1 /*
2 * Copyright (c) 2005, 2018, 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, "sha1_implCompress") == 0 ||
1023 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
1024 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
1025 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
1026 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
1027 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
1028 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
1029 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
1030 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
1031 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
1032 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
1033 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0)
1034 ))) {
1035 call->dump();
1036 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
1037 }
1038 #endif
1039 // Always process arraycopy's destination object since
1040 // we need to add all possible edges to references in
1041 // source object.
1042 if (arg_esc >= PointsToNode::ArgEscape &&
1043 !arg_is_arraycopy_dest) {
1044 continue;
1045 }
1046 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
1047 if (call->is_ArrayCopy()) {
1048 ArrayCopyNode* ac = call->as_ArrayCopy();
1049 if (ac->is_clonebasic() ||
1050 ac->is_arraycopy_validated() ||
1051 ac->is_copyof_validated() ||
1052 ac->is_copyofrange_validated()) {
1053 es = PointsToNode::NoEscape;
1054 }
1055 }
1056 set_escape_state(arg_ptn, es);
1057 if (arg_is_arraycopy_dest) {
1058 Node* src = call->in(TypeFunc::Parms);
1059 if (src->is_AddP()) {
1060 src = get_addp_base(src);
1061 }
1062 PointsToNode* src_ptn = ptnode_adr(src->_idx);
1063 assert(src_ptn != NULL, "should be registered");
1064 if (arg_ptn != src_ptn) {
1065 // Special arraycopy edge:
1066 // A destination object's field can't have the source object
1067 // as base since objects escape states are not related.
1068 // Only escape state of destination object's fields affects
1069 // escape state of fields in source object.
1070 add_arraycopy(call, es, src_ptn, arg_ptn);
1071 }
1072 }
1073 }
1074 }
1075 break;
1076 }
1077 case Op_CallStaticJava: {
1078 // For a static call, we know exactly what method is being called.
1079 // Use bytecode estimator to record the call's escape affects
1080 #ifdef ASSERT
1081 const char* name = call->as_CallStaticJava()->_name;
1082 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1083 #endif
1084 ciMethod* meth = call->as_CallJava()->method();
1085 if ((meth != NULL) && meth->is_boxing_method()) {
1086 break; // Boxing methods do not modify any oops.
1087 }
1088 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1089 // fall-through if not a Java method or no analyzer information
1090 if (call_analyzer != NULL) {
1091 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1092 const TypeTuple* d = call->tf()->domain();
1093 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1094 const Type* at = d->field_at(i);
1095 int k = i - TypeFunc::Parms;
1096 Node* arg = call->in(i);
1097 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1098 if (at->isa_ptr() != NULL &&
1099 call_analyzer->is_arg_returned(k)) {
1100 // The call returns arguments.
1101 if (call_ptn != NULL) { // Is call's result used?
1102 assert(call_ptn->is_LocalVar(), "node should be registered");
1103 assert(arg_ptn != NULL, "node should be registered");
1104 add_edge(call_ptn, arg_ptn);
1105 }
1106 }
1107 if (at->isa_oopptr() != NULL &&
1108 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1109 if (!call_analyzer->is_arg_stack(k)) {
1110 // The argument global escapes
1111 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1112 } else {
1113 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1114 if (!call_analyzer->is_arg_local(k)) {
1115 // The argument itself doesn't escape, but any fields might
1116 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1117 }
1118 }
1119 }
1120 }
1121 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1122 // The call returns arguments.
1123 assert(call_ptn->edge_count() > 0, "sanity");
1124 if (!call_analyzer->is_return_local()) {
1125 // Returns also unknown object.
1126 add_edge(call_ptn, phantom_obj);
1127 }
1128 }
1129 break;
1130 }
1131 }
1132 default: {
1133 // Fall-through here if not a Java method or no analyzer information
1134 // or some other type of call, assume the worst case: all arguments
1135 // globally escape.
1136 const TypeTuple* d = call->tf()->domain();
1137 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1138 const Type* at = d->field_at(i);
1139 if (at->isa_oopptr() != NULL) {
1140 Node* arg = call->in(i);
1141 if (arg->is_AddP()) {
1142 arg = get_addp_base(arg);
1143 }
1144 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1145 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1146 }
1147 }
1148 }
1149 }
1150 }
1151
1152
1153 // Finish Graph construction.
1154 bool ConnectionGraph::complete_connection_graph(
1155 GrowableArray<PointsToNode*>& ptnodes_worklist,
1156 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1157 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1158 GrowableArray<FieldNode*>& oop_fields_worklist) {
1159 // Normally only 1-3 passes needed to build Connection Graph depending
1160 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1161 // Set limit to 20 to catch situation when something did go wrong and
1162 // bailout Escape Analysis.
1163 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1164 #define CG_BUILD_ITER_LIMIT 20
1165
1166 // Propagate GlobalEscape and ArgEscape escape states and check that
1167 // we still have non-escaping objects. The method pushs on _worklist
1168 // Field nodes which reference phantom_object.
1169 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1170 return false; // Nothing to do.
1171 }
1172 // Now propagate references to all JavaObject nodes.
1173 int java_objects_length = java_objects_worklist.length();
1174 elapsedTimer time;
1175 bool timeout = false;
1176 int new_edges = 1;
1177 int iterations = 0;
1178 do {
1179 while ((new_edges > 0) &&
1180 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1181 double start_time = time.seconds();
1182 time.start();
1183 new_edges = 0;
1184 // Propagate references to phantom_object for nodes pushed on _worklist
1185 // by find_non_escaped_objects() and find_field_value().
1186 new_edges += add_java_object_edges(phantom_obj, false);
1187 for (int next = 0; next < java_objects_length; ++next) {
1188 JavaObjectNode* ptn = java_objects_worklist.at(next);
1189 new_edges += add_java_object_edges(ptn, true);
1190
1191 #define SAMPLE_SIZE 4
1192 if ((next % SAMPLE_SIZE) == 0) {
1193 // Each 4 iterations calculate how much time it will take
1194 // to complete graph construction.
1195 time.stop();
1196 // Poll for requests from shutdown mechanism to quiesce compiler
1197 // because Connection graph construction may take long time.
1198 CompileBroker::maybe_block();
1199 double stop_time = time.seconds();
1200 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1201 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1202 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1203 timeout = true;
1204 break; // Timeout
1205 }
1206 start_time = stop_time;
1207 time.start();
1208 }
1209 #undef SAMPLE_SIZE
1210
1211 }
1212 if (timeout) break;
1213 if (new_edges > 0) {
1214 // Update escape states on each iteration if graph was updated.
1215 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1216 return false; // Nothing to do.
1217 }
1218 }
1219 time.stop();
1220 if (time.seconds() >= EscapeAnalysisTimeout) {
1221 timeout = true;
1222 break;
1223 }
1224 }
1225 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1226 time.start();
1227 // Find fields which have unknown value.
1228 int fields_length = oop_fields_worklist.length();
1229 for (int next = 0; next < fields_length; next++) {
1230 FieldNode* field = oop_fields_worklist.at(next);
1231 if (field->edge_count() == 0) {
1232 new_edges += find_field_value(field);
1233 // This code may added new edges to phantom_object.
1234 // Need an other cycle to propagate references to phantom_object.
1235 }
1236 }
1237 time.stop();
1238 if (time.seconds() >= EscapeAnalysisTimeout) {
1239 timeout = true;
1240 break;
1241 }
1242 } else {
1243 new_edges = 0; // Bailout
1244 }
1245 } while (new_edges > 0);
1246
1247 // Bailout if passed limits.
1248 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1249 Compile* C = _compile;
1250 if (C->log() != NULL) {
1251 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1252 C->log()->text("%s", timeout ? "time" : "iterations");
1253 C->log()->end_elem(" limit'");
1254 }
1255 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1256 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length());
1257 // Possible infinite build_connection_graph loop,
1258 // bailout (no changes to ideal graph were made).
1259 return false;
1260 }
1261 #ifdef ASSERT
1262 if (Verbose && PrintEscapeAnalysis) {
1263 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1264 iterations, nodes_size(), ptnodes_worklist.length());
1265 }
1266 #endif
1267
1268 #undef CG_BUILD_ITER_LIMIT
1269
1270 // Find fields initialized by NULL for non-escaping Allocations.
1271 int non_escaped_length = non_escaped_worklist.length();
1272 for (int next = 0; next < non_escaped_length; next++) {
1273 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1274 PointsToNode::EscapeState es = ptn->escape_state();
1275 assert(es <= PointsToNode::ArgEscape, "sanity");
1276 if (es == PointsToNode::NoEscape) {
1277 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1278 // Adding references to NULL object does not change escape states
1279 // since it does not escape. Also no fields are added to NULL object.
1280 add_java_object_edges(null_obj, false);
1281 }
1282 }
1283 Node* n = ptn->ideal_node();
1284 if (n->is_Allocate()) {
1285 // The object allocated by this Allocate node will never be
1286 // seen by an other thread. Mark it so that when it is
1287 // expanded no MemBarStoreStore is added.
1288 InitializeNode* ini = n->as_Allocate()->initialization();
1289 if (ini != NULL)
1290 ini->set_does_not_escape();
1291 }
1292 }
1293 return true; // Finished graph construction.
1294 }
1295
1296 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1297 // and check that we still have non-escaping java objects.
1298 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1299 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1300 GrowableArray<PointsToNode*> escape_worklist;
1301 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1302 int ptnodes_length = ptnodes_worklist.length();
1303 for (int next = 0; next < ptnodes_length; ++next) {
1304 PointsToNode* ptn = ptnodes_worklist.at(next);
1305 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1306 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1307 escape_worklist.push(ptn);
1308 }
1309 }
1310 // Set escape states to referenced nodes (edges list).
1311 while (escape_worklist.length() > 0) {
1312 PointsToNode* ptn = escape_worklist.pop();
1313 PointsToNode::EscapeState es = ptn->escape_state();
1314 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1315 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1316 es >= PointsToNode::ArgEscape) {
1317 // GlobalEscape or ArgEscape state of field means it has unknown value.
1318 if (add_edge(ptn, phantom_obj)) {
1319 // New edge was added
1320 add_field_uses_to_worklist(ptn->as_Field());
1321 }
1322 }
1323 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1324 PointsToNode* e = i.get();
1325 if (e->is_Arraycopy()) {
1326 assert(ptn->arraycopy_dst(), "sanity");
1327 // Propagate only fields escape state through arraycopy edge.
1328 if (e->fields_escape_state() < field_es) {
1329 set_fields_escape_state(e, field_es);
1330 escape_worklist.push(e);
1331 }
1332 } else if (es >= field_es) {
1333 // fields_escape_state is also set to 'es' if it is less than 'es'.
1334 if (e->escape_state() < es) {
1335 set_escape_state(e, es);
1336 escape_worklist.push(e);
1337 }
1338 } else {
1339 // Propagate field escape state.
