1 /*
2 * Copyright (c) 2005, 2019, 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 */
24
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/c2compiler.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/escape.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #if INCLUDE_G1GC
43 #include "gc/g1/g1ThreadLocalData.hpp"
44 #endif // INCLUDE_G1GC
45 #if INCLUDE_ZGC
46 #include "gc/z/c2/zBarrierSetC2.hpp"
47 #endif
48
49 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
50 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
51 _in_worklist(C->comp_arena()),
52 _next_pidx(0),
53 _collecting(true),
54 _verify(false),
55 _compile(C),
56 _igvn(igvn),
57 _node_map(C->comp_arena()) {
58 // Add unknown java object.
59 add_java_object(C->top(), PointsToNode::GlobalEscape);
60 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
61 // Add ConP(#NULL) and ConN(#NULL) nodes.
62 Node* oop_null = igvn->zerocon(T_OBJECT);
63 assert(oop_null->_idx < nodes_size(), "should be created already");
64 add_java_object(oop_null, PointsToNode::NoEscape);
65 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
66 if (UseCompressedOops) {
67 Node* noop_null = igvn->zerocon(T_NARROWOOP);
68 assert(noop_null->_idx < nodes_size(), "should be created already");
69 map_ideal_node(noop_null, null_obj);
70 }
71 _pcmp_neq = NULL; // Should be initialized
72 _pcmp_eq = NULL;
73 }
74
75 bool ConnectionGraph::has_candidates(Compile *C) {
76 // EA brings benefits only when the code has allocations and/or locks which
77 // are represented by ideal Macro nodes.
78 int cnt = C->macro_count();
79 for (int i = 0; i < cnt; i++) {
80 Node *n = C->macro_node(i);
81 if (n->is_Allocate())
82 return true;
83 if (n->is_Lock()) {
84 Node* obj = n->as_Lock()->obj_node()->uncast();
85 if (!(obj->is_Parm() || obj->is_Con()))
86 return true;
87 }
88 if (n->is_CallStaticJava() &&
89 n->as_CallStaticJava()->is_boxing_method()) {
90 return true;
91 }
92 }
93 return false;
94 }
95
96 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
97 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]);
98 ResourceMark rm;
99
100 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
101 // to create space for them in ConnectionGraph::_nodes[].
102 Node* oop_null = igvn->zerocon(T_OBJECT);
103 Node* noop_null = igvn->zerocon(T_NARROWOOP);
104 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
105 // Perform escape analysis
106 if (congraph->compute_escape()) {
107 // There are non escaping objects.
108 C->set_congraph(congraph);
109 }
110 // Cleanup.
111 if (oop_null->outcnt() == 0)
112 igvn->hash_delete(oop_null);
113 if (noop_null->outcnt() == 0)
114 igvn->hash_delete(noop_null);
115 }
116
117 bool ConnectionGraph::compute_escape() {
118 Compile* C = _compile;
119 PhaseGVN* igvn = _igvn;
120
121 // Worklists used by EA.
122 Unique_Node_List delayed_worklist;
123 GrowableArray<Node*> alloc_worklist;
124 GrowableArray<Node*> ptr_cmp_worklist;
125 GrowableArray<Node*> storestore_worklist;
126 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
127 GrowableArray<PointsToNode*> ptnodes_worklist;
128 GrowableArray<JavaObjectNode*> java_objects_worklist;
129 GrowableArray<JavaObjectNode*> non_escaped_worklist;
130 GrowableArray<FieldNode*> oop_fields_worklist;
131 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
132
133 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]);
134
135 // 1. Populate Connection Graph (CG) with PointsTo nodes.
136 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
137 // Initialize worklist
138 if (C->root() != NULL) {
139 ideal_nodes.push(C->root());
140 }
141 // Processed ideal nodes are unique on ideal_nodes list
142 // but several ideal nodes are mapped to the phantom_obj.
143 // To avoid duplicated entries on the following worklists
144 // add the phantom_obj only once to them.
145 ptnodes_worklist.append(phantom_obj);
146 java_objects_worklist.append(phantom_obj);
147 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
148 Node* n = ideal_nodes.at(next);
149 // Create PointsTo nodes and add them to Connection Graph. Called
150 // only once per ideal node since ideal_nodes is Unique_Node list.
151 add_node_to_connection_graph(n, &delayed_worklist);
152 PointsToNode* ptn = ptnode_adr(n->_idx);
153 if (ptn != NULL && ptn != phantom_obj) {
154 ptnodes_worklist.append(ptn);
155 if (ptn->is_JavaObject()) {
156 java_objects_worklist.append(ptn->as_JavaObject());
157 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
158 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
159 // Only allocations and java static calls results are interesting.
160 non_escaped_worklist.append(ptn->as_JavaObject());
161 }
162 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
163 oop_fields_worklist.append(ptn->as_Field());
164 }
165 }
166 if (n->is_MergeMem()) {
167 // Collect all MergeMem nodes to add memory slices for
168 // scalar replaceable objects in split_unique_types().
169 _mergemem_worklist.append(n->as_MergeMem());
170 } else if (OptimizePtrCompare && n->is_Cmp() &&
171 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
172 // Collect compare pointers nodes.
173 ptr_cmp_worklist.append(n);
174 } else if (n->is_MemBarStoreStore()) {
175 // Collect all MemBarStoreStore nodes so that depending on the
176 // escape status of the associated Allocate node some of them
177 // may be eliminated.
178 storestore_worklist.append(n);
179 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
180 (n->req() > MemBarNode::Precedent)) {
181 record_for_optimizer(n);
182 #ifdef ASSERT
183 } else if (n->is_AddP()) {
184 // Collect address nodes for graph verification.
185 addp_worklist.append(n);
186 #endif
187 } else if (n->is_ArrayCopy()) {
188 // Keep a list of ArrayCopy nodes so if one of its input is non
189 // escaping, we can record a unique type
190 arraycopy_worklist.append(n->as_ArrayCopy());
191 }
192 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
193 Node* m = n->fast_out(i); // Get user
194 ideal_nodes.push(m);
195 }
196 }
197 if (non_escaped_worklist.length() == 0) {
198 _collecting = false;
199 return false; // Nothing to do.
200 }
201 // Add final simple edges to graph.
202 while(delayed_worklist.size() > 0) {
203 Node* n = delayed_worklist.pop();
204 add_final_edges(n);
205 }
206 int ptnodes_length = ptnodes_worklist.length();
207
208 #ifdef ASSERT
209 if (VerifyConnectionGraph) {
210 // Verify that no new simple edges could be created and all
211 // local vars has edges.
212 _verify = true;
213 for (int next = 0; next < ptnodes_length; ++next) {
214 PointsToNode* ptn = ptnodes_worklist.at(next);
215 add_final_edges(ptn->ideal_node());
216 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
217 ptn->dump();
218 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
219 }
220 }
221 _verify = false;
222 }
223 #endif
224 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
225 // processing, calls to CI to resolve symbols (types, fields, methods)
226 // referenced in bytecode. During symbol resolution VM may throw
227 // an exception which CI cleans and converts to compilation failure.
228 if (C->failing()) return false;
229
230 // 2. Finish Graph construction by propagating references to all
231 // java objects through graph.
232 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
233 java_objects_worklist, oop_fields_worklist)) {
234 // All objects escaped or hit time or iterations limits.
235 _collecting = false;
236 return false;
237 }
238
239 // 3. Adjust scalar_replaceable state of nonescaping objects and push
240 // scalar replaceable allocations on alloc_worklist for processing
241 // in split_unique_types().
242 int non_escaped_length = non_escaped_worklist.length();
243 for (int next = 0; next < non_escaped_length; next++) {
244 JavaObjectNode* ptn = non_escaped_worklist.at(next);
245 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
246 Node* n = ptn->ideal_node();
247 if (n->is_Allocate()) {
248 n->as_Allocate()->_is_non_escaping = noescape;
249 }
250 if (n->is_CallStaticJava()) {
251 n->as_CallStaticJava()->_is_non_escaping = noescape;
252 }
253 if (noescape && ptn->scalar_replaceable()) {
254 adjust_scalar_replaceable_state(ptn);
255 if (ptn->scalar_replaceable()) {
256 alloc_worklist.append(ptn->ideal_node());
257 }
258 }
259 }
260
261 #ifdef ASSERT
262 if (VerifyConnectionGraph) {
263 // Verify that graph is complete - no new edges could be added or needed.
264 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
265 java_objects_worklist, addp_worklist);
266 }
267 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
268 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
269 null_obj->edge_count() == 0 &&
270 !null_obj->arraycopy_src() &&
271 !null_obj->arraycopy_dst(), "sanity");
272 #endif
273
274 _collecting = false;
275
276 } // TracePhase t3("connectionGraph")
277
278 // 4. Optimize ideal graph based on EA information.
279 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
280 if (has_non_escaping_obj) {
281 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
282 }
283
284 #ifndef PRODUCT
285 if (PrintEscapeAnalysis) {
286 dump(ptnodes_worklist); // Dump ConnectionGraph
287 }
288 #endif
289
290 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
291 #ifdef ASSERT
292 if (VerifyConnectionGraph) {
293 int alloc_length = alloc_worklist.length();
294 for (int next = 0; next < alloc_length; ++next) {
295 Node* n = alloc_worklist.at(next);
296 PointsToNode* ptn = ptnode_adr(n->_idx);
297 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
298 }
299 }
300 #endif
301
302 // 5. Separate memory graph for scalar replaceable allcations.
303 if (has_scalar_replaceable_candidates &&
304 C->AliasLevel() >= 3 && EliminateAllocations) {
305 // Now use the escape information to create unique types for
306 // scalar replaceable objects.
307 split_unique_types(alloc_worklist, arraycopy_worklist);
308 if (C->failing()) return false;
309 C->print_method(PHASE_AFTER_EA, 2);
310
311 #ifdef ASSERT
312 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
313 tty->print("=== No allocations eliminated for ");
314 C->method()->print_short_name();
315 if(!EliminateAllocations) {
316 tty->print(" since EliminateAllocations is off ===");
317 } else if(!has_scalar_replaceable_candidates) {
318 tty->print(" since there are no scalar replaceable candidates ===");
319 } else if(C->AliasLevel() < 3) {
320 tty->print(" since AliasLevel < 3 ===");
321 }
322 tty->cr();
323 #endif
324 }
325 return has_non_escaping_obj;
326 }
327
328 // Utility function for nodes that load an object
329 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
330 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
331 // ThreadLocal has RawPtr type.
332 const Type* t = _igvn->type(n);
333 if (t->make_ptr() != NULL) {
334 Node* adr = n->in(MemNode::Address);
335 #ifdef ASSERT
336 if (!adr->is_AddP()) {
337 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
338 } else {
339 assert((ptnode_adr(adr->_idx) == NULL ||
340 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
341 }
342 #endif
343 add_local_var_and_edge(n, PointsToNode::NoEscape,
344 adr, delayed_worklist);
345 }
346 }
347
348 // Populate Connection Graph with PointsTo nodes and create simple
349 // connection graph edges.
350 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
351 assert(!_verify, "this method should not be called for verification");
352 PhaseGVN* igvn = _igvn;
353 uint n_idx = n->_idx;
354 PointsToNode* n_ptn = ptnode_adr(n_idx);
355 if (n_ptn != NULL)
356 return; // No need to redefine PointsTo node during first iteration.
357
358 if (n->is_Call()) {
359 // Arguments to allocation and locking don't escape.
360 if (n->is_AbstractLock()) {
361 // Put Lock and Unlock nodes on IGVN worklist to process them during
362 // first IGVN optimization when escape information is still available.
363 record_for_optimizer(n);
364 } else if (n->is_Allocate()) {
365 add_call_node(n->as_Call());
366 record_for_optimizer(n);
367 } else {
368 if (n->is_CallStaticJava()) {
369 const char* name = n->as_CallStaticJava()->_name;
370 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
371 return; // Skip uncommon traps
372 }
373 // Don't mark as processed since call's arguments have to be processed.
374 delayed_worklist->push(n);
375 // Check if a call returns an object.
376 if ((n->as_Call()->returns_pointer() &&
377 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) ||
378 (n->is_CallStaticJava() &&
379 n->as_CallStaticJava()->is_boxing_method())) {
380 add_call_node(n->as_Call());
381 }
382 }
383 return;
384 }
385 // Put this check here to process call arguments since some call nodes
386 // point to phantom_obj.
387 if (n_ptn == phantom_obj || n_ptn == null_obj)
388 return; // Skip predefined nodes.
389
390 int opcode = n->Opcode();
391 switch (opcode) {
392 case Op_AddP: {
393 Node* base = get_addp_base(n);
394 PointsToNode* ptn_base = ptnode_adr(base->_idx);
395 // Field nodes are created for all field types. They are used in
396 // adjust_scalar_replaceable_state() and split_unique_types().
397 // Note, non-oop fields will have only base edges in Connection
398 // Graph because such fields are not used for oop loads and stores.
399 int offset = address_offset(n, igvn);
400 add_field(n, PointsToNode::NoEscape, offset);
401 if (ptn_base == NULL) {
402 delayed_worklist->push(n); // Process it later.
403 } else {
404 n_ptn = ptnode_adr(n_idx);
405 add_base(n_ptn->as_Field(), ptn_base);
406 }
407 break;
408 }
409 case Op_CastX2P: {
410 map_ideal_node(n, phantom_obj);
411 break;
412 }
413 case Op_CastPP:
414 case Op_CheckCastPP:
415 case Op_EncodeP:
416 case Op_DecodeN:
417 case Op_EncodePKlass:
418 case Op_DecodeNKlass: {
419 add_local_var_and_edge(n, PointsToNode::NoEscape,
420 n->in(1), delayed_worklist);
421 break;
422 }
423 case Op_CMoveP: {
424 add_local_var(n, PointsToNode::NoEscape);
425 // Do not add edges during first iteration because some could be
426 // not defined yet.
427 delayed_worklist->push(n);
428 break;
429 }
430 case Op_ConP:
431 case Op_ConN:
432 case Op_ConNKlass: {
433 // assume all oop constants globally escape except for null
434 PointsToNode::EscapeState es;
435 const Type* t = igvn->type(n);
436 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
437 es = PointsToNode::NoEscape;
438 } else {
439 es = PointsToNode::GlobalEscape;
440 }
441 add_java_object(n, es);
442 break;
443 }
444 case Op_CreateEx: {
445 // assume that all exception objects globally escape
446 map_ideal_node(n, phantom_obj);
447 break;
448 }
449 case Op_LoadKlass:
450 case Op_LoadNKlass: {
451 // Unknown class is loaded
452 map_ideal_node(n, phantom_obj);
453 break;
454 }
455 case Op_LoadP:
456 #if INCLUDE_ZGC
457 case Op_LoadBarrierSlowReg:
458 case Op_LoadBarrierWeakSlowReg:
459 #endif
460 case Op_LoadN:
461 case Op_LoadPLocked: {
462 add_objload_to_connection_graph(n, delayed_worklist);
463 break;
464 }
465 case Op_Parm: {
466 map_ideal_node(n, phantom_obj);
467 break;
468 }
469 case Op_PartialSubtypeCheck: {
470 // Produces Null or notNull and is used in only in CmpP so
471 // phantom_obj could be used.
