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, "encodeBlock") == 0 ||
1023 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
1024 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
1025 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
1026 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
1027 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
1028 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
1029 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
1030 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
1031 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
1032 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
1033 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
1034 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0)
1035 ))) {
1036 call->dump();
1037 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
1038 }
1039 #endif
1040 // Always process arraycopy's destination object since
1041 // we need to add all possible edges to references in
1042 // source object.
1043 if (arg_esc >= PointsToNode::ArgEscape &&
1044 !arg_is_arraycopy_dest) {
1045 continue;
1046 }
1047 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
1048 if (call->is_ArrayCopy()) {
1049 ArrayCopyNode* ac = call->as_ArrayCopy();
1050 if (ac->is_clonebasic() ||
1051 ac->is_arraycopy_validated() ||
1052 ac->is_copyof_validated() ||
1053 ac->is_copyofrange_validated()) {
1054 es = PointsToNode::NoEscape;
1055 }
1056 }
1057 set_escape_state(arg_ptn, es);
1058 if (arg_is_arraycopy_dest) {
1059 Node* src = call->in(TypeFunc::Parms);
1060 if (src->is_AddP()) {
1061 src = get_addp_base(src);
1062 }
1063 PointsToNode* src_ptn = ptnode_adr(src->_idx);
1064 assert(src_ptn != NULL, "should be registered");
1065 if (arg_ptn != src_ptn) {
1066 // Special arraycopy edge:
1067 // A destination object's field can't have the source object
1068 // as base since objects escape states are not related.
1069 // Only escape state of destination object's fields affects
1070 // escape state of fields in source object.
1071 add_arraycopy(call, es, src_ptn, arg_ptn);
1072 }
1073 }
1074 }
1075 }
1076 break;
1077 }
1078 case Op_CallStaticJava: {
1079 // For a static call, we know exactly what method is being called.
1080 // Use bytecode estimator to record the call's escape affects
1081 #ifdef ASSERT
1082 const char* name = call->as_CallStaticJava()->_name;
1083 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1084 #endif
1085 ciMethod* meth = call->as_CallJava()->method();
1086 if ((meth != NULL) && meth->is_boxing_method()) {
1087 break; // Boxing methods do not modify any oops.
1088 }
1089 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1090 // fall-through if not a Java method or no analyzer information
1091 if (call_analyzer != NULL) {
1092 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1093 const TypeTuple* d = call->tf()->domain();
1094 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1095 const Type* at = d->field_at(i);
1096 int k = i - TypeFunc::Parms;
1097 Node* arg = call->in(i);
1098 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1099 if (at->isa_ptr() != NULL &&
1100 call_analyzer->is_arg_returned(k)) {
1101 // The call returns arguments.
1102 if (call_ptn != NULL) { // Is call's result used?
1103 assert(call_ptn->is_LocalVar(), "node should be registered");
1104 assert(arg_ptn != NULL, "node should be registered");
1105 add_edge(call_ptn, arg_ptn);
1106 }
1107 }
1108 if (at->isa_oopptr() != NULL &&
1109 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1110 if (!call_analyzer->is_arg_stack(k)) {
1111 // The argument global escapes
1112 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1113 } else {
1114 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1115 if (!call_analyzer->is_arg_local(k)) {
1116 // The argument itself doesn't escape, but any fields might
1117 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1118 }
1119 }
1120 }
1121 }
1122 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1123 // The call returns arguments.
1124 assert(call_ptn->edge_count() > 0, "sanity");
1125 if (!call_analyzer->is_return_local()) {
1126 // Returns also unknown object.
1127 add_edge(call_ptn, phantom_obj);
1128 }
1129 }
1130 break;
1131 }
1132 }
1133 default: {
1134 // Fall-through here if not a Java method or no analyzer information
1135 // or some other type of call, assume the worst case: all arguments
1136 // globally escape.
1137 const TypeTuple* d = call->tf()->domain();
1138 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1139 const Type* at = d->field_at(i);
1140 if (at->isa_oopptr() != NULL) {
1141 Node* arg = call->in(i);
1142 if (arg->is_AddP()) {
1143 arg = get_addp_base(arg);
1144 }
1145 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1146 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1147 }
1148 }
1149 }
1150 }
1151 }
1152
1153
1154 // Finish Graph construction.
1155 bool ConnectionGraph::complete_connection_graph(
1156 GrowableArray<PointsToNode*>& ptnodes_worklist,
1157 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1158 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1159 GrowableArray<FieldNode*>& oop_fields_worklist) {
1160 // Normally only 1-3 passes needed to build Connection Graph depending
1161 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1162 // Set limit to 20 to catch situation when something did go wrong and
1163 // bailout Escape Analysis.
1164 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1165 #define CG_BUILD_ITER_LIMIT 20
1166
1167 // Propagate GlobalEscape and ArgEscape escape states and check that
1168 // we still have non-escaping objects. The method pushs on _worklist
1169 // Field nodes which reference phantom_object.
1170 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1171 return false; // Nothing to do.
1172 }
1173 // Now propagate references to all JavaObject nodes.
1174 int java_objects_length = java_objects_worklist.length();
1175 elapsedTimer time;
1176 bool timeout = false;
1177 int new_edges = 1;
1178 int iterations = 0;
1179 do {
1180 while ((new_edges > 0) &&
1181 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1182 double start_time = time.seconds();
1183 time.start();
1184 new_edges = 0;
1185 // Propagate references to phantom_object for nodes pushed on _worklist
1186 // by find_non_escaped_objects() and find_field_value().
1187 new_edges += add_java_object_edges(phantom_obj, false);
1188 for (int next = 0; next < java_objects_length; ++next) {
1189 JavaObjectNode* ptn = java_objects_worklist.at(next);
1190 new_edges += add_java_object_edges(ptn, true);
1191
1192 #define SAMPLE_SIZE 4
1193 if ((next % SAMPLE_SIZE) == 0) {
1194 // Each 4 iterations calculate how much time it will take
1195 // to complete graph construction.
1196 time.stop();
1197 // Poll for requests from shutdown mechanism to quiesce compiler
1198 // because Connection graph construction may take long time.
1199 CompileBroker::maybe_block();
1200 double stop_time = time.seconds();
1201 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1202 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1203 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1204 timeout = true;
1205 break; // Timeout
1206 }
1207 start_time = stop_time;
1208 time.start();
1209 }
1210 #undef SAMPLE_SIZE
1211
1212 }
1213 if (timeout) break;
1214 if (new_edges > 0) {
1215 // Update escape states on each iteration if graph was updated.
1216 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1217 return false; // Nothing to do.
1218 }
1219 }
1220 time.stop();
1221 if (time.seconds() >= EscapeAnalysisTimeout) {
1222 timeout = true;
1223 break;
1224 }
1225 }
1226 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1227 time.start();
1228 // Find fields which have unknown value.
1229 int fields_length = oop_fields_worklist.length();
1230 for (int next = 0; next < fields_length; next++) {
1231 FieldNode* field = oop_fields_worklist.at(next);
1232 if (field->edge_count() == 0) {
1233 new_edges += find_field_value(field);
1234 // This code may added new edges to phantom_object.
1235 // Need an other cycle to propagate references to phantom_object.
1236 }
1237 }
1238 time.stop();
1239 if (time.seconds() >= EscapeAnalysisTimeout) {
1240 timeout = true;
1241 break;
1242 }
1243 } else {
1244 new_edges = 0; // Bailout
1245 }
1246 } while (new_edges > 0);
1247
1248 // Bailout if passed limits.
1249 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1250 Compile* C = _compile;
1251 if (C->log() != NULL) {
1252 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1253 C->log()->text("%s", timeout ? "time" : "iterations");
1254 C->log()->end_elem(" limit'");
1255 }
1256 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1257 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length());
1258 // Possible infinite build_connection_graph loop,
1259 // bailout (no changes to ideal graph were made).
1260 return false;
1261 }
1262 #ifdef ASSERT
1263 if (Verbose && PrintEscapeAnalysis) {
1264 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1265 iterations, nodes_size(), ptnodes_worklist.length());
1266 }
1267 #endif
1268
1269 #undef CG_BUILD_ITER_LIMIT
1270
1271 // Find fields initialized by NULL for non-escaping Allocations.
1272 int non_escaped_length = non_escaped_worklist.length();
1273 for (int next = 0; next < non_escaped_length; next++) {
1274 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1275 PointsToNode::EscapeState es = ptn->escape_state();
1276 assert(es <= PointsToNode::ArgEscape, "sanity");
1277 if (es == PointsToNode::NoEscape) {
1278 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1279 // Adding references to NULL object does not change escape states
1280 // since it does not escape. Also no fields are added to NULL object.
1281 add_java_object_edges(null_obj, false);
1282 }
1283 }
1284 Node* n = ptn->ideal_node();
1285 if (n->is_Allocate()) {
1286 // The object allocated by this Allocate node will never be
1287 // seen by an other thread. Mark it so that when it is
1288 // expanded no MemBarStoreStore is added.
1289 InitializeNode* ini = n->as_Allocate()->initialization();
1290 if (ini != NULL)
1291 ini->set_does_not_escape();
1292 }
1293 }
1294 return true; // Finished graph construction.
1295 }
1296
1297 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1298 // and check that we still have non-escaping java objects.
1299 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1300 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1301 GrowableArray<PointsToNode*> escape_worklist;
1302 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1303 int ptnodes_length = ptnodes_worklist.length();
1304 for (int next = 0; next < ptnodes_length; ++next) {
1305 PointsToNode* ptn = ptnodes_worklist.at(next);
1306 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1307 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1308 escape_worklist.push(ptn);
1309 }
1310 }
1311 // Set escape states to referenced nodes (edges list).
1312 while (escape_worklist.length() > 0) {
1313 PointsToNode* ptn = escape_worklist.pop();
1314 PointsToNode::EscapeState es = ptn->escape_state();
1315 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1316 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1317 es >= PointsToNode::ArgEscape) {
1318 // GlobalEscape or ArgEscape state of field means it has unknown value.
1319 if (add_edge(ptn, phantom_obj)) {
1320 // New edge was added
1321 add_field_uses_to_worklist(ptn->as_Field());
1322 }
1323 }
1324 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1325 PointsToNode* e = i.get();
1326 if (e->is_Arraycopy()) {
1327 assert(ptn->arraycopy_dst(), "sanity");
1328 // Propagate only fields escape state through arraycopy edge.
1329 if (e->fields_escape_state() < field_es) {
1330 set_fields_escape_state(e, field_es);
1331 escape_worklist.push(e);
1332 }
1333 } else if (es >= field_es) {
1334 // fields_escape_state is also set to 'es' if it is less than 'es'.
1335 if (e->escape_state() < es) {
1336 set_escape_state(e, es);
1337 escape_worklist.push(e);
1338 }
1339 } else {
1340 // Propagate field escape state.