1340 bool es_changed = false;
1341 if (e->fields_escape_state() < field_es) {
1342 set_fields_escape_state(e, field_es);
1343 es_changed = true;
1344 }
1345 if ((e->escape_state() < field_es) &&
1346 e->is_Field() && ptn->is_JavaObject() &&
1347 e->as_Field()->is_oop()) {
1348 // Change escape state of referenced fields.
1349 set_escape_state(e, field_es);
1350 es_changed = true;
1351 } else if (e->escape_state() < es) {
1352 set_escape_state(e, es);
1353 es_changed = true;
1354 }
1355 if (es_changed) {
1356 escape_worklist.push(e);
1357 }
1358 }
1359 }
1360 }
1361 // Remove escaped objects from non_escaped list.
1362 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1363 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1364 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1365 non_escaped_worklist.delete_at(next);
1366 }
1367 if (ptn->escape_state() == PointsToNode::NoEscape) {
1368 // Find fields in non-escaped allocations which have unknown value.
1369 find_init_values(ptn, phantom_obj, NULL);
1370 }
1371 }
1372 return (non_escaped_worklist.length() > 0);
1373 }
1374
1375 // Add all references to JavaObject node by walking over all uses.
1376 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1377 int new_edges = 0;
1378 if (populate_worklist) {
1379 // Populate _worklist by uses of jobj's uses.
1380 for (UseIterator i(jobj); i.has_next(); i.next()) {
1381 PointsToNode* use = i.get();
1382 if (use->is_Arraycopy())
1383 continue;
1384 add_uses_to_worklist(use);
1385 if (use->is_Field() && use->as_Field()->is_oop()) {
1386 // Put on worklist all field's uses (loads) and
1387 // related field nodes (same base and offset).
1388 add_field_uses_to_worklist(use->as_Field());
1389 }
1390 }
1391 }
1392 for (int l = 0; l < _worklist.length(); l++) {
1393 PointsToNode* use = _worklist.at(l);
1394 if (PointsToNode::is_base_use(use)) {
1395 // Add reference from jobj to field and from field to jobj (field's base).
1396 use = PointsToNode::get_use_node(use)->as_Field();
1397 if (add_base(use->as_Field(), jobj)) {
1398 new_edges++;
1399 }
1400 continue;
1401 }
1402 assert(!use->is_JavaObject(), "sanity");
1403 if (use->is_Arraycopy()) {
1404 if (jobj == null_obj) // NULL object does not have field edges
1405 continue;
1406 // Added edge from Arraycopy node to arraycopy's source java object
1407 if (add_edge(use, jobj)) {
1408 jobj->set_arraycopy_src();
1409 new_edges++;
1410 }
1411 // and stop here.
1412 continue;
1413 }
1414 if (!add_edge(use, jobj))
1415 continue; // No new edge added, there was such edge already.
1416 new_edges++;
1417 if (use->is_LocalVar()) {
1418 add_uses_to_worklist(use);
1419 if (use->arraycopy_dst()) {
1420 for (EdgeIterator i(use); i.has_next(); i.next()) {
1421 PointsToNode* e = i.get();
1422 if (e->is_Arraycopy()) {
1423 if (jobj == null_obj) // NULL object does not have field edges
1424 continue;
1425 // Add edge from arraycopy's destination java object to Arraycopy node.
1426 if (add_edge(jobj, e)) {
1427 new_edges++;
1428 jobj->set_arraycopy_dst();
1429 }
1430 }
1431 }
1432 }
1433 } else {
1434 // Added new edge to stored in field values.
1435 // Put on worklist all field's uses (loads) and
1436 // related field nodes (same base and offset).
1437 add_field_uses_to_worklist(use->as_Field());
1438 }
1439 }
1440 _worklist.clear();
1441 _in_worklist.Reset();
1442 return new_edges;
1443 }
1444
1445 // Put on worklist all related field nodes.
1446 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1447 assert(field->is_oop(), "sanity");
1448 int offset = field->offset();
1449 add_uses_to_worklist(field);
1450 // Loop over all bases of this field and push on worklist Field nodes
1451 // with the same offset and base (since they may reference the same field).
1452 for (BaseIterator i(field); i.has_next(); i.next()) {
1453 PointsToNode* base = i.get();
1454 add_fields_to_worklist(field, base);
1455 // Check if the base was source object of arraycopy and go over arraycopy's
1456 // destination objects since values stored to a field of source object are
1457 // accessable by uses (loads) of fields of destination objects.
1458 if (base->arraycopy_src()) {
1459 for (UseIterator j(base); j.has_next(); j.next()) {
1460 PointsToNode* arycp = j.get();
1461 if (arycp->is_Arraycopy()) {
1462 for (UseIterator k(arycp); k.has_next(); k.next()) {
1463 PointsToNode* abase = k.get();
1464 if (abase->arraycopy_dst() && abase != base) {
1465 // Look for the same arraycopy reference.
1466 add_fields_to_worklist(field, abase);
1467 }
1468 }
1469 }
1470 }
1471 }
1472 }
1473 }
1474
1475 // Put on worklist all related field nodes.
1476 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1477 int offset = field->offset();
1478 if (base->is_LocalVar()) {
1479 for (UseIterator j(base); j.has_next(); j.next()) {
1480 PointsToNode* f = j.get();
1481 if (PointsToNode::is_base_use(f)) { // Field
1482 f = PointsToNode::get_use_node(f);
1483 if (f == field || !f->as_Field()->is_oop())
1484 continue;
1485 int offs = f->as_Field()->offset();
1486 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1487 add_to_worklist(f);
1488 }
1489 }
1490 }
1491 } else {
1492 assert(base->is_JavaObject(), "sanity");
1493 if (// Skip phantom_object since it is only used to indicate that
1494 // this field's content globally escapes.
1495 (base != phantom_obj) &&
1496 // NULL object node does not have fields.
1497 (base != null_obj)) {
1498 for (EdgeIterator i(base); i.has_next(); i.next()) {
1499 PointsToNode* f = i.get();
1500 // Skip arraycopy edge since store to destination object field
1501 // does not update value in source object field.
1502 if (f->is_Arraycopy()) {
1503 assert(base->arraycopy_dst(), "sanity");
1504 continue;
1505 }
1506 if (f == field || !f->as_Field()->is_oop())
1507 continue;
1508 int offs = f->as_Field()->offset();
1509 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1510 add_to_worklist(f);
1511 }
1512 }
1513 }
1514 }
1515 }
1516
1517 // Find fields which have unknown value.
1518 int ConnectionGraph::find_field_value(FieldNode* field) {
1519 // Escaped fields should have init value already.
1520 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1521 int new_edges = 0;
1522 for (BaseIterator i(field); i.has_next(); i.next()) {
1523 PointsToNode* base = i.get();
1524 if (base->is_JavaObject()) {
1525 // Skip Allocate's fields which will be processed later.
1526 if (base->ideal_node()->is_Allocate())
1527 return 0;
1528 assert(base == null_obj, "only NULL ptr base expected here");
1529 }
1530 }
1531 if (add_edge(field, phantom_obj)) {
1532 // New edge was added
1533 new_edges++;
1534 add_field_uses_to_worklist(field);
1535 }
1536 return new_edges;
1537 }
1538
1539 // Find fields initializing values for allocations.
1540 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1541 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1542 int new_edges = 0;
1543 Node* alloc = pta->ideal_node();
1544 if (init_val == phantom_obj) {
1545 // Do nothing for Allocate nodes since its fields values are
1546 // "known" unless they are initialized by arraycopy/clone.
1547 if (alloc->is_Allocate() && !pta->arraycopy_dst())
1548 return 0;
1549 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
1550 #ifdef ASSERT
1551 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) {
1552 const char* name = alloc->as_CallStaticJava()->_name;
1553 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1554 }
1555 #endif
1556 // Non-escaped allocation returned from Java or runtime call have
1557 // unknown values in fields.
1558 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1559 PointsToNode* field = i.get();
1560 if (field->is_Field() && field->as_Field()->is_oop()) {
1561 if (add_edge(field, phantom_obj)) {
1562 // New edge was added
1563 new_edges++;
1564 add_field_uses_to_worklist(field->as_Field());
1565 }
1566 }
1567 }
1568 return new_edges;
1569 }
1570 assert(init_val == null_obj, "sanity");
1571 // Do nothing for Call nodes since its fields values are unknown.
1572 if (!alloc->is_Allocate())
1573 return 0;
1574
1575 InitializeNode* ini = alloc->as_Allocate()->initialization();
1576 bool visited_bottom_offset = false;
1577 GrowableArray<int> offsets_worklist;
1578
1579 // Check if an oop field's initializing value is recorded and add
1580 // a corresponding NULL if field's value if it is not recorded.
1581 // Connection Graph does not record a default initialization by NULL
1582 // captured by Initialize node.
1583 //
1584 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1585 PointsToNode* field = i.get(); // Field (AddP)
1586 if (!field->is_Field() || !field->as_Field()->is_oop())
1587 continue; // Not oop field
1588 int offset = field->as_Field()->offset();
1589 if (offset == Type::OffsetBot) {
1590 if (!visited_bottom_offset) {
1591 // OffsetBot is used to reference array's element,
1592 // always add reference to NULL to all Field nodes since we don't
1593 // known which element is referenced.
1594 if (add_edge(field, null_obj)) {
1595 // New edge was added
1596 new_edges++;
1597 add_field_uses_to_worklist(field->as_Field());
1598 visited_bottom_offset = true;
1599 }
1600 }
1601 } else {
1602 // Check only oop fields.
1603 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1604 if (adr_type->isa_rawptr()) {
1605 #ifdef ASSERT
1606 // Raw pointers are used for initializing stores so skip it
1607 // since it should be recorded already
1608 Node* base = get_addp_base(field->ideal_node());
1609 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1610 (base->in(0) == alloc),"unexpected pointer type");
1611 #endif
1612 continue;
1613 }
1614 if (!offsets_worklist.contains(offset)) {
1615 offsets_worklist.append(offset);
1616 Node* value = NULL;
1617 if (ini != NULL) {
1618 // StoreP::memory_type() == T_ADDRESS
1619 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1620 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1621 // Make sure initializing store has the same type as this AddP.
1622 // This AddP may reference non existing field because it is on a
1623 // dead branch of bimorphic call which is not eliminated yet.
1624 if (store != NULL && store->is_Store() &&
1625 store->as_Store()->memory_type() == ft) {
1626 value = store->in(MemNode::ValueIn);
1627 #ifdef ASSERT
1628 if (VerifyConnectionGraph) {
1629 // Verify that AddP already points to all objects the value points to.