472 map_ideal_node(n, phantom_obj); // Result is unknown
473 break;
474 }
475 case Op_Phi: {
476 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
477 // ThreadLocal has RawPtr type.
478 const Type* t = n->as_Phi()->type();
479 if (t->make_ptr() != NULL) {
480 add_local_var(n, PointsToNode::NoEscape);
481 // Do not add edges during first iteration because some could be
482 // not defined yet.
483 delayed_worklist->push(n);
484 }
485 break;
486 }
487 case Op_Proj: {
488 // we are only interested in the oop result projection from a call
489 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
490 n->in(0)->as_Call()->returns_pointer()) {
491 add_local_var_and_edge(n, PointsToNode::NoEscape,
492 n->in(0), delayed_worklist);
493 }
494 #if INCLUDE_ZGC
495 else if (UseZGC) {
496 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) {
497 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), delayed_worklist);
498 }
499 }
500 #endif
501 break;
502 }
503 case Op_Rethrow: // Exception object escapes
504 case Op_Return: {
505 if (n->req() > TypeFunc::Parms &&
506 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
507 // Treat Return value as LocalVar with GlobalEscape escape state.
508 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
509 n->in(TypeFunc::Parms), delayed_worklist);
510 }
511 break;
512 }
513 case Op_CompareAndExchangeP:
514 case Op_CompareAndExchangeN:
515 case Op_GetAndSetP:
516 case Op_GetAndSetN: {
517 add_objload_to_connection_graph(n, delayed_worklist);
518 // fallthrough
519 }
520 case Op_StoreP:
521 case Op_StoreN:
522 case Op_StoreNKlass:
523 case Op_StorePConditional:
524 case Op_WeakCompareAndSwapP:
525 case Op_WeakCompareAndSwapN:
526 case Op_CompareAndSwapP:
527 case Op_CompareAndSwapN: {
528 Node* adr = n->in(MemNode::Address);
529 const Type *adr_type = igvn->type(adr);
530 adr_type = adr_type->make_ptr();
531 if (adr_type == NULL) {
532 break; // skip dead nodes
533 }
534 if ( adr_type->isa_oopptr()
535 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
536 && adr_type == TypeRawPtr::NOTNULL
537 && adr->in(AddPNode::Address)->is_Proj()
538 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
539 delayed_worklist->push(n); // Process it later.
540 #ifdef ASSERT
541 assert(adr->is_AddP(), "expecting an AddP");
542 if (adr_type == TypeRawPtr::NOTNULL) {
543 // Verify a raw address for a store captured by Initialize node.
544 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
545 assert(offs != Type::OffsetBot, "offset must be a constant");
546 }
547 #endif
548 } else {
549 // Ignore copy the displaced header to the BoxNode (OSR compilation).
550 if (adr->is_BoxLock())
551 break;
552 // Stored value escapes in unsafe access.
553 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
554 // Pointer stores in G1 barriers looks like unsafe access.
555 // Ignore such stores to be able scalar replace non-escaping
556 // allocations.
557 #if INCLUDE_G1GC
558 if (UseG1GC && adr->is_AddP()) {
559 Node* base = get_addp_base(adr);
560 if (base->Opcode() == Op_LoadP &&
561 base->in(MemNode::Address)->is_AddP()) {
562 adr = base->in(MemNode::Address);
563 Node* tls = get_addp_base(adr);
564 if (tls->Opcode() == Op_ThreadLocal) {
565 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
566 if (offs == in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset())) {
567 break; // G1 pre barrier previous oop value store.
568 }
569 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) {
570 break; // G1 post barrier card address store.
571 }
572 }
573 }
574 }
575 #endif
576 delayed_worklist->push(n); // Process unsafe access later.
577 break;
578 }
579 #ifdef ASSERT
580 n->dump(1);
581 assert(false, "not unsafe or G1 barrier raw StoreP");
582 #endif
583 }
584 break;
585 }
586 case Op_AryEq:
587 case Op_HasNegatives:
588 case Op_StrComp:
589 case Op_StrEquals:
590 case Op_StrIndexOf:
591 case Op_StrIndexOfChar:
592 case Op_StrInflatedCopy:
593 case Op_StrCompressedCopy:
594 case Op_EncodeISOArray: {
595 add_local_var(n, PointsToNode::ArgEscape);
596 delayed_worklist->push(n); // Process it later.
597 break;
598 }
599 case Op_ThreadLocal: {
600 add_java_object(n, PointsToNode::ArgEscape);
601 break;
602 }
603 default:
604 ; // Do nothing for nodes not related to EA.
605 }
606 return;
607 }
608
609 #ifdef ASSERT
610 #define ELSE_FAIL(name) \
611 /* Should not be called for not pointer type. */ \
612 n->dump(1); \
613 assert(false, name); \
614 break;
615 #else
616 #define ELSE_FAIL(name) \
617 break;
618 #endif
619
620 // Add final simple edges to graph.
621 void ConnectionGraph::add_final_edges(Node *n) {
622 PointsToNode* n_ptn = ptnode_adr(n->_idx);
623 #ifdef ASSERT
624 if (_verify && n_ptn->is_JavaObject())
625 return; // This method does not change graph for JavaObject.
626 #endif
627
628 if (n->is_Call()) {
629 process_call_arguments(n->as_Call());
630 return;
631 }
632 assert(n->is_Store() || n->is_LoadStore() ||
633 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
634 "node should be registered already");
635 int opcode = n->Opcode();
636 switch (opcode) {
637 case Op_AddP: {
638 Node* base = get_addp_base(n);
639 PointsToNode* ptn_base = ptnode_adr(base->_idx);
640 assert(ptn_base != NULL, "field's base should be registered");
641 add_base(n_ptn->as_Field(), ptn_base);
642 break;
643 }
644 case Op_CastPP:
645 case Op_CheckCastPP:
646 case Op_EncodeP:
647 case Op_DecodeN:
648 case Op_EncodePKlass:
649 case Op_DecodeNKlass: {
650 add_local_var_and_edge(n, PointsToNode::NoEscape,
651 n->in(1), NULL);
652 break;
653 }
654 case Op_CMoveP: {
655 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
656 Node* in = n->in(i);
657 if (in == NULL)
658 continue; // ignore NULL
659 Node* uncast_in = in->uncast();
660 if (uncast_in->is_top() || uncast_in == n)
661 continue; // ignore top or inputs which go back this node
662 PointsToNode* ptn = ptnode_adr(in->_idx);
663 assert(ptn != NULL, "node should be registered");
664 add_edge(n_ptn, ptn);
665 }
666 break;
667 }
668 case Op_LoadP:
669 #if INCLUDE_ZGC
670 case Op_LoadBarrierSlowReg:
671 case Op_LoadBarrierWeakSlowReg:
672 #endif
673 case Op_LoadN:
674 case Op_LoadPLocked: {
675 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
676 // ThreadLocal has RawPtr type.
677 const Type* t = _igvn->type(n);
678 if (t->make_ptr() != NULL) {
679 Node* adr = n->in(MemNode::Address);
680 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
681 break;
682 }
683 ELSE_FAIL("Op_LoadP");
684 }
685 case Op_Phi: {
686 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
687 // ThreadLocal has RawPtr type.
688 const Type* t = n->as_Phi()->type();
689 if (t->make_ptr() != NULL) {
690 for (uint i = 1; i < n->req(); i++) {
691 Node* in = n->in(i);
692 if (in == NULL)
693 continue; // ignore NULL
694 Node* uncast_in = in->uncast();
695 if (uncast_in->is_top() || uncast_in == n)
696 continue; // ignore top or inputs which go back this node
697 PointsToNode* ptn = ptnode_adr(in->_idx);
698 assert(ptn != NULL, "node should be registered");
699 add_edge(n_ptn, ptn);
700 }
701 break;
702 }
703 ELSE_FAIL("Op_Phi");
704 }
705 case Op_Proj: {
706 // we are only interested in the oop result projection from a call
707 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
708 n->in(0)->as_Call()->returns_pointer()) {
709 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
710 break;
711 }
712 #if INCLUDE_ZGC
713 else if (UseZGC) {
714 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) {
715 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), NULL);
716 break;
717 }
718 }
719 #endif
720 ELSE_FAIL("Op_Proj");
721 }
722 case Op_Rethrow: // Exception object escapes
723 case Op_Return: {
724 if (n->req() > TypeFunc::Parms &&
725 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
726 // Treat Return value as LocalVar with GlobalEscape escape state.
727 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
728 n->in(TypeFunc::Parms), NULL);
729 break;
730 }
731 ELSE_FAIL("Op_Return");
732 }
733 case Op_StoreP:
734 case Op_StoreN:
735 case Op_StoreNKlass:
736 case Op_StorePConditional:
737 case Op_CompareAndExchangeP:
738 case Op_CompareAndExchangeN:
739 case Op_CompareAndSwapP:
740 case Op_CompareAndSwapN:
741 case Op_WeakCompareAndSwapP:
742 case Op_WeakCompareAndSwapN:
743 case Op_GetAndSetP:
744 case Op_GetAndSetN: {
745 Node* adr = n->in(MemNode::Address);
746 const Type *adr_type = _igvn->type(adr);
747 adr_type = adr_type->make_ptr();
748 #ifdef ASSERT
749 if (adr_type == NULL) {
750 n->dump(1);
751 assert(adr_type != NULL, "dead node should not be on list");
752 break;
753 }
754 #endif
755 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN ||
756 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) {
757 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
758 }
759 if ( adr_type->isa_oopptr()
760 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
761 && adr_type == TypeRawPtr::NOTNULL
762 && adr->in(AddPNode::Address)->is_Proj()
763 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
764 // Point Address to Value
765 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
766 assert(adr_ptn != NULL &&
767 adr_ptn->as_Field()->is_oop(), "node should be registered");
768 Node *val = n->in(MemNode::ValueIn);
769 PointsToNode* ptn = ptnode_adr(val->_idx);
770 assert(ptn != NULL, "node should be registered");
771 add_edge(adr_ptn, ptn);
772 break;
773 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
774 // Stored value escapes in unsafe access.
775 Node *val = n->in(MemNode::ValueIn);
776 PointsToNode* ptn = ptnode_adr(val->_idx);
777 assert(ptn != NULL, "node should be registered");
778 set_escape_state(ptn, PointsToNode::GlobalEscape);
779 // Add edge to object for unsafe access with offset.
780 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
781 assert(adr_ptn != NULL, "node should be registered");
782 if (adr_ptn->is_Field()) {
783 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
784 add_edge(adr_ptn, ptn);
785 }
786 break;
787 }
788 ELSE_FAIL("Op_StoreP");
789 }
790 case Op_AryEq:
791 case Op_HasNegatives:
792 case Op_StrComp:
793 case Op_StrEquals:
794 case Op_StrIndexOf:
795 case Op_StrIndexOfChar:
796 case Op_StrInflatedCopy:
797 case Op_StrCompressedCopy:
798 case Op_EncodeISOArray: {
799 // char[]/byte[] arrays passed to string intrinsic do not escape but
800 // they are not scalar replaceable. Adjust escape state for them.
801 // Start from in(2) edge since in(1) is memory edge.
802 for (uint i = 2; i < n->req(); i++) {
803 Node* adr = n->in(i);
804 const Type* at = _igvn->type(adr);
805 if (!adr->is_top() && at->isa_ptr()) {
806 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
807 at->isa_ptr() != NULL, "expecting a pointer");
808 if (adr->is_AddP()) {
809 adr = get_addp_base(adr);
810 }
811 PointsToNode* ptn = ptnode_adr(adr->_idx);
812 assert(ptn != NULL, "node should be registered");
813 add_edge(n_ptn, ptn);
814 }
815 }
816 break;
817 }
818 default: {
819 // This method should be called only for EA specific nodes which may
820 // miss some edges when they were created.
821 #ifdef ASSERT
822 n->dump(1);
823 #endif
824 guarantee(false, "unknown node");
825 }
826 }
827 return;
828 }
829
830 void ConnectionGraph::add_call_node(CallNode* call) {
831 assert(call->returns_pointer(), "only for call which returns pointer");
832 uint call_idx = call->_idx;
833 if (call->is_Allocate()) {
834 Node* k = call->in(AllocateNode::KlassNode);
835 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
836 assert(kt != NULL, "TypeKlassPtr required.");
837 ciKlass* cik = kt->klass();
838 PointsToNode::EscapeState es = PointsToNode::NoEscape;
839 bool scalar_replaceable = true;
840 if (call->is_AllocateArray()) {
841 if (!cik->is_array_klass()) { // StressReflectiveCode
842 es = PointsToNode::GlobalEscape;
843 } else {
844 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
845 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
846 // Not scalar replaceable if the length is not constant or too big.
847 scalar_replaceable = false;
848 }
849 }
850 } else { // Allocate instance
851 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
852 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
853 !cik->is_instance_klass() || // StressReflectiveCode
854 !cik->as_instance_klass()->can_be_instantiated() ||
855 cik->as_instance_klass()->has_finalizer()) {
856 es = PointsToNode::GlobalEscape;
857 }
858 }
859 add_java_object(call, es);
860 PointsToNode* ptn = ptnode_adr(call_idx);
861 if (!scalar_replaceable && ptn->scalar_replaceable()) {
862 ptn->set_scalar_replaceable(false);
863 }
864 } else if (call->is_CallStaticJava()) {
865 // Call nodes could be different types:
866 //
867 // 1. CallDynamicJavaNode (what happened during call is unknown):
868 //
869 // - mapped to GlobalEscape JavaObject node if oop is returned;
870 //
871 // - all oop arguments are escaping globally;
872 //
873 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
874 //
875 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
876 //
877 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
878 // - mapped to NoEscape JavaObject node if non-escaping object allocated
879 // during call is returned;
880 // - mapped to ArgEscape LocalVar node pointed to object arguments
881 // which are returned and does not escape during call;
882 //
883 // - oop arguments escaping status is defined by bytecode analysis;
884 //
885 // For a static call, we know exactly what method is being called.
886 // Use bytecode estimator to record whether the call's return value escapes.
887 ciMethod* meth = call->as_CallJava()->method();
888 if (meth == NULL) {
889 const char* name = call->as_CallStaticJava()->_name;
890 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
891 // Returns a newly allocated unescaped object.
892 add_java_object(call, PointsToNode::NoEscape);
893 ptnode_adr(call_idx)->set_scalar_replaceable(false);
894 } else if (meth->is_boxing_method()) {
895 // Returns boxing object
896 PointsToNode::EscapeState es;
897 vmIntrinsics::ID intr = meth->intrinsic_id();
898 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
899 // It does not escape if object is always allocated.
900 es = PointsToNode::NoEscape;
901 } else {
902 // It escapes globally if object could be loaded from cache.
903 es = PointsToNode::GlobalEscape;
904 }
905 add_java_object(call, es);
906 } else {
907 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
908 call_analyzer->copy_dependencies(_compile->dependencies());
909 if (call_analyzer->is_return_allocated()) {
910 // Returns a newly allocated unescaped object, simply
911 // update dependency information.
912 // Mark it as NoEscape so that objects referenced by
913 // it's fields will be marked as NoEscape at least.
914 add_java_object(call, PointsToNode::NoEscape);
915 ptnode_adr(call_idx)->set_scalar_replaceable(false);
916 } else {
917 // Determine whether any arguments are returned.