1341 bool es_changed = false;
1342 if (e->fields_escape_state() < field_es) {
1343 set_fields_escape_state(e, field_es);
1344 es_changed = true;
1345 }
1346 if ((e->escape_state() < field_es) &&
1347 e->is_Field() && ptn->is_JavaObject() &&
1348 e->as_Field()->is_oop()) {
1349 // Change escape state of referenced fields.
1350 set_escape_state(e, field_es);
1351 es_changed = true;
1352 } else if (e->escape_state() < es) {
1353 set_escape_state(e, es);
1354 es_changed = true;
1355 }
1356 if (es_changed) {
1357 escape_worklist.push(e);
1358 }
1359 }
1360 }
1361 }
1362 // Remove escaped objects from non_escaped list.
1363 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1364 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1365 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1366 non_escaped_worklist.delete_at(next);
1367 }
1368 if (ptn->escape_state() == PointsToNode::NoEscape) {
1369 // Find fields in non-escaped allocations which have unknown value.
1370 find_init_values(ptn, phantom_obj, NULL);
1371 }
1372 }
1373 return (non_escaped_worklist.length() > 0);
1374 }
1375
1376 // Add all references to JavaObject node by walking over all uses.
1377 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1378 int new_edges = 0;
1379 if (populate_worklist) {
1380 // Populate _worklist by uses of jobj's uses.
1381 for (UseIterator i(jobj); i.has_next(); i.next()) {
1382 PointsToNode* use = i.get();
1383 if (use->is_Arraycopy())
1384 continue;
1385 add_uses_to_worklist(use);
1386 if (use->is_Field() && use->as_Field()->is_oop()) {
1387 // Put on worklist all field's uses (loads) and
1388 // related field nodes (same base and offset).
1389 add_field_uses_to_worklist(use->as_Field());
1390 }
1391 }
1392 }
1393 for (int l = 0; l < _worklist.length(); l++) {
1394 PointsToNode* use = _worklist.at(l);
1395 if (PointsToNode::is_base_use(use)) {
1396 // Add reference from jobj to field and from field to jobj (field's base).
1397 use = PointsToNode::get_use_node(use)->as_Field();
1398 if (add_base(use->as_Field(), jobj)) {
1399 new_edges++;
1400 }
1401 continue;
1402 }
1403 assert(!use->is_JavaObject(), "sanity");
1404 if (use->is_Arraycopy()) {
1405 if (jobj == null_obj) // NULL object does not have field edges
1406 continue;
1407 // Added edge from Arraycopy node to arraycopy's source java object
1408 if (add_edge(use, jobj)) {
1409 jobj->set_arraycopy_src();
1410 new_edges++;
1411 }
1412 // and stop here.
1413 continue;
1414 }
1415 if (!add_edge(use, jobj))
1416 continue; // No new edge added, there was such edge already.
1417 new_edges++;
1418 if (use->is_LocalVar()) {
1419 add_uses_to_worklist(use);
1420 if (use->arraycopy_dst()) {
1421 for (EdgeIterator i(use); i.has_next(); i.next()) {
1422 PointsToNode* e = i.get();
1423 if (e->is_Arraycopy()) {
1424 if (jobj == null_obj) // NULL object does not have field edges
1425 continue;
1426 // Add edge from arraycopy's destination java object to Arraycopy node.
1427 if (add_edge(jobj, e)) {
1428 new_edges++;
1429 jobj->set_arraycopy_dst();
1430 }
1431 }
1432 }
1433 }
1434 } else {
1435 // Added new edge to stored in field values.
1436 // Put on worklist all field's uses (loads) and
1437 // related field nodes (same base and offset).
1438 add_field_uses_to_worklist(use->as_Field());
1439 }
1440 }
1441 _worklist.clear();
1442 _in_worklist.Reset();
1443 return new_edges;
1444 }
1445
1446 // Put on worklist all related field nodes.
1447 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1448 assert(field->is_oop(), "sanity");
1449 int offset = field->offset();
1450 add_uses_to_worklist(field);
1451 // Loop over all bases of this field and push on worklist Field nodes
1452 // with the same offset and base (since they may reference the same field).
1453 for (BaseIterator i(field); i.has_next(); i.next()) {
1454 PointsToNode* base = i.get();
1455 add_fields_to_worklist(field, base);
1456 // Check if the base was source object of arraycopy and go over arraycopy's
1457 // destination objects since values stored to a field of source object are
1458 // accessable by uses (loads) of fields of destination objects.
1459 if (base->arraycopy_src()) {
1460 for (UseIterator j(base); j.has_next(); j.next()) {
1461 PointsToNode* arycp = j.get();
1462 if (arycp->is_Arraycopy()) {
1463 for (UseIterator k(arycp); k.has_next(); k.next()) {
1464 PointsToNode* abase = k.get();
1465 if (abase->arraycopy_dst() && abase != base) {
1466 // Look for the same arraycopy reference.
1467 add_fields_to_worklist(field, abase);
1468 }
1469 }
1470 }
1471 }
1472 }
1473 }
1474 }
1475
1476 // Put on worklist all related field nodes.
1477 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1478 int offset = field->offset();
1479 if (base->is_LocalVar()) {
1480 for (UseIterator j(base); j.has_next(); j.next()) {
1481 PointsToNode* f = j.get();
1482 if (PointsToNode::is_base_use(f)) { // Field
1483 f = PointsToNode::get_use_node(f);
1484 if (f == field || !f->as_Field()->is_oop())
1485 continue;
1486 int offs = f->as_Field()->offset();
1487 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1488 add_to_worklist(f);
1489 }
1490 }
1491 }
1492 } else {
1493 assert(base->is_JavaObject(), "sanity");
1494 if (// Skip phantom_object since it is only used to indicate that
1495 // this field's content globally escapes.
1496 (base != phantom_obj) &&
1497 // NULL object node does not have fields.
1498 (base != null_obj)) {
1499 for (EdgeIterator i(base); i.has_next(); i.next()) {
1500 PointsToNode* f = i.get();
1501 // Skip arraycopy edge since store to destination object field
1502 // does not update value in source object field.
1503 if (f->is_Arraycopy()) {
1504 assert(base->arraycopy_dst(), "sanity");
1505 continue;
1506 }
1507 if (f == field || !f->as_Field()->is_oop())
1508 continue;
1509 int offs = f->as_Field()->offset();
1510 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1511 add_to_worklist(f);
1512 }
1513 }
1514 }
1515 }
1516 }
1517
1518 // Find fields which have unknown value.
1519 int ConnectionGraph::find_field_value(FieldNode* field) {
1520 // Escaped fields should have init value already.
1521 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1522 int new_edges = 0;
1523 for (BaseIterator i(field); i.has_next(); i.next()) {
1524 PointsToNode* base = i.get();
1525 if (base->is_JavaObject()) {
1526 // Skip Allocate's fields which will be processed later.
1527 if (base->ideal_node()->is_Allocate())
1528 return 0;
1529 assert(base == null_obj, "only NULL ptr base expected here");
1530 }
1531 }
1532 if (add_edge(field, phantom_obj)) {
1533 // New edge was added
1534 new_edges++;
1535 add_field_uses_to_worklist(field);
1536 }
1537 return new_edges;
1538 }
1539
1540 // Find fields initializing values for allocations.
1541 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1542 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1543 int new_edges = 0;
1544 Node* alloc = pta->ideal_node();
1545 if (init_val == phantom_obj) {
1546 // Do nothing for Allocate nodes since its fields values are
1547 // "known" unless they are initialized by arraycopy/clone.
1548 if (alloc->is_Allocate() && !pta->arraycopy_dst())
1549 return 0;
1550 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
1551 #ifdef ASSERT
1552 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) {
1553 const char* name = alloc->as_CallStaticJava()->_name;
1554 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1555 }
1556 #endif
1557 // Non-escaped allocation returned from Java or runtime call have
1558 // unknown values in fields.
1559 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1560 PointsToNode* field = i.get();
1561 if (field->is_Field() && field->as_Field()->is_oop()) {
1562 if (add_edge(field, phantom_obj)) {
1563 // New edge was added
1564 new_edges++;
1565 add_field_uses_to_worklist(field->as_Field());
1566 }
1567 }
1568 }
1569 return new_edges;
1570 }
1571 assert(init_val == null_obj, "sanity");
1572 // Do nothing for Call nodes since its fields values are unknown.
1573 if (!alloc->is_Allocate())
1574 return 0;
1575
1576 InitializeNode* ini = alloc->as_Allocate()->initialization();
1577 bool visited_bottom_offset = false;
1578 GrowableArray<int> offsets_worklist;
1579
1580 // Check if an oop field's initializing value is recorded and add
1581 // a corresponding NULL if field's value if it is not recorded.
1582 // Connection Graph does not record a default initialization by NULL
1583 // captured by Initialize node.
1584 //
1585 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1586 PointsToNode* field = i.get(); // Field (AddP)
1587 if (!field->is_Field() || !field->as_Field()->is_oop())
1588 continue; // Not oop field
1589 int offset = field->as_Field()->offset();
1590 if (offset == Type::OffsetBot) {
1591 if (!visited_bottom_offset) {
1592 // OffsetBot is used to reference array's element,
1593 // always add reference to NULL to all Field nodes since we don't
1594 // known which element is referenced.
1595 if (add_edge(field, null_obj)) {
1596 // New edge was added
1597 new_edges++;
1598 add_field_uses_to_worklist(field->as_Field());
1599 visited_bottom_offset = true;
1600 }
1601 }
1602 } else {
1603 // Check only oop fields.
1604 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1605 if (adr_type->isa_rawptr()) {
1606 #ifdef ASSERT
1607 // Raw pointers are used for initializing stores so skip it
1608 // since it should be recorded already
1609 Node* base = get_addp_base(field->ideal_node());
1610 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1611 (base->in(0) == alloc),"unexpected pointer type");
1612 #endif
1613 continue;
1614 }
1615 if (!offsets_worklist.contains(offset)) {
1616 offsets_worklist.append(offset);
1617 Node* value = NULL;
1618 if (ini != NULL) {
1619 // StoreP::memory_type() == T_ADDRESS
1620 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1621 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1622 // Make sure initializing store has the same type as this AddP.
1623 // This AddP may reference non existing field because it is on a
1624 // dead branch of bimorphic call which is not eliminated yet.
1625 if (store != NULL && store->is_Store() &&
1626 store->as_Store()->memory_type() == ft) {
1627 value = store->in(MemNode::ValueIn);
1628 #ifdef ASSERT
1629 if (VerifyConnectionGraph) {
1630 // Verify that AddP already points to all objects the value points to.