1630 PointsToNode* val = ptnode_adr(value->_idx);
1631 assert((val != NULL), "should be processed already");
1632 PointsToNode* missed_obj = NULL;
1633 if (val->is_JavaObject()) {
1634 if (!field->points_to(val->as_JavaObject())) {
1635 missed_obj = val;
1636 }
1637 } else {
1638 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1639 tty->print_cr("----------init store has invalid value -----");
1640 store->dump();
1641 val->dump();
1642 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1643 }
1644 for (EdgeIterator j(val); j.has_next(); j.next()) {
1645 PointsToNode* obj = j.get();
1646 if (obj->is_JavaObject()) {
1647 if (!field->points_to(obj->as_JavaObject())) {
1648 missed_obj = obj;
1649 break;
1650 }
1651 }
1652 }
1653 }
1654 if (missed_obj != NULL) {
1655 tty->print_cr("----------field---------------------------------");
1656 field->dump();
1657 tty->print_cr("----------missed referernce to object-----------");
1658 missed_obj->dump();
1659 tty->print_cr("----------object referernced by init store -----");
1660 store->dump();
1661 val->dump();
1662 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1663 }
1664 }
1665 #endif
1666 } else {
1667 // There could be initializing stores which follow allocation.
1668 // For example, a volatile field store is not collected
1669 // by Initialize node.
1670 //
1671 // Need to check for dependent loads to separate such stores from
1672 // stores which follow loads. For now, add initial value NULL so
1673 // that compare pointers optimization works correctly.
1674 }
1675 }
1676 if (value == NULL) {
1677 // A field's initializing value was not recorded. Add NULL.
1678 if (add_edge(field, null_obj)) {
1679 // New edge was added
1680 new_edges++;
1681 add_field_uses_to_worklist(field->as_Field());
1682 }
1683 }
1684 }
1685 }
1686 }
1687 return new_edges;
1688 }
1689
1690 // Adjust scalar_replaceable state after Connection Graph is built.
1691 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1692 // Search for non-escaping objects which are not scalar replaceable
1693 // and mark them to propagate the state to referenced objects.
1694
1695 // 1. An object is not scalar replaceable if the field into which it is
1696 // stored has unknown offset (stored into unknown element of an array).
1697 //
1698 for (UseIterator i(jobj); i.has_next(); i.next()) {
1699 PointsToNode* use = i.get();
1700 if (use->is_Arraycopy()) {
1701 continue;
1702 }
1703 if (use->is_Field()) {
1704 FieldNode* field = use->as_Field();
1705 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
1706 if (field->offset() == Type::OffsetBot) {
1707 jobj->set_scalar_replaceable(false);
1708 return;
1709 }
1710 // 2. An object is not scalar replaceable if the field into which it is
1711 // stored has multiple bases one of which is null.
1712 if (field->base_count() > 1) {
1713 for (BaseIterator i(field); i.has_next(); i.next()) {
1714 PointsToNode* base = i.get();
1715 if (base == null_obj) {
1716 jobj->set_scalar_replaceable(false);
1717 return;
1718 }
1719 }
1720 }
1721 }
1722 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1723 // 3. An object is not scalar replaceable if it is merged with other objects.
1724 for (EdgeIterator j(use); j.has_next(); j.next()) {
1725 PointsToNode* ptn = j.get();
1726 if (ptn->is_JavaObject() && ptn != jobj) {
1727 // Mark all objects.
1728 jobj->set_scalar_replaceable(false);
1729 ptn->set_scalar_replaceable(false);
1730 }
1731 }
1732 if (!jobj->scalar_replaceable()) {
1733 return;
1734 }
1735 }
1736
1737 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1738 if (j.get()->is_Arraycopy()) {
1739 continue;
1740 }
1741
1742 // Non-escaping object node should point only to field nodes.
1743 FieldNode* field = j.get()->as_Field();
1744 int offset = field->as_Field()->offset();
1745
1746 // 4. An object is not scalar replaceable if it has a field with unknown
1747 // offset (array's element is accessed in loop).
1748 if (offset == Type::OffsetBot) {
1749 jobj->set_scalar_replaceable(false);
1750 return;
1751 }
1752 // 5. Currently an object is not scalar replaceable if a LoadStore node
1753 // access its field since the field value is unknown after it.
1754 //
1755 Node* n = field->ideal_node();
1756 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1757 if (n->fast_out(i)->is_LoadStore()) {
1758 jobj->set_scalar_replaceable(false);
1759 return;
1760 }
1761 }
1762
1763 // 6. Or the address may point to more then one object. This may produce
1764 // the false positive result (set not scalar replaceable)
1765 // since the flow-insensitive escape analysis can't separate
1766 // the case when stores overwrite the field's value from the case
1767 // when stores happened on different control branches.
1768 //
1769 // Note: it will disable scalar replacement in some cases:
1770 //
1771 // Point p[] = new Point[1];
1772 // p[0] = new Point(); // Will be not scalar replaced
1773 //
1774 // but it will save us from incorrect optimizations in next cases:
1775 //
1776 // Point p[] = new Point[1];
1777 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1778 //
1779 if (field->base_count() > 1) {
1780 for (BaseIterator i(field); i.has_next(); i.next()) {
1781 PointsToNode* base = i.get();
1782 // Don't take into account LocalVar nodes which
1783 // may point to only one object which should be also
1784 // this field's base by now.
1785 if (base->is_JavaObject() && base != jobj) {
1786 // Mark all bases.
1787 jobj->set_scalar_replaceable(false);
1788 base->set_scalar_replaceable(false);
1789 }
1790 }
1791 }
1792 }
1793 }
1794
1795 #ifdef ASSERT
1796 void ConnectionGraph::verify_connection_graph(
1797 GrowableArray<PointsToNode*>& ptnodes_worklist,
1798 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1799 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1800 GrowableArray<Node*>& addp_worklist) {
1801 // Verify that graph is complete - no new edges could be added.
1802 int java_objects_length = java_objects_worklist.length();
1803 int non_escaped_length = non_escaped_worklist.length();
1804 int new_edges = 0;
1805 for (int next = 0; next < java_objects_length; ++next) {
1806 JavaObjectNode* ptn = java_objects_worklist.at(next);
1807 new_edges += add_java_object_edges(ptn, true);
1808 }
1809 assert(new_edges == 0, "graph was not complete");
1810 // Verify that escape state is final.
1811 int length = non_escaped_worklist.length();
1812 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1813 assert((non_escaped_length == non_escaped_worklist.length()) &&
1814 (non_escaped_length == length) &&
1815 (_worklist.length() == 0), "escape state was not final");
1816
1817 // Verify fields information.
1818 int addp_length = addp_worklist.length();
1819 for (int next = 0; next < addp_length; ++next ) {
1820 Node* n = addp_worklist.at(next);
1821 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1822 if (field->is_oop()) {
1823 // Verify that field has all bases
1824 Node* base = get_addp_base(n);
1825 PointsToNode* ptn = ptnode_adr(base->_idx);
1826 if (ptn->is_JavaObject()) {
1827 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1828 } else {
1829 assert(ptn->is_LocalVar(), "sanity");
1830 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1831 PointsToNode* e = i.get();
1832 if (e->is_JavaObject()) {
1833 assert(field->has_base(e->as_JavaObject()), "sanity");
1834 }
1835 }
1836 }
1837 // Verify that all fields have initializing values.
1838 if (field->edge_count() == 0) {
1839 tty->print_cr("----------field does not have references----------");
1840 field->dump();
1841 for (BaseIterator i(field); i.has_next(); i.next()) {
1842 PointsToNode* base = i.get();
1843 tty->print_cr("----------field has next base---------------------");
1844 base->dump();
1845 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1846 tty->print_cr("----------base has fields-------------------------");
1847 for (EdgeIterator j(base); j.has_next(); j.next()) {
1848 j.get()->dump();
1849 }
1850 tty->print_cr("----------base has references---------------------");
1851 for (UseIterator j(base); j.has_next(); j.next()) {
1852 j.get()->dump();
1853 }
1854 }
1855 }
1856 for (UseIterator i(field); i.has_next(); i.next()) {
1857 i.get()->dump();
1858 }
1859 assert(field->edge_count() > 0, "sanity");
1860 }
1861 }
1862 }
1863 }
1864 #endif
1865
1866 // Optimize ideal graph.
1867 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1868 GrowableArray<Node*>& storestore_worklist) {
1869 Compile* C = _compile;
1870 PhaseIterGVN* igvn = _igvn;
1871 if (EliminateLocks) {
1872 // Mark locks before changing ideal graph.
1873 int cnt = C->macro_count();
1874 for( int i=0; i < cnt; i++ ) {
1875 Node *n = C->macro_node(i);
1876 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1877 AbstractLockNode* alock = n->as_AbstractLock();
1878 if (!alock->is_non_esc_obj()) {
1879 if (not_global_escape(alock->obj_node())) {
1880 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1881 // The lock could be marked eliminated by lock coarsening
1882 // code during first IGVN before EA. Replace coarsened flag
1883 // to eliminate all associated locks/unlocks.
1884 #ifdef ASSERT
1885 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1886 #endif
1887 alock->set_non_esc_obj();
1888 }
1889 }
1890 }
1891 }
1892 }
1893
1894 if (OptimizePtrCompare) {
1895 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1896 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1897 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1898 // Optimize objects compare.
1899 while (ptr_cmp_worklist.length() != 0) {
1900 Node *n = ptr_cmp_worklist.pop();
1901 Node *res = optimize_ptr_compare(n);
1902 if (res != NULL) {
1903 #ifndef PRODUCT
1904 if (PrintOptimizePtrCompare) {
1905 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"));
1906 if (Verbose) {
1907 n->dump(1);
1908 }
1909 }
1910 #endif
1911 igvn->replace_node(n, res);
1912 }
1913 }
1914 // cleanup
1915 if (_pcmp_neq->outcnt() == 0)
1916 igvn->hash_delete(_pcmp_neq);
1917 if (_pcmp_eq->outcnt() == 0)
1918 igvn->hash_delete(_pcmp_eq);
1919 }
1920
1921 // For MemBarStoreStore nodes added in library_call.cpp, check
1922 // escape status of associated AllocateNode and optimize out
1923 // MemBarStoreStore node if the allocated object never escapes.
1924 while (storestore_worklist.length() != 0) {
1925 Node *n = storestore_worklist.pop();
1926 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1927 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1928 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1929 if (not_global_escape(alloc)) {
1930 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1931 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1932 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1933 igvn->register_new_node_with_optimizer(mb);
1934 igvn->replace_node(storestore, mb);
1935 }
1936 }
1937 }
1938
1939 // Optimize objects compare.
1940 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1941 assert(OptimizePtrCompare, "sanity");
1942 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1943 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1944 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1945 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1946 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1947 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1948
1949 // Check simple cases first.
1950 if (jobj1 != NULL) {
1951 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1952 if (jobj1 == jobj2) {
1953 // Comparing the same not escaping object.
1954 return _pcmp_eq;
1955 }
1956 Node* obj = jobj1->ideal_node();
1957 // Comparing not escaping allocation.
1958 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1959 !ptn2->points_to(jobj1)) {
1960 return _pcmp_neq; // This includes nullness check.
