1 /*
2 * Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/c2compiler.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/rootnode.hpp"
42 #include "utilities/macros.hpp"
43
44 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
45 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
46 _in_worklist(C->comp_arena()),
47 _next_pidx(0),
48 _collecting(true),
49 _verify(false),
50 _compile(C),
51 _igvn(igvn),
52 _node_map(C->comp_arena()) {
53 // Add unknown java object.
54 add_java_object(C->top(), PointsToNode::GlobalEscape);
55 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
56 // Add ConP(#NULL) and ConN(#NULL) nodes.
57 Node* oop_null = igvn->zerocon(T_OBJECT);
58 assert(oop_null->_idx < nodes_size(), "should be created already");
59 add_java_object(oop_null, PointsToNode::NoEscape);
60 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
61 if (UseCompressedOops) {
62 Node* noop_null = igvn->zerocon(T_NARROWOOP);
63 assert(noop_null->_idx < nodes_size(), "should be created already");
64 map_ideal_node(noop_null, null_obj);
65 }
66 _pcmp_neq = NULL; // Should be initialized
67 _pcmp_eq = NULL;
68 }
69
70 bool ConnectionGraph::has_candidates(Compile *C) {
71 // EA brings benefits only when the code has allocations and/or locks which
72 // are represented by ideal Macro nodes.
73 int cnt = C->macro_count();
74 for (int i = 0; i < cnt; i++) {
75 Node *n = C->macro_node(i);
76 if (n->is_Allocate())
77 return true;
78 if (n->is_Lock()) {
79 Node* obj = n->as_Lock()->obj_node()->uncast();
80 if (!(obj->is_Parm() || obj->is_Con()))
81 return true;
82 }
83 if (n->is_CallStaticJava() &&
84 n->as_CallStaticJava()->is_boxing_method()) {
85 return true;
86 }
87 }
88 return false;
89 }
90
91 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
92 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]);
93 ResourceMark rm;
94
95 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
96 // to create space for them in ConnectionGraph::_nodes[].
97 Node* oop_null = igvn->zerocon(T_OBJECT);
98 Node* noop_null = igvn->zerocon(T_NARROWOOP);
99 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
100 // Perform escape analysis
101 if (congraph->compute_escape()) {
102 // There are non escaping objects.
103 C->set_congraph(congraph);
104 }
105 // Cleanup.
106 if (oop_null->outcnt() == 0)
107 igvn->hash_delete(oop_null);
108 if (noop_null->outcnt() == 0)
109 igvn->hash_delete(noop_null);
110 }
111
112 bool ConnectionGraph::compute_escape() {
113 Compile* C = _compile;
114 PhaseGVN* igvn = _igvn;
115
116 // Worklists used by EA.
117 Unique_Node_List delayed_worklist;
118 GrowableArray<Node*> alloc_worklist;
119 GrowableArray<Node*> ptr_cmp_worklist;
120 GrowableArray<Node*> storestore_worklist;
121 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
122 GrowableArray<PointsToNode*> ptnodes_worklist;
123 GrowableArray<JavaObjectNode*> java_objects_worklist;
124 GrowableArray<JavaObjectNode*> non_escaped_worklist;
125 GrowableArray<FieldNode*> oop_fields_worklist;
126 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
127
128 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]);
129
130 // 1. Populate Connection Graph (CG) with PointsTo nodes.
131 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
132 // Initialize worklist
133 if (C->root() != NULL) {
134 ideal_nodes.push(C->root());
135 }
136 // Processed ideal nodes are unique on ideal_nodes list
137 // but several ideal nodes are mapped to the phantom_obj.
138 // To avoid duplicated entries on the following worklists
139 // add the phantom_obj only once to them.
140 ptnodes_worklist.append(phantom_obj);
141 java_objects_worklist.append(phantom_obj);
142 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
143 Node* n = ideal_nodes.at(next);
144 // Create PointsTo nodes and add them to Connection Graph. Called
145 // only once per ideal node since ideal_nodes is Unique_Node list.
146 add_node_to_connection_graph(n, &delayed_worklist);
147 PointsToNode* ptn = ptnode_adr(n->_idx);
148 if (ptn != NULL && ptn != phantom_obj) {
149 ptnodes_worklist.append(ptn);
150 if (ptn->is_JavaObject()) {
151 java_objects_worklist.append(ptn->as_JavaObject());
152 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
153 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
154 // Only allocations and java static calls results are interesting.
155 non_escaped_worklist.append(ptn->as_JavaObject());
156 }
157 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
158 oop_fields_worklist.append(ptn->as_Field());
159 }
160 }
161 if (n->is_MergeMem()) {
162 // Collect all MergeMem nodes to add memory slices for
163 // scalar replaceable objects in split_unique_types().