918 const TypeTuple* d = call->tf()->domain();
919 bool ret_arg = false;
920 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
921 if (d->field_at(i)->isa_ptr() != NULL &&
922 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
923 ret_arg = true;
924 break;
925 }
926 }
927 if (ret_arg) {
928 add_local_var(call, PointsToNode::ArgEscape);
929 } else {
930 // Returns unknown object.
931 map_ideal_node(call, phantom_obj);
932 }
933 }
934 }
935 } else {
936 // An other type of call, assume the worst case:
937 // returned value is unknown and globally escapes.
938 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
939 map_ideal_node(call, phantom_obj);
940 }
941 }
942
943 void ConnectionGraph::process_call_arguments(CallNode *call) {
944 bool is_arraycopy = false;
945 switch (call->Opcode()) {
946 #ifdef ASSERT
947 case Op_Allocate:
948 case Op_AllocateArray:
949 case Op_Lock:
950 case Op_Unlock:
951 assert(false, "should be done already");
952 break;
953 #endif
954 case Op_ArrayCopy:
955 case Op_CallLeafNoFP:
956 // Most array copies are ArrayCopy nodes at this point but there
957 // are still a few direct calls to the copy subroutines (See
958 // PhaseStringOpts::copy_string())
959 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
960 call->as_CallLeaf()->is_call_to_arraycopystub();
961 // fall through
962 case Op_CallLeaf: {
963 // Stub calls, objects do not escape but they are not scale replaceable.
964 // Adjust escape state for outgoing arguments.
965 const TypeTuple * d = call->tf()->domain();
966 bool src_has_oops = false;
967 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
968 const Type* at = d->field_at(i);
969 Node *arg = call->in(i);
970 if (arg == NULL) {
971 continue;
972 }
973 const Type *aat = _igvn->type(arg);
974 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
975 continue;
976 if (arg->is_AddP()) {
977 //
978 // The inline_native_clone() case when the arraycopy stub is called
979 // after the allocation before Initialize and CheckCastPP nodes.
980 // Or normal arraycopy for object arrays case.
981 //
982 // Set AddP's base (Allocate) as not scalar replaceable since
983 // pointer to the base (with offset) is passed as argument.
984 //
985 arg = get_addp_base(arg);
986 }
987 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
988 assert(arg_ptn != NULL, "should be registered");
989 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
990 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
991 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
992 aat->isa_ptr() != NULL, "expecting an Ptr");
993 bool arg_has_oops = aat->isa_oopptr() &&
994 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
995 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
996 if (i == TypeFunc::Parms) {
997 src_has_oops = arg_has_oops;
998 }
999 //
1000 // src or dst could be j.l.Object when other is basic type array:
1001 //
1002 // arraycopy(char[],0,Object*,0,size);
1003 // arraycopy(Object*,0,char[],0,size);
1004 //
1005 // Don't add edges in such cases.
1006 //
1007 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
1008 arg_has_oops && (i > TypeFunc::Parms);
1009 #ifdef ASSERT
1010 if (!(is_arraycopy ||
1011 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
1012 (call->as_CallLeaf()->_name != NULL &&
1013 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
1014 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
1015 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
1016 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
1017 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
1018 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
1019 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
1020 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
1021 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
1022 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
1023 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
1024 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
1025 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
1026 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
1027 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
1028 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
1029 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
1030 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
1031 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
1032 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
1033 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
1034 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0)
1035 ))) {
1036 call->dump();
1037 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
1038 }
1039 #endif
1040 // Always process arraycopy's destination object since
1041 // we need to add all possible edges to references in
1042 // source object.
1043 if (arg_esc >= PointsToNode::ArgEscape &&
1044 !arg_is_arraycopy_dest) {
1045 continue;
1046 }
1047 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
1048 if (call->is_ArrayCopy()) {
1049 ArrayCopyNode* ac = call->as_ArrayCopy();
1050 if (ac->is_clonebasic() ||
1051 ac->is_arraycopy_validated() ||
1052 ac->is_copyof_validated() ||
1053 ac->is_copyofrange_validated()) {
1054 es = PointsToNode::NoEscape;
1055 }
1056 }
1057 set_escape_state(arg_ptn, es);
1058 if (arg_is_arraycopy_dest) {
1059 Node* src = call->in(TypeFunc::Parms);
1060 if (src->is_AddP()) {
1061 src = get_addp_base(src);
1062 }
1063 PointsToNode* src_ptn = ptnode_adr(src->_idx);
1064 assert(src_ptn != NULL, "should be registered");
1065 if (arg_ptn != src_ptn) {
1066 // Special arraycopy edge:
1067 // A destination object's field can't have the source object
1068 // as base since objects escape states are not related.
1069 // Only escape state of destination object's fields affects
1070 // escape state of fields in source object.
1071 add_arraycopy(call, es, src_ptn, arg_ptn);
1072 }
1073 }
1074 }
1075 }
1076 break;
1077 }
1078 case Op_CallStaticJava: {
1079 // For a static call, we know exactly what method is being called.
1080 // Use bytecode estimator to record the call's escape affects
1081 #ifdef ASSERT
1082 const char* name = call->as_CallStaticJava()->_name;
1083 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1084 #endif
1085 ciMethod* meth = call->as_CallJava()->method();
1086 if ((meth != NULL) && meth->is_boxing_method()) {
1087 break; // Boxing methods do not modify any oops.
1088 }
1089 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1090 // fall-through if not a Java method or no analyzer information
1091 if (call_analyzer != NULL) {
1092 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1093 const TypeTuple* d = call->tf()->domain();
1094 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1095 const Type* at = d->field_at(i);
1096 int k = i - TypeFunc::Parms;
1097 Node* arg = call->in(i);
1098 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1099 if (at->isa_ptr() != NULL &&
1100 call_analyzer->is_arg_returned(k)) {
1101 // The call returns arguments.
1102 if (call_ptn != NULL) { // Is call's result used?
1103 assert(call_ptn->is_LocalVar(), "node should be registered");
1104 assert(arg_ptn != NULL, "node should be registered");
1105 add_edge(call_ptn, arg_ptn);
1106 }
1107 }
1108 if (at->isa_oopptr() != NULL &&
1109 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1110 if (!call_analyzer->is_arg_stack(k)) {
1111 // The argument global escapes
1112 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1113 } else {
1114 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1115 if (!call_analyzer->is_arg_local(k)) {
1116 // The argument itself doesn't escape, but any fields might
1117 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1118 }
1119 }
1120 }
1121 }
1122 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1123 // The call returns arguments.
1124 assert(call_ptn->edge_count() > 0, "sanity");
1125 if (!call_analyzer->is_return_local()) {
1126 // Returns also unknown object.
1127 add_edge(call_ptn, phantom_obj);
1128 }
1129 }
1130 break;
1131 }
1132 }
1133 default: {
1134 // Fall-through here if not a Java method or no analyzer information
1135 // or some other type of call, assume the worst case: all arguments
1136 // globally escape.
1137 const TypeTuple* d = call->tf()->domain();
1138 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1139 const Type* at = d->field_at(i);
1140 if (at->isa_oopptr() != NULL) {
1141 Node* arg = call->in(i);
1142 if (arg->is_AddP()) {
1143 arg = get_addp_base(arg);
1144 }
1145 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1146 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1147 }
1148 }
1149 }
1150 }
1151 }
1152
1153
1154 // Finish Graph construction.
1155 bool ConnectionGraph::complete_connection_graph(
1156 GrowableArray<PointsToNode*>& ptnodes_worklist,
1157 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1158 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1159 GrowableArray<FieldNode*>& oop_fields_worklist) {
1160 // Normally only 1-3 passes needed to build Connection Graph depending
1161 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1162 // Set limit to 20 to catch situation when something did go wrong and
1163 // bailout Escape Analysis.
1164 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1165 #define CG_BUILD_ITER_LIMIT 20
1166
1167 // Propagate GlobalEscape and ArgEscape escape states and check that
1168 // we still have non-escaping objects. The method pushs on _worklist
1169 // Field nodes which reference phantom_object.
1170 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1171 return false; // Nothing to do.
1172 }
1173 // Now propagate references to all JavaObject nodes.
1174 int java_objects_length = java_objects_worklist.length();
1175 elapsedTimer time;
1176 bool timeout = false;
1177 int new_edges = 1;
1178 int iterations = 0;
1179 do {
1180 while ((new_edges > 0) &&
1181 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1182 double start_time = time.seconds();
1183 time.start();
1184 new_edges = 0;
1185 // Propagate references to phantom_object for nodes pushed on _worklist
1186 // by find_non_escaped_objects() and find_field_value().
1187 new_edges += add_java_object_edges(phantom_obj, false);
1188 for (int next = 0; next < java_objects_length; ++next) {
1189 JavaObjectNode* ptn = java_objects_worklist.at(next);
1190 new_edges += add_java_object_edges(ptn, true);
1191
1192 #define SAMPLE_SIZE 4
1193 if ((next % SAMPLE_SIZE) == 0) {
1194 // Each 4 iterations calculate how much time it will take
1195 // to complete graph construction.
1196 time.stop();
1197 // Poll for requests from shutdown mechanism to quiesce compiler
1198 // because Connection graph construction may take long time.
1199 CompileBroker::maybe_block();
1200 double stop_time = time.seconds();
1201 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1202 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1203 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1204 timeout = true;
1205 break; // Timeout
1206 }
1207 start_time = stop_time;
1208 time.start();
1209 }
1210 #undef SAMPLE_SIZE
1211
1212 }
1213 if (timeout) break;
1214 if (new_edges > 0) {
1215 // Update escape states on each iteration if graph was updated.
1216 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1217 return false; // Nothing to do.
1218 }
1219 }
1220 time.stop();
1221 if (time.seconds() >= EscapeAnalysisTimeout) {
1222 timeout = true;
1223 break;
1224 }
1225 }
1226 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1227 time.start();
1228 // Find fields which have unknown value.
1229 int fields_length = oop_fields_worklist.length();
1230 for (int next = 0; next < fields_length; next++) {
1231 FieldNode* field = oop_fields_worklist.at(next);
1232 if (field->edge_count() == 0) {
1233 new_edges += find_field_value(field);
1234 // This code may added new edges to phantom_object.
1235 // Need an other cycle to propagate references to phantom_object.
1236 }
1237 }
1238 time.stop();
1239 if (time.seconds() >= EscapeAnalysisTimeout) {
1240 timeout = true;
1241 break;
1242 }
1243 } else {
1244 new_edges = 0; // Bailout
1245 }
1246 } while (new_edges > 0);
1247
1248 // Bailout if passed limits.
1249 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1250 Compile* C = _compile;
1251 if (C->log() != NULL) {
1252 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1253 C->log()->text("%s", timeout ? "time" : "iterations");
1254 C->log()->end_elem(" limit'");
1255 }
1256 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1257 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length());
1258 // Possible infinite build_connection_graph loop,
1259 // bailout (no changes to ideal graph were made).
1260 return false;
1261 }
1262 #ifdef ASSERT
1263 if (Verbose && PrintEscapeAnalysis) {
1264 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1265 iterations, nodes_size(), ptnodes_worklist.length());
1266 }
1267 #endif
1268
1269 #undef CG_BUILD_ITER_LIMIT
1270
1271 // Find fields initialized by NULL for non-escaping Allocations.
1272 int non_escaped_length = non_escaped_worklist.length();
1273 for (int next = 0; next < non_escaped_length; next++) {
1274 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1275 PointsToNode::EscapeState es = ptn->escape_state();
1276 assert(es <= PointsToNode::ArgEscape, "sanity");
1277 if (es == PointsToNode::NoEscape) {
1278 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1279 // Adding references to NULL object does not change escape states
1280 // since it does not escape. Also no fields are added to NULL object.
1281 add_java_object_edges(null_obj, false);
1282 }
1283 }
1284 Node* n = ptn->ideal_node();
1285 if (n->is_Allocate()) {
1286 // The object allocated by this Allocate node will never be
1287 // seen by an other thread. Mark it so that when it is
1288 // expanded no MemBarStoreStore is added.
1289 InitializeNode* ini = n->as_Allocate()->initialization();
1290 if (ini != NULL)
1291 ini->set_does_not_escape();
1292 }
1293 }
1294 return true; // Finished graph construction.
1295 }
1296
1297 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1298 // and check that we still have non-escaping java objects.
1299 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1300 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1301 GrowableArray<PointsToNode*> escape_worklist;
1302 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1303 int ptnodes_length = ptnodes_worklist.length();
1304 for (int next = 0; next < ptnodes_length; ++next) {
1305 PointsToNode* ptn = ptnodes_worklist.at(next);
1306 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1307 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1308 escape_worklist.push(ptn);
1309 }
1310 }
1311 // Set escape states to referenced nodes (edges list).
1312 while (escape_worklist.length() > 0) {
1313 PointsToNode* ptn = escape_worklist.pop();
1314 PointsToNode::EscapeState es = ptn->escape_state();
1315 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1316 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1317 es >= PointsToNode::ArgEscape) {
1318 // GlobalEscape or ArgEscape state of field means it has unknown value.
1319 if (add_edge(ptn, phantom_obj)) {
1320 // New edge was added
1321 add_field_uses_to_worklist(ptn->as_Field());
1322 }
1323 }
1324 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1325 PointsToNode* e = i.get();
1326 if (e->is_Arraycopy()) {
1327 assert(ptn->arraycopy_dst(), "sanity");
1328 // Propagate only fields escape state through arraycopy edge.
1329 if (e->fields_escape_state() < field_es) {
1330 set_fields_escape_state(e, field_es);
1331 escape_worklist.push(e);
1332 }
1333 } else if (es >= field_es) {
1334 // fields_escape_state is also set to 'es' if it is less than 'es'.
1335 if (e->escape_state() < es) {
1336 set_escape_state(e, es);
1337 escape_worklist.push(e);
1338 }
1339 } else {
1340 // Propagate field escape state.
1341 bool es_changed = false;
1342 if (e->fields_escape_state() < field_es) {
1343 set_fields_escape_state(e, field_es);
1344 es_changed = true;
1345 }
1346 if ((e->escape_state() < field_es) &&
1347 e->is_Field() && ptn->is_JavaObject() &&
1348 e->as_Field()->is_oop()) {
1349 // Change escape state of referenced fields.
1350 set_escape_state(e, field_es);
1351 es_changed = true;
1352 } else if (e->escape_state() < es) {
1353 set_escape_state(e, es);
1354 es_changed = true;
1355 }
1356 if (es_changed) {
1357 escape_worklist.push(e);
1358 }
1359 }
1360 }
1361 }
1362 // Remove escaped objects from non_escaped list.
1363 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1364 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1365 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1366 non_escaped_worklist.delete_at(next);
1367 }
1368 if (ptn->escape_state() == PointsToNode::NoEscape) {
1369 // Find fields in non-escaped allocations which have unknown value.
1370 find_init_values(ptn, phantom_obj, NULL);
1371 }
1372 }
1373 return (non_escaped_worklist.length() > 0);
1374 }
1375
1376 // Add all references to JavaObject node by walking over all uses.
1377 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1378 int new_edges = 0;
1379 if (populate_worklist) {
1380 // Populate _worklist by uses of jobj's uses.