1961 }
1962 }
1963 }
1964 if (jobj2 != NULL) {
1965 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1966 Node* obj = jobj2->ideal_node();
1967 // Comparing not escaping allocation.
1968 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1969 !ptn1->points_to(jobj2)) {
1970 return _pcmp_neq; // This includes nullness check.
1971 }
1972 }
1973 }
1974 if (jobj1 != NULL && jobj1 != phantom_obj &&
1975 jobj2 != NULL && jobj2 != phantom_obj &&
1976 jobj1->ideal_node()->is_Con() &&
1977 jobj2->ideal_node()->is_Con()) {
1978 // Klass or String constants compare. Need to be careful with
1979 // compressed pointers - compare types of ConN and ConP instead of nodes.
1980 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1981 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1982 if (t1->make_ptr() == t2->make_ptr()) {
1983 return _pcmp_eq;
1984 } else {
1985 return _pcmp_neq;
1986 }
1987 }
1988 if (ptn1->meet(ptn2)) {
1989 return NULL; // Sets are not disjoint
1990 }
1991
1992 // Sets are disjoint.
1993 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1994 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1995 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1996 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1997 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
1998 (set2_has_unknown_ptr && set1_has_null_ptr)) {
1999 // Check nullness of unknown object.
2000 return NULL;
2001 }
2002
2003 // Disjointness by itself is not sufficient since
2004 // alias analysis is not complete for escaped objects.
2005 // Disjoint sets are definitely unrelated only when
2006 // at least one set has only not escaping allocations.
2007 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
2008 if (ptn1->non_escaping_allocation()) {
2009 return _pcmp_neq;
2010 }
2011 }
2012 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2013 if (ptn2->non_escaping_allocation()) {
2014 return _pcmp_neq;
2015 }
2016 }
2017 return NULL;
2018 }
2019
2020 // Connection Graph constuction functions.
2021
2022 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
2023 PointsToNode* ptadr = _nodes.at(n->_idx);
2024 if (ptadr != NULL) {
2025 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
2026 return;
2027 }
2028 Compile* C = _compile;
2029 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
2030 _nodes.at_put(n->_idx, ptadr);
2031 }
2032
2033 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
2034 PointsToNode* ptadr = _nodes.at(n->_idx);
2035 if (ptadr != NULL) {
2036 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
2037 return;
2038 }
2039 Compile* C = _compile;
2040 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
2041 _nodes.at_put(n->_idx, ptadr);
2042 }
2043
2044 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
2045 PointsToNode* ptadr = _nodes.at(n->_idx);
2046 if (ptadr != NULL) {
2047 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
2048 return;
2049 }
2050 bool unsafe = false;
2051 bool is_oop = is_oop_field(n, offset, &unsafe);
2052 if (unsafe) {
2053 es = PointsToNode::GlobalEscape;
2054 }
2055 Compile* C = _compile;
2056 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
2057 _nodes.at_put(n->_idx, field);
2058 }
2059
2060 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
2061 PointsToNode* src, PointsToNode* dst) {
2062 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
2063 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
2064 PointsToNode* ptadr = _nodes.at(n->_idx);
2065 if (ptadr != NULL) {
2066 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
2067 return;
2068 }
2069 Compile* C = _compile;
2070 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
2071 _nodes.at_put(n->_idx, ptadr);
2072 // Add edge from arraycopy node to source object.
2073 (void)add_edge(ptadr, src);
2074 src->set_arraycopy_src();
2075 // Add edge from destination object to arraycopy node.
2076 (void)add_edge(dst, ptadr);
2077 dst->set_arraycopy_dst();
2078 }
2079
2080 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
2081 const Type* adr_type = n->as_AddP()->bottom_type();
2082 BasicType bt = T_INT;
2083 if (offset == Type::OffsetBot) {
2084 // Check only oop fields.
2085 if (!adr_type->isa_aryptr() ||
2086 (adr_type->isa_aryptr()->klass() == NULL) ||
2087 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2088 // OffsetBot is used to reference array's element. Ignore first AddP.
2089 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2090 bt = T_OBJECT;
2091 }
2092 }
2093 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2094 if (adr_type->isa_instptr()) {
2095 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
2096 if (field != NULL) {
2097 bt = field->layout_type();
2098 } else {
2099 // Check for unsafe oop field access
2100 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2101 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2102 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2103 bt = T_OBJECT;
2104 (*unsafe) = true;
2105 }
2106 }
2107 } else if (adr_type->isa_aryptr()) {
2108 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2109 // Ignore array length load.
2110 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2111 // Ignore first AddP.
2112 } else {
2113 const Type* elemtype = adr_type->isa_aryptr()->elem();
2114 bt = elemtype->array_element_basic_type();
2115 }
2116 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2117 // Allocation initialization, ThreadLocal field access, unsafe access
2118 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2119 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2120 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2121 bt = T_OBJECT;
2122 }
2123 }
2124 }
2125 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2126 }
2127
2128 // Returns unique pointed java object or NULL.
2129 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2130 assert(!_collecting, "should not call when contructed graph");
2131 // If the node was created after the escape computation we can't answer.
2132 uint idx = n->_idx;
2133 if (idx >= nodes_size()) {
2134 return NULL;
2135 }
2136 PointsToNode* ptn = ptnode_adr(idx);
2137 if (ptn->is_JavaObject()) {
2138 return ptn->as_JavaObject();
2139 }
2140 assert(ptn->is_LocalVar(), "sanity");
2141 // Check all java objects it points to.
2142 JavaObjectNode* jobj = NULL;
2143 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2144 PointsToNode* e = i.get();
2145 if (e->is_JavaObject()) {
2146 if (jobj == NULL) {
2147 jobj = e->as_JavaObject();
2148 } else if (jobj != e) {
2149 return NULL;
2150 }
2151 }
2152 }
2153 return jobj;
2154 }
2155
2156 // Return true if this node points only to non-escaping allocations.
2157 bool PointsToNode::non_escaping_allocation() {
2158 if (is_JavaObject()) {
2159 Node* n = ideal_node();
2160 if (n->is_Allocate() || n->is_CallStaticJava()) {
2161 return (escape_state() == PointsToNode::NoEscape);
2162 } else {
2163 return false;
2164 }
2165 }
2166 assert(is_LocalVar(), "sanity");
2167 // Check all java objects it points to.
2168 for (EdgeIterator i(this); i.has_next(); i.next()) {
2169 PointsToNode* e = i.get();
2170 if (e->is_JavaObject()) {
2171 Node* n = e->ideal_node();
2172 if ((e->escape_state() != PointsToNode::NoEscape) ||
2173 !(n->is_Allocate() || n->is_CallStaticJava())) {
2174 return false;
2175 }
2176 }
2177 }
2178 return true;
2179 }
2180
2181 // Return true if we know the node does not escape globally.
2182 bool ConnectionGraph::not_global_escape(Node *n) {
2183 assert(!_collecting, "should not call during graph construction");
2184 // If the node was created after the escape computation we can't answer.
2185 uint idx = n->_idx;
2186 if (idx >= nodes_size()) {
2187 return false;
2188 }
2189 PointsToNode* ptn = ptnode_adr(idx);
2190 PointsToNode::EscapeState es = ptn->escape_state();
2191 // If we have already computed a value, return it.
2192 if (es >= PointsToNode::GlobalEscape)
2193 return false;
2194 if (ptn->is_JavaObject()) {
2195 return true; // (es < PointsToNode::GlobalEscape);
2196 }
2197 assert(ptn->is_LocalVar(), "sanity");
2198 // Check all java objects it points to.
2199 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2200 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2201 return false;
2202 }
2203 return true;
2204 }
2205
2206
2207 // Helper functions
2208
2209 // Return true if this node points to specified node or nodes it points to.
2210 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2211 if (is_JavaObject()) {
2212 return (this == ptn);
2213 }
2214 assert(is_LocalVar() || is_Field(), "sanity");
2215 for (EdgeIterator i(this); i.has_next(); i.next()) {
2216 if (i.get() == ptn)
2217 return true;
2218 }
2219 return false;
2220 }
2221
2222 // Return true if one node points to an other.
2223 bool PointsToNode::meet(PointsToNode* ptn) {
2224 if (this == ptn) {
2225 return true;
2226 } else if (ptn->is_JavaObject()) {
2227 return this->points_to(ptn->as_JavaObject());
2228 } else if (this->is_JavaObject()) {
2229 return ptn->points_to(this->as_JavaObject());
2230 }
2231 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2232 int ptn_count = ptn->edge_count();
2233 for (EdgeIterator i(this); i.has_next(); i.next()) {
2234 PointsToNode* this_e = i.get();
2235 for (int j = 0; j < ptn_count; j++) {
2236 if (this_e == ptn->edge(j))
2237 return true;
2238 }
2239 }
2240 return false;
2241 }
2242
2243 #ifdef ASSERT
2244 // Return true if bases point to this java object.
2245 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2246 for (BaseIterator i(this); i.has_next(); i.next()) {
2247 if (i.get() == jobj)
2248 return true;
2249 }
2250 return false;
2251 }
2252 #endif
2253
2254 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2255 const Type *adr_type = phase->type(adr);
2256 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2257 adr->in(AddPNode::Address)->is_Proj() &&
2258 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2259 // We are computing a raw address for a store captured by an Initialize
2260 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2261 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2262 assert(offs != Type::OffsetBot ||
2263 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2264 "offset must be a constant or it is initialization of array");
2265 return offs;
2266 }
2267 const TypePtr *t_ptr = adr_type->isa_ptr();
2268 assert(t_ptr != NULL, "must be a pointer type");
2269 return t_ptr->offset();
2270 }
2271
2272 Node* ConnectionGraph::get_addp_base(Node *addp) {
2273 assert(addp->is_AddP(), "must be AddP");
2274 //
2275 // AddP cases for Base and Address inputs:
2276 // case #1. Direct object's field reference:
2277 // Allocate
2278 // |
2279 // Proj #5 ( oop result )
2280 // |
2281 // CheckCastPP (cast to instance type)
2282 // | |
2283 // AddP ( base == address )
2284 //
2285 // case #2. Indirect object's field reference:
2286 // Phi
2287 // |
2288 // CastPP (cast to instance type)
2289 // | |
2290 // AddP ( base == address )
2291 //
2292 // case #3. Raw object's field reference for Initialize node:
2293 // Allocate
2294 // |
2295 // Proj #5 ( oop result )
2296 // top |
2297 // \ |
2298 // AddP ( base == top )
2299 //
2300 // case #4. Array's element reference:
2301 // {CheckCastPP | CastPP}
2302 // | | |
2303 // | AddP ( array's element offset )
2304 // | |
2305 // AddP ( array's offset )
2306 //
2307 // case #5. Raw object's field reference for arraycopy stub call:
2308 // The inline_native_clone() case when the arraycopy stub is called
2309 // after the allocation before Initialize and CheckCastPP nodes.