1631 PointsToNode* val = ptnode_adr(value->_idx);
1632 assert((val != NULL), "should be processed already");
1633 PointsToNode* missed_obj = NULL;
1634 if (val->is_JavaObject()) {
1635 if (!field->points_to(val->as_JavaObject())) {
1636 missed_obj = val;
1637 }
1638 } else {
1639 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1640 tty->print_cr("----------init store has invalid value -----");
1641 store->dump();
1642 val->dump();
1643 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1644 }
1645 for (EdgeIterator j(val); j.has_next(); j.next()) {
1646 PointsToNode* obj = j.get();
1647 if (obj->is_JavaObject()) {
1648 if (!field->points_to(obj->as_JavaObject())) {
1649 missed_obj = obj;
1650 break;
1651 }
1652 }
1653 }
1654 }
1655 if (missed_obj != NULL) {
1656 tty->print_cr("----------field---------------------------------");
1657 field->dump();
1658 tty->print_cr("----------missed referernce to object-----------");
1659 missed_obj->dump();
1660 tty->print_cr("----------object referernced by init store -----");
1661 store->dump();
1662 val->dump();
1663 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1664 }
1665 }
1666 #endif
1667 } else {
1668 // There could be initializing stores which follow allocation.
1669 // For example, a volatile field store is not collected
1670 // by Initialize node.
1671 //
1672 // Need to check for dependent loads to separate such stores from
1673 // stores which follow loads. For now, add initial value NULL so
1674 // that compare pointers optimization works correctly.
1675 }
1676 }
1677 if (value == NULL) {
1678 // A field's initializing value was not recorded. Add NULL.
1679 if (add_edge(field, null_obj)) {
1680 // New edge was added
1681 new_edges++;
1682 add_field_uses_to_worklist(field->as_Field());
1683 }
1684 }
1685 }
1686 }
1687 }
1688 return new_edges;
1689 }
1690
1691 // Adjust scalar_replaceable state after Connection Graph is built.
1692 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1693 // Search for non-escaping objects which are not scalar replaceable
1694 // and mark them to propagate the state to referenced objects.
1695
1696 // 1. An object is not scalar replaceable if the field into which it is
1697 // stored has unknown offset (stored into unknown element of an array).
1698 //
1699 for (UseIterator i(jobj); i.has_next(); i.next()) {
1700 PointsToNode* use = i.get();
1701 if (use->is_Arraycopy()) {
1702 continue;
1703 }
1704 if (use->is_Field()) {
1705 FieldNode* field = use->as_Field();
1706 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
1707 if (field->offset() == Type::OffsetBot) {
1708 jobj->set_scalar_replaceable(false);
1709 return;
1710 }
1711 // 2. An object is not scalar replaceable if the field into which it is
1712 // stored has multiple bases one of which is null.
1713 if (field->base_count() > 1) {
1714 for (BaseIterator i(field); i.has_next(); i.next()) {
1715 PointsToNode* base = i.get();
1716 if (base == null_obj) {
1717 jobj->set_scalar_replaceable(false);
1718 return;
1719 }
1720 }
1721 }
1722 }
1723 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1724 // 3. An object is not scalar replaceable if it is merged with other objects.
1725 for (EdgeIterator j(use); j.has_next(); j.next()) {
1726 PointsToNode* ptn = j.get();
1727 if (ptn->is_JavaObject() && ptn != jobj) {
1728 // Mark all objects.
1729 jobj->set_scalar_replaceable(false);
1730 ptn->set_scalar_replaceable(false);
1731 }
1732 }
1733 if (!jobj->scalar_replaceable()) {
1734 return;
1735 }
1736 }
1737
1738 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1739 if (j.get()->is_Arraycopy()) {
1740 continue;
1741 }
1742
1743 // Non-escaping object node should point only to field nodes.
1744 FieldNode* field = j.get()->as_Field();
1745 int offset = field->as_Field()->offset();
1746
1747 // 4. An object is not scalar replaceable if it has a field with unknown
1748 // offset (array's element is accessed in loop).
1749 if (offset == Type::OffsetBot) {
1750 jobj->set_scalar_replaceable(false);
1751 return;
1752 }
1753 // 5. Currently an object is not scalar replaceable if a LoadStore node
1754 // access its field since the field value is unknown after it.
1755 //
1756 Node* n = field->ideal_node();
1757 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1758 if (n->fast_out(i)->is_LoadStore()) {
1759 jobj->set_scalar_replaceable(false);
1760 return;
1761 }
1762 }
1763
1764 // 6. Or the address may point to more then one object. This may produce
1765 // the false positive result (set not scalar replaceable)
1766 // since the flow-insensitive escape analysis can't separate
1767 // the case when stores overwrite the field's value from the case
1768 // when stores happened on different control branches.
1769 //
1770 // Note: it will disable scalar replacement in some cases:
1771 //
1772 // Point p[] = new Point[1];
1773 // p[0] = new Point(); // Will be not scalar replaced
1774 //
1775 // but it will save us from incorrect optimizations in next cases:
1776 //
1777 // Point p[] = new Point[1];
1778 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1779 //
1780 if (field->base_count() > 1) {
1781 for (BaseIterator i(field); i.has_next(); i.next()) {
1782 PointsToNode* base = i.get();
1783 // Don't take into account LocalVar nodes which
1784 // may point to only one object which should be also
1785 // this field's base by now.
1786 if (base->is_JavaObject() && base != jobj) {
1787 // Mark all bases.
1788 jobj->set_scalar_replaceable(false);
1789 base->set_scalar_replaceable(false);
1790 }
1791 }
1792 }
1793 }
1794 }
1795
1796 #ifdef ASSERT
1797 void ConnectionGraph::verify_connection_graph(
1798 GrowableArray<PointsToNode*>& ptnodes_worklist,
1799 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1800 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1801 GrowableArray<Node*>& addp_worklist) {
1802 // Verify that graph is complete - no new edges could be added.
1803 int java_objects_length = java_objects_worklist.length();
1804 int non_escaped_length = non_escaped_worklist.length();
1805 int new_edges = 0;
1806 for (int next = 0; next < java_objects_length; ++next) {
1807 JavaObjectNode* ptn = java_objects_worklist.at(next);
1808 new_edges += add_java_object_edges(ptn, true);
1809 }
1810 assert(new_edges == 0, "graph was not complete");
1811 // Verify that escape state is final.
1812 int length = non_escaped_worklist.length();
1813 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1814 assert((non_escaped_length == non_escaped_worklist.length()) &&
1815 (non_escaped_length == length) &&
1816 (_worklist.length() == 0), "escape state was not final");
1817
1818 // Verify fields information.
1819 int addp_length = addp_worklist.length();
1820 for (int next = 0; next < addp_length; ++next ) {
1821 Node* n = addp_worklist.at(next);
1822 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1823 if (field->is_oop()) {
1824 // Verify that field has all bases
1825 Node* base = get_addp_base(n);
1826 PointsToNode* ptn = ptnode_adr(base->_idx);
1827 if (ptn->is_JavaObject()) {
1828 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1829 } else {
1830 assert(ptn->is_LocalVar(), "sanity");
1831 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1832 PointsToNode* e = i.get();
1833 if (e->is_JavaObject()) {
1834 assert(field->has_base(e->as_JavaObject()), "sanity");
1835 }
1836 }
1837 }
1838 // Verify that all fields have initializing values.
1839 if (field->edge_count() == 0) {
1840 tty->print_cr("----------field does not have references----------");
1841 field->dump();
1842 for (BaseIterator i(field); i.has_next(); i.next()) {
1843 PointsToNode* base = i.get();
1844 tty->print_cr("----------field has next base---------------------");
1845 base->dump();
1846 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1847 tty->print_cr("----------base has fields-------------------------");
1848 for (EdgeIterator j(base); j.has_next(); j.next()) {
1849 j.get()->dump();
1850 }
1851 tty->print_cr("----------base has references---------------------");
1852 for (UseIterator j(base); j.has_next(); j.next()) {
1853 j.get()->dump();
1854 }
1855 }
1856 }
1857 for (UseIterator i(field); i.has_next(); i.next()) {
1858 i.get()->dump();
1859 }
1860 assert(field->edge_count() > 0, "sanity");
1861 }
1862 }
1863 }
1864 }
1865 #endif
1866
1867 // Optimize ideal graph.
1868 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1869 GrowableArray<Node*>& storestore_worklist) {
1870 Compile* C = _compile;
1871 PhaseIterGVN* igvn = _igvn;
1872 if (EliminateLocks) {
1873 // Mark locks before changing ideal graph.
1874 int cnt = C->macro_count();
1875 for( int i=0; i < cnt; i++ ) {
1876 Node *n = C->macro_node(i);
1877 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1878 AbstractLockNode* alock = n->as_AbstractLock();
1879 if (!alock->is_non_esc_obj()) {
1880 if (not_global_escape(alock->obj_node())) {
1881 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1882 // The lock could be marked eliminated by lock coarsening
1883 // code during first IGVN before EA. Replace coarsened flag
1884 // to eliminate all associated locks/unlocks.
1885 #ifdef ASSERT
1886 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1887 #endif
1888 alock->set_non_esc_obj();
1889 }
1890 }
1891 }
1892 }
1893 }
1894
1895 if (OptimizePtrCompare) {
1896 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1897 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1898 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1899 // Optimize objects compare.
1900 while (ptr_cmp_worklist.length() != 0) {
1901 Node *n = ptr_cmp_worklist.pop();
1902 Node *res = optimize_ptr_compare(n);
1903 if (res != NULL) {
1904 #ifndef PRODUCT
1905 if (PrintOptimizePtrCompare) {
1906 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"));
1907 if (Verbose) {
1908 n->dump(1);
1909 }
1910 }
1911 #endif
1912 igvn->replace_node(n, res);
1913 }
1914 }
1915 // cleanup
1916 if (_pcmp_neq->outcnt() == 0)
1917 igvn->hash_delete(_pcmp_neq);
1918 if (_pcmp_eq->outcnt() == 0)
1919 igvn->hash_delete(_pcmp_eq);
1920 }
1921
1922 // For MemBarStoreStore nodes added in library_call.cpp, check
1923 // escape status of associated AllocateNode and optimize out
1924 // MemBarStoreStore node if the allocated object never escapes.
1925 while (storestore_worklist.length() != 0) {
1926 Node *n = storestore_worklist.pop();
1927 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1928 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1929 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1930 if (not_global_escape(alloc)) {
1931 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1932 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1933 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1934 igvn->register_new_node_with_optimizer(mb);
1935 igvn->replace_node(storestore, mb);
1936 }
1937 }
1938 }
1939
1940 // Optimize objects compare.
1941 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1942 assert(OptimizePtrCompare, "sanity");
1943 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1944 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1945 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1946 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1947 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1948 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1949
1950 // Check simple cases first.
1951 if (jobj1 != NULL) {
1952 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1953 if (jobj1 == jobj2) {
1954 // Comparing the same not escaping object.
1955 return _pcmp_eq;
1956 }
1957 Node* obj = jobj1->ideal_node();
1958 // Comparing not escaping allocation.
1959 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1960 !ptn2->points_to(jobj1)) {
1961 return _pcmp_neq; // This includes nullness check.
1962 }
1963 }
1964 }
1965 if (jobj2 != NULL) {
1966 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1967 Node* obj = jobj2->ideal_node();
1968 // Comparing not escaping allocation.
1969 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1970 !ptn1->points_to(jobj2)) {
1971 return _pcmp_neq; // This includes nullness check.