1381 for (UseIterator i(jobj); i.has_next(); i.next()) {
1382 PointsToNode* use = i.get();
1383 if (use->is_Arraycopy())
1384 continue;
1385 add_uses_to_worklist(use);
1386 if (use->is_Field() && use->as_Field()->is_oop()) {
1387 // Put on worklist all field's uses (loads) and
1388 // related field nodes (same base and offset).
1389 add_field_uses_to_worklist(use->as_Field());
1390 }
1391 }
1392 }
1393 for (int l = 0; l < _worklist.length(); l++) {
1394 PointsToNode* use = _worklist.at(l);
1395 if (PointsToNode::is_base_use(use)) {
1396 // Add reference from jobj to field and from field to jobj (field's base).
1397 use = PointsToNode::get_use_node(use)->as_Field();
1398 if (add_base(use->as_Field(), jobj)) {
1399 new_edges++;
1400 }
1401 continue;
1402 }
1403 assert(!use->is_JavaObject(), "sanity");
1404 if (use->is_Arraycopy()) {
1405 if (jobj == null_obj) // NULL object does not have field edges
1406 continue;
1407 // Added edge from Arraycopy node to arraycopy's source java object
1408 if (add_edge(use, jobj)) {
1409 jobj->set_arraycopy_src();
1410 new_edges++;
1411 }
1412 // and stop here.
1413 continue;
1414 }
1415 if (!add_edge(use, jobj))
1416 continue; // No new edge added, there was such edge already.
1417 new_edges++;
1418 if (use->is_LocalVar()) {
1419 add_uses_to_worklist(use);
1420 if (use->arraycopy_dst()) {
1421 for (EdgeIterator i(use); i.has_next(); i.next()) {
1422 PointsToNode* e = i.get();
1423 if (e->is_Arraycopy()) {
1424 if (jobj == null_obj) // NULL object does not have field edges
1425 continue;
1426 // Add edge from arraycopy's destination java object to Arraycopy node.
1427 if (add_edge(jobj, e)) {
1428 new_edges++;
1429 jobj->set_arraycopy_dst();
1430 }
1431 }
1432 }
1433 }
1434 } else {
1435 // Added new edge to stored in field values.
1436 // Put on worklist all field's uses (loads) and
1437 // related field nodes (same base and offset).
1438 add_field_uses_to_worklist(use->as_Field());
1439 }
1440 }
1441 _worklist.clear();
1442 _in_worklist.Reset();
1443 return new_edges;
1444 }
1445
1446 // Put on worklist all related field nodes.
1447 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1448 assert(field->is_oop(), "sanity");
1449 int offset = field->offset();
1450 add_uses_to_worklist(field);
1451 // Loop over all bases of this field and push on worklist Field nodes
1452 // with the same offset and base (since they may reference the same field).
1453 for (BaseIterator i(field); i.has_next(); i.next()) {
1454 PointsToNode* base = i.get();
1455 add_fields_to_worklist(field, base);
1456 // Check if the base was source object of arraycopy and go over arraycopy's
1457 // destination objects since values stored to a field of source object are
1458 // accessable by uses (loads) of fields of destination objects.
1459 if (base->arraycopy_src()) {
1460 for (UseIterator j(base); j.has_next(); j.next()) {
1461 PointsToNode* arycp = j.get();
1462 if (arycp->is_Arraycopy()) {
1463 for (UseIterator k(arycp); k.has_next(); k.next()) {
1464 PointsToNode* abase = k.get();
1465 if (abase->arraycopy_dst() && abase != base) {
1466 // Look for the same arraycopy reference.
1467 add_fields_to_worklist(field, abase);
1468 }
1469 }
1470 }
1471 }
1472 }
1473 }
1474 }
1475
1476 // Put on worklist all related field nodes.
1477 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1478 int offset = field->offset();
1479 if (base->is_LocalVar()) {
1480 for (UseIterator j(base); j.has_next(); j.next()) {
1481 PointsToNode* f = j.get();
1482 if (PointsToNode::is_base_use(f)) { // Field
1483 f = PointsToNode::get_use_node(f);
1484 if (f == field || !f->as_Field()->is_oop())
1485 continue;
1486 int offs = f->as_Field()->offset();
1487 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1488 add_to_worklist(f);
1489 }
1490 }
1491 }
1492 } else {
1493 assert(base->is_JavaObject(), "sanity");
1494 if (// Skip phantom_object since it is only used to indicate that
1495 // this field's content globally escapes.
1496 (base != phantom_obj) &&
1497 // NULL object node does not have fields.
1498 (base != null_obj)) {
1499 for (EdgeIterator i(base); i.has_next(); i.next()) {
1500 PointsToNode* f = i.get();
1501 // Skip arraycopy edge since store to destination object field
1502 // does not update value in source object field.
1503 if (f->is_Arraycopy()) {
1504 assert(base->arraycopy_dst(), "sanity");
1505 continue;
1506 }
1507 if (f == field || !f->as_Field()->is_oop())
1508 continue;
1509 int offs = f->as_Field()->offset();
1510 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1511 add_to_worklist(f);
1512 }
1513 }
1514 }
1515 }
1516 }
1517
1518 // Find fields which have unknown value.
1519 int ConnectionGraph::find_field_value(FieldNode* field) {
1520 // Escaped fields should have init value already.
1521 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1522 int new_edges = 0;
1523 for (BaseIterator i(field); i.has_next(); i.next()) {
1524 PointsToNode* base = i.get();
1525 if (base->is_JavaObject()) {
1526 // Skip Allocate's fields which will be processed later.
1527 if (base->ideal_node()->is_Allocate())
1528 return 0;
1529 assert(base == null_obj, "only NULL ptr base expected here");
1530 }
1531 }
1532 if (add_edge(field, phantom_obj)) {
1533 // New edge was added
1534 new_edges++;
1535 add_field_uses_to_worklist(field);
1536 }
1537 return new_edges;
1538 }
1539
1540 // Find fields initializing values for allocations.
1541 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1542 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1543 int new_edges = 0;
1544 Node* alloc = pta->ideal_node();
1545 if (init_val == phantom_obj) {
1546 // Do nothing for Allocate nodes since its fields values are
1547 // "known" unless they are initialized by arraycopy/clone.
1548 if (alloc->is_Allocate() && !pta->arraycopy_dst())
1549 return 0;
1550 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
1551 #ifdef ASSERT
1552 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) {
1553 const char* name = alloc->as_CallStaticJava()->_name;
1554 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1555 }
1556 #endif
1557 // Non-escaped allocation returned from Java or runtime call have
1558 // unknown values in fields.
1559 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1560 PointsToNode* field = i.get();
1561 if (field->is_Field() && field->as_Field()->is_oop()) {
1562 if (add_edge(field, phantom_obj)) {
1563 // New edge was added
1564 new_edges++;
1565 add_field_uses_to_worklist(field->as_Field());
1566 }
1567 }
1568 }
1569 return new_edges;
1570 }
1571 assert(init_val == null_obj, "sanity");
1572 // Do nothing for Call nodes since its fields values are unknown.
1573 if (!alloc->is_Allocate())
1574 return 0;
1575
1576 InitializeNode* ini = alloc->as_Allocate()->initialization();
1577 bool visited_bottom_offset = false;
1578 GrowableArray<int> offsets_worklist;
1579
1580 // Check if an oop field's initializing value is recorded and add
1581 // a corresponding NULL if field's value if it is not recorded.
1582 // Connection Graph does not record a default initialization by NULL
1583 // captured by Initialize node.
1584 //
1585 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1586 PointsToNode* field = i.get(); // Field (AddP)
1587 if (!field->is_Field() || !field->as_Field()->is_oop())
1588 continue; // Not oop field
1589 int offset = field->as_Field()->offset();
1590 if (offset == Type::OffsetBot) {
1591 if (!visited_bottom_offset) {
1592 // OffsetBot is used to reference array's element,
1593 // always add reference to NULL to all Field nodes since we don't
1594 // known which element is referenced.
1595 if (add_edge(field, null_obj)) {
1596 // New edge was added
1597 new_edges++;
1598 add_field_uses_to_worklist(field->as_Field());
1599 visited_bottom_offset = true;
1600 }
1601 }
1602 } else {
1603 // Check only oop fields.
1604 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1605 if (adr_type->isa_rawptr()) {
1606 #ifdef ASSERT
1607 // Raw pointers are used for initializing stores so skip it
1608 // since it should be recorded already
1609 Node* base = get_addp_base(field->ideal_node());
1610 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1611 (base->in(0) == alloc),"unexpected pointer type");
1612 #endif
1613 continue;
1614 }
1615 if (!offsets_worklist.contains(offset)) {
1616 offsets_worklist.append(offset);
1617 Node* value = NULL;
1618 if (ini != NULL) {
1619 // StoreP::memory_type() == T_ADDRESS
1620 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1621 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1622 // Make sure initializing store has the same type as this AddP.
1623 // This AddP may reference non existing field because it is on a
1624 // dead branch of bimorphic call which is not eliminated yet.
1625 if (store != NULL && store->is_Store() &&
1626 store->as_Store()->memory_type() == ft) {
1627 value = store->in(MemNode::ValueIn);
1628 #ifdef ASSERT
1629 if (VerifyConnectionGraph) {
1630 // Verify that AddP already points to all objects the value points to.
1631 PointsToNode* val = ptnode_adr(value->_idx);
1632 assert((val != NULL), "should be processed already");
1633 PointsToNode* missed_obj = NULL;
1634 if (val->is_JavaObject()) {
1635 if (!field->points_to(val->as_JavaObject())) {
1636 missed_obj = val;
1637 }
1638 } else {
1639 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1640 tty->print_cr("----------init store has invalid value -----");
1641 store->dump();
1642 val->dump();
1643 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1644 }
1645 for (EdgeIterator j(val); j.has_next(); j.next()) {
1646 PointsToNode* obj = j.get();
1647 if (obj->is_JavaObject()) {
1648 if (!field->points_to(obj->as_JavaObject())) {
1649 missed_obj = obj;
1650 break;
1651 }
1652 }
1653 }
1654 }
1655 if (missed_obj != NULL) {
1656 tty->print_cr("----------field---------------------------------");
1657 field->dump();
1658 tty->print_cr("----------missed referernce to object-----------");
1659 missed_obj->dump();
1660 tty->print_cr("----------object referernced by init store -----");
1661 store->dump();
1662 val->dump();
1663 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1664 }
1665 }
1666 #endif
1667 } else {
1668 // There could be initializing stores which follow allocation.
1669 // For example, a volatile field store is not collected
1670 // by Initialize node.
1671 //
1672 // Need to check for dependent loads to separate such stores from
1673 // stores which follow loads. For now, add initial value NULL so
1674 // that compare pointers optimization works correctly.
1675 }
1676 }
1677 if (value == NULL) {
1678 // A field's initializing value was not recorded. Add NULL.
1679 if (add_edge(field, null_obj)) {
1680 // New edge was added
1681 new_edges++;
1682 add_field_uses_to_worklist(field->as_Field());
1683 }
1684 }
1685 }
1686 }
1687 }
1688 return new_edges;
1689 }
1690
1691 // Adjust scalar_replaceable state after Connection Graph is built.
1692 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1693 // Search for non-escaping objects which are not scalar replaceable
1694 // and mark them to propagate the state to referenced objects.
1695
1696 // 1. An object is not scalar replaceable if the field into which it is
1697 // stored has unknown offset (stored into unknown element of an array).
1698 //
1699 for (UseIterator i(jobj); i.has_next(); i.next()) {
1700 PointsToNode* use = i.get();
1701 if (use->is_Arraycopy()) {
1702 continue;
1703 }
1704 if (use->is_Field()) {
1705 FieldNode* field = use->as_Field();
1706 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
1707 if (field->offset() == Type::OffsetBot) {
1708 jobj->set_scalar_replaceable(false);
1709 return;
1710 }
1711 // 2. An object is not scalar replaceable if the field into which it is
1712 // stored has multiple bases one of which is null.
1713 if (field->base_count() > 1) {
1714 for (BaseIterator i(field); i.has_next(); i.next()) {
1715 PointsToNode* base = i.get();
1716 if (base == null_obj) {
1717 jobj->set_scalar_replaceable(false);
1718 return;
1719 }
1720 }
1721 }
1722 }
1723 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1724 // 3. An object is not scalar replaceable if it is merged with other objects.
1725 for (EdgeIterator j(use); j.has_next(); j.next()) {
1726 PointsToNode* ptn = j.get();
1727 if (ptn->is_JavaObject() && ptn != jobj) {
1728 // Mark all objects.
1729 jobj->set_scalar_replaceable(false);
1730 ptn->set_scalar_replaceable(false);
1731 }
1732 }
1733 if (!jobj->scalar_replaceable()) {
1734 return;
1735 }
1736 }
1737
1738 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1739 if (j.get()->is_Arraycopy()) {
1740 continue;
1741 }
1742
1743 // Non-escaping object node should point only to field nodes.
1744 FieldNode* field = j.get()->as_Field();
1745 int offset = field->as_Field()->offset();
1746
1747 // 4. An object is not scalar replaceable if it has a field with unknown
1748 // offset (array's element is accessed in loop).
1749 if (offset == Type::OffsetBot) {
1750 jobj->set_scalar_replaceable(false);
1751 return;
1752 }
1753 // 5. Currently an object is not scalar replaceable if a LoadStore node
1754 // access its field since the field value is unknown after it.
1755 //
1756 Node* n = field->ideal_node();
1757
1758 // Test for an unsafe access that was parsed as maybe off heap
1759 // (with a CheckCastPP to raw memory).
1760 assert(n->is_AddP(), "expect an address computation");
1761 if (n->in(AddPNode::Base)->is_top() &&
1762 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
1763 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
1764 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
1765 jobj->set_scalar_replaceable(false);
1766 return;
1767 }
1768
1769 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1770 Node* u = n->fast_out(i);
1771 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
1772 jobj->set_scalar_replaceable(false);
1773 return;
1774 }
1775 }
1776
1777 // 6. Or the address may point to more then one object. This may produce
1778 // the false positive result (set not scalar replaceable)
1779 // since the flow-insensitive escape analysis can't separate
1780 // the case when stores overwrite the field's value from the case
1781 // when stores happened on different control branches.
1782 //
1783 // Note: it will disable scalar replacement in some cases:
1784 //
1785 // Point p[] = new Point[1];
1786 // p[0] = new Point(); // Will be not scalar replaced
1787 //
1788 // but it will save us from incorrect optimizations in next cases:
1789 //
1790 // Point p[] = new Point[1];
1791 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1792 //
1793 if (field->base_count() > 1) {
1794 for (BaseIterator i(field); i.has_next(); i.next()) {
1795 PointsToNode* base = i.get();
1796 // Don't take into account LocalVar nodes which
1797 // may point to only one object which should be also
1798 // this field's base by now.
1799 if (base->is_JavaObject() && base != jobj) {
1800 // Mark all bases.
1801 jobj->set_scalar_replaceable(false);
1802 base->set_scalar_replaceable(false);
1803 }
1804 }
1805 }
1806 }
1807 }
1808
1809 #ifdef ASSERT
1810 void ConnectionGraph::verify_connection_graph(
1811 GrowableArray<PointsToNode*>& ptnodes_worklist,
1812 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1813 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1814 GrowableArray<Node*>& addp_worklist) {
1815 // Verify that graph is complete - no new edges could be added.