2310 // Allocate
2311 // |
2312 // Proj #5 ( oop result )
2313 // | |
2314 // AddP ( base == address )
2315 //
2316 // case #6. Constant Pool, ThreadLocal, CastX2P or
2317 // Raw object's field reference:
2318 // {ConP, ThreadLocal, CastX2P, raw Load}
2319 // top |
2320 // \ |
2321 // AddP ( base == top )
2322 //
2323 // case #7. Klass's field reference.
2324 // LoadKlass
2325 // | |
2326 // AddP ( base == address )
2327 //
2328 // case #8. narrow Klass's field reference.
2329 // LoadNKlass
2330 // |
2331 // DecodeN
2332 // | |
2333 // AddP ( base == address )
2334 //
2335 // case #9. Mixed unsafe access
2336 // {instance}
2337 // |
2338 // CheckCastPP (raw)
2339 // top |
2340 // \ |
2341 // AddP ( base == top )
2342 //
2343 Node *base = addp->in(AddPNode::Base);
2344 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
2345 base = addp->in(AddPNode::Address);
2346 while (base->is_AddP()) {
2347 // Case #6 (unsafe access) may have several chained AddP nodes.
2348 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2349 base = base->in(AddPNode::Address);
2350 }
2351 if (base->Opcode() == Op_CheckCastPP &&
2352 base->bottom_type()->isa_rawptr() &&
2353 _igvn->type(base->in(1))->isa_oopptr()) {
2354 base = base->in(1); // Case #9
2355 } else {
2356 Node* uncast_base = base->uncast();
2357 int opcode = uncast_base->Opcode();
2358 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2359 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2360 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2361 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2362 }
2363 }
2364 return base;
2365 }
2366
2367 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2368 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2369 Node* addp2 = addp->raw_out(0);
2370 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2371 addp2->in(AddPNode::Base) == n &&
2372 addp2->in(AddPNode::Address) == addp) {
2373 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2374 //
2375 // Find array's offset to push it on worklist first and
2376 // as result process an array's element offset first (pushed second)
2377 // to avoid CastPP for the array's offset.
2378 // Otherwise the inserted CastPP (LocalVar) will point to what
2379 // the AddP (Field) points to. Which would be wrong since
2380 // the algorithm expects the CastPP has the same point as
2381 // as AddP's base CheckCastPP (LocalVar).
2382 //
2383 // ArrayAllocation
2384 // |
2385 // CheckCastPP
2386 // |
2387 // memProj (from ArrayAllocation CheckCastPP)
2388 // | ||
2389 // | || Int (element index)
2390 // | || | ConI (log(element size))
2391 // | || | /
2392 // | || LShift
2393 // | || /
2394 // | AddP (array's element offset)
2395 // | |
2396 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2397 // | / /
2398 // AddP (array's offset)
2399 // |
2400 // Load/Store (memory operation on array's element)
2401 //
2402 return addp2;
2403 }
2404 return NULL;
2405 }
2406
2407 //
2408 // Adjust the type and inputs of an AddP which computes the
2409 // address of a field of an instance
2410 //
2411 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2412 PhaseGVN* igvn = _igvn;
2413 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2414 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2415 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2416 if (t == NULL) {
2417 // We are computing a raw address for a store captured by an Initialize
2418 // compute an appropriate address type (cases #3 and #5).
2419 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2420 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2421 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2422 assert(offs != Type::OffsetBot, "offset must be a constant");
2423 t = base_t->add_offset(offs)->is_oopptr();
2424 }
2425 int inst_id = base_t->instance_id();
2426 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2427 "old type must be non-instance or match new type");
2428
2429 // The type 't' could be subclass of 'base_t'.
2430 // As result t->offset() could be large then base_t's size and it will
2431 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2432 // constructor verifies correctness of the offset.
2433 //
2434 // It could happened on subclass's branch (from the type profiling
2435 // inlining) which was not eliminated during parsing since the exactness
2436 // of the allocation type was not propagated to the subclass type check.
2437 //
2438 // Or the type 't' could be not related to 'base_t' at all.
2439 // It could happened when CHA type is different from MDO type on a dead path
2440 // (for example, from instanceof check) which is not collapsed during parsing.
2441 //
2442 // Do nothing for such AddP node and don't process its users since
2443 // this code branch will go away.
2444 //
2445 if (!t->is_known_instance() &&
2446 !base_t->klass()->is_subtype_of(t->klass())) {
2447 return false; // bail out
2448 }
2449 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2450 // Do NOT remove the next line: ensure a new alias index is allocated
2451 // for the instance type. Note: C++ will not remove it since the call
2452 // has side effect.
2453 int alias_idx = _compile->get_alias_index(tinst);
2454 igvn->set_type(addp, tinst);
2455 // record the allocation in the node map
2456 set_map(addp, get_map(base->_idx));
2457 // Set addp's Base and Address to 'base'.
2458 Node *abase = addp->in(AddPNode::Base);
2459 Node *adr = addp->in(AddPNode::Address);
2460 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2461 adr->in(0)->_idx == (uint)inst_id) {
2462 // Skip AddP cases #3 and #5.
2463 } else {
2464 assert(!abase->is_top(), "sanity"); // AddP case #3
2465 if (abase != base) {
2466 igvn->hash_delete(addp);
2467 addp->set_req(AddPNode::Base, base);
2468 if (abase == adr) {
2469 addp->set_req(AddPNode::Address, base);
2470 } else {
2471 // AddP case #4 (adr is array's element offset AddP node)
2472 #ifdef ASSERT
2473 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2474 assert(adr->is_AddP() && atype != NULL &&
2475 atype->instance_id() == inst_id, "array's element offset should be processed first");
2476 #endif
2477 }
2478 igvn->hash_insert(addp);
2479 }
2480 }
2481 // Put on IGVN worklist since at least addp's type was changed above.
2482 record_for_optimizer(addp);
2483 return true;
2484 }
2485
2486 //
2487 // Create a new version of orig_phi if necessary. Returns either the newly
2488 // created phi or an existing phi. Sets create_new to indicate whether a new
2489 // phi was created. Cache the last newly created phi in the node map.
2490 //
2491 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2492 Compile *C = _compile;
2493 PhaseGVN* igvn = _igvn;
2494 new_created = false;
2495 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2496 // nothing to do if orig_phi is bottom memory or matches alias_idx
2497 if (phi_alias_idx == alias_idx) {
2498 return orig_phi;
2499 }
2500 // Have we recently created a Phi for this alias index?
2501 PhiNode *result = get_map_phi(orig_phi->_idx);
2502 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2503 return result;
2504 }
2505 // Previous check may fail when the same wide memory Phi was split into Phis
2506 // for different memory slices. Search all Phis for this region.
2507 if (result != NULL) {
2508 Node* region = orig_phi->in(0);
2509 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2510 Node* phi = region->fast_out(i);
2511 if (phi->is_Phi() &&
2512 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2513 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2514 return phi->as_Phi();
2515 }
2516 }
2517 }
2518 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2519 if (C->do_escape_analysis() == true && !C->failing()) {
2520 // Retry compilation without escape analysis.
2521 // If this is the first failure, the sentinel string will "stick"
2522 // to the Compile object, and the C2Compiler will see it and retry.
2523 C->record_failure(C2Compiler::retry_no_escape_analysis());
2524 }
2525 return NULL;
2526 }
2527 orig_phi_worklist.append_if_missing(orig_phi);
2528 const TypePtr *atype = C->get_adr_type(alias_idx);
2529 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2530 C->copy_node_notes_to(result, orig_phi);
2531 igvn->set_type(result, result->bottom_type());
2532 record_for_optimizer(result);
2533 set_map(orig_phi, result);
2534 new_created = true;
2535 return result;
2536 }
2537
2538 //
2539 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2540 // specified alias index.
2541 //
2542 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2543 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2544 Compile *C = _compile;
2545 PhaseGVN* igvn = _igvn;
2546 bool new_phi_created;
2547 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2548 if (!new_phi_created) {
2549 return result;
2550 }
2551 GrowableArray<PhiNode *> phi_list;
2552 GrowableArray<uint> cur_input;
2553 PhiNode *phi = orig_phi;
2554 uint idx = 1;
2555 bool finished = false;
2556 while(!finished) {
2557 while (idx < phi->req()) {
2558 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2559 if (mem != NULL && mem->is_Phi()) {
2560 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2561 if (new_phi_created) {
2562 // found an phi for which we created a new split, push current one on worklist and begin
2563 // processing new one
2564 phi_list.push(phi);
2565 cur_input.push(idx);
2566 phi = mem->as_Phi();
2567 result = newphi;
2568 idx = 1;
2569 continue;
2570 } else {
2571 mem = newphi;
2572 }
2573 }
2574 if (C->failing()) {
2575 return NULL;
2576 }
2577 result->set_req(idx++, mem);
2578 }
2579 #ifdef ASSERT
2580 // verify that the new Phi has an input for each input of the original
2581 assert( phi->req() == result->req(), "must have same number of inputs.");
2582 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2583 #endif
2584 // Check if all new phi's inputs have specified alias index.
2585 // Otherwise use old phi.
2586 for (uint i = 1; i < phi->req(); i++) {
2587 Node* in = result->in(i);
2588 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2589 }
2590 // we have finished processing a Phi, see if there are any more to do
2591 finished = (phi_list.length() == 0 );
2592 if (!finished) {
2593 phi = phi_list.pop();
2594 idx = cur_input.pop();
2595 PhiNode *prev_result = get_map_phi(phi->_idx);
2596 prev_result->set_req(idx++, result);
2597 result = prev_result;
2598 }
2599 }
2600 return result;
2601 }
2602
2603 //
2604 // The next methods are derived from methods in MemNode.
2605 //
2606 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2607 Node *mem = mmem;
2608 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2609 // means an array I have not precisely typed yet. Do not do any
2610 // alias stuff with it any time soon.
2611 if (toop->base() != Type::AnyPtr &&
2612 !(toop->klass() != NULL &&
2613 toop->klass()->is_java_lang_Object() &&
2614 toop->offset() == Type::OffsetBot)) {
2615 mem = mmem->memory_at(alias_idx);
2616 // Update input if it is progress over what we have now
2617 }
2618 return mem;
2619 }
2620
2621 //
2622 // Move memory users to their memory slices.
2623 //
2624 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2625 Compile* C = _compile;
2626 PhaseGVN* igvn = _igvn;
2627 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2628 assert(tp != NULL, "ptr type");
2629 int alias_idx = C->get_alias_index(tp);
2630 int general_idx = C->get_general_index(alias_idx);
2631
2632 // Move users first
2633 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2634 Node* use = n->fast_out(i);
2635 if (use->is_MergeMem()) {
2636 MergeMemNode* mmem = use->as_MergeMem();
2637 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2638 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2639 continue; // Nothing to do
2640 }
2641 // Replace previous general reference to mem node.