1972 }
1973 }
1974 }
1975 if (jobj1 != NULL && jobj1 != phantom_obj &&
1976 jobj2 != NULL && jobj2 != phantom_obj &&
1977 jobj1->ideal_node()->is_Con() &&
1978 jobj2->ideal_node()->is_Con()) {
1979 // Klass or String constants compare. Need to be careful with
1980 // compressed pointers - compare types of ConN and ConP instead of nodes.
1981 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1982 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1983 if (t1->make_ptr() == t2->make_ptr()) {
1984 return _pcmp_eq;
1985 } else {
1986 return _pcmp_neq;
1987 }
1988 }
1989 if (ptn1->meet(ptn2)) {
1990 return NULL; // Sets are not disjoint
1991 }
1992
1993 // Sets are disjoint.
1994 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1995 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1996 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1997 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1998 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
1999 (set2_has_unknown_ptr && set1_has_null_ptr)) {
2000 // Check nullness of unknown object.
2001 return NULL;
2002 }
2003
2004 // Disjointness by itself is not sufficient since
2005 // alias analysis is not complete for escaped objects.
2006 // Disjoint sets are definitely unrelated only when
2007 // at least one set has only not escaping allocations.
2008 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
2009 if (ptn1->non_escaping_allocation()) {
2010 return _pcmp_neq;
2011 }
2012 }
2013 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2014 if (ptn2->non_escaping_allocation()) {
2015 return _pcmp_neq;
2016 }
2017 }
2018 return NULL;
2019 }
2020
2021 // Connection Graph constuction functions.
2022
2023 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
2024 PointsToNode* ptadr = _nodes.at(n->_idx);
2025 if (ptadr != NULL) {
2026 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
2027 return;
2028 }
2029 Compile* C = _compile;
2030 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
2031 _nodes.at_put(n->_idx, ptadr);
2032 }
2033
2034 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
2035 PointsToNode* ptadr = _nodes.at(n->_idx);
2036 if (ptadr != NULL) {
2037 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
2038 return;
2039 }
2040 Compile* C = _compile;
2041 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
2042 _nodes.at_put(n->_idx, ptadr);
2043 }
2044
2045 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
2046 PointsToNode* ptadr = _nodes.at(n->_idx);
2047 if (ptadr != NULL) {
2048 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
2049 return;
2050 }
2051 bool unsafe = false;
2052 bool is_oop = is_oop_field(n, offset, &unsafe);
2053 if (unsafe) {
2054 es = PointsToNode::GlobalEscape;
2055 }
2056 Compile* C = _compile;
2057 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
2058 _nodes.at_put(n->_idx, field);
2059 }
2060
2061 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
2062 PointsToNode* src, PointsToNode* dst) {
2063 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
2064 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
2065 PointsToNode* ptadr = _nodes.at(n->_idx);
2066 if (ptadr != NULL) {
2067 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
2068 return;
2069 }
2070 Compile* C = _compile;
2071 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
2072 _nodes.at_put(n->_idx, ptadr);
2073 // Add edge from arraycopy node to source object.
2074 (void)add_edge(ptadr, src);
2075 src->set_arraycopy_src();
2076 // Add edge from destination object to arraycopy node.
2077 (void)add_edge(dst, ptadr);
2078 dst->set_arraycopy_dst();
2079 }
2080
2081 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
2082 const Type* adr_type = n->as_AddP()->bottom_type();
2083 BasicType bt = T_INT;
2084 if (offset == Type::OffsetBot) {
2085 // Check only oop fields.
2086 if (!adr_type->isa_aryptr() ||
2087 (adr_type->isa_aryptr()->klass() == NULL) ||
2088 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2089 // OffsetBot is used to reference array's element. Ignore first AddP.
2090 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2091 bt = T_OBJECT;
2092 }
2093 }
2094 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2095 if (adr_type->isa_instptr()) {
2096 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
2097 if (field != NULL) {
2098 bt = field->layout_type();
2099 } else {
2100 // Check for unsafe oop field access
2101 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2102 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2103 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2104 bt = T_OBJECT;
2105 (*unsafe) = true;
2106 }
2107 }
2108 } else if (adr_type->isa_aryptr()) {
2109 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2110 // Ignore array length load.
2111 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2112 // Ignore first AddP.
2113 } else {
2114 const Type* elemtype = adr_type->isa_aryptr()->elem();
2115 bt = elemtype->array_element_basic_type();
2116 }
2117 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2118 // Allocation initialization, ThreadLocal field access, unsafe access
2119 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2120 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2121 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2122 bt = T_OBJECT;
2123 }
2124 }
2125 }
2126 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
2127 }
2128
2129 // Returns unique pointed java object or NULL.
2130 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2131 assert(!_collecting, "should not call when contructed graph");
2132 // If the node was created after the escape computation we can't answer.
2133 uint idx = n->_idx;
2134 if (idx >= nodes_size()) {
2135 return NULL;
2136 }
2137 PointsToNode* ptn = ptnode_adr(idx);
2138 if (ptn->is_JavaObject()) {
2139 return ptn->as_JavaObject();
2140 }
2141 assert(ptn->is_LocalVar(), "sanity");
2142 // Check all java objects it points to.
2143 JavaObjectNode* jobj = NULL;
2144 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2145 PointsToNode* e = i.get();
2146 if (e->is_JavaObject()) {
2147 if (jobj == NULL) {
2148 jobj = e->as_JavaObject();
2149 } else if (jobj != e) {
2150 return NULL;
2151 }
2152 }
2153 }
2154 return jobj;
2155 }
2156
2157 // Return true if this node points only to non-escaping allocations.
2158 bool PointsToNode::non_escaping_allocation() {
2159 if (is_JavaObject()) {
2160 Node* n = ideal_node();
2161 if (n->is_Allocate() || n->is_CallStaticJava()) {
2162 return (escape_state() == PointsToNode::NoEscape);
2163 } else {
2164 return false;
2165 }
2166 }
2167 assert(is_LocalVar(), "sanity");
2168 // Check all java objects it points to.
2169 for (EdgeIterator i(this); i.has_next(); i.next()) {
2170 PointsToNode* e = i.get();
2171 if (e->is_JavaObject()) {
2172 Node* n = e->ideal_node();
2173 if ((e->escape_state() != PointsToNode::NoEscape) ||
2174 !(n->is_Allocate() || n->is_CallStaticJava())) {
2175 return false;
2176 }
2177 }
2178 }
2179 return true;
2180 }
2181
2182 // Return true if we know the node does not escape globally.
2183 bool ConnectionGraph::not_global_escape(Node *n) {
2184 assert(!_collecting, "should not call during graph construction");
2185 // If the node was created after the escape computation we can't answer.
2186 uint idx = n->_idx;
2187 if (idx >= nodes_size()) {
2188 return false;
2189 }
2190 PointsToNode* ptn = ptnode_adr(idx);
2191 PointsToNode::EscapeState es = ptn->escape_state();
2192 // If we have already computed a value, return it.
2193 if (es >= PointsToNode::GlobalEscape)
2194 return false;
2195 if (ptn->is_JavaObject()) {
2196 return true; // (es < PointsToNode::GlobalEscape);
2197 }
2198 assert(ptn->is_LocalVar(), "sanity");
2199 // Check all java objects it points to.
2200 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2201 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2202 return false;
2203 }
2204 return true;
2205 }
2206
2207
2208 // Helper functions
2209
2210 // Return true if this node points to specified node or nodes it points to.
2211 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2212 if (is_JavaObject()) {
2213 return (this == ptn);
2214 }
2215 assert(is_LocalVar() || is_Field(), "sanity");
2216 for (EdgeIterator i(this); i.has_next(); i.next()) {
2217 if (i.get() == ptn)
2218 return true;
2219 }
2220 return false;
2221 }
2222
2223 // Return true if one node points to an other.
2224 bool PointsToNode::meet(PointsToNode* ptn) {
2225 if (this == ptn) {
2226 return true;
2227 } else if (ptn->is_JavaObject()) {
2228 return this->points_to(ptn->as_JavaObject());
2229 } else if (this->is_JavaObject()) {
2230 return ptn->points_to(this->as_JavaObject());
2231 }
2232 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2233 int ptn_count = ptn->edge_count();
2234 for (EdgeIterator i(this); i.has_next(); i.next()) {
2235 PointsToNode* this_e = i.get();
2236 for (int j = 0; j < ptn_count; j++) {
2237 if (this_e == ptn->edge(j))
2238 return true;
2239 }
2240 }
2241 return false;
2242 }
2243
2244 #ifdef ASSERT
2245 // Return true if bases point to this java object.
2246 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2247 for (BaseIterator i(this); i.has_next(); i.next()) {
2248 if (i.get() == jobj)
2249 return true;
2250 }
2251 return false;
2252 }
2253 #endif
2254
2255 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2256 const Type *adr_type = phase->type(adr);
2257 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2258 adr->in(AddPNode::Address)->is_Proj() &&
2259 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2260 // We are computing a raw address for a store captured by an Initialize
2261 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2262 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2263 assert(offs != Type::OffsetBot ||
2264 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2265 "offset must be a constant or it is initialization of array");
2266 return offs;
2267 }
2268 const TypePtr *t_ptr = adr_type->isa_ptr();
2269 assert(t_ptr != NULL, "must be a pointer type");
2270 return t_ptr->offset();
2271 }
2272
2273 Node* ConnectionGraph::get_addp_base(Node *addp) {
2274 assert(addp->is_AddP(), "must be AddP");
2275 //
2276 // AddP cases for Base and Address inputs:
2277 // case #1. Direct object's field reference:
2278 // Allocate
2279 // |
2280 // Proj #5 ( oop result )
2281 // |
2282 // CheckCastPP (cast to instance type)
2283 // | |
2284 // AddP ( base == address )
2285 //
2286 // case #2. Indirect object's field reference:
2287 // Phi
2288 // |
2289 // CastPP (cast to instance type)
2290 // | |
2291 // AddP ( base == address )
2292 //
2293 // case #3. Raw object's field reference for Initialize node:
2294 // Allocate
2295 // |
2296 // Proj #5 ( oop result )
2297 // top |
2298 // \ |
2299 // AddP ( base == top )
2300 //
2301 // case #4. Array's element reference:
2302 // {CheckCastPP | CastPP}
2303 // | | |
2304 // | AddP ( array's element offset )
2305 // | |
2306 // AddP ( array's offset )
2307 //
2308 // case #5. Raw object's field reference for arraycopy stub call:
2309 // The inline_native_clone() case when the arraycopy stub is called
2310 // after the allocation before Initialize and CheckCastPP nodes.
2311 // Allocate
2312 // |
2313 // Proj #5 ( oop result )
2314 // | |
2315 // AddP ( base == address )
2316 //
2317 // case #6. Constant Pool, ThreadLocal, CastX2P or
2318 // Raw object's field reference:
2319 // {ConP, ThreadLocal, CastX2P, raw Load}
2320 // top |
2321 // \ |
2322 // AddP ( base == top )
2323 //
2324 // case #7. Klass's field reference.
