_alias_types();
3010 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3012 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3013 // compute new values for Memory inputs (the Memory inputs are not
3014 // actually updated until phase 4.)
3015 if (memnode_worklist.length() == 0)
3016 return; // nothing to do
3017 while (memnode_worklist.length() != 0) {
3018 Node *n = memnode_worklist.pop();
3019 if (visited.test_set(n->_idx))
3020 continue;
3021 if (n->is_Phi() || n->is_ClearArray()) {
3022 // we don't need to do anything, but the users must be pushed
3023 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3024 // we don't need to do anything, but the users must be pushed
3025 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3026 if (n == NULL)
3027 continue;
3028 } else if (n->Opcode() == Op_EncodeISOArray) {
3029 // get the memory projection
3030 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3031 Node *use = n->fast_out(i);
3032 if (use->Opcode() == Op_SCMemProj) {
3033 n = use;
3034 break;
3035 }
3036 }
3037 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3038 } else {
3039 assert(n->is_Mem(), "memory node required.");
3040 Node *addr = n->in(MemNode::Address);
3041 const Type *addr_t = igvn->type(addr);
3042 if (addr_t == Type::TOP)
3043 continue;
3044 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3045 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3046 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3047 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3048 if (_compile->failing()) {
3049 return;
3050 }
3051 if (mem != n->in(MemNode::Memory)) {
3052 // We delay the memory edge update since we need old one in
3053 // MergeMem code below when instances memory slices are separated.
3054 set_map(n, mem);
3055 }
3056 if (n->is_Load()) {
3057 continue; // don't push users
3058 } else if (n->is_LoadStore()) {
3059 // get the memory projection
3060 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3061 Node *use = n->fast_out(i);
3062 if (use->Opcode() == Op_SCMemProj) {
3063 n = use;
3064 break;
3065 }
3066 }
3067 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3068 }
3069 }
3070 // push user on appropriate worklist
3071 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3072 Node *use = n->fast_out(i);
3073 if (use->is_Phi() || use->is_ClearArray()) {
3074 memnode_worklist.append_if_missing(use);
3075 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3076 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3077 continue;
3078 memnode_worklist.append_if_missing(use);
3079 } else if (use->is_MemBar()) {
3080 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3081 memnode_worklist.append_if_missing(use);
3082 }
3083 #ifdef ASSERT
3084 } else if(use->is_Mem()) {
3085 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3086 } else if (use->is_MergeMem()) {
3087 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3088 } else if (use->Opcode() == Op_EncodeISOArray) {
3089 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3090 // EncodeISOArray overwrites destination array
3091 memnode_worklist.append_if_missing(use);
3092 }
3093 } else {
3094 uint op = use->Opcode();
3095 if (!(op == Op_StoreCM ||
3096 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3097 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3098 op == Op_AryEq || op == Op_StrComp ||
3099 op == Op_StrEquals || op == Op_StrIndexOf)) {
3100 n->dump();
3101 use->dump();
3102 assert(false, "EA: missing memory path");
3103 }
3104 #endif
3105 }
3106 }
3107 }
3109 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3110 // Walk each memory slice moving the first node encountered of each
3111 // instance type to the the input corresponding to its alias index.
3112 uint length = _mergemem_worklist.length();
3113 for( uint next = 0; next < length; ++next ) {
3114 MergeMemNode* nmm = _mergemem_worklist.at(next);
3115 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3116 // Note: we don't want to use MergeMemStream here because we only want to
3117 // scan inputs which exist at the start, not ones we add during processing.
3118 // Note 2: MergeMem may already contains instance memory slices added
3119 // during find_inst_mem() call when memory nodes were processed above.
3120 igvn->hash_delete(nmm);
3121 uint nslices = nmm->req();
3122 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3123 Node* mem = nmm->in(i);
3124 Node* cur = NULL;
3125 if (mem == NULL || mem->is_top())
3126 continue;
3127 // First, update mergemem by moving memory nodes to corresponding slices
3128 // if their type became more precise since this mergemem was created.
3129 while (mem->is_Mem()) {
3130 const Type *at = igvn->type(mem->in(MemNode::Address));
3131 if (at != Type::TOP) {
3132 assert (at->isa_ptr() != NULL, "pointer type required.");
3133 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3134 if (idx == i) {
3135 if (cur == NULL)
3136 cur = mem;
3137 } else {
3138 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3139 nmm->set_memory_at(idx, mem);
3140 }
3141 }
3142 }
3143 mem = mem->in(MemNode::Memory);
3144 }
3145 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3146 // Find any instance of the current type if we haven't encountered
3147 // already a memory slice of the instance along the memory chain.
3148 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3149 if((uint)_compile->get_general_index(ni) == i) {
3150 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3151 if (nmm->is_empty_memory(m)) {
3152 Node* result = find_inst_mem(mem, ni, orig_phis);
3153 if (_compile->failing()) {
3154 return;
3155 }
3156 nmm->set_memory_at(ni, result);
3157 }
3158 }
3159 }
3160 }
3161 // Find the rest of instances values
3162 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3163 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3164 Node* result = step_through_mergemem(nmm, ni, tinst);
3165 if (result == nmm->base_memory()) {
3166 // Didn't find instance memory, search through general slice recursively.
3167 result = nmm->memory_at(_compile->get_general_index(ni));
3168 result = find_inst_mem(result, ni, orig_phis);
3169 if (_compile->failing()) {
3170 return;
3171 }
3172 nmm->set_memory_at(ni, result);
3173 }
3174 }
3175 igvn->hash_insert(nmm);
3176 record_for_optimizer(nmm);
3177 }
3179 // Phase 4: Update the inputs of non-instance memory Phis and
3180 // the Memory input of memnodes
3181 // First update the inputs of any non-instance Phi's from
3182 // which we split out an instance Phi. Note we don't have
3183 // to recursively process Phi's encounted on the input memory
3184 // chains as is done in split_memory_phi() since they will
3185 // also be processed here.