2642 uint orig_uniq = C->unique();
2643 Node* m = find_inst_mem(n, general_idx, orig_phis);
2644 assert(orig_uniq == C->unique(), "no new nodes");
2645 mmem->set_memory_at(general_idx, m);
2646 --imax;
2647 --i;
2648 } else if (use->is_MemBar()) {
2649 assert(!use->is_Initialize(), "initializing stores should not be moved");
2650 if (use->req() > MemBarNode::Precedent &&
2651 use->in(MemBarNode::Precedent) == n) {
2652 // Don't move related membars.
2653 record_for_optimizer(use);
2654 continue;
2655 }
2656 tp = use->as_MemBar()->adr_type()->isa_ptr();
2657 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) ||
2658 alias_idx == general_idx) {
2659 continue; // Nothing to do
2660 }
2661 // Move to general memory slice.
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 igvn->hash_delete(use);
2666 imax -= use->replace_edge(n, m);
2667 igvn->hash_insert(use);
2668 record_for_optimizer(use);
2669 --i;
2670 #ifdef ASSERT
2671 } else if (use->is_Mem()) {
2672 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2673 // Don't move related cardmark.
2674 continue;
2675 }
2676 // Memory nodes should have new memory input.
2677 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2678 assert(tp != NULL, "ptr type");
2679 int idx = C->get_alias_index(tp);
2680 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2681 "Following memory nodes should have new memory input or be on the same memory slice");
2682 } else if (use->is_Phi()) {
2683 // Phi nodes should be split and moved already.
2684 tp = use->as_Phi()->adr_type()->isa_ptr();
2685 assert(tp != NULL, "ptr type");
2686 int idx = C->get_alias_index(tp);
2687 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2688 } else {
2689 use->dump();
2690 assert(false, "should not be here");
2691 #endif
2692 }
2693 }
2694 }
2695
2696 //
2697 // Search memory chain of "mem" to find a MemNode whose address
2698 // is the specified alias index.
2699 //
2700 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2701 if (orig_mem == NULL)
2702 return orig_mem;
2703 Compile* C = _compile;
2704 PhaseGVN* igvn = _igvn;
2705 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2706 bool is_instance = (toop != NULL) && toop->is_known_instance();
2707 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
2708 Node *prev = NULL;
2709 Node *result = orig_mem;
2710 while (prev != result) {
2711 prev = result;
2712 if (result == start_mem)
2713 break; // hit one of our sentinels
2714 if (result->is_Mem()) {
2715 const Type *at = igvn->type(result->in(MemNode::Address));
2716 if (at == Type::TOP)
2717 break; // Dead
2718 assert (at->isa_ptr() != NULL, "pointer type required.");
2719 int idx = C->get_alias_index(at->is_ptr());
2720 if (idx == alias_idx)
2721 break; // Found
2722 if (!is_instance && (at->isa_oopptr() == NULL ||
2723 !at->is_oopptr()->is_known_instance())) {
2724 break; // Do not skip store to general memory slice.
2725 }
2726 result = result->in(MemNode::Memory);
2727 }
2728 if (!is_instance)
2729 continue; // don't search further for non-instance types
2730 // skip over a call which does not affect this memory slice
2731 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2732 Node *proj_in = result->in(0);
2733 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2734 break; // hit one of our sentinels
2735 } else if (proj_in->is_Call()) {
2736 // ArrayCopy node processed here as well
2737 CallNode *call = proj_in->as_Call();
2738 if (!call->may_modify(toop, igvn)) {
2739 result = call->in(TypeFunc::Memory);
2740 }
2741 } else if (proj_in->is_Initialize()) {
2742 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2743 // Stop if this is the initialization for the object instance which
2744 // which contains this memory slice, otherwise skip over it.
2745 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2746 result = proj_in->in(TypeFunc::Memory);
2747 }
2748 } else if (proj_in->is_MemBar()) {
2749 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() &&
2750 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() &&
2751 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) {
2752 // clone
2753 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy();
2754 if (ac->may_modify(toop, igvn)) {
2755 break;
2756 }
2757 }
2758 result = proj_in->in(TypeFunc::Memory);
2759 }
2760 } else if (result->is_MergeMem()) {
2761 MergeMemNode *mmem = result->as_MergeMem();
2762 result = step_through_mergemem(mmem, alias_idx, toop);
2763 if (result == mmem->base_memory()) {
2764 // Didn't find instance memory, search through general slice recursively.
2765 result = mmem->memory_at(C->get_general_index(alias_idx));
2766 result = find_inst_mem(result, alias_idx, orig_phis);
2767 if (C->failing()) {
2768 return NULL;
2769 }
2770 mmem->set_memory_at(alias_idx, result);
2771 }
2772 } else if (result->is_Phi() &&
2773 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2774 Node *un = result->as_Phi()->unique_input(igvn);
2775 if (un != NULL) {
2776 orig_phis.append_if_missing(result->as_Phi());
2777 result = un;
2778 } else {
2779 break;
2780 }
2781 } else if (result->is_ClearArray()) {
2782 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2783 // Can not bypass initialization of the instance
2784 // we are looking for.
2785 break;
2786 }
2787 // Otherwise skip it (the call updated 'result' value).
2788 } else if (result->Opcode() == Op_SCMemProj) {
2789 Node* mem = result->in(0);
2790 Node* adr = NULL;
2791 if (mem->is_LoadStore()) {
2792 adr = mem->in(MemNode::Address);
2793 } else {
2794 assert(mem->Opcode() == Op_EncodeISOArray ||
2795 mem->Opcode() == Op_StrCompressedCopy, "sanity");
2796 adr = mem->in(3); // Memory edge corresponds to destination array
2797 }
2798 const Type *at = igvn->type(adr);
2799 if (at != Type::TOP) {
2800 assert(at->isa_ptr() != NULL, "pointer type required.");
2801 int idx = C->get_alias_index(at->is_ptr());
2802 if (idx == alias_idx) {
2803 // Assert in debug mode
2804 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
2805 break; // In product mode return SCMemProj node
2806 }
2807 }
2808 result = mem->in(MemNode::Memory);
2809 } else if (result->Opcode() == Op_StrInflatedCopy) {
2810 Node* adr = result->in(3); // Memory edge corresponds to destination array
2811 const Type *at = igvn->type(adr);
2812 if (at != Type::TOP) {
2813 assert(at->isa_ptr() != NULL, "pointer type required.");
2814 int idx = C->get_alias_index(at->is_ptr());
2815 if (idx == alias_idx) {
2816 // Assert in debug mode
2817 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
2818 break; // In product mode return SCMemProj node
2819 }
2820 }
2821 result = result->in(MemNode::Memory);
2822 }
2823 }
2824 if (result->is_Phi()) {
2825 PhiNode *mphi = result->as_Phi();
2826 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2827 const TypePtr *t = mphi->adr_type();
2828 if (!is_instance) {
2829 // Push all non-instance Phis on the orig_phis worklist to update inputs
2830 // during Phase 4 if needed.
2831 orig_phis.append_if_missing(mphi);
2832 } else if (C->get_alias_index(t) != alias_idx) {
2833 // Create a new Phi with the specified alias index type.
2834 result = split_memory_phi(mphi, alias_idx, orig_phis);
2835 }
2836 }
2837 // the result is either MemNode, PhiNode, InitializeNode.
2838 return result;
2839 }
2840
2841 //
2842 // Convert the types of unescaped object to instance types where possible,
2843 // propagate the new type information through the graph, and update memory
2844 // edges and MergeMem inputs to reflect the new type.
2845 //
2846 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2847 // The processing is done in 4 phases:
2848 //
2849 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2850 // types for the CheckCastPP for allocations where possible.
2851 // Propagate the new types through users as follows:
2852 // casts and Phi: push users on alloc_worklist
2853 // AddP: cast Base and Address inputs to the instance type
2854 // push any AddP users on alloc_worklist and push any memnode
2855 // users onto memnode_worklist.
2856 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2857 // search the Memory chain for a store with the appropriate type
2858 // address type. If a Phi is found, create a new version with
2859 // the appropriate memory slices from each of the Phi inputs.
2860 // For stores, process the users as follows:
2861 // MemNode: push on memnode_worklist
2862 // MergeMem: push on mergemem_worklist
2863 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2864 // moving the first node encountered of each instance type to the
2865 // the input corresponding to its alias index.
2866 // appropriate memory slice.
2867 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2868 //
2869 // In the following example, the CheckCastPP nodes are the cast of allocation
2870 // results and the allocation of node 29 is unescaped and eligible to be an
2871 // instance type.
2872 //
2873 // We start with:
2874 //
2875 // 7 Parm #memory
2876 // 10 ConI "12"
2877 // 19 CheckCastPP "Foo"
2878 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2879 // 29 CheckCastPP "Foo"
2880 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2881 //
2882 // 40 StoreP 25 7 20 ... alias_index=4
2883 // 50 StoreP 35 40 30 ... alias_index=4
2884 // 60 StoreP 45 50 20 ... alias_index=4
2885 // 70 LoadP _ 60 30 ... alias_index=4
2886 // 80 Phi 75 50 60 Memory alias_index=4
2887 // 90 LoadP _ 80 30 ... alias_index=4
2888 // 100 LoadP _ 80 20 ... alias_index=4
2889 //
2890 //
2891 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2892 // and creating a new alias index for node 30. This gives:
2893 //
2894 // 7 Parm #memory
2895 // 10 ConI "12"
2896 // 19 CheckCastPP "Foo"
2897 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2898 // 29 CheckCastPP "Foo" iid=24
2899 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2900 //
2901 // 40 StoreP 25 7 20 ... alias_index=4
2902 // 50 StoreP 35 40 30 ... alias_index=6
2903 // 60 StoreP 45 50 20 ... alias_index=4
2904 // 70 LoadP _ 60 30 ... alias_index=6
2905 // 80 Phi 75 50 60 Memory alias_index=4
2906 // 90 LoadP _ 80 30 ... alias_index=6
2907 // 100 LoadP _ 80 20 ... alias_index=4
2908 //
2909 // In phase 2, new memory inputs are computed for the loads and stores,
2910 // And a new version of the phi is created. In phase 4, the inputs to
2911 // node 80 are updated and then the memory nodes are updated with the
2912 // values computed in phase 2. This results in:
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 7 30 ... alias_index=6
2923 // 60 StoreP 45 40 20 ... alias_index=4
2924 // 70 LoadP _ 50 30 ... alias_index=6
2925 // 80 Phi 75 40 60 Memory alias_index=4
2926 // 120 Phi 75 50 50 Memory alias_index=6
2927 // 90 LoadP _ 120 30 ... alias_index=6
2928 // 100 LoadP _ 80 20 ... alias_index=4
2929 //
2930 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) {
2931 GrowableArray<Node *> memnode_worklist;
2932 GrowableArray<PhiNode *> orig_phis;
2933 PhaseIterGVN *igvn = _igvn;
2934 uint new_index_start = (uint) _compile->num_alias_types();
2935 Arena* arena = Thread::current()->resource_area();
2936 VectorSet visited(arena);
2937 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2938 uint unique_old = _compile->unique();
2939
2940 // Phase 1: Process possible allocations from alloc_worklist.