164 _mergemem_worklist.append(n->as_MergeMem());
165 } else if (OptimizePtrCompare && n->is_Cmp() &&
166 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
167 // Collect compare pointers nodes.
168 ptr_cmp_worklist.append(n);
169 } else if (n->is_MemBarStoreStore()) {
170 // Collect all MemBarStoreStore nodes so that depending on the
171 // escape status of the associated Allocate node some of them
172 // may be eliminated.
173 storestore_worklist.append(n);
174 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
175 (n->req() > MemBarNode::Precedent)) {
176 record_for_optimizer(n);
177 #ifdef ASSERT
178 } else if (n->is_AddP()) {
179 // Collect address nodes for graph verification.
180 addp_worklist.append(n);
181 #endif
182 } else if (n->is_ArrayCopy()) {
183 // Keep a list of ArrayCopy nodes so if one of its input is non
184 // escaping, we can record a unique type
185 arraycopy_worklist.append(n->as_ArrayCopy());
186 }
187 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
188 Node* m = n->fast_out(i); // Get user
189 ideal_nodes.push(m);
190 }
191 }
192 if (non_escaped_worklist.length() == 0) {
193 _collecting = false;
194 return false; // Nothing to do.
195 }
196 // Add final simple edges to graph.
197 while(delayed_worklist.size() > 0) {
198 Node* n = delayed_worklist.pop();
199 add_final_edges(n);
200 }
201 int ptnodes_length = ptnodes_worklist.length();
202
203 #ifdef ASSERT
204 if (VerifyConnectionGraph) {
205 // Verify that no new simple edges could be created and all
206 // local vars has edges.
207 _verify = true;
208 for (int next = 0; next < ptnodes_length; ++next) {
209 PointsToNode* ptn = ptnodes_worklist.at(next);
210 add_final_edges(ptn->ideal_node());
211 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
212 ptn->dump();
213 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
214 }
215 }
216 _verify = false;
217 }
218 #endif
219 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
220 // processing, calls to CI to resolve symbols (types, fields, methods)
221 // referenced in bytecode. During symbol resolution VM may throw
222 // an exception which CI cleans and converts to compilation failure.
223 if (C->failing()) return false;
224
225 // 2. Finish Graph construction by propagating references to all
226 // java objects through graph.
227 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
228 java_objects_worklist, oop_fields_worklist)) {
229 // All objects escaped or hit time or iterations limits.
230 _collecting = false;
231 return false;
232 }
233
234 // 3. Adjust scalar_replaceable state of nonescaping objects and push
235 // scalar replaceable allocations on alloc_worklist for processing
236 // in split_unique_types().
237 int non_escaped_length = non_escaped_worklist.length();
238 for (int next = 0; next < non_escaped_length; next++) {
239 JavaObjectNode* ptn = non_escaped_worklist.at(next);
240 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
241 Node* n = ptn->ideal_node();
242 if (n->is_Allocate()) {
243 n->as_Allocate()->_is_non_escaping = noescape;
244 }
245 if (n->is_CallStaticJava()) {
246 n->as_CallStaticJava()->_is_non_escaping = noescape;
247 }
248 if (noescape && ptn->scalar_replaceable()) {
249 adjust_scalar_replaceable_state(ptn);
250 if (ptn->scalar_replaceable()) {
251 alloc_worklist.append(ptn->ideal_node());
252 }
253 }
254 }
255
256 #ifdef ASSERT
257 if (VerifyConnectionGraph) {
258 // Verify that graph is complete - no new edges could be added or needed.
259 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
260 java_objects_worklist, addp_worklist);
261 }
262 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
263 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
264 null_obj->edge_count() == 0 &&
265 !null_obj->arraycopy_src() &&
266 !null_obj->arraycopy_dst(), "sanity");
267 #endif
268
269 _collecting = false;
270
271 } // TracePhase t3("connectionGraph")
272
273 // 4. Optimize ideal graph based on EA information.
274 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
275 if (has_non_escaping_obj) {
276 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
277 }
278
279 #ifndef PRODUCT
280 if (PrintEscapeAnalysis) {
281 dump(ptnodes_worklist); // Dump ConnectionGraph
282 }
283 #endif
284
285 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
286 #ifdef ASSERT
287 if (VerifyConnectionGraph) {
288 int alloc_length = alloc_worklist.length();
289 for (int next = 0; next < alloc_length; ++next) {
290 Node* n = alloc_worklist.at(next);
291 PointsToNode* ptn = ptnode_adr(n->_idx);
292 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
293 }
294 }
295 #endif
296
297 // 5. Separate memory graph for scalar replaceable allcations.
298 if (has_scalar_replaceable_candidates &&
299 C->AliasLevel() >= 3 && EliminateAllocations) {
300 // Now use the escape information to create unique types for
301 // scalar replaceable objects.