3186 for (int j = 0; j < orig_phis.length(); j++) {
3187 PhiNode *phi = orig_phis.at(j);
3188 int alias_idx = _compile->get_alias_index(phi->adr_type());
3189 igvn->hash_delete(phi);
3190 for (uint i = 1; i < phi->req(); i++) {
3191 Node *mem = phi->in(i);
3192 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3193 if (_compile->failing()) {
3194 return;
3195 }
3196 if (mem != new_mem) {
3197 phi->set_req(i, new_mem);
3198 }
3199 }
3200 igvn->hash_insert(phi);
3201 record_for_optimizer(phi);
3202 }
3204 // Update the memory inputs of MemNodes with the value we computed
3205 // in Phase 2 and move stores memory users to corresponding memory slices.
3206 // Disable memory split verification code until the fix for 6984348.
3207 // Currently it produces false negative results since it does not cover all cases.
3208 #if 0 // ifdef ASSERT
3209 visited.Reset();
3210 Node_Stack old_mems(arena, _compile->unique() >> 2);
3211 #endif
3212 for (uint i = 0; i < ideal_nodes.size(); i++) {
3213 Node* n = ideal_nodes.at(i);
3214 Node* nmem = get_map(n->_idx);
3215 assert(nmem != NULL, "sanity");
3216 if (n->is_Mem()) {
3217 #if 0 // ifdef ASSERT
3218 Node* old_mem = n->in(MemNode::Memory);
3219 if (!visited.test_set(old_mem->_idx)) {
3220 old_mems.push(old_mem, old_mem->outcnt());
3221 }
3222 #endif
3223 assert(n->in(MemNode::Memory) != nmem, "sanity");
3224 if (!n->is_Load()) {
3225 // Move memory users of a store first.
3226 move_inst_mem(n, orig_phis);
3227 }
3228 // Now update memory input
3229 igvn->hash_delete(n);
3230 n->set_req(MemNode::Memory, nmem);
3231 igvn->hash_insert(n);
3232 record_for_optimizer(n);
3233 } else {
3234 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3235 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3236 }
3237 }
3238 #if 0 // ifdef ASSERT
3239 // Verify that memory was split correctly
3240 while (old_mems.is_nonempty()) {
3241 Node* old_mem = old_mems.node();
3242 uint old_cnt = old_mems.index();
3243 old_mems.pop();
3244 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3245 }
3246 #endif
3247 }
3249 #ifndef PRODUCT
3250 static const char *node_type_names[] = {
3251 "UnknownType",
3252 "JavaObject",
3253 "LocalVar",
3254 "Field",
3255 "Arraycopy"
3256 };
3258 static const char *esc_names[] = {
3259 "UnknownEscape",
3260 "NoEscape",
3261 "ArgEscape",
3262 "GlobalEscape"
3263 };
3265 void PointsToNode::dump(bool print_state) const {
3266 NodeType nt = node_type();
3267 tty->print("%s ", node_type_names[(int) nt]);
3268 if (print_state) {
3269 EscapeState es = escape_state();
3270 EscapeState fields_es = fields_escape_state();
3271 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3272 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3273 tty->print("NSR ");
3274 }
3275 if (is_Field()) {
3276 FieldNode* f = (FieldNode*)this;
3277 if (f->is_oop())
3278 tty->print("oop ");
3279 if (f->offset() > 0)
3280 tty->print("+%d ", f->offset());
3281 tty->print("(");
3282 for (BaseIterator i(f); i.has_next(); i.next()) {
3283 PointsToNode* b = i.get();
3284 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3285 }
3286 tty->print(" )");
3287 }
3288 tty->print("[");
3289 for (EdgeIterator i(this); i.has_next(); i.next()) {
3290 PointsToNode* e = i.get();
3291 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3292 }
3293 tty->print(" [");
3294 for (UseIterator i(this); i.has_next(); i.next()) {
3295 PointsToNode* u = i.get();
3296 bool is_base = false;
3297 if (PointsToNode::is_base_use(u)) {
3298 is_base = true;
3299 u = PointsToNode::get_use_node(u)->as_Field();
3300 }
3301 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3302 }
3303 tty->print(" ]] ");
3304 if (_node == NULL)
3305 tty->print_cr("<null>");
3306 else
3307 _node->dump();
3308 }
3310 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3311 bool first = true;
3312 int ptnodes_length = ptnodes_worklist.length();
3313 for (int i = 0; i < ptnodes_length; i++) {
3314 PointsToNode *ptn = ptnodes_worklist.at(i);
3315 if (ptn == NULL || !ptn->is_JavaObject())
3316 continue;
3317 PointsToNode::EscapeState es = ptn->escape_state();
3318 if ((es != PointsToNode::NoEscape) && !Verbose) {
3319 continue;
3320 }
3321 Node* n = ptn->ideal_node();
3322 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3323 n->as_CallStaticJava()->is_boxing_method())) {
3324 if (first) {
3325 tty->cr();
3326 tty->print("======== Connection graph for ");
3327 _compile->method()->print_short_name();
3328 tty->cr();
3329 first = false;
3330 }
3331 ptn->dump();
3332 // Print all locals and fields which reference this allocation
3333 for (UseIterator j(ptn); j.has_next(); j.next()) {
3334 PointsToNode* use = j.get();
3335 if (use->is_LocalVar()) {
3336 use->dump(Verbose);
3337 } else if (Verbose) {
3338 use->dump();
3339 }
3340 }
3341 tty->cr();
3342 }
3343 }
3344 }
3345 #endif