2941 // Create instance types for the CheckCastPP for allocations where possible.
2942 //
2943 // (Note: don't forget to change the order of the second AddP node on
2944 // the alloc_worklist if the order of the worklist processing is changed,
2945 // see the comment in find_second_addp().)
2946 //
2947 while (alloc_worklist.length() != 0) {
2948 Node *n = alloc_worklist.pop();
2949 uint ni = n->_idx;
2950 if (n->is_Call()) {
2951 CallNode *alloc = n->as_Call();
2952 // copy escape information to call node
2953 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2954 PointsToNode::EscapeState es = ptn->escape_state();
2955 // We have an allocation or call which returns a Java object,
2956 // see if it is unescaped.
2957 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2958 continue;
2959 // Find CheckCastPP for the allocate or for the return value of a call
2960 n = alloc->result_cast();
2961 if (n == NULL) { // No uses except Initialize node
2962 if (alloc->is_Allocate()) {
2963 // Set the scalar_replaceable flag for allocation
2964 // so it could be eliminated if it has no uses.
2965 alloc->as_Allocate()->_is_scalar_replaceable = true;
2966 }
2967 if (alloc->is_CallStaticJava()) {
2968 // Set the scalar_replaceable flag for boxing method
2969 // so it could be eliminated if it has no uses.
2970 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2971 }
2972 continue;
2973 }
2974 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2975 assert(!alloc->is_Allocate(), "allocation should have unique type");
2976 continue;
2977 }
2978
2979 // The inline code for Object.clone() casts the allocation result to
2980 // java.lang.Object and then to the actual type of the allocated
2981 // object. Detect this case and use the second cast.
2982 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2983 // the allocation result is cast to java.lang.Object and then
2984 // to the actual Array type.
2985 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2986 && (alloc->is_AllocateArray() ||
2987 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2988 Node *cast2 = NULL;
2989 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2990 Node *use = n->fast_out(i);
2991 if (use->is_CheckCastPP()) {
2992 cast2 = use;
2993 break;
2994 }
2995 }
2996 if (cast2 != NULL) {
2997 n = cast2;
2998 } else {
2999 // Non-scalar replaceable if the allocation type is unknown statically
3000 // (reflection allocation), the object can't be restored during
3001 // deoptimization without precise type.
3002 continue;
3003 }
3004 }
3005
3006 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
3007 if (t == NULL)
3008 continue; // not a TypeOopPtr
3009 if (!t->klass_is_exact())
3010 continue; // not an unique type
3011
3012 if (alloc->is_Allocate()) {
3013 // Set the scalar_replaceable flag for allocation
3014 // so it could be eliminated.
3015 alloc->as_Allocate()->_is_scalar_replaceable = true;
3016 }
3017 if (alloc->is_CallStaticJava()) {
3018 // Set the scalar_replaceable flag for boxing method
3019 // so it could be eliminated.
3020 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3021 }
3022 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
3023 // in order for an object to be scalar-replaceable, it must be:
3024 // - a direct allocation (not a call returning an object)
3025 // - non-escaping
3026 // - eligible to be a unique type
3027 // - not determined to be ineligible by escape analysis
3028 set_map(alloc, n);
3029 set_map(n, alloc);
3030 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
3031 igvn->hash_delete(n);
3032 igvn->set_type(n, tinst);
3033 n->raise_bottom_type(tinst);
3034 igvn->hash_insert(n);
3035 record_for_optimizer(n);
3036 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
3037
3038 // First, put on the worklist all Field edges from Connection Graph
3039 // which is more accurate than putting immediate users from Ideal Graph.
3040 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
3041 PointsToNode* tgt = e.get();
3042 if (tgt->is_Arraycopy()) {
3043 continue;
3044 }
3045 Node* use = tgt->ideal_node();
3046 assert(tgt->is_Field() && use->is_AddP(),
3047 "only AddP nodes are Field edges in CG");
3048 if (use->outcnt() > 0) { // Don't process dead nodes
3049 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
3050 if (addp2 != NULL) {
3051 assert(alloc->is_AllocateArray(),"array allocation was expected");
3052 alloc_worklist.append_if_missing(addp2);
3053 }
3054 alloc_worklist.append_if_missing(use);
3055 }
3056 }
3057
3058 // An allocation may have an Initialize which has raw stores. Scan
3059 // the users of the raw allocation result and push AddP users
3060 // on alloc_worklist.
3061 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
3062 assert (raw_result != NULL, "must have an allocation result");
3063 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
3064 Node *use = raw_result->fast_out(i);
3065 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
3066 Node* addp2 = find_second_addp(use, raw_result);
3067 if (addp2 != NULL) {
3068 assert(alloc->is_AllocateArray(),"array allocation was expected");
3069 alloc_worklist.append_if_missing(addp2);
3070 }
3071 alloc_worklist.append_if_missing(use);
3072 } else if (use->is_MemBar()) {
3073 memnode_worklist.append_if_missing(use);
3074 }
3075 }
3076 }
3077 } else if (n->is_AddP()) {
3078 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
3079 if (jobj == NULL || jobj == phantom_obj) {
3080 #ifdef ASSERT
3081 ptnode_adr(get_addp_base(n)->_idx)->dump();
3082 ptnode_adr(n->_idx)->dump();
3083 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3084 #endif
3085 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3086 return;
3087 }
3088 Node *base = get_map(jobj->idx()); // CheckCastPP node
3089 if (!split_AddP(n, base)) continue; // wrong type from dead path
3090 } else if (n->is_Phi() ||
3091 n->is_CheckCastPP() ||
3092 n->is_EncodeP() ||
3093 n->is_DecodeN() ||
3094 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
3095 if (visited.test_set(n->_idx)) {
3096 assert(n->is_Phi(), "loops only through Phi's");
3097 continue; // already processed
3098 }
3099 JavaObjectNode* jobj = unique_java_object(n);
3100 if (jobj == NULL || jobj == phantom_obj) {
3101 #ifdef ASSERT
3102 ptnode_adr(n->_idx)->dump();
3103 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3104 #endif
3105 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3106 return;
3107 } else {
3108 Node *val = get_map(jobj->idx()); // CheckCastPP node
3109 TypeNode *tn = n->as_Type();
3110 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
3111 assert(tinst != NULL && tinst->is_known_instance() &&
3112 tinst->instance_id() == jobj->idx() , "instance type expected.");
3113
3114 const Type *tn_type = igvn->type(tn);
3115 const TypeOopPtr *tn_t;
3116 if (tn_type->isa_narrowoop()) {
3117 tn_t = tn_type->make_ptr()->isa_oopptr();
3118 } else {
3119 tn_t = tn_type->isa_oopptr();
3120 }
3121 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
3122 if (tn_type->isa_narrowoop()) {
3123 tn_type = tinst->make_narrowoop();
3124 } else {
3125 tn_type = tinst;
3126 }
3127 igvn->hash_delete(tn);
3128 igvn->set_type(tn, tn_type);
3129 tn->set_type(tn_type);
3130 igvn->hash_insert(tn);
3131 record_for_optimizer(n);
3132 } else {
3133 assert(tn_type == TypePtr::NULL_PTR ||
3134 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3135 "unexpected type");
3136 continue; // Skip dead path with different type
3137 }
3138 }
3139 } else {
3140 debug_only(n->dump();)
3141 assert(false, "EA: unexpected node");
3142 continue;
3143 }
3144 // push allocation's users on appropriate worklist
3145 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3146 Node *use = n->fast_out(i);
3147 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3148 // Load/store to instance's field
3149 memnode_worklist.append_if_missing(use);
3150 } else if (use->is_MemBar()) {
3151 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3152 memnode_worklist.append_if_missing(use);
3153 }
3154 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3155 Node* addp2 = find_second_addp(use, n);
3156 if (addp2 != NULL) {
3157 alloc_worklist.append_if_missing(addp2);
3158 }
3159 alloc_worklist.append_if_missing(use);
3160 } else if (use->is_Phi() ||
3161 use->is_CheckCastPP() ||
3162 use->is_EncodeNarrowPtr() ||
3163 use->is_DecodeNarrowPtr() ||
3164 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3165 alloc_worklist.append_if_missing(use);
3166 #ifdef ASSERT
3167 } else if (use->is_Mem()) {
3168 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3169 } else if (use->is_MergeMem()) {
3170 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3171 } else if (use->is_SafePoint()) {
3172 // Look for MergeMem nodes for calls which reference unique allocation
3173 // (through CheckCastPP nodes) even for debug info.
3174 Node* m = use->in(TypeFunc::Memory);
3175 if (m->is_MergeMem()) {
3176 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3177 }
3178 } else if (use->Opcode() == Op_EncodeISOArray) {
3179 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3180 // EncodeISOArray overwrites destination array
3181 memnode_worklist.append_if_missing(use);
3182 }
3183 } else {
3184 uint op = use->Opcode();
3185 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
3186 (use->in(MemNode::Memory) == n)) {
3187 // They overwrite memory edge corresponding to destination array,
3188 memnode_worklist.append_if_missing(use);
3189 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
3190 op == Op_CastP2X || op == Op_StoreCM ||
3191 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3192 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3193 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
3194 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
3195 n->dump();
3196 use->dump();
3197 assert(false, "EA: missing allocation reference path");
3198 }
3199 #endif
3200 }
3201 }
3202
3203 }
3204
3205 // Go over all ArrayCopy nodes and if one of the inputs has a unique
3206 // type, record it in the ArrayCopy node so we know what memory this
3207 // node uses/modified.
3208 for (int next = 0; next < arraycopy_worklist.length(); next++) {
3209 ArrayCopyNode* ac = arraycopy_worklist.at(next);
3210 Node* dest = ac->in(ArrayCopyNode::Dest);
3211 if (dest->is_AddP()) {
3212 dest = get_addp_base(dest);
3213 }
3214 JavaObjectNode* jobj = unique_java_object(dest);
3215 if (jobj != NULL) {
3216 Node *base = get_map(jobj->idx());
3217 if (base != NULL) {
3218 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3219 ac->_dest_type = base_t;
3220 }
3221 }
3222 Node* src = ac->in(ArrayCopyNode::Src);
3223 if (src->is_AddP()) {
3224 src = get_addp_base(src);
3225 }
3226 jobj = unique_java_object(src);
3227 if (jobj != NULL) {
3228 Node* base = get_map(jobj->idx());
3229 if (base != NULL) {
3230 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3231 ac->_src_type = base_t;
3232 }
3233 }
3234 }
3235
3236 // New alias types were created in split_AddP().
3237 uint new_index_end = (uint) _compile->num_alias_types();
3238 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3239
3240 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3241 // compute new values for Memory inputs (the Memory inputs are not
3242 // actually updated until phase 4.)