2325 // LoadKlass
2326 // | |
2327 // AddP ( base == address )
2328 //
2329 // case #8. narrow Klass's field reference.
2330 // LoadNKlass
2331 // |
2332 // DecodeN
2333 // | |
2334 // AddP ( base == address )
2335 //
2336 // case #9. Mixed unsafe access
2337 // {instance}
2338 // |
2339 // CheckCastPP (raw)
2340 // top |
2341 // \ |
2342 // AddP ( base == top )
2343 //
2344 Node *base = addp->in(AddPNode::Base);
2345 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
2346 base = addp->in(AddPNode::Address);
2347 while (base->is_AddP()) {
2348 // Case #6 (unsafe access) may have several chained AddP nodes.
2349 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2350 base = base->in(AddPNode::Address);
2351 }
2352 if (base->Opcode() == Op_CheckCastPP &&
2353 base->bottom_type()->isa_rawptr() &&
2354 _igvn->type(base->in(1))->isa_oopptr()) {
2355 base = base->in(1); // Case #9
2356 } else {
2357 Node* uncast_base = base->uncast();
2358 int opcode = uncast_base->Opcode();
2359 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2360 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2361 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2362 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2363 }
2364 }
2365 return base;
2366 }
2367
2368 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2369 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2370 Node* addp2 = addp->raw_out(0);
2371 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2372 addp2->in(AddPNode::Base) == n &&
2373 addp2->in(AddPNode::Address) == addp) {
2374 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2375 //
2376 // Find array's offset to push it on worklist first and
2377 // as result process an array's element offset first (pushed second)
2378 // to avoid CastPP for the array's offset.
2379 // Otherwise the inserted CastPP (LocalVar) will point to what
2380 // the AddP (Field) points to. Which would be wrong since
2381 // the algorithm expects the CastPP has the same point as
2382 // as AddP's base CheckCastPP (LocalVar).
2383 //
2384 // ArrayAllocation
2385 // |
2386 // CheckCastPP
2387 // |
2388 // memProj (from ArrayAllocation CheckCastPP)
2389 // | ||
2390 // | || Int (element index)
2391 // | || | ConI (log(element size))
2392 // | || | /
2393 // | || LShift
2394 // | || /
2395 // | AddP (array's element offset)
2396 // | |
2397 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2398 // | / /
2399 // AddP (array's offset)
2400 // |
2401 // Load/Store (memory operation on array's element)
2402 //
2403 return addp2;
2404 }
2405 return NULL;
2406 }
2407
2408 //
2409 // Adjust the type and inputs of an AddP which computes the
2410 // address of a field of an instance
2411 //
2412 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2413 PhaseGVN* igvn = _igvn;
2414 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2415 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2416 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2417 if (t == NULL) {
2418 // We are computing a raw address for a store captured by an Initialize
2419 // compute an appropriate address type (cases #3 and #5).
2420 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2421 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2422 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2423 assert(offs != Type::OffsetBot, "offset must be a constant");
2424 t = base_t->add_offset(offs)->is_oopptr();
2425 }
2426 int inst_id = base_t->instance_id();
2427 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2428 "old type must be non-instance or match new type");
2429
2430 // The type 't' could be subclass of 'base_t'.
2431 // As result t->offset() could be large then base_t's size and it will
2432 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2433 // constructor verifies correctness of the offset.
2434 //
2435 // It could happened on subclass's branch (from the type profiling
2436 // inlining) which was not eliminated during parsing since the exactness
2437 // of the allocation type was not propagated to the subclass type check.
2438 //
2439 // Or the type 't' could be not related to 'base_t' at all.
2440 // It could happened when CHA type is different from MDO type on a dead path
2441 // (for example, from instanceof check) which is not collapsed during parsing.
2442 //
2443 // Do nothing for such AddP node and don't process its users since
2444 // this code branch will go away.
2445 //
2446 if (!t->is_known_instance() &&
2447 !base_t->klass()->is_subtype_of(t->klass())) {
2448 return false; // bail out
2449 }
2450 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2451 // Do NOT remove the next line: ensure a new alias index is allocated
2452 // for the instance type. Note: C++ will not remove it since the call
2453 // has side effect.
2454 int alias_idx = _compile->get_alias_index(tinst);
2455 igvn->set_type(addp, tinst);
2456 // record the allocation in the node map
2457 set_map(addp, get_map(base->_idx));
2458 // Set addp's Base and Address to 'base'.
2459 Node *abase = addp->in(AddPNode::Base);
2460 Node *adr = addp->in(AddPNode::Address);
2461 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2462 adr->in(0)->_idx == (uint)inst_id) {
2463 // Skip AddP cases #3 and #5.
2464 } else {
2465 assert(!abase->is_top(), "sanity"); // AddP case #3
2466 if (abase != base) {
2467 igvn->hash_delete(addp);
2468 addp->set_req(AddPNode::Base, base);
2469 if (abase == adr) {
2470 addp->set_req(AddPNode::Address, base);
2471 } else {
2472 // AddP case #4 (adr is array's element offset AddP node)
2473 #ifdef ASSERT
2474 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2475 assert(adr->is_AddP() && atype != NULL &&
2476 atype->instance_id() == inst_id, "array's element offset should be processed first");
2477 #endif
2478 }
2479 igvn->hash_insert(addp);
2480 }
2481 }
2482 // Put on IGVN worklist since at least addp's type was changed above.
2483 record_for_optimizer(addp);
2484 return true;
2485 }
2486
2487 //
2488 // Create a new version of orig_phi if necessary. Returns either the newly
2489 // created phi or an existing phi. Sets create_new to indicate whether a new
2490 // phi was created. Cache the last newly created phi in the node map.
2491 //
2492 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2493 Compile *C = _compile;
2494 PhaseGVN* igvn = _igvn;
2495 new_created = false;
2496 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2497 // nothing to do if orig_phi is bottom memory or matches alias_idx
2498 if (phi_alias_idx == alias_idx) {
2499 return orig_phi;
2500 }
2501 // Have we recently created a Phi for this alias index?
2502 PhiNode *result = get_map_phi(orig_phi->_idx);
2503 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2504 return result;
2505 }
2506 // Previous check may fail when the same wide memory Phi was split into Phis
2507 // for different memory slices. Search all Phis for this region.
2508 if (result != NULL) {
2509 Node* region = orig_phi->in(0);
2510 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2511 Node* phi = region->fast_out(i);
2512 if (phi->is_Phi() &&
2513 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2514 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2515 return phi->as_Phi();
2516 }
2517 }
2518 }
2519 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2520 if (C->do_escape_analysis() == true && !C->failing()) {
2521 // Retry compilation without escape analysis.
2522 // If this is the first failure, the sentinel string will "stick"
2523 // to the Compile object, and the C2Compiler will see it and retry.
2524 C->record_failure(C2Compiler::retry_no_escape_analysis());
2525 }
2526 return NULL;
2527 }
2528 orig_phi_worklist.append_if_missing(orig_phi);
2529 const TypePtr *atype = C->get_adr_type(alias_idx);
2530 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2531 C->copy_node_notes_to(result, orig_phi);
2532 igvn->set_type(result, result->bottom_type());
2533 record_for_optimizer(result);
2534 set_map(orig_phi, result);
2535 new_created = true;
2536 return result;
2537 }
2538
2539 //
2540 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2541 // specified alias index.
2542 //
2543 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2544 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2545 Compile *C = _compile;
2546 PhaseGVN* igvn = _igvn;
2547 bool new_phi_created;
2548 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2549 if (!new_phi_created) {
2550 return result;
2551 }
2552 GrowableArray<PhiNode *> phi_list;
2553 GrowableArray<uint> cur_input;
2554 PhiNode *phi = orig_phi;
2555 uint idx = 1;
2556 bool finished = false;
2557 while(!finished) {
2558 while (idx < phi->req()) {
2559 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2560 if (mem != NULL && mem->is_Phi()) {
2561 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2562 if (new_phi_created) {
2563 // found an phi for which we created a new split, push current one on worklist and begin
2564 // processing new one
2565 phi_list.push(phi);
2566 cur_input.push(idx);
2567 phi = mem->as_Phi();
2568 result = newphi;
2569 idx = 1;
2570 continue;
2571 } else {
2572 mem = newphi;
2573 }
2574 }
2575 if (C->failing()) {
2576 return NULL;
2577 }
2578 result->set_req(idx++, mem);
2579 }
2580 #ifdef ASSERT
2581 // verify that the new Phi has an input for each input of the original
2582 assert( phi->req() == result->req(), "must have same number of inputs.");
2583 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2584 #endif
2585 // Check if all new phi's inputs have specified alias index.
2586 // Otherwise use old phi.
2587 for (uint i = 1; i < phi->req(); i++) {
2588 Node* in = result->in(i);
2589 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2590 }
2591 // we have finished processing a Phi, see if there are any more to do
2592 finished = (phi_list.length() == 0 );
2593 if (!finished) {
2594 phi = phi_list.pop();
2595 idx = cur_input.pop();
2596 PhiNode *prev_result = get_map_phi(phi->_idx);
2597 prev_result->set_req(idx++, result);
2598 result = prev_result;
2599 }
2600 }
2601 return result;
2602 }
2603
2604 //
2605 // The next methods are derived from methods in MemNode.
2606 //
2607 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2608 Node *mem = mmem;
2609 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2610 // means an array I have not precisely typed yet. Do not do any
2611 // alias stuff with it any time soon.
2612 if (toop->base() != Type::AnyPtr &&
2613 !(toop->klass() != NULL &&
2614 toop->klass()->is_java_lang_Object() &&
2615 toop->offset() == Type::OffsetBot)) {
2616 mem = mmem->memory_at(alias_idx);
2617 // Update input if it is progress over what we have now
2618 }
2619 return mem;
2620 }
2621
2622 //
2623 // Move memory users to their memory slices.
2624 //
2625 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2626 Compile* C = _compile;
2627 PhaseGVN* igvn = _igvn;
2628 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2629 assert(tp != NULL, "ptr type");
2630 int alias_idx = C->get_alias_index(tp);
2631 int general_idx = C->get_general_index(alias_idx);
2632
2633 // Move users first
2634 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2635 Node* use = n->fast_out(i);
2636 if (use->is_MergeMem()) {
2637 MergeMemNode* mmem = use->as_MergeMem();
2638 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2639 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2640 continue; // Nothing to do
2641 }
2642 // Replace previous general reference to mem node.
2643 uint orig_uniq = C->unique();
2644 Node* m = find_inst_mem(n, general_idx, orig_phis);
2645 assert(orig_uniq == C->unique(), "no new nodes");
2646 mmem->set_memory_at(general_idx, m);
2647 --imax;
2648 --i;
2649 } else if (use->is_MemBar()) {
2650 assert(!use->is_Initialize(), "initializing stores should not be moved");
2651 if (use->req() > MemBarNode::Precedent &&
2652 use->in(MemBarNode::Precedent) == n) {
2653 // Don't move related membars.