1816 int java_objects_length = java_objects_worklist.length();
1817 int non_escaped_length = non_escaped_worklist.length();
1818 int new_edges = 0;
1819 for (int next = 0; next < java_objects_length; ++next) {
1820 JavaObjectNode* ptn = java_objects_worklist.at(next);
1821 new_edges += add_java_object_edges(ptn, true);
1822 }
1823 assert(new_edges == 0, "graph was not complete");
1824 // Verify that escape state is final.
1825 int length = non_escaped_worklist.length();
1826 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1827 assert((non_escaped_length == non_escaped_worklist.length()) &&
1828 (non_escaped_length == length) &&
1829 (_worklist.length() == 0), "escape state was not final");
1830
1831 // Verify fields information.
1832 int addp_length = addp_worklist.length();
1833 for (int next = 0; next < addp_length; ++next ) {
1834 Node* n = addp_worklist.at(next);
1835 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1836 if (field->is_oop()) {
1837 // Verify that field has all bases
1838 Node* base = get_addp_base(n);
1839 PointsToNode* ptn = ptnode_adr(base->_idx);
1840 if (ptn->is_JavaObject()) {
1841 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1842 } else {
1843 assert(ptn->is_LocalVar(), "sanity");
1844 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1845 PointsToNode* e = i.get();
1846 if (e->is_JavaObject()) {
1847 assert(field->has_base(e->as_JavaObject()), "sanity");
1848 }
1849 }
1850 }
1851 // Verify that all fields have initializing values.
1852 if (field->edge_count() == 0) {
1853 tty->print_cr("----------field does not have references----------");
1854 field->dump();
1855 for (BaseIterator i(field); i.has_next(); i.next()) {
1856 PointsToNode* base = i.get();
1857 tty->print_cr("----------field has next base---------------------");
1858 base->dump();
1859 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1860 tty->print_cr("----------base has fields-------------------------");
1861 for (EdgeIterator j(base); j.has_next(); j.next()) {
1862 j.get()->dump();
1863 }
1864 tty->print_cr("----------base has references---------------------");
1865 for (UseIterator j(base); j.has_next(); j.next()) {
1866 j.get()->dump();
1867 }
1868 }
1869 }
1870 for (UseIterator i(field); i.has_next(); i.next()) {
1871 i.get()->dump();
1872 }
1873 assert(field->edge_count() > 0, "sanity");
1874 }
1875 }
1876 }
1877 }
1878 #endif
1879
1880 // Optimize ideal graph.
1881 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1882 GrowableArray<Node*>& storestore_worklist) {
1883 Compile* C = _compile;
1884 PhaseIterGVN* igvn = _igvn;
1885 if (EliminateLocks) {
1886 // Mark locks before changing ideal graph.
1887 int cnt = C->macro_count();
1888 for( int i=0; i < cnt; i++ ) {
1889 Node *n = C->macro_node(i);
1890 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1891 AbstractLockNode* alock = n->as_AbstractLock();
1892 if (!alock->is_non_esc_obj()) {
1893 if (not_global_escape(alock->obj_node())) {
1894 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1895 // The lock could be marked eliminated by lock coarsening
1896 // code during first IGVN before EA. Replace coarsened flag
1897 // to eliminate all associated locks/unlocks.
1898 #ifdef ASSERT
1899 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1900 #endif
1901 alock->set_non_esc_obj();
1902 }
1903 }
1904 }
1905 }
1906 }
1907
1908 if (OptimizePtrCompare) {
1909 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1910 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1911 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1912 // Optimize objects compare.
1913 while (ptr_cmp_worklist.length() != 0) {
1914 Node *n = ptr_cmp_worklist.pop();
1915 Node *res = optimize_ptr_compare(n);
1916 if (res != NULL) {
1917 #ifndef PRODUCT
1918 if (PrintOptimizePtrCompare) {
1919 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"));
1920 if (Verbose) {
1921 n->dump(1);
1922 }
1923 }
1924 #endif
1925 igvn->replace_node(n, res);
1926 }
1927 }
1928 // cleanup
1929 if (_pcmp_neq->outcnt() == 0)
1930 igvn->hash_delete(_pcmp_neq);
1931 if (_pcmp_eq->outcnt() == 0)
1932 igvn->hash_delete(_pcmp_eq);
1933 }
1934
1935 // For MemBarStoreStore nodes added in library_call.cpp, check
1936 // escape status of associated AllocateNode and optimize out
1937 // MemBarStoreStore node if the allocated object never escapes.
1938 while (storestore_worklist.length() != 0) {
1939 Node *n = storestore_worklist.pop();
1940 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1941 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1942 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1943 if (not_global_escape(alloc)) {
1944 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1945 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1946 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1947 igvn->register_new_node_with_optimizer(mb);
1948 igvn->replace_node(storestore, mb);
1949 }
1950 }
1951 }
1952
1953 // Optimize objects compare.
1954 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1955 assert(OptimizePtrCompare, "sanity");
1956 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1957 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1958 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1959 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1960 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1961 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1962
1963 // Check simple cases first.
1964 if (jobj1 != NULL) {
1965 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1966 if (jobj1 == jobj2) {
1967 // Comparing the same not escaping object.
1968 return _pcmp_eq;
1969 }
1970 Node* obj = jobj1->ideal_node();
1971 // Comparing not escaping allocation.
1972 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1973 !ptn2->points_to(jobj1)) {
1974 return _pcmp_neq; // This includes nullness check.
1975 }
1976 }
1977 }
1978 if (jobj2 != NULL) {
1979 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1980 Node* obj = jobj2->ideal_node();
1981 // Comparing not escaping allocation.
1982 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1983 !ptn1->points_to(jobj2)) {
1984 return _pcmp_neq; // This includes nullness check.
1985 }
1986 }
1987 }
1988 if (jobj1 != NULL && jobj1 != phantom_obj &&
1989 jobj2 != NULL && jobj2 != phantom_obj &&
1990 jobj1->ideal_node()->is_Con() &&
1991 jobj2->ideal_node()->is_Con()) {
1992 // Klass or String constants compare. Need to be careful with
1993 // compressed pointers - compare types of ConN and ConP instead of nodes.
1994 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1995 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1996 if (t1->make_ptr() == t2->make_ptr()) {
1997 return _pcmp_eq;
1998 } else {
1999 return _pcmp_neq;
2000 }
2001 }
2002 if (ptn1->meet(ptn2)) {
2003 return NULL; // Sets are not disjoint
2004 }
2005
2006 // Sets are disjoint.
2007 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
2008 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
2009 bool set1_has_null_ptr = ptn1->points_to(null_obj);
2010 bool set2_has_null_ptr = ptn2->points_to(null_obj);
2011 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
2012 (set2_has_unknown_ptr && set1_has_null_ptr)) {
2013 // Check nullness of unknown object.
2014 return NULL;
2015 }
2016
2017 // Disjointness by itself is not sufficient since
2018 // alias analysis is not complete for escaped objects.
2019 // Disjoint sets are definitely unrelated only when
2020 // at least one set has only not escaping allocations.
2021 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
2022 if (ptn1->non_escaping_allocation()) {
2023 return _pcmp_neq;
2024 }
2025 }
2026 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2027 if (ptn2->non_escaping_allocation()) {
2028 return _pcmp_neq;
2029 }
2030 }
2031 return NULL;
2032 }
2033
2034 // Connection Graph constuction functions.
2035
2036 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
2037 PointsToNode* ptadr = _nodes.at(n->_idx);
2038 if (ptadr != NULL) {
2039 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
2040 return;
2041 }
2042 Compile* C = _compile;
2043 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
2044 _nodes.at_put(n->_idx, ptadr);
2045 }
2046
2047 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
2048 PointsToNode* ptadr = _nodes.at(n->_idx);
2049 if (ptadr != NULL) {
2050 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
2051 return;
2052 }
2053 Compile* C = _compile;
2054 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
2055 _nodes.at_put(n->_idx, ptadr);
2056 }
2057
2058 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
2059 PointsToNode* ptadr = _nodes.at(n->_idx);
2060 if (ptadr != NULL) {
2061 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
2062 return;
2063 }
2064 bool unsafe = false;
2065 bool is_oop = is_oop_field(n, offset, &unsafe);
2066 if (unsafe) {
2067 es = PointsToNode::GlobalEscape;
2068 }
2069 Compile* C = _compile;
2070 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
2071 _nodes.at_put(n->_idx, field);
2072 }
2073
2074 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
2075 PointsToNode* src, PointsToNode* dst) {
2076 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
2077 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
2078 PointsToNode* ptadr = _nodes.at(n->_idx);
2079 if (ptadr != NULL) {
2080 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
2081 return;
2082 }
2083 Compile* C = _compile;
2084 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
2085 _nodes.at_put(n->_idx, ptadr);
2086 // Add edge from arraycopy node to source object.
2087 (void)add_edge(ptadr, src);
2088 src->set_arraycopy_src();
2089 // Add edge from destination object to arraycopy node.
2090 (void)add_edge(dst, ptadr);
2091 dst->set_arraycopy_dst();
2092 }
2093
2094 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
2095 const Type* adr_type = n->as_AddP()->bottom_type();
2096 BasicType bt = T_INT;
2097 if (offset == Type::OffsetBot) {
2098 // Check only oop fields.
2099 if (!adr_type->isa_aryptr() ||
2100 (adr_type->isa_aryptr()->klass() == NULL) ||
2101 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2102 // OffsetBot is used to reference array's element. Ignore first AddP.
2103 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2104 bt = T_OBJECT;
2105 }
2106 }
2107 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2108 if (adr_type->isa_instptr()) {
2109 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
2110 if (field != NULL) {
2111 bt = field->layout_type();
2112 } else {
2113 // Check for unsafe oop field access
2114 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2115 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2116 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2117 bt = T_OBJECT;
2118 (*unsafe) = true;
2119 }
2120 }
2121 } else if (adr_type->isa_aryptr()) {
2122 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2123 // Ignore array length load.
2124 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2125 // Ignore first AddP.
2126 } else {
2127 const Type* elemtype = adr_type->isa_aryptr()->elem();
2128 bt = elemtype->array_element_basic_type();
2129 }
2130 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2131 // Allocation initialization, ThreadLocal field access, unsafe access
2132 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2133 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2134 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2135 bt = T_OBJECT;
2136 }
2137 }
2138 }
2139 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2140 }
2141
2142 // Returns unique pointed java object or NULL.
2143 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2144 assert(!_collecting, "should not call when contructed graph");
2145 // If the node was created after the escape computation we can't answer.
2146 uint idx = n->_idx;
2147 if (idx >= nodes_size()) {
2148 return NULL;
2149 }
2150 PointsToNode* ptn = ptnode_adr(idx);
2151 if (ptn == NULL) {
2152 return NULL;
2153 }
2154 if (ptn->is_JavaObject()) {
2155 return ptn->as_JavaObject();
2156 }
2157 assert(ptn->is_LocalVar(), "sanity");
2158 // Check all java objects it points to.
2159 JavaObjectNode* jobj = NULL;
2160 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2161 PointsToNode* e = i.get();
2162 if (e->is_JavaObject()) {
2163 if (jobj == NULL) {
2164 jobj = e->as_JavaObject();
2165 } else if (jobj != e) {
2166 return NULL;
2167 }
2168 }
2169 }
2170 return jobj;
2171 }
2172
2173 // Return true if this node points only to non-escaping allocations.
2174 bool PointsToNode::non_escaping_allocation() {
2175 if (is_JavaObject()) {
2176 Node* n = ideal_node();
2177 if (n->is_Allocate() || n->is_CallStaticJava()) {
2178 return (escape_state() == PointsToNode::NoEscape);
2179 } else {
2180 return false;
2181 }
2182 }
2183 assert(is_LocalVar(), "sanity");
2184 // Check all java objects it points to.
2185 for (EdgeIterator i(this); i.has_next(); i.next()) {
2186 PointsToNode* e = i.get();
2187 if (e->is_JavaObject()) {
2188 Node* n = e->ideal_node();
2189 if ((e->escape_state() != PointsToNode::NoEscape) ||
2190 !(n->is_Allocate() || n->is_CallStaticJava())) {
2191 return false;
2192 }
2193 }
2194 }
2195 return true;
2196 }
2197
2198 // Return true if we know the node does not escape globally.
2199 bool ConnectionGraph::not_global_escape(Node *n) {
2200 assert(!_collecting, "should not call during graph construction");
2201 // If the node was created after the escape computation we can't answer.
2202 uint idx = n->_idx;
2203 if (idx >= nodes_size()) {
2204 return false;
2205 }
2206 PointsToNode* ptn = ptnode_adr(idx);
2207 if (ptn == NULL) {
2208 return false; // not in congraph (e.g. ConI)
2209 }
2210 PointsToNode::EscapeState es = ptn->escape_state();
2211 // If we have already computed a value, return it.
2212 if (es >= PointsToNode::GlobalEscape)
2213 return false;
2214 if (ptn->is_JavaObject()) {
2215 return true; // (es < PointsToNode::GlobalEscape);
2216 }
2217 assert(ptn->is_LocalVar(), "sanity");
2218 // Check all java objects it points to.
2219 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2220 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2221 return false;
2222 }
2223 return true;
2224 }
2225
2226
2227 // Helper functions
2228
2229 // Return true if this node points to specified node or nodes it points to.
2230 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2231 if (is_JavaObject()) {
2232 return (this == ptn);
2233 }
2234 assert(is_LocalVar() || is_Field(), "sanity");
2235 for (EdgeIterator i(this); i.has_next(); i.next()) {
2236 if (i.get() == ptn)
2237 return true;
2238 }
2239 return false;
2240 }
2241
2242 // Return true if one node points to an other.
2243 bool PointsToNode::meet(PointsToNode* ptn) {
2244 if (this == ptn) {
2245 return true;
2246 } else if (ptn->is_JavaObject()) {
2247 return this->points_to(ptn->as_JavaObject());
2248 } else if (this->is_JavaObject()) {
2249 return ptn->points_to(this->as_JavaObject());
2250 }
2251 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2252 int ptn_count = ptn->edge_count();
2253 for (EdgeIterator i(this); i.has_next(); i.next()) {
2254 PointsToNode* this_e = i.get();
2255 for (int j = 0; j < ptn_count; j++) {
2256 if (this_e == ptn->edge(j))
2257 return true;
2258 }
2259 }
2260 return false;
2261 }
2262
2263 #ifdef ASSERT
2264 // Return true if bases point to this java object.