3243 if (memnode_worklist.length() == 0)
3244 return; // nothing to do
3245 while (memnode_worklist.length() != 0) {
3246 Node *n = memnode_worklist.pop();
3247 if (visited.test_set(n->_idx))
3248 continue;
3249 if (n->is_Phi() || n->is_ClearArray()) {
3250 // we don't need to do anything, but the users must be pushed
3251 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3252 // we don't need to do anything, but the users must be pushed
3253 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
3254 if (n == NULL)
3255 continue;
3256 } else if (n->Opcode() == Op_StrCompressedCopy ||
3257 n->Opcode() == Op_EncodeISOArray) {
3258 // get the memory projection
3259 n = n->find_out_with(Op_SCMemProj);
3260 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3261 } else {
3262 assert(n->is_Mem(), "memory node required.");
3263 Node *addr = n->in(MemNode::Address);
3264 const Type *addr_t = igvn->type(addr);
3265 if (addr_t == Type::TOP)
3266 continue;
3267 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3268 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3269 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3270 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3271 if (_compile->failing()) {
3272 return;
3273 }
3274 if (mem != n->in(MemNode::Memory)) {
3275 // We delay the memory edge update since we need old one in
3276 // MergeMem code below when instances memory slices are separated.
3277 set_map(n, mem);
3278 }
3279 if (n->is_Load()) {
3280 continue; // don't push users
3281 } else if (n->is_LoadStore()) {
3282 // get the memory projection
3283 n = n->find_out_with(Op_SCMemProj);
3284 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3285 }
3286 }
3287 // push user on appropriate worklist
3288 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3289 Node *use = n->fast_out(i);
3290 if (use->is_Phi() || use->is_ClearArray()) {
3291 memnode_worklist.append_if_missing(use);
3292 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3293 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3294 continue;
3295 memnode_worklist.append_if_missing(use);
3296 } else if (use->is_MemBar()) {
3297 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3298 memnode_worklist.append_if_missing(use);
3299 }
3300 #ifdef ASSERT
3301 } else if(use->is_Mem()) {
3302 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3303 } else if (use->is_MergeMem()) {
3304 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3305 } else if (use->Opcode() == Op_EncodeISOArray) {
3306 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3307 // EncodeISOArray overwrites destination array
3308 memnode_worklist.append_if_missing(use);
3309 }
3310 } else {
3311 uint op = use->Opcode();
3312 if ((use->in(MemNode::Memory) == n) &&
3313 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
3314 // They overwrite memory edge corresponding to destination array,
3315 memnode_worklist.append_if_missing(use);
3316 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3317 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3318 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3319 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
3320 n->dump();
3321 use->dump();
3322 assert(false, "EA: missing memory path");
3323 }
3324 #endif
3325 }
3326 }
3327 }
3328
3329 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3330 // Walk each memory slice moving the first node encountered of each
3331 // instance type to the the input corresponding to its alias index.
3332 uint length = _mergemem_worklist.length();
3333 for( uint next = 0; next < length; ++next ) {
3334 MergeMemNode* nmm = _mergemem_worklist.at(next);
3335 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3336 // Note: we don't want to use MergeMemStream here because we only want to
3337 // scan inputs which exist at the start, not ones we add during processing.
3338 // Note 2: MergeMem may already contains instance memory slices added
3339 // during find_inst_mem() call when memory nodes were processed above.
3340 igvn->hash_delete(nmm);
3341 uint nslices = MIN2(nmm->req(), new_index_start);
3342 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3343 Node* mem = nmm->in(i);
3344 Node* cur = NULL;
3345 if (mem == NULL || mem->is_top())
3346 continue;
3347 // First, update mergemem by moving memory nodes to corresponding slices
3348 // if their type became more precise since this mergemem was created.
3349 while (mem->is_Mem()) {
3350 const Type *at = igvn->type(mem->in(MemNode::Address));
3351 if (at != Type::TOP) {
3352 assert (at->isa_ptr() != NULL, "pointer type required.");
3353 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3354 if (idx == i) {
3355 if (cur == NULL)
3356 cur = mem;
3357 } else {
3358 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3359 nmm->set_memory_at(idx, mem);
3360 }
3361 }
3362 }
3363 mem = mem->in(MemNode::Memory);
3364 }
3365 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3366 // Find any instance of the current type if we haven't encountered
3367 // already a memory slice of the instance along the memory chain.
3368 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3369 if((uint)_compile->get_general_index(ni) == i) {
3370 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3371 if (nmm->is_empty_memory(m)) {
3372 Node* result = find_inst_mem(mem, ni, orig_phis);
3373 if (_compile->failing()) {
3374 return;
3375 }
3376 nmm->set_memory_at(ni, result);
3377 }
3378 }
3379 }
3380 }
3381 // Find the rest of instances values
3382 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3383 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3384 Node* result = step_through_mergemem(nmm, ni, tinst);
3385 if (result == nmm->base_memory()) {
3386 // Didn't find instance memory, search through general slice recursively.
3387 result = nmm->memory_at(_compile->get_general_index(ni));
3388 result = find_inst_mem(result, ni, orig_phis);
3389 if (_compile->failing()) {
3390 return;
3391 }
3392 nmm->set_memory_at(ni, result);
3393 }
3394 }
3395 igvn->hash_insert(nmm);
3396 record_for_optimizer(nmm);
3397 }
3398
3399 // Phase 4: Update the inputs of non-instance memory Phis and
3400 // the Memory input of memnodes
3401 // First update the inputs of any non-instance Phi's from
3402 // which we split out an instance Phi. Note we don't have
3403 // to recursively process Phi's encounted on the input memory
3404 // chains as is done in split_memory_phi() since they will
3405 // also be processed here.
3406 for (int j = 0; j < orig_phis.length(); j++) {
3407 PhiNode *phi = orig_phis.at(j);
3408 int alias_idx = _compile->get_alias_index(phi->adr_type());
3409 igvn->hash_delete(phi);
3410 for (uint i = 1; i < phi->req(); i++) {
3411 Node *mem = phi->in(i);
3412 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3413 if (_compile->failing()) {
3414 return;
3415 }
3416 if (mem != new_mem) {
3417 phi->set_req(i, new_mem);
3418 }
3419 }
3420 igvn->hash_insert(phi);
3421 record_for_optimizer(phi);
3422 }
3423
3424 // Update the memory inputs of MemNodes with the value we computed
3425 // in Phase 2 and move stores memory users to corresponding memory slices.
3426 // Disable memory split verification code until the fix for 6984348.
3427 // Currently it produces false negative results since it does not cover all cases.
3428 #if 0 // ifdef ASSERT
3429 visited.Reset();
3430 Node_Stack old_mems(arena, _compile->unique() >> 2);
3431 #endif
3432 for (uint i = 0; i < ideal_nodes.size(); i++) {
3433 Node* n = ideal_nodes.at(i);
3434 Node* nmem = get_map(n->_idx);
3435 assert(nmem != NULL, "sanity");
3436 if (n->is_Mem()) {
3437 #if 0 // ifdef ASSERT
3438 Node* old_mem = n->in(MemNode::Memory);
3439 if (!visited.test_set(old_mem->_idx)) {
3440 old_mems.push(old_mem, old_mem->outcnt());
3441 }
3442 #endif
3443 assert(n->in(MemNode::Memory) != nmem, "sanity");
3444 if (!n->is_Load()) {
3445 // Move memory users of a store first.
3446 move_inst_mem(n, orig_phis);
3447 }
3448 // Now update memory input
3449 igvn->hash_delete(n);
3450 n->set_req(MemNode::Memory, nmem);
3451 igvn->hash_insert(n);
3452 record_for_optimizer(n);
3453 } else {
3454 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3455 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3456 }
3457 }
3458 #if 0 // ifdef ASSERT
3459 // Verify that memory was split correctly
3460 while (old_mems.is_nonempty()) {
3461 Node* old_mem = old_mems.node();
3462 uint old_cnt = old_mems.index();
3463 old_mems.pop();
3464 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3465 }
3466 #endif
3467 }
3468
3469 #ifndef PRODUCT
3470 static const char *node_type_names[] = {
3471 "UnknownType",
3472 "JavaObject",
3473 "LocalVar",
3474 "Field",
3475 "Arraycopy"
3476 };
3477
3478 static const char *esc_names[] = {
3479 "UnknownEscape",
3480 "NoEscape",
3481 "ArgEscape",
3482 "GlobalEscape"
3483 };
3484
3485 void PointsToNode::dump(bool print_state) const {
3486 NodeType nt = node_type();
3487 tty->print("%s ", node_type_names[(int) nt]);
3488 if (print_state) {
3489 EscapeState es = escape_state();
3490 EscapeState fields_es = fields_escape_state();
3491 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3492 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3493 tty->print("NSR ");
3494 }
3495 if (is_Field()) {
3496 FieldNode* f = (FieldNode*)this;
3497 if (f->is_oop())
3498 tty->print("oop ");
3499 if (f->offset() > 0)
3500 tty->print("+%d ", f->offset());
3501 tty->print("(");
3502 for (BaseIterator i(f); i.has_next(); i.next()) {
3503 PointsToNode* b = i.get();
3504 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3505 }
3506 tty->print(" )");
3507 }
3508 tty->print("[");
3509 for (EdgeIterator i(this); i.has_next(); i.next()) {
3510 PointsToNode* e = i.get();
3511 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3512 }
3513 tty->print(" [");
3514 for (UseIterator i(this); i.has_next(); i.next()) {
3515 PointsToNode* u = i.get();
3516 bool is_base = false;
3517 if (PointsToNode::is_base_use(u)) {
3518 is_base = true;
3519 u = PointsToNode::get_use_node(u)->as_Field();
3520 }
3521 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3522 }
3523 tty->print(" ]] ");
3524 if (_node == NULL)
3525 tty->print_cr("<null>");
3526 else
3527 _node->dump();
3528 }
3529
3530 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3531 bool first = true;
3532 int ptnodes_length = ptnodes_worklist.length();
3533 for (int i = 0; i < ptnodes_length; i++) {
3534 PointsToNode *ptn = ptnodes_worklist.at(i);
3535 if (ptn == NULL || !ptn->is_JavaObject())
3536 continue;
3537 PointsToNode::EscapeState es = ptn->escape_state();
3538 if ((es != PointsToNode::NoEscape) && !Verbose) {
3539 continue;
3540 }
3541 Node* n = ptn->ideal_node();
3542 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3543 n->as_CallStaticJava()->is_boxing_method())) {
3544 if (first) {
3545 tty->cr();
3546 tty->print("======== Connection graph for ");
3547 _compile->method()->print_short_name();
3548 tty->cr();
3549 first = false;
3550 }
3551 ptn->dump();
3552 // Print all locals and fields which reference this allocation
3553 for (UseIterator j(ptn); j.has_next(); j.next()) {
3554 PointsToNode* use = j.get();
3555 if (use->is_LocalVar()) {
3556 use->dump(Verbose);
3557 } else if (Verbose) {
3558 use->dump();
3559 }
3560 }
3561 tty->cr();
3562 }
3563 }
3564 }
3565 #endif