2654 record_for_optimizer(use);
2655 continue;
2656 }
2657 tp = use->as_MemBar()->adr_type()->isa_ptr();
2658 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) ||
2659 alias_idx == general_idx) {
2660 continue; // Nothing to do
2661 }
2662 // Move to general memory slice.
2663 uint orig_uniq = C->unique();
2664 Node* m = find_inst_mem(n, general_idx, orig_phis);
2665 assert(orig_uniq == C->unique(), "no new nodes");
2666 igvn->hash_delete(use);
2667 imax -= use->replace_edge(n, m);
2668 igvn->hash_insert(use);
2669 record_for_optimizer(use);
2670 --i;
2671 #ifdef ASSERT
2672 } else if (use->is_Mem()) {
2673 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2674 // Don't move related cardmark.
2675 continue;
2676 }
2677 // Memory nodes should have new memory input.
2678 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2679 assert(tp != NULL, "ptr type");
2680 int idx = C->get_alias_index(tp);
2681 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2682 "Following memory nodes should have new memory input or be on the same memory slice");
2683 } else if (use->is_Phi()) {
2684 // Phi nodes should be split and moved already.
2685 tp = use->as_Phi()->adr_type()->isa_ptr();
2686 assert(tp != NULL, "ptr type");
2687 int idx = C->get_alias_index(tp);
2688 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2689 } else {
2690 use->dump();
2691 assert(false, "should not be here");
2692 #endif
2693 }
2694 }
2695 }
2696
2697 //
2698 // Search memory chain of "mem" to find a MemNode whose address
2699 // is the specified alias index.
2700 //
2701 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2702 if (orig_mem == NULL)
2703 return orig_mem;
2704 Compile* C = _compile;
2705 PhaseGVN* igvn = _igvn;
2706 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2707 bool is_instance = (toop != NULL) && toop->is_known_instance();
2708 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
2709 Node *prev = NULL;
2710 Node *result = orig_mem;
2711 while (prev != result) {
2712 prev = result;
2713 if (result == start_mem)
2714 break; // hit one of our sentinels
2715 if (result->is_Mem()) {
2716 const Type *at = igvn->type(result->in(MemNode::Address));
2717 if (at == Type::TOP)
2718 break; // Dead
2719 assert (at->isa_ptr() != NULL, "pointer type required.");
2720 int idx = C->get_alias_index(at->is_ptr());
2721 if (idx == alias_idx)
2722 break; // Found
2723 if (!is_instance && (at->isa_oopptr() == NULL ||
2724 !at->is_oopptr()->is_known_instance())) {
2725 break; // Do not skip store to general memory slice.
2726 }
2727 result = result->in(MemNode::Memory);
2728 }
2729 if (!is_instance)
2730 continue; // don't search further for non-instance types
2731 // skip over a call which does not affect this memory slice
2732 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2733 Node *proj_in = result->in(0);
2734 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2735 break; // hit one of our sentinels
2736 } else if (proj_in->is_Call()) {
2737 // ArrayCopy node processed here as well
2738 CallNode *call = proj_in->as_Call();
2739 if (!call->may_modify(toop, igvn)) {
2740 result = call->in(TypeFunc::Memory);
2741 }
2742 } else if (proj_in->is_Initialize()) {
2743 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2744 // Stop if this is the initialization for the object instance which
2745 // which contains this memory slice, otherwise skip over it.
2746 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2747 result = proj_in->in(TypeFunc::Memory);
2748 }
2749 } else if (proj_in->is_MemBar()) {
2750 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() &&
2751 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() &&
2752 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) {
2753 // clone
2754 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy();
2755 if (ac->may_modify(toop, igvn)) {
2756 break;
2757 }
2758 }
2759 result = proj_in->in(TypeFunc::Memory);
2760 }
2761 } else if (result->is_MergeMem()) {
2762 MergeMemNode *mmem = result->as_MergeMem();
2763 result = step_through_mergemem(mmem, alias_idx, toop);
2764 if (result == mmem->base_memory()) {
2765 // Didn't find instance memory, search through general slice recursively.
2766 result = mmem->memory_at(C->get_general_index(alias_idx));
2767 result = find_inst_mem(result, alias_idx, orig_phis);
2768 if (C->failing()) {
2769 return NULL;
2770 }
2771 mmem->set_memory_at(alias_idx, result);
2772 }
2773 } else if (result->is_Phi() &&
2774 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2775 Node *un = result->as_Phi()->unique_input(igvn);
2776 if (un != NULL) {
2777 orig_phis.append_if_missing(result->as_Phi());
2778 result = un;
2779 } else {
2780 break;
2781 }
2782 } else if (result->is_ClearArray()) {
2783 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2784 // Can not bypass initialization of the instance
2785 // we are looking for.
2786 break;
2787 }
2788 // Otherwise skip it (the call updated 'result' value).
2789 } else if (result->Opcode() == Op_SCMemProj) {
2790 Node* mem = result->in(0);
2791 Node* adr = NULL;
2792 if (mem->is_LoadStore()) {
2793 adr = mem->in(MemNode::Address);
2794 } else {
2795 assert(mem->Opcode() == Op_EncodeISOArray ||
2796 mem->Opcode() == Op_StrCompressedCopy, "sanity");
2797 adr = mem->in(3); // Memory edge corresponds to destination array
2798 }
2799 const Type *at = igvn->type(adr);
2800 if (at != Type::TOP) {
2801 assert(at->isa_ptr() != NULL, "pointer type required.");
2802 int idx = C->get_alias_index(at->is_ptr());
2803 if (idx == alias_idx) {
2804 // Assert in debug mode
2805 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
2806 break; // In product mode return SCMemProj node
2807 }
2808 }
2809 result = mem->in(MemNode::Memory);
2810 } else if (result->Opcode() == Op_StrInflatedCopy) {
2811 Node* adr = result->in(3); // Memory edge corresponds to destination array
2812 const Type *at = igvn->type(adr);
2813 if (at != Type::TOP) {
2814 assert(at->isa_ptr() != NULL, "pointer type required.");
2815 int idx = C->get_alias_index(at->is_ptr());
2816 if (idx == alias_idx) {
2817 // Assert in debug mode
2818 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
2819 break; // In product mode return SCMemProj node
2820 }
2821 }
2822 result = result->in(MemNode::Memory);
2823 }
2824 }
2825 if (result->is_Phi()) {
2826 PhiNode *mphi = result->as_Phi();
2827 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2828 const TypePtr *t = mphi->adr_type();
2829 if (!is_instance) {
2830 // Push all non-instance Phis on the orig_phis worklist to update inputs
2831 // during Phase 4 if needed.
2832 orig_phis.append_if_missing(mphi);
2833 } else if (C->get_alias_index(t) != alias_idx) {
2834 // Create a new Phi with the specified alias index type.
2835 result = split_memory_phi(mphi, alias_idx, orig_phis);
2836 }
2837 }
2838 // the result is either MemNode, PhiNode, InitializeNode.
2839 return result;
2840 }
2841
2842 //
2843 // Convert the types of unescaped object to instance types where possible,
2844 // propagate the new type information through the graph, and update memory
2845 // edges and MergeMem inputs to reflect the new type.
2846 //
2847 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2848 // The processing is done in 4 phases:
2849 //
2850 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2851 // types for the CheckCastPP for allocations where possible.
2852 // Propagate the new types through users as follows:
2853 // casts and Phi: push users on alloc_worklist
2854 // AddP: cast Base and Address inputs to the instance type
2855 // push any AddP users on alloc_worklist and push any memnode
2856 // users onto memnode_worklist.
2857 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2858 // search the Memory chain for a store with the appropriate type
2859 // address type. If a Phi is found, create a new version with
2860 // the appropriate memory slices from each of the Phi inputs.
2861 // For stores, process the users as follows:
2862 // MemNode: push on memnode_worklist
2863 // MergeMem: push on mergemem_worklist
2864 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2865 // moving the first node encountered of each instance type to the
2866 // the input corresponding to its alias index.
2867 // appropriate memory slice.
2868 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2869 //
2870 // In the following example, the CheckCastPP nodes are the cast of allocation
2871 // results and the allocation of node 29 is unescaped and eligible to be an
2872 // instance type.
2873 //
2874 // We start with:
2875 //
2876 // 7 Parm #memory
2877 // 10 ConI "12"
2878 // 19 CheckCastPP "Foo"
2879 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2880 // 29 CheckCastPP "Foo"
2881 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2882 //
2883 // 40 StoreP 25 7 20 ... alias_index=4
2884 // 50 StoreP 35 40 30 ... alias_index=4
2885 // 60 StoreP 45 50 20 ... alias_index=4
2886 // 70 LoadP _ 60 30 ... alias_index=4
2887 // 80 Phi 75 50 60 Memory alias_index=4
2888 // 90 LoadP _ 80 30 ... alias_index=4
2889 // 100 LoadP _ 80 20 ... alias_index=4
2890 //
2891 //
2892 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2893 // and creating a new alias index for node 30. This gives:
2894 //
2895 // 7 Parm #memory
2896 // 10 ConI "12"
2897 // 19 CheckCastPP "Foo"
2898 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2899 // 29 CheckCastPP "Foo" iid=24
2900 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2901 //
2902 // 40 StoreP 25 7 20 ... alias_index=4
2903 // 50 StoreP 35 40 30 ... alias_index=6
2904 // 60 StoreP 45 50 20 ... alias_index=4
2905 // 70 LoadP _ 60 30 ... alias_index=6
2906 // 80 Phi 75 50 60 Memory alias_index=4
2907 // 90 LoadP _ 80 30 ... alias_index=6
2908 // 100 LoadP _ 80 20 ... alias_index=4
2909 //
2910 // In phase 2, new memory inputs are computed for the loads and stores,
2911 // And a new version of the phi is created. In phase 4, the inputs to
2912 // node 80 are updated and then the memory nodes are updated with the
2913 // values computed in phase 2. This results in:
2914 //
2915 // 7 Parm #memory
2916 // 10 ConI "12"
2917 // 19 CheckCastPP "Foo"
2918 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2919 // 29 CheckCastPP "Foo" iid=24
2920 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2921 //
2922 // 40 StoreP 25 7 20 ... alias_index=4
2923 // 50 StoreP 35 7 30 ... alias_index=6
2924 // 60 StoreP 45 40 20 ... alias_index=4
2925 // 70 LoadP _ 50 30 ... alias_index=6
2926 // 80 Phi 75 40 60 Memory alias_index=4
2927 // 120 Phi 75 50 50 Memory alias_index=6
2928 // 90 LoadP _ 120 30 ... alias_index=6
2929 // 100 LoadP _ 80 20 ... alias_index=4
2930 //
2931 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) {
2932 GrowableArray<Node *> memnode_worklist;
2933 GrowableArray<PhiNode *> orig_phis;
2934 PhaseIterGVN *igvn = _igvn;
2935 uint new_index_start = (uint) _compile->num_alias_types();
2936 Arena* arena = Thread::current()->resource_area();
2937 VectorSet visited(arena);
2938 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2939 uint unique_old = _compile->unique();
2940
2941 // Phase 1: Process possible allocations from alloc_worklist.