2265 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2266 for (BaseIterator i(this); i.has_next(); i.next()) {
2267 if (i.get() == jobj)
2268 return true;
2269 }
2270 return false;
2271 }
2272 #endif
2273
2274 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2275 const Type *adr_type = phase->type(adr);
2276 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2277 adr->in(AddPNode::Address)->is_Proj() &&
2278 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2279 // We are computing a raw address for a store captured by an Initialize
2280 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2281 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2282 assert(offs != Type::OffsetBot ||
2283 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2284 "offset must be a constant or it is initialization of array");
2285 return offs;
2286 }
2287 const TypePtr *t_ptr = adr_type->isa_ptr();
2288 assert(t_ptr != NULL, "must be a pointer type");
2289 return t_ptr->offset();
2290 }
2291
2292 Node* ConnectionGraph::get_addp_base(Node *addp) {
2293 assert(addp->is_AddP(), "must be AddP");
2294 //
2295 // AddP cases for Base and Address inputs:
2296 // case #1. Direct object's field reference:
2297 // Allocate
2298 // |
2299 // Proj #5 ( oop result )
2300 // |
2301 // CheckCastPP (cast to instance type)
2302 // | |
2303 // AddP ( base == address )
2304 //
2305 // case #2. Indirect object's field reference:
2306 // Phi
2307 // |
2308 // CastPP (cast to instance type)
2309 // | |
2310 // AddP ( base == address )
2311 //
2312 // case #3. Raw object's field reference for Initialize node:
2313 // Allocate
2314 // |
2315 // Proj #5 ( oop result )
2316 // top |
2317 // \ |
2318 // AddP ( base == top )
2319 //
2320 // case #4. Array's element reference:
2321 // {CheckCastPP | CastPP}
2322 // | | |
2323 // | AddP ( array's element offset )
2324 // | |
2325 // AddP ( array's offset )
2326 //
2327 // case #5. Raw object's field reference for arraycopy stub call:
2328 // The inline_native_clone() case when the arraycopy stub is called
2329 // after the allocation before Initialize and CheckCastPP nodes.
2330 // Allocate
2331 // |
2332 // Proj #5 ( oop result )
2333 // | |
2334 // AddP ( base == address )
2335 //
2336 // case #6. Constant Pool, ThreadLocal, CastX2P or
2337 // Raw object's field reference:
2338 // {ConP, ThreadLocal, CastX2P, raw Load}
2339 // top |
2340 // \ |
2341 // AddP ( base == top )
2342 //
2343 // case #7. Klass's field reference.
2344 // LoadKlass
2345 // | |
2346 // AddP ( base == address )
2347 //
2348 // case #8. narrow Klass's field reference.
2349 // LoadNKlass
2350 // |
2351 // DecodeN
2352 // | |
2353 // AddP ( base == address )
2354 //
2355 // case #9. Mixed unsafe access
2356 // {instance}
2357 // |
2358 // CheckCastPP (raw)
2359 // top |
2360 // \ |
2361 // AddP ( base == top )
2362 //
2363 Node *base = addp->in(AddPNode::Base);
2364 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
2365 base = addp->in(AddPNode::Address);
2366 while (base->is_AddP()) {
2367 // Case #6 (unsafe access) may have several chained AddP nodes.
2368 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2369 base = base->in(AddPNode::Address);
2370 }
2371 if (base->Opcode() == Op_CheckCastPP &&
2372 base->bottom_type()->isa_rawptr() &&
2373 _igvn->type(base->in(1))->isa_oopptr()) {
2374 base = base->in(1); // Case #9
2375 } else {
2376 Node* uncast_base = base->uncast();
2377 int opcode = uncast_base->Opcode();
2378 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2379 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2380 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2381 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2382 }
2383 }
2384 return base;
2385 }
2386
2387 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2388 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2389 Node* addp2 = addp->raw_out(0);
2390 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2391 addp2->in(AddPNode::Base) == n &&
2392 addp2->in(AddPNode::Address) == addp) {
2393 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2394 //
2395 // Find array's offset to push it on worklist first and
2396 // as result process an array's element offset first (pushed second)
2397 // to avoid CastPP for the array's offset.
2398 // Otherwise the inserted CastPP (LocalVar) will point to what
2399 // the AddP (Field) points to. Which would be wrong since
2400 // the algorithm expects the CastPP has the same point as
2401 // as AddP's base CheckCastPP (LocalVar).
2402 //
2403 // ArrayAllocation
2404 // |
2405 // CheckCastPP
2406 // |
2407 // memProj (from ArrayAllocation CheckCastPP)
2408 // | ||
2409 // | || Int (element index)
2410 // | || | ConI (log(element size))
2411 // | || | /
2412 // | || LShift
2413 // | || /
2414 // | AddP (array's element offset)
2415 // | |
2416 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2417 // | / /
2418 // AddP (array's offset)
2419 // |
2420 // Load/Store (memory operation on array's element)
2421 //
2422 return addp2;
2423 }
2424 return NULL;
2425 }
2426
2427 //
2428 // Adjust the type and inputs of an AddP which computes the
2429 // address of a field of an instance
2430 //
2431 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2432 PhaseGVN* igvn = _igvn;
2433 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2434 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2435 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2436 if (t == NULL) {
2437 // We are computing a raw address for a store captured by an Initialize
2438 // compute an appropriate address type (cases #3 and #5).
2439 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2440 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2441 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2442 assert(offs != Type::OffsetBot, "offset must be a constant");
2443 t = base_t->add_offset(offs)->is_oopptr();
2444 }
2445 int inst_id = base_t->instance_id();
2446 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2447 "old type must be non-instance or match new type");
2448
2449 // The type 't' could be subclass of 'base_t'.
2450 // As result t->offset() could be large then base_t's size and it will
2451 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2452 // constructor verifies correctness of the offset.
2453 //
2454 // It could happened on subclass's branch (from the type profiling
2455 // inlining) which was not eliminated during parsing since the exactness
2456 // of the allocation type was not propagated to the subclass type check.
2457 //
2458 // Or the type 't' could be not related to 'base_t' at all.
2459 // It could happened when CHA type is different from MDO type on a dead path
2460 // (for example, from instanceof check) which is not collapsed during parsing.
2461 //
2462 // Do nothing for such AddP node and don't process its users since
2463 // this code branch will go away.
2464 //
2465 if (!t->is_known_instance() &&
2466 !base_t->klass()->is_subtype_of(t->klass())) {
2467 return false; // bail out
2468 }
2469 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2470 // Do NOT remove the next line: ensure a new alias index is allocated
2471 // for the instance type. Note: C++ will not remove it since the call
2472 // has side effect.
2473 int alias_idx = _compile->get_alias_index(tinst);
2474 igvn->set_type(addp, tinst);
2475 // record the allocation in the node map
2476 set_map(addp, get_map(base->_idx));
2477 // Set addp's Base and Address to 'base'.
2478 Node *abase = addp->in(AddPNode::Base);
2479 Node *adr = addp->in(AddPNode::Address);
2480 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2481 adr->in(0)->_idx == (uint)inst_id) {
2482 // Skip AddP cases #3 and #5.
2483 } else {
2484 assert(!abase->is_top(), "sanity"); // AddP case #3
2485 if (abase != base) {
2486 igvn->hash_delete(addp);
2487 addp->set_req(AddPNode::Base, base);
2488 if (abase == adr) {
2489 addp->set_req(AddPNode::Address, base);
2490 } else {
2491 // AddP case #4 (adr is array's element offset AddP node)
2492 #ifdef ASSERT
2493 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2494 assert(adr->is_AddP() && atype != NULL &&
2495 atype->instance_id() == inst_id, "array's element offset should be processed first");
2496 #endif
2497 }
2498 igvn->hash_insert(addp);
2499 }
2500 }
2501 // Put on IGVN worklist since at least addp's type was changed above.
2502 record_for_optimizer(addp);
2503 return true;
2504 }
2505
2506 //
2507 // Create a new version of orig_phi if necessary. Returns either the newly
2508 // created phi or an existing phi. Sets create_new to indicate whether a new
2509 // phi was created. Cache the last newly created phi in the node map.
2510 //
2511 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2512 Compile *C = _compile;
2513 PhaseGVN* igvn = _igvn;
2514 new_created = false;
2515 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2516 // nothing to do if orig_phi is bottom memory or matches alias_idx
2517 if (phi_alias_idx == alias_idx) {
2518 return orig_phi;
2519 }
2520 // Have we recently created a Phi for this alias index?
2521 PhiNode *result = get_map_phi(orig_phi->_idx);
2522 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2523 return result;
2524 }
2525 // Previous check may fail when the same wide memory Phi was split into Phis
2526 // for different memory slices. Search all Phis for this region.
2527 if (result != NULL) {
2528 Node* region = orig_phi->in(0);
2529 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2530 Node* phi = region->fast_out(i);
2531 if (phi->is_Phi() &&
2532 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2533 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2534 return phi->as_Phi();
2535 }
2536 }
2537 }
2538 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2539 if (C->do_escape_analysis() == true && !C->failing()) {
2540 // Retry compilation without escape analysis.
2541 // If this is the first failure, the sentinel string will "stick"
2542 // to the Compile object, and the C2Compiler will see it and retry.
2543 C->record_failure(C2Compiler::retry_no_escape_analysis());
2544 }
2545 return NULL;
2546 }
2547 orig_phi_worklist.append_if_missing(orig_phi);
2548 const TypePtr *atype = C->get_adr_type(alias_idx);
2549 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2550 C->copy_node_notes_to(result, orig_phi);
2551 igvn->set_type(result, result->bottom_type());
2552 record_for_optimizer(result);
2553 set_map(orig_phi, result);
2554 new_created = true;
2555 return result;
2556 }
2557
2558 //
2559 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2560 // specified alias index.
2561 //
2562 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2563 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2564 Compile *C = _compile;
2565 PhaseGVN* igvn = _igvn;
2566 bool new_phi_created;
2567 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2568 if (!new_phi_created) {
2569 return result;
2570 }
2571 GrowableArray<PhiNode *> phi_list;
2572 GrowableArray<uint> cur_input;
2573 PhiNode *phi = orig_phi;
2574 uint idx = 1;
2575 bool finished = false;
2576 while(!finished) {
2577 while (idx < phi->req()) {
2578 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2579 if (mem != NULL && mem->is_Phi()) {
2580 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2581 if (new_phi_created) {
2582 // found an phi for which we created a new split, push current one on worklist and begin
2583 // processing new one
2584 phi_list.push(phi);
2585 cur_input.push(idx);
2586 phi = mem->as_Phi();
2587 result = newphi;
2588 idx = 1;
2589 continue;
2590 } else {
2591 mem = newphi;
2592 }
2593 }
2594 if (C->failing()) {
2595 return NULL;
2596 }
2597 result->set_req(idx++, mem);
2598 }
2599 #ifdef ASSERT
2600 // verify that the new Phi has an input for each input of the original
2601 assert( phi->req() == result->req(), "must have same number of inputs.");
2602 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2603 #endif
2604 // Check if all new phi's inputs have specified alias index.
2605 // Otherwise use old phi.
2606 for (uint i = 1; i < phi->req(); i++) {
2607 Node* in = result->in(i);
2608 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2609 }
2610 // we have finished processing a Phi, see if there are any more to do
2611 finished = (phi_list.length() == 0 );
2612 if (!finished) {
2613 phi = phi_list.pop();
2614 idx = cur_input.pop();
2615 PhiNode *prev_result = get_map_phi(phi->_idx);
2616 prev_result->set_req(idx++, result);
2617 result = prev_result;
2618 }
2619 }
2620 return result;
2621 }
2622
2623 //
2624 // The next methods are derived from methods in MemNode.
2625 //
2626 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2627 Node *mem = mmem;
2628 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2629 // means an array I have not precisely typed yet. Do not do any
2630 // alias stuff with it any time soon.
2631 if (toop->base() != Type::AnyPtr &&
2632 !(toop->klass() != NULL &&
2633 toop->klass()->is_java_lang_Object() &&
2634 toop->offset() == Type::OffsetBot)) {
2635 mem = mmem->memory_at(alias_idx);
2636 // Update input if it is progress over what we have now
2637 }
2638 return mem;
2639 }
2640
2641 //
2642 // Move memory users to their memory slices.
2643 //
2644 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2645 Compile* C = _compile;
2646 PhaseGVN* igvn = _igvn;
2647 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2648 assert(tp != NULL, "ptr type");
2649 int alias_idx = C->get_alias_index(tp);
2650 int general_idx = C->get_general_index(alias_idx);
2651
2652 // Move users first
2653 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2654 Node* use = n->fast_out(i);
2655 if (use->is_MergeMem()) {
2656 MergeMemNode* mmem = use->as_MergeMem();
2657 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2658 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2659 continue; // Nothing to do
2660 }
2661 // Replace previous general reference to mem node.
2662 uint orig_uniq = C->unique();
2663 Node* m = find_inst_mem(n, general_idx, orig_phis);
2664 assert(orig_uniq == C->unique(), "no new nodes");
2665 mmem->set_memory_at(general_idx, m);
2666 --imax;
2667 --i;
2668 } else if (use->is_MemBar()) {
2669 assert(!use->is_Initialize(), "initializing stores should not be moved");
2670 if (use->req() > MemBarNode::Precedent &&
2671 use->in(MemBarNode::Precedent) == n) {
2672 // Don't move related membars.
2673 record_for_optimizer(use);
2674 continue;
2675 }
2676 tp = use->as_MemBar()->adr_type()->isa_ptr();
2677 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) ||
2678 alias_idx == general_idx) {
2679 continue; // Nothing to do
2680 }
2681 // Move to general memory slice.
2682 uint orig_uniq = C->unique();
2683 Node* m = find_inst_mem(n, general_idx, orig_phis);
2684 assert(orig_uniq == C->unique(), "no new nodes");
2685 igvn->hash_delete(use);
2686 imax -= use->replace_edge(n, m);
2687 igvn->hash_insert(use);
2688 record_for_optimizer(use);
2689 --i;
2690 #ifdef ASSERT
2691 } else if (use->is_Mem()) {
2692 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2693 // Don't move related cardmark.
2694 continue;
2695 }
2696 // Memory nodes should have new memory input.
2697 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2698 assert(tp != NULL, "ptr type");
2699 int idx = C->get_alias_index(tp);
2700 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2701 "Following memory nodes should have new memory input or be on the same memory slice");
2702 } else if (use->is_Phi()) {
2703 // Phi nodes should be split and moved already.
2704 tp = use->as_Phi()->adr_type()->isa_ptr();
2705 assert(tp != NULL, "ptr type");
2706 int idx = C->get_alias_index(tp);
2707 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2708 } else {
2709 use->dump();
2710 assert(false, "should not be here");
2711 #endif
2712 }
2713 }
2714 }
2715
2716 //
2717 // Search memory chain of "mem" to find a MemNode whose address
2718 // is the specified alias index.
2719 //
2720 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2721 if (orig_mem == NULL)
2722 return orig_mem;
2723 Compile* C = _compile;
2724 PhaseGVN* igvn = _igvn;
2725 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2726 bool is_instance = (toop != NULL) && toop->is_known_instance();
2727 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
2728 Node *prev = NULL;
2729 Node *result = orig_mem;
2730 while (prev != result) {
2731 prev = result;
2732 if (result == start_mem)
2733 break; // hit one of our sentinels
2734 if (result->is_Mem()) {
2735 const Type *at = igvn->type(result->in(MemNode::Address));
2736 if (at == Type::TOP)
2737 break; // Dead
2738 assert (at->isa_ptr() != NULL, "pointer type required.");
2739 int idx = C->get_alias_index(at->is_ptr());
2740 if (idx == alias_idx)
2741 break; // Found
2742 if (!is_instance && (at->isa_oopptr() == NULL ||
2743 !at->is_oopptr()->is_known_instance())) {
2744 break; // Do not skip store to general memory slice.