2942 // Create instance types for the CheckCastPP for allocations where possible.
2943 //
2944 // (Note: don't forget to change the order of the second AddP node on
2945 // the alloc_worklist if the order of the worklist processing is changed,
2946 // see the comment in find_second_addp().)
2947 //
2948 while (alloc_worklist.length() != 0) {
2949 Node *n = alloc_worklist.pop();
2950 uint ni = n->_idx;
2951 if (n->is_Call()) {
2952 CallNode *alloc = n->as_Call();
2953 // copy escape information to call node
2954 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2955 PointsToNode::EscapeState es = ptn->escape_state();
2956 // We have an allocation or call which returns a Java object,
2957 // see if it is unescaped.
2958 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2959 continue;
2960 // Find CheckCastPP for the allocate or for the return value of a call
2961 n = alloc->result_cast();
2962 if (n == NULL) { // No uses except Initialize node
2963 if (alloc->is_Allocate()) {
2964 // Set the scalar_replaceable flag for allocation
2965 // so it could be eliminated if it has no uses.
2966 alloc->as_Allocate()->_is_scalar_replaceable = true;
2967 }
2968 if (alloc->is_CallStaticJava()) {
2969 // Set the scalar_replaceable flag for boxing method
2970 // so it could be eliminated if it has no uses.
2971 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2972 }
2973 continue;
2974 }
2975 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2976 assert(!alloc->is_Allocate(), "allocation should have unique type");
2977 continue;
2978 }
2979
2980 // The inline code for Object.clone() casts the allocation result to
2981 // java.lang.Object and then to the actual type of the allocated
2982 // object. Detect this case and use the second cast.
2983 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2984 // the allocation result is cast to java.lang.Object and then
2985 // to the actual Array type.
2986 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2987 && (alloc->is_AllocateArray() ||
2988 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2989 Node *cast2 = NULL;
2990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2991 Node *use = n->fast_out(i);
2992 if (use->is_CheckCastPP()) {
2993 cast2 = use;
2994 break;
2995 }
2996 }
2997 if (cast2 != NULL) {
2998 n = cast2;
2999 } else {
3000 // Non-scalar replaceable if the allocation type is unknown statically
3001 // (reflection allocation), the object can't be restored during
3002 // deoptimization without precise type.
3003 continue;
3004 }
3005 }
3006
3007 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
3008 if (t == NULL)
3009 continue; // not a TypeOopPtr
3010 if (!t->klass_is_exact())
3011 continue; // not an unique type
3012
3013 if (alloc->is_Allocate()) {
3014 // Set the scalar_replaceable flag for allocation
3015 // so it could be eliminated.
3016 alloc->as_Allocate()->_is_scalar_replaceable = true;
3017 }
3018 if (alloc->is_CallStaticJava()) {
3019 // Set the scalar_replaceable flag for boxing method
3020 // so it could be eliminated.
3021 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3022 }
3023 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
3024 // in order for an object to be scalar-replaceable, it must be:
3025 // - a direct allocation (not a call returning an object)
3026 // - non-escaping
3027 // - eligible to be a unique type
3028 // - not determined to be ineligible by escape analysis
3029 set_map(alloc, n);
3030 set_map(n, alloc);
3031 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
3032 igvn->hash_delete(n);
3033 igvn->set_type(n, tinst);
3034 n->raise_bottom_type(tinst);
3035 igvn->hash_insert(n);
3036 record_for_optimizer(n);
3037 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
3038
3039 // First, put on the worklist all Field edges from Connection Graph
3040 // which is more accurate than putting immediate users from Ideal Graph.
3041 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
3042 PointsToNode* tgt = e.get();
3043 if (tgt->is_Arraycopy()) {
3044 continue;
3045 }
3046 Node* use = tgt->ideal_node();
3047 assert(tgt->is_Field() && use->is_AddP(),
3048 "only AddP nodes are Field edges in CG");
3049 if (use->outcnt() > 0) { // Don't process dead nodes
3050 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
3051 if (addp2 != NULL) {
3052 assert(alloc->is_AllocateArray(),"array allocation was expected");
3053 alloc_worklist.append_if_missing(addp2);
3054 }
3055 alloc_worklist.append_if_missing(use);
3056 }
3057 }
3058
3059 // An allocation may have an Initialize which has raw stores. Scan
3060 // the users of the raw allocation result and push AddP users
3061 // on alloc_worklist.
3062 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
3063 assert (raw_result != NULL, "must have an allocation result");
3064 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
3065 Node *use = raw_result->fast_out(i);
3066 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
3067 Node* addp2 = find_second_addp(use, raw_result);
3068 if (addp2 != NULL) {
3069 assert(alloc->is_AllocateArray(),"array allocation was expected");
3070 alloc_worklist.append_if_missing(addp2);
3071 }
3072 alloc_worklist.append_if_missing(use);
3073 } else if (use->is_MemBar()) {
3074 memnode_worklist.append_if_missing(use);
3075 }
3076 }
3077 }
3078 } else if (n->is_AddP()) {
3079 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
3080 if (jobj == NULL || jobj == phantom_obj) {
3081 #ifdef ASSERT
3082 ptnode_adr(get_addp_base(n)->_idx)->dump();
3083 ptnode_adr(n->_idx)->dump();
3084 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3085 #endif
3086 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3087 return;
3088 }
3089 Node *base = get_map(jobj->idx()); // CheckCastPP node
3090 if (!split_AddP(n, base)) continue; // wrong type from dead path
3091 } else if (n->is_Phi() ||
3092 n->is_CheckCastPP() ||
3093 n->is_EncodeP() ||
3094 n->is_DecodeN() ||
3095 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
3096 if (visited.test_set(n->_idx)) {
3097 assert(n->is_Phi(), "loops only through Phi's");
3098 continue; // already processed
3099 }
3100 JavaObjectNode* jobj = unique_java_object(n);
3101 if (jobj == NULL || jobj == phantom_obj) {
3102 #ifdef ASSERT
3103 ptnode_adr(n->_idx)->dump();
3104 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3105 #endif
3106 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3107 return;
3108 } else {
3109 Node *val = get_map(jobj->idx()); // CheckCastPP node
3110 TypeNode *tn = n->as_Type();
3111 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
3112 assert(tinst != NULL && tinst->is_known_instance() &&
3113 tinst->instance_id() == jobj->idx() , "instance type expected.");
3114
3115 const Type *tn_type = igvn->type(tn);
3116 const TypeOopPtr *tn_t;
3117 if (tn_type->isa_narrowoop()) {
3118 tn_t = tn_type->make_ptr()->isa_oopptr();
3119 } else {
3120 tn_t = tn_type->isa_oopptr();
3121 }
3122 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
3123 if (tn_type->isa_narrowoop()) {
3124 tn_type = tinst->make_narrowoop();
3125 } else {
3126 tn_type = tinst;
3127 }
3128 igvn->hash_delete(tn);
3129 igvn->set_type(tn, tn_type);
3130 tn->set_type(tn_type);
3131 igvn->hash_insert(tn);
3132 record_for_optimizer(n);
3133 } else {
3134 assert(tn_type == TypePtr::NULL_PTR ||
3135 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3136 "unexpected type");
3137 continue; // Skip dead path with different type
3138 }
3139 }
3140 } else {
3141 debug_only(n->dump();)
3142 assert(false, "EA: unexpected node");
3143 continue;
3144 }
3145 // push allocation's users on appropriate worklist
3146 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3147 Node *use = n->fast_out(i);
3148 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3149 // Load/store to instance's field
3150 memnode_worklist.append_if_missing(use);
3151 } else if (use->is_MemBar()) {
3152 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3153 memnode_worklist.append_if_missing(use);
3154 }
3155 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3156 Node* addp2 = find_second_addp(use, n);
3157 if (addp2 != NULL) {
3158 alloc_worklist.append_if_missing(addp2);
3159 }
3160 alloc_worklist.append_if_missing(use);
3161 } else if (use->is_Phi() ||
3162 use->is_CheckCastPP() ||
3163 use->is_EncodeNarrowPtr() ||
3164 use->is_DecodeNarrowPtr() ||
3165 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3166 alloc_worklist.append_if_missing(use);
3167 #ifdef ASSERT
3168 } else if (use->is_Mem()) {
3169 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3170 } else if (use->is_MergeMem()) {
3171 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3172 } else if (use->is_SafePoint()) {
3173 // Look for MergeMem nodes for calls which reference unique allocation
3174 // (through CheckCastPP nodes) even for debug info.
3175 Node* m = use->in(TypeFunc::Memory);
3176 if (m->is_MergeMem()) {
3177 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3178 }
3179 } else if (use->Opcode() == Op_EncodeISOArray) {
3180 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3181 // EncodeISOArray overwrites destination array
3182 memnode_worklist.append_if_missing(use);
3183 }
3184 } else {
3185 uint op = use->Opcode();
3186 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
3187 (use->in(MemNode::Memory) == n)) {
3188 // They overwrite memory edge corresponding to destination array,
3189 memnode_worklist.append_if_missing(use);
3190 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
3191 op == Op_CastP2X || op == Op_StoreCM ||
3192 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3193 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3194 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
3195 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
3196 n->dump();
3197 use->dump();
3198 assert(false, "EA: missing allocation reference path");
3199 }
3200 #endif
3201 }
3202 }
3203
3204 }
3205
3206 // Go over all ArrayCopy nodes and if one of the inputs has a unique
3207 // type, record it in the ArrayCopy node so we know what memory this
3208 // node uses/modified.
3209 for (int next = 0; next < arraycopy_worklist.length(); next++) {
3210 ArrayCopyNode* ac = arraycopy_worklist.at(next);
3211 Node* dest = ac->in(ArrayCopyNode::Dest);
3212 if (dest->is_AddP()) {
3213 dest = get_addp_base(dest);
3214 }
3215 JavaObjectNode* jobj = unique_java_object(dest);
3216 if (jobj != NULL) {
3217 Node *base = get_map(jobj->idx());
3218 if (base != NULL) {
3219 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3220 ac->_dest_type = base_t;
3221 }
3222 }
3223 Node* src = ac->in(ArrayCopyNode::Src);
3224 if (src->is_AddP()) {
3225 src = get_addp_base(src);
3226 }
3227 jobj = unique_java_object(src);
3228 if (jobj != NULL) {
3229 Node* base = get_map(jobj->idx());
3230 if (base != NULL) {
3231 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3232 ac->_src_type = base_t;
3233 }
3234 }
3235 }
3236
3237 // New alias types were created in split_AddP().
3238 uint new_index_end = (uint) _compile->num_alias_types();
3239 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3240
3241 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3242 // compute new values for Memory inputs (the Memory inputs are not
3243 // actually updated until phase 4.)