2745 }
2746 result = result->in(MemNode::Memory);
2747 }
2748 if (!is_instance)
2749 continue; // don't search further for non-instance types
2750 // skip over a call which does not affect this memory slice
2751 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2752 Node *proj_in = result->in(0);
2753 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2754 break; // hit one of our sentinels
2755 } else if (proj_in->is_Call()) {
2756 // ArrayCopy node processed here as well
2757 CallNode *call = proj_in->as_Call();
2758 if (!call->may_modify(toop, igvn)) {
2759 result = call->in(TypeFunc::Memory);
2760 }
2761 } else if (proj_in->is_Initialize()) {
2762 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2763 // Stop if this is the initialization for the object instance which
2764 // which contains this memory slice, otherwise skip over it.
2765 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2766 result = proj_in->in(TypeFunc::Memory);
2767 }
2768 } else if (proj_in->is_MemBar()) {
2769 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() &&
2770 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() &&
2771 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) {
2772 // clone
2773 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy();
2774 if (ac->may_modify(toop, igvn)) {
2775 break;
2776 }
2777 }
2778 result = proj_in->in(TypeFunc::Memory);
2779 }
2780 } else if (result->is_MergeMem()) {
2781 MergeMemNode *mmem = result->as_MergeMem();
2782 result = step_through_mergemem(mmem, alias_idx, toop);
2783 if (result == mmem->base_memory()) {
2784 // Didn't find instance memory, search through general slice recursively.
2785 result = mmem->memory_at(C->get_general_index(alias_idx));
2786 result = find_inst_mem(result, alias_idx, orig_phis);
2787 if (C->failing()) {
2788 return NULL;
2789 }
2790 mmem->set_memory_at(alias_idx, result);
2791 }
2792 } else if (result->is_Phi() &&
2793 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2794 Node *un = result->as_Phi()->unique_input(igvn);
2795 if (un != NULL) {
2796 orig_phis.append_if_missing(result->as_Phi());
2797 result = un;
2798 } else {
2799 break;
2800 }
2801 } else if (result->is_ClearArray()) {
2802 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2803 // Can not bypass initialization of the instance
2804 // we are looking for.
2805 break;
2806 }
2807 // Otherwise skip it (the call updated 'result' value).
2808 } else if (result->Opcode() == Op_SCMemProj) {
2809 Node* mem = result->in(0);
2810 Node* adr = NULL;
2811 if (mem->is_LoadStore()) {
2812 adr = mem->in(MemNode::Address);
2813 } else {
2814 assert(mem->Opcode() == Op_EncodeISOArray ||
2815 mem->Opcode() == Op_StrCompressedCopy, "sanity");
2816 adr = mem->in(3); // Memory edge corresponds to destination array
2817 }
2818 const Type *at = igvn->type(adr);
2819 if (at != Type::TOP) {
2820 assert(at->isa_ptr() != NULL, "pointer type required.");
2821 int idx = C->get_alias_index(at->is_ptr());
2822 if (idx == alias_idx) {
2823 // Assert in debug mode
2824 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
2825 break; // In product mode return SCMemProj node
2826 }
2827 }
2828 result = mem->in(MemNode::Memory);
2829 } else if (result->Opcode() == Op_StrInflatedCopy) {
2830 Node* adr = result->in(3); // Memory edge corresponds to destination array
2831 const Type *at = igvn->type(adr);
2832 if (at != Type::TOP) {
2833 assert(at->isa_ptr() != NULL, "pointer type required.");
2834 int idx = C->get_alias_index(at->is_ptr());
2835 if (idx == alias_idx) {
2836 // Assert in debug mode
2837 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
2838 break; // In product mode return SCMemProj node
2839 }
2840 }
2841 result = result->in(MemNode::Memory);
2842 }
2843 }
2844 if (result->is_Phi()) {
2845 PhiNode *mphi = result->as_Phi();
2846 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2847 const TypePtr *t = mphi->adr_type();
2848 if (!is_instance) {
2849 // Push all non-instance Phis on the orig_phis worklist to update inputs
2850 // during Phase 4 if needed.
2851 orig_phis.append_if_missing(mphi);
2852 } else if (C->get_alias_index(t) != alias_idx) {
2853 // Create a new Phi with the specified alias index type.
2854 result = split_memory_phi(mphi, alias_idx, orig_phis);
2855 }
2856 }
2857 // the result is either MemNode, PhiNode, InitializeNode.
2858 return result;
2859 }
2860
2861 //
2862 // Convert the types of unescaped object to instance types where possible,
2863 // propagate the new type information through the graph, and update memory
2864 // edges and MergeMem inputs to reflect the new type.
2865 //
2866 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2867 // The processing is done in 4 phases:
2868 //
2869 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2870 // types for the CheckCastPP for allocations where possible.
2871 // Propagate the new types through users as follows:
2872 // casts and Phi: push users on alloc_worklist
2873 // AddP: cast Base and Address inputs to the instance type
2874 // push any AddP users on alloc_worklist and push any memnode
2875 // users onto memnode_worklist.
2876 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2877 // search the Memory chain for a store with the appropriate type
2878 // address type. If a Phi is found, create a new version with
2879 // the appropriate memory slices from each of the Phi inputs.
2880 // For stores, process the users as follows:
2881 // MemNode: push on memnode_worklist
2882 // MergeMem: push on mergemem_worklist
2883 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2884 // moving the first node encountered of each instance type to the
2885 // the input corresponding to its alias index.
2886 // appropriate memory slice.
2887 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2888 //
2889 // In the following example, the CheckCastPP nodes are the cast of allocation
2890 // results and the allocation of node 29 is unescaped and eligible to be an
2891 // instance type.
2892 //
2893 // We start with:
2894 //
2895 // 7 Parm #memory
2896 // 10 ConI "12"
2897 // 19 CheckCastPP "Foo"
2898 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2899 // 29 CheckCastPP "Foo"
2900 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2901 //
2902 // 40 StoreP 25 7 20 ... alias_index=4
2903 // 50 StoreP 35 40 30 ... alias_index=4
2904 // 60 StoreP 45 50 20 ... alias_index=4
2905 // 70 LoadP _ 60 30 ... alias_index=4
2906 // 80 Phi 75 50 60 Memory alias_index=4
2907 // 90 LoadP _ 80 30 ... alias_index=4
2908 // 100 LoadP _ 80 20 ... alias_index=4
2909 //
2910 //
2911 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2912 // and creating a new alias index for node 30. This gives:
2913 //
2914 // 7 Parm #memory
2915 // 10 ConI "12"
2916 // 19 CheckCastPP "Foo"
2917 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2918 // 29 CheckCastPP "Foo" iid=24
2919 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2920 //
2921 // 40 StoreP 25 7 20 ... alias_index=4
2922 // 50 StoreP 35 40 30 ... alias_index=6
2923 // 60 StoreP 45 50 20 ... alias_index=4
2924 // 70 LoadP _ 60 30 ... alias_index=6
2925 // 80 Phi 75 50 60 Memory alias_index=4
2926 // 90 LoadP _ 80 30 ... alias_index=6
2927 // 100 LoadP _ 80 20 ... alias_index=4
2928 //
2929 // In phase 2, new memory inputs are computed for the loads and stores,
2930 // And a new version of the phi is created. In phase 4, the inputs to
2931 // node 80 are updated and then the memory nodes are updated with the
2932 // values computed in phase 2. This results in:
2933 //
2934 // 7 Parm #memory
2935 // 10 ConI "12"
2936 // 19 CheckCastPP "Foo"
2937 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2938 // 29 CheckCastPP "Foo" iid=24
2939 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2940 //
2941 // 40 StoreP 25 7 20 ... alias_index=4
2942 // 50 StoreP 35 7 30 ... alias_index=6
2943 // 60 StoreP 45 40 20 ... alias_index=4
2944 // 70 LoadP _ 50 30 ... alias_index=6
2945 // 80 Phi 75 40 60 Memory alias_index=4
2946 // 120 Phi 75 50 50 Memory alias_index=6
2947 // 90 LoadP _ 120 30 ... alias_index=6
2948 // 100 LoadP _ 80 20 ... alias_index=4
2949 //
2950 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) {
2951 GrowableArray<Node *> memnode_worklist;
2952 GrowableArray<PhiNode *> orig_phis;
2953 PhaseIterGVN *igvn = _igvn;
2954 uint new_index_start = (uint) _compile->num_alias_types();
2955 Arena* arena = Thread::current()->resource_area();
2956 VectorSet visited(arena);
2957 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2958 uint unique_old = _compile->unique();
2959
2960 // Phase 1: Process possible allocations from alloc_worklist.
2961 // Create instance types for the CheckCastPP for allocations where possible.
2962 //
2963 // (Note: don't forget to change the order of the second AddP node on
2964 // the alloc_worklist if the order of the worklist processing is changed,
2965 // see the comment in find_second_addp().)
2966 //
2967 while (alloc_worklist.length() != 0) {
2968 Node *n = alloc_worklist.pop();
2969 uint ni = n->_idx;
2970 if (n->is_Call()) {
2971 CallNode *alloc = n->as_Call();
2972 // copy escape information to call node
2973 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2974 PointsToNode::EscapeState es = ptn->escape_state();
2975 // We have an allocation or call which returns a Java object,
2976 // see if it is unescaped.
2977 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2978 continue;
2979 // Find CheckCastPP for the allocate or for the return value of a call
2980 n = alloc->result_cast();
2981 if (n == NULL) { // No uses except Initialize node
2982 if (alloc->is_Allocate()) {
2983 // Set the scalar_replaceable flag for allocation
2984 // so it could be eliminated if it has no uses.
2985 alloc->as_Allocate()->_is_scalar_replaceable = true;
2986 }
2987 if (alloc->is_CallStaticJava()) {
2988 // Set the scalar_replaceable flag for boxing method
2989 // so it could be eliminated if it has no uses.
2990 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2991 }
2992 continue;
2993 }
2994 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2995 assert(!alloc->is_Allocate(), "allocation should have unique type");
2996 continue;
2997 }
2998
2999 // The inline code for Object.clone() casts the allocation result to
3000 // java.lang.Object and then to the actual type of the allocated
3001 // object. Detect this case and use the second cast.
3002 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
3003 // the allocation result is cast to java.lang.Object and then
3004 // to the actual Array type.
3005 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
3006 && (alloc->is_AllocateArray() ||
3007 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
3008 Node *cast2 = NULL;
3009 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3010 Node *use = n->fast_out(i);
3011 if (use->is_CheckCastPP()) {
3012 cast2 = use;
3013 break;
3014 }
3015 }
3016 if (cast2 != NULL) {
3017 n = cast2;
3018 } else {
3019 // Non-scalar replaceable if the allocation type is unknown statically
3020 // (reflection allocation), the object can't be restored during
3021 // deoptimization without precise type.
3022 continue;
3023 }
3024 }
3025
3026 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
3027 if (t == NULL)
3028 continue; // not a TypeOopPtr
3029 if (!t->klass_is_exact())
3030 continue; // not an unique type
3031
3032 if (alloc->is_Allocate()) {
3033 // Set the scalar_replaceable flag for allocation
3034 // so it could be eliminated.
3035 alloc->as_Allocate()->_is_scalar_replaceable = true;
3036 }
3037 if (alloc->is_CallStaticJava()) {
3038 // Set the scalar_replaceable flag for boxing method
3039 // so it could be eliminated.
3040 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3041 }
3042 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
3043 // in order for an object to be scalar-replaceable, it must be:
3044 // - a direct allocation (not a call returning an object)
3045 // - non-escaping
3046 // - eligible to be a unique type
3047 // - not determined to be ineligible by escape analysis
3048 set_map(alloc, n);
3049 set_map(n, alloc);
3050 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
3051 igvn->hash_delete(n);
3052 igvn->set_type(n, tinst);
3053 n->raise_bottom_type(tinst);
3054 igvn->hash_insert(n);
3055 record_for_optimizer(n);
3056 // Allocate an alias index for the header fields. Accesses to
3057 // the header emitted during macro expansion wouldn't have
3058 // correct memory state otherwise.
3059 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
3060 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
3061 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
3062
3063 // First, put on the worklist all Field edges from Connection Graph
3064 // which is more accurate than putting immediate users from Ideal Graph.
3065 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
3066 PointsToNode* tgt = e.get();
3067 if (tgt->is_Arraycopy()) {
3068 continue;
3069 }
3070 Node* use = tgt->ideal_node();
3071 assert(tgt->is_Field() && use->is_AddP(),
3072 "only AddP nodes are Field edges in CG");
3073 if (use->outcnt() > 0) { // Don't process dead nodes
3074 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
3075 if (addp2 != NULL) {
3076 assert(alloc->is_AllocateArray(),"array allocation was expected");
3077 alloc_worklist.append_if_missing(addp2);
3078 }
3079 alloc_worklist.append_if_missing(use);
3080 }
3081 }
3082
3083 // An allocation may have an Initialize which has raw stores. Scan
3084 // the users of the raw allocation result and push AddP users
3085 // on alloc_worklist.
3086 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
3087 assert (raw_result != NULL, "must have an allocation result");
3088 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
3089 Node *use = raw_result->fast_out(i);
3090 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
3091 Node* addp2 = find_second_addp(use, raw_result);
3092 if (addp2 != NULL) {
3093 assert(alloc->is_AllocateArray(),"array allocation was expected");
3094 alloc_worklist.append_if_missing(addp2);
3095 }
3096 alloc_worklist.append_if_missing(use);
3097 } else if (use->is_MemBar()) {
3098 memnode_worklist.append_if_missing(use);
3099 }
3100 }
3101 }
3102 } else if (n->is_AddP()) {
3103 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
3104 if (jobj == NULL || jobj == phantom_obj) {
3105 #ifdef ASSERT
3106 ptnode_adr(get_addp_base(n)->_idx)->dump();
3107 ptnode_adr(n->_idx)->dump();
3108 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3109 #endif
3110 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3111 return;
3112 }
3113 Node *base = get_map(jobj->idx()); // CheckCastPP node
3114 if (!split_AddP(n, base)) continue; // wrong type from dead path
3115 } else if (n->is_Phi() ||
3116 n->is_CheckCastPP() ||
3117 n->is_EncodeP() ||
3118 n->is_DecodeN() ||
3119 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
3120 if (visited.test_set(n->_idx)) {
3121 assert(n->is_Phi(), "loops only through Phi's");
3122 continue; // already processed
3123 }
3124 JavaObjectNode* jobj = unique_java_object(n);
3125 if (jobj == NULL || jobj == phantom_obj) {
3126 #ifdef ASSERT
3127 ptnode_adr(n->_idx)->dump();
3128 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3129 #endif
3130 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3131 return;
3132 } else {
3133 Node *val = get_map(jobj->idx()); // CheckCastPP node
3134 TypeNode *tn = n->as_Type();
3135 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
3136 assert(tinst != NULL && tinst->is_known_instance() &&
3137 tinst->instance_id() == jobj->idx() , "instance type expected.");
3138
3139 const Type *tn_type = igvn->type(tn);
3140 const TypeOopPtr *tn_t;
3141 if (tn_type->isa_narrowoop()) {
3142 tn_t = tn_type->make_ptr()->isa_oopptr();
3143 } else {
3144 tn_t = tn_type->isa_oopptr();
3145 }
3146 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
3147 if (tn_type->isa_narrowoop()) {
3148 tn_type = tinst->make_narrowoop();
3149 } else {
3150 tn_type = tinst;
3151 }
3152 igvn->hash_delete(tn);
3153 igvn->set_type(tn, tn_type);
3154 tn->set_type(tn_type);
3155 igvn->hash_insert(tn);
3156 record_for_optimizer(n);
3157 } else {
3158 assert(tn_type == TypePtr::NULL_PTR ||
3159 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3160 "unexpected type");
3161 continue; // Skip dead path with different type
3162 }
3163 }
3164 } else {
3165 debug_only(n->dump();)
3166 assert(false, "EA: unexpected node");
3167 continue;
3168 }
3169 // push allocation's users on appropriate worklist
3170 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3171 Node *use = n->fast_out(i);
3172 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3173 // Load/store to instance's field
3174 memnode_worklist.append_if_missing(use);
3175 } else if (use->is_MemBar()) {
3176 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3177 memnode_worklist.append_if_missing(use);
3178 }
3179 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3180 Node* addp2 = find_second_addp(use, n);
3181 if (addp2 != NULL) {
3182 alloc_worklist.append_if_missing(addp2);
3183 }
3184 alloc_worklist.append_if_missing(use);
3185 } else if (use->is_Phi() ||
3186 use->is_CheckCastPP() ||
3187 use->is_EncodeNarrowPtr() ||
3188 use->is_DecodeNarrowPtr() ||
3189 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3190 alloc_worklist.append_if_missing(use);
3191 #ifdef ASSERT
3192 } else if (use->is_Mem()) {
3193 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3194 } else if (use->is_MergeMem()) {
3195 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3196 } else if (use->is_SafePoint()) {
3197 // Look for MergeMem nodes for calls which reference unique allocation
3198 // (through CheckCastPP nodes) even for debug info.