3244 if (memnode_worklist.length() == 0)
3245 return; // nothing to do
3246 while (memnode_worklist.length() != 0) {
3247 Node *n = memnode_worklist.pop();
3248 if (visited.test_set(n->_idx))
3249 continue;
3250 if (n->is_Phi() || n->is_ClearArray()) {
3251 // we don't need to do anything, but the users must be pushed
3252 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3253 // we don't need to do anything, but the users must be pushed
3254 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
3255 if (n == NULL)
3256 continue;
3257 } else if (n->Opcode() == Op_StrCompressedCopy ||
3258 n->Opcode() == Op_EncodeISOArray) {
3259 // get the memory projection
3260 n = n->find_out_with(Op_SCMemProj);
3261 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3262 } else {
3263 assert(n->is_Mem(), "memory node required.");
3264 Node *addr = n->in(MemNode::Address);
3265 const Type *addr_t = igvn->type(addr);
3266 if (addr_t == Type::TOP)
3267 continue;
3268 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3269 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3270 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3271 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3272 if (_compile->failing()) {
3273 return;
3274 }
3275 if (mem != n->in(MemNode::Memory)) {
3276 // We delay the memory edge update since we need old one in
3277 // MergeMem code below when instances memory slices are separated.
3278 set_map(n, mem);
3279 }
3280 if (n->is_Load()) {
3281 continue; // don't push users
3282 } else if (n->is_LoadStore()) {
3283 // get the memory projection
3284 n = n->find_out_with(Op_SCMemProj);
3285 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3286 }
3287 }
3288 // push user on appropriate worklist
3289 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3290 Node *use = n->fast_out(i);
3291 if (use->is_Phi() || use->is_ClearArray()) {
3292 memnode_worklist.append_if_missing(use);
3293 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3294 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3295 continue;
3296 memnode_worklist.append_if_missing(use);
3297 } else if (use->is_MemBar()) {
3298 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3299 memnode_worklist.append_if_missing(use);
3300 }
3301 #ifdef ASSERT
3302 } else if(use->is_Mem()) {
3303 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3304 } else if (use->is_MergeMem()) {
3305 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3306 } else if (use->Opcode() == Op_EncodeISOArray) {
3307 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3308 // EncodeISOArray overwrites destination array
3309 memnode_worklist.append_if_missing(use);
3310 }
3311 } else {
3312 uint op = use->Opcode();
3313 if ((use->in(MemNode::Memory) == n) &&
3314 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
3315 // They overwrite memory edge corresponding to destination array,
3316 memnode_worklist.append_if_missing(use);
3317 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3318 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3319 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3320 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
3321 n->dump();
3322 use->dump();
3323 assert(false, "EA: missing memory path");
3324 }
3325 #endif
3326 }
3327 }
3328 }
3329
3330 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3331 // Walk each memory slice moving the first node encountered of each
3332 // instance type to the the input corresponding to its alias index.
3333 uint length = _mergemem_worklist.length();
3334 for( uint next = 0; next < length; ++next ) {
3335 MergeMemNode* nmm = _mergemem_worklist.at(next);
3336 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3337 // Note: we don't want to use MergeMemStream here because we only want to
3338 // scan inputs which exist at the start, not ones we add during processing.
3339 // Note 2: MergeMem may already contains instance memory slices added
3340 // during find_inst_mem() call when memory nodes were processed above.
3341 igvn->hash_delete(nmm);
3342 uint nslices = MIN2(nmm->req(), new_index_start);
3343 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3344 Node* mem = nmm->in(i);
3345 Node* cur = NULL;
3346 if (mem == NULL || mem->is_top())
3347 continue;
3348 // First, update mergemem by moving memory nodes to corresponding slices
3349 // if their type became more precise since this mergemem was created.
3350 while (mem->is_Mem()) {
3351 const Type *at = igvn->type(mem->in(MemNode::Address));
3352 if (at != Type::TOP) {
3353 assert (at->isa_ptr() != NULL, "pointer type required.");
3354 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3355 if (idx == i) {
3356 if (cur == NULL)
3357 cur = mem;
3358 } else {
3359 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3360 nmm->set_memory_at(idx, mem);
3361 }
3362 }
3363 }
3364 mem = mem->in(MemNode::Memory);
3365 }
3366 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3367 // Find any instance of the current type if we haven't encountered
3368 // already a memory slice of the instance along the memory chain.
3369 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3370 if((uint)_compile->get_general_index(ni) == i) {
3371 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3372 if (nmm->is_empty_memory(m)) {
3373 Node* result = find_inst_mem(mem, ni, orig_phis);
3374 if (_compile->failing()) {
3375 return;
3376 }
3377 nmm->set_memory_at(ni, result);
3378 }
3379 }
3380 }
3381 }
3382 // Find the rest of instances values
3383 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3384 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3385 Node* result = step_through_mergemem(nmm, ni, tinst);
3386 if (result == nmm->base_memory()) {
3387 // Didn't find instance memory, search through general slice recursively.
3388 result = nmm->memory_at(_compile->get_general_index(ni));
3389 result = find_inst_mem(result, ni, orig_phis);
3390 if (_compile->failing()) {
3391 return;
3392 }
3393 nmm->set_memory_at(ni, result);
3394 }
3395 }
3396 igvn->hash_insert(nmm);
3397 record_for_optimizer(nmm);
3398 }
3399
3400 // Phase 4: Update the inputs of non-instance memory Phis and
3401 // the Memory input of memnodes
3402 // First update the inputs of any non-instance Phi's from
3403 // which we split out an instance Phi. Note we don't have
3404 // to recursively process Phi's encounted on the input memory
3405 // chains as is done in split_memory_phi() since they will
3406 // also be processed here.
3407 for (int j = 0; j < orig_phis.length(); j++) {
3408 PhiNode *phi = orig_phis.at(j);
3409 int alias_idx = _compile->get_alias_index(phi->adr_type());
3410 igvn->hash_delete(phi);
3411 for (uint i = 1; i < phi->req(); i++) {
3412 Node *mem = phi->in(i);
3413 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3414 if (_compile->failing()) {
3415 return;
3416 }
3417 if (mem != new_mem) {
3418 phi->set_req(i, new_mem);
3419 }
3420 }
3421 igvn->hash_insert(phi);
3422 record_for_optimizer(phi);
3423 }
3424
3425 // Update the memory inputs of MemNodes with the value we computed
3426 // in Phase 2 and move stores memory users to corresponding memory slices.
3427 // Disable memory split verification code until the fix for 6984348.
3428 // Currently it produces false negative results since it does not cover all cases.
3429 #if 0 // ifdef ASSERT
3430 visited.Reset();
3431 Node_Stack old_mems(arena, _compile->unique() >> 2);
3432 #endif
3433 for (uint i = 0; i < ideal_nodes.size(); i++) {
3434 Node* n = ideal_nodes.at(i);
3435 Node* nmem = get_map(n->_idx);
3436 assert(nmem != NULL, "sanity");
3437 if (n->is_Mem()) {
3438 #if 0 // ifdef ASSERT
3439 Node* old_mem = n->in(MemNode::Memory);
3440 if (!visited.test_set(old_mem->_idx)) {
3441 old_mems.push(old_mem, old_mem->outcnt());
3442 }
3443 #endif
3444 assert(n->in(MemNode::Memory) != nmem, "sanity");
3445 if (!n->is_Load()) {
3446 // Move memory users of a store first.
3447 move_inst_mem(n, orig_phis);
3448 }
3449 // Now update memory input
3450 igvn->hash_delete(n);
3451 n->set_req(MemNode::Memory, nmem);
3452 igvn->hash_insert(n);
3453 record_for_optimizer(n);
3454 } else {
3455 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3456 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3457 }
3458 }
3459 #if 0 // ifdef ASSERT
3460 // Verify that memory was split correctly
3461 while (old_mems.is_nonempty()) {
3462 Node* old_mem = old_mems.node();
3463 uint old_cnt = old_mems.index();
3464 old_mems.pop();
3465 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3466 }
3467 #endif
3468 }
3469
3470 #ifndef PRODUCT
3471 static const char *node_type_names[] = {
3472 "UnknownType",
3473 "JavaObject",
3474 "LocalVar",
3475 "Field",
3476 "Arraycopy"
3477 };
3478
3479 static const char *esc_names[] = {
3480 "UnknownEscape",
3481 "NoEscape",
3482 "ArgEscape",
3483 "GlobalEscape"
3484 };
3485
3486 void PointsToNode::dump(bool print_state) const {
3487 NodeType nt = node_type();
3488 tty->print("%s ", node_type_names[(int) nt]);
3489 if (print_state) {
3490 EscapeState es = escape_state();
3491 EscapeState fields_es = fields_escape_state();
3492 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3493 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3494 tty->print("NSR ");
3495 }
3496 if (is_Field()) {
3497 FieldNode* f = (FieldNode*)this;
3498 if (f->is_oop())
3499 tty->print("oop ");
3500 if (f->offset() > 0)
3501 tty->print("+%d ", f->offset());
3502 tty->print("(");
3503 for (BaseIterator i(f); i.has_next(); i.next()) {
3504 PointsToNode* b = i.get();
3505 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3506 }
3507 tty->print(" )");
3508 }
3509 tty->print("[");
3510 for (EdgeIterator i(this); i.has_next(); i.next()) {
3511 PointsToNode* e = i.get();
3512 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3513 }
3514 tty->print(" [");
3515 for (UseIterator i(this); i.has_next(); i.next()) {
3516 PointsToNode* u = i.get();
3517 bool is_base = false;
3518 if (PointsToNode::is_base_use(u)) {
3519 is_base = true;
3520 u = PointsToNode::get_use_node(u)->as_Field();
3521 }
3522 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3523 }
3524 tty->print(" ]] ");
3525 if (_node == NULL)
3526 tty->print_cr("<null>");
3527 else
3528 _node->dump();
3529 }
3530
3531 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3532 bool first = true;
3533 int ptnodes_length = ptnodes_worklist.length();
3534 for (int i = 0; i < ptnodes_length; i++) {
3535 PointsToNode *ptn = ptnodes_worklist.at(i);
3536 if (ptn == NULL || !ptn->is_JavaObject())
3537 continue;
3538 PointsToNode::EscapeState es = ptn->escape_state();
3539 if ((es != PointsToNode::NoEscape) && !Verbose) {
3540 continue;
3541 }
3542 Node* n = ptn->ideal_node();
3543 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3544 n->as_CallStaticJava()->is_boxing_method())) {
3545 if (first) {
3546 tty->cr();
3547 tty->print("======== Connection graph for ");
3548 _compile->method()->print_short_name();
3549 tty->cr();
3550 first = false;
3551 }
3552 ptn->dump();
3553 // Print all locals and fields which reference this allocation
3554 for (UseIterator j(ptn); j.has_next(); j.next()) {
3555 PointsToNode* use = j.get();
3556 if (use->is_LocalVar()) {
3557 use->dump(Verbose);
3558 } else if (Verbose) {
3559 use->dump();
3560 }
3561 }
3562 tty->cr();
3563 }
3564 }
3565 }
3566 #endif