3199 Node* m = use->in(TypeFunc::Memory);
3200 if (m->is_MergeMem()) {
3201 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3202 }
3203 } else if (use->Opcode() == Op_EncodeISOArray) {
3204 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3205 // EncodeISOArray overwrites destination array
3206 memnode_worklist.append_if_missing(use);
3207 }
3208 } else {
3209 uint op = use->Opcode();
3210 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
3211 (use->in(MemNode::Memory) == n)) {
3212 // They overwrite memory edge corresponding to destination array,
3213 memnode_worklist.append_if_missing(use);
3214 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
3215 op == Op_CastP2X || op == Op_StoreCM ||
3216 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3217 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3218 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
3219 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
3220 n->dump();
3221 use->dump();
3222 assert(false, "EA: missing allocation reference path");
3223 }
3224 #endif
3225 }
3226 }
3227
3228 }
3229
3230 // Go over all ArrayCopy nodes and if one of the inputs has a unique
3231 // type, record it in the ArrayCopy node so we know what memory this
3232 // node uses/modified.
3233 for (int next = 0; next < arraycopy_worklist.length(); next++) {
3234 ArrayCopyNode* ac = arraycopy_worklist.at(next);
3235 Node* dest = ac->in(ArrayCopyNode::Dest);
3236 if (dest->is_AddP()) {
3237 dest = get_addp_base(dest);
3238 }
3239 JavaObjectNode* jobj = unique_java_object(dest);
3240 if (jobj != NULL) {
3241 Node *base = get_map(jobj->idx());
3242 if (base != NULL) {
3243 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3244 ac->_dest_type = base_t;
3245 }
3246 }
3247 Node* src = ac->in(ArrayCopyNode::Src);
3248 if (src->is_AddP()) {
3249 src = get_addp_base(src);
3250 }
3251 jobj = unique_java_object(src);
3252 if (jobj != NULL) {
3253 Node* base = get_map(jobj->idx());
3254 if (base != NULL) {
3255 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3256 ac->_src_type = base_t;
3257 }
3258 }
3259 }
3260
3261 // New alias types were created in split_AddP().
3262 uint new_index_end = (uint) _compile->num_alias_types();
3263 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3264
3265 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3266 // compute new values for Memory inputs (the Memory inputs are not
3267 // actually updated until phase 4.)
3268 if (memnode_worklist.length() == 0)
3269 return; // nothing to do
3270 while (memnode_worklist.length() != 0) {
3271 Node *n = memnode_worklist.pop();
3272 if (visited.test_set(n->_idx))
3273 continue;
3274 if (n->is_Phi() || n->is_ClearArray()) {
3275 // we don't need to do anything, but the users must be pushed
3276 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3277 // we don't need to do anything, but the users must be pushed
3278 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
3279 if (n == NULL)
3280 continue;
3281 } else if (n->Opcode() == Op_StrCompressedCopy ||
3282 n->Opcode() == Op_EncodeISOArray) {
3283 // get the memory projection
3284 n = n->find_out_with(Op_SCMemProj);
3285 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3286 } else {
3287 assert(n->is_Mem(), "memory node required.");
3288 Node *addr = n->in(MemNode::Address);
3289 const Type *addr_t = igvn->type(addr);
3290 if (addr_t == Type::TOP)
3291 continue;
3292 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3293 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3294 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3295 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3296 if (_compile->failing()) {
3297 return;
3298 }
3299 if (mem != n->in(MemNode::Memory)) {
3300 // We delay the memory edge update since we need old one in
3301 // MergeMem code below when instances memory slices are separated.
3302 set_map(n, mem);
3303 }
3304 if (n->is_Load()) {
3305 continue; // don't push users
3306 } else if (n->is_LoadStore()) {
3307 // get the memory projection
3308 n = n->find_out_with(Op_SCMemProj);
3309 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3310 }
3311 }
3312 // push user on appropriate worklist
3313 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3314 Node *use = n->fast_out(i);
3315 if (use->is_Phi() || use->is_ClearArray()) {
3316 memnode_worklist.append_if_missing(use);
3317 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3318 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3319 continue;
3320 memnode_worklist.append_if_missing(use);
3321 } else if (use->is_MemBar()) {
3322 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3323 memnode_worklist.append_if_missing(use);
3324 }
3325 #ifdef ASSERT
3326 } else if(use->is_Mem()) {
3327 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3328 } else if (use->is_MergeMem()) {
3329 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3330 } else if (use->Opcode() == Op_EncodeISOArray) {
3331 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3332 // EncodeISOArray overwrites destination array
3333 memnode_worklist.append_if_missing(use);
3334 }
3335 } else {
3336 uint op = use->Opcode();
3337 if ((use->in(MemNode::Memory) == n) &&
3338 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
3339 // They overwrite memory edge corresponding to destination array,
3340 memnode_worklist.append_if_missing(use);
3341 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3342 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3343 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3344 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
3345 n->dump();
3346 use->dump();
3347 assert(false, "EA: missing memory path");
3348 }
3349 #endif
3350 }
3351 }
3352 }
3353
3354 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3355 // Walk each memory slice moving the first node encountered of each
3356 // instance type to the the input corresponding to its alias index.
3357 uint length = _mergemem_worklist.length();
3358 for( uint next = 0; next < length; ++next ) {
3359 MergeMemNode* nmm = _mergemem_worklist.at(next);
3360 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3361 // Note: we don't want to use MergeMemStream here because we only want to
3362 // scan inputs which exist at the start, not ones we add during processing.
3363 // Note 2: MergeMem may already contains instance memory slices added
3364 // during find_inst_mem() call when memory nodes were processed above.
3365 igvn->hash_delete(nmm);
3366 uint nslices = MIN2(nmm->req(), new_index_start);
3367 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3368 Node* mem = nmm->in(i);
3369 Node* cur = NULL;
3370 if (mem == NULL || mem->is_top())
3371 continue;
3372 // First, update mergemem by moving memory nodes to corresponding slices
3373 // if their type became more precise since this mergemem was created.
3374 while (mem->is_Mem()) {
3375 const Type *at = igvn->type(mem->in(MemNode::Address));
3376 if (at != Type::TOP) {
3377 assert (at->isa_ptr() != NULL, "pointer type required.");
3378 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3379 if (idx == i) {
3380 if (cur == NULL)
3381 cur = mem;
3382 } else {
3383 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3384 nmm->set_memory_at(idx, mem);
3385 }
3386 }
3387 }
3388 mem = mem->in(MemNode::Memory);
3389 }
3390 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3391 // Find any instance of the current type if we haven't encountered
3392 // already a memory slice of the instance along the memory chain.
3393 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3394 if((uint)_compile->get_general_index(ni) == i) {
3395 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3396 if (nmm->is_empty_memory(m)) {
3397 Node* result = find_inst_mem(mem, ni, orig_phis);
3398 if (_compile->failing()) {
3399 return;
3400 }
3401 nmm->set_memory_at(ni, result);
3402 }
3403 }
3404 }
3405 }
3406 // Find the rest of instances values
3407 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3408 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3409 Node* result = step_through_mergemem(nmm, ni, tinst);
3410 if (result == nmm->base_memory()) {
3411 // Didn't find instance memory, search through general slice recursively.
3412 result = nmm->memory_at(_compile->get_general_index(ni));
3413 result = find_inst_mem(result, ni, orig_phis);
3414 if (_compile->failing()) {
3415 return;
3416 }
3417 nmm->set_memory_at(ni, result);
3418 }
3419 }
3420 igvn->hash_insert(nmm);
3421 record_for_optimizer(nmm);
3422 }
3423
3424 // Phase 4: Update the inputs of non-instance memory Phis and
3425 // the Memory input of memnodes
3426 // First update the inputs of any non-instance Phi's from
3427 // which we split out an instance Phi. Note we don't have
3428 // to recursively process Phi's encounted on the input memory
3429 // chains as is done in split_memory_phi() since they will
3430 // also be processed here.
3431 for (int j = 0; j < orig_phis.length(); j++) {
3432 PhiNode *phi = orig_phis.at(j);
3433 int alias_idx = _compile->get_alias_index(phi->adr_type());
3434 igvn->hash_delete(phi);
3435 for (uint i = 1; i < phi->req(); i++) {
3436 Node *mem = phi->in(i);
3437 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3438 if (_compile->failing()) {
3439 return;
3440 }
3441 if (mem != new_mem) {
3442 phi->set_req(i, new_mem);
3443 }
3444 }
3445 igvn->hash_insert(phi);
3446 record_for_optimizer(phi);
3447 }
3448
3449 // Update the memory inputs of MemNodes with the value we computed
3450 // in Phase 2 and move stores memory users to corresponding memory slices.
3451 // Disable memory split verification code until the fix for 6984348.
3452 // Currently it produces false negative results since it does not cover all cases.
3453 #if 0 // ifdef ASSERT
3454 visited.Reset();
3455 Node_Stack old_mems(arena, _compile->unique() >> 2);
3456 #endif
3457 for (uint i = 0; i < ideal_nodes.size(); i++) {
3458 Node* n = ideal_nodes.at(i);
3459 Node* nmem = get_map(n->_idx);
3460 assert(nmem != NULL, "sanity");
3461 if (n->is_Mem()) {
3462 #if 0 // ifdef ASSERT
3463 Node* old_mem = n->in(MemNode::Memory);
3464 if (!visited.test_set(old_mem->_idx)) {
3465 old_mems.push(old_mem, old_mem->outcnt());
3466 }
3467 #endif
3468 assert(n->in(MemNode::Memory) != nmem, "sanity");
3469 if (!n->is_Load()) {
3470 // Move memory users of a store first.
3471 move_inst_mem(n, orig_phis);
3472 }
3473 // Now update memory input
3474 igvn->hash_delete(n);
3475 n->set_req(MemNode::Memory, nmem);
3476 igvn->hash_insert(n);
3477 record_for_optimizer(n);
3478 } else {
3479 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3480 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3481 }
3482 }
3483 #if 0 // ifdef ASSERT
3484 // Verify that memory was split correctly
3485 while (old_mems.is_nonempty()) {
3486 Node* old_mem = old_mems.node();
3487 uint old_cnt = old_mems.index();
3488 old_mems.pop();
3489 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3490 }
3491 #endif
3492 }
3493
3494 #ifndef PRODUCT
3495 static const char *node_type_names[] = {
3496 "UnknownType",
3497 "JavaObject",
3498 "LocalVar",
3499 "Field",
3500 "Arraycopy"
3501 };
3502
3503 static const char *esc_names[] = {
3504 "UnknownEscape",
3505 "NoEscape",
3506 "ArgEscape",
3507 "GlobalEscape"
3508 };
3509
3510 void PointsToNode::dump(bool print_state) const {
3511 NodeType nt = node_type();
3512 tty->print("%s ", node_type_names[(int) nt]);
3513 if (print_state) {
3514 EscapeState es = escape_state();
3515 EscapeState fields_es = fields_escape_state();
3516 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3517 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3518 tty->print("NSR ");
3519 }
3520 if (is_Field()) {
3521 FieldNode* f = (FieldNode*)this;
3522 if (f->is_oop())
3523 tty->print("oop ");
3524 if (f->offset() > 0)
3525 tty->print("+%d ", f->offset());
3526 tty->print("(");
3527 for (BaseIterator i(f); i.has_next(); i.next()) {
3528 PointsToNode* b = i.get();
3529 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3530 }
3531 tty->print(" )");
3532 }
3533 tty->print("[");
3534 for (EdgeIterator i(this); i.has_next(); i.next()) {
3535 PointsToNode* e = i.get();
3536 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3537 }
3538 tty->print(" [");
3539 for (UseIterator i(this); i.has_next(); i.next()) {
3540 PointsToNode* u = i.get();
3541 bool is_base = false;
3542 if (PointsToNode::is_base_use(u)) {
3543 is_base = true;
3544 u = PointsToNode::get_use_node(u)->as_Field();
3545 }
3546 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3547 }
3548 tty->print(" ]] ");
3549 if (_node == NULL)
3550 tty->print_cr("<null>");
3551 else
3552 _node->dump();
3553 }
3554
3555 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3556 bool first = true;
3557 int ptnodes_length = ptnodes_worklist.length();
3558 for (int i = 0; i < ptnodes_length; i++) {
3559 PointsToNode *ptn = ptnodes_worklist.at(i);
3560 if (ptn == NULL || !ptn->is_JavaObject())
3561 continue;
3562 PointsToNode::EscapeState es = ptn->escape_state();
3563 if ((es != PointsToNode::NoEscape) && !Verbose) {
3564 continue;
3565 }
3566 Node* n = ptn->ideal_node();
3567 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3568 n->as_CallStaticJava()->is_boxing_method())) {
3569 if (first) {
3570 tty->cr();
3571 tty->print("======== Connection graph for ");
3572 _compile->method()->print_short_name();
3573 tty->cr();
3574 first = false;
3575 }
3576 ptn->dump();
3577 // Print all locals and fields which reference this allocation
3578 for (UseIterator j(ptn); j.has_next(); j.next()) {
3579 PointsToNode* use = j.get();
3580 if (use->is_LocalVar()) {
3581 use->dump(Verbose);
3582 } else if (Verbose) {
3583 use->dump();
3584 }
3585 }
3586 tty->cr();
3587 }
3588 }
3589 }
3590 #endif