302 split_unique_types(alloc_worklist, arraycopy_worklist);
303 if (C->failing()) return false;
304 C->print_method(PHASE_AFTER_EA, 2);
305
306 #ifdef ASSERT
307 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
308 tty->print("=== No allocations eliminated for ");
309 C->method()->print_short_name();
310 if(!EliminateAllocations) {
311 tty->print(" since EliminateAllocations is off ===");
312 } else if(!has_scalar_replaceable_candidates) {
313 tty->print(" since there are no scalar replaceable candidates ===");
314 } else if(C->AliasLevel() < 3) {
315 tty->print(" since AliasLevel < 3 ===");
316 }
317 tty->cr();
318 #endif
319 }
320 return has_non_escaping_obj;
321 }
322
323 // Utility function for nodes that load an object
324 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
325 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
326 // ThreadLocal has RawPtr type.
327 const Type* t = _igvn->type(n);
328 if (t->make_ptr() != NULL) {
329 Node* adr = n->in(MemNode::Address);
330 #ifdef ASSERT
331 if (!adr->is_AddP()) {
332 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
333 } else {
334 assert((ptnode_adr(adr->_idx) == NULL ||
335 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
336 }
337 #endif
338 add_local_var_and_edge(n, PointsToNode::NoEscape,
339 adr, delayed_worklist);
340 }
341 }
342
343 // Populate Connection Graph with PointsTo nodes and create simple
344 // connection graph edges.
345 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
346 assert(!_verify, "this method should not be called for verification");
347 PhaseGVN* igvn = _igvn;
348 uint n_idx = n->_idx;
349 PointsToNode* n_ptn = ptnode_adr(n_idx);
350 if (n_ptn != NULL)
351 return; // No need to redefine PointsTo node during first iteration.
352
353 int opcode = n->Opcode();
354 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
355 if (gc_handled) {
356 return; // Ignore node if already handled by GC.
357 }
358
359 if (n->is_Call()) {
360 // Arguments to allocation and locking don't escape.
361 if (n->is_AbstractLock()) {
362 // Put Lock and Unlock nodes on IGVN worklist to process them during
363 // first IGVN optimization when escape information is still available.
364 record_for_optimizer(n);
365 } else if (n->is_Allocate()) {
366 add_call_node(n->as_Call());
367 record_for_optimizer(n);
368 } else {
369 if (n->is_CallStaticJava()) {
370 const char* name = n->as_CallStaticJava()->_name;
371 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
372 return; // Skip uncommon traps
373 }
374 // Don't mark as processed since call's arguments have to be processed.
375 delayed_worklist->push(n);
376 // Check if a call returns an object.
377 if ((n->as_Call()->returns_pointer() &&
378 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != NULL) ||
379 (n->is_CallStaticJava() &&
380 n->as_CallStaticJava()->is_boxing_method())) {
381 add_call_node(n->as_Call());
382 }
383 }
384 return;
385 }
386 // Put this check here to process call arguments since some call nodes
387 // point to phantom_obj.
388 if (n_ptn == phantom_obj || n_ptn == null_obj)
389 return; // Skip predefined nodes.
390
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 case Op_LoadN:
457 case Op_LoadPLocked: {
458 add_objload_to_connection_graph(n, delayed_worklist);
459 break;
460 }
461 case Op_Parm: {
462 map_ideal_node(n, phantom_obj);
463 break;
464 }
465 case Op_PartialSubtypeCheck: {
466 // Produces Null or notNull and is used in only in CmpP so
467 // phantom_obj could be used.
468 map_ideal_node(n, phantom_obj); // Result is unknown
469 break;
470 }
471 case Op_Phi: {
472 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
473 // ThreadLocal has RawPtr type.
474 const Type* t = n->as_Phi()->type();
475 if (t->make_ptr() != NULL) {
476 add_local_var(n, PointsToNode::NoEscape);
477 // Do not add edges during first iteration because some could be
478 // not defined yet.
479 delayed_worklist->push(n);
480 }
481 break;
482 }
483 case Op_Proj: {
484 // we are only interested in the oop result projection from a call
485 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
486 n->in(0)->as_Call()->returns_pointer()) {
487 add_local_var_and_edge(n, PointsToNode::NoEscape,
488 n->in(0), delayed_worklist);
489 }
490 break;
491 }
492 case Op_Rethrow: // Exception object escapes
493 case Op_Return: {
494 if (n->req() > TypeFunc::Parms &&
495 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
496 // Treat Return value as LocalVar with GlobalEscape escape state.
497 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
498 n->in(TypeFunc::Parms), delayed_worklist);
499 }
500 break;
501 }
502 case Op_CompareAndExchangeP:
503 case Op_CompareAndExchangeN:
504 case Op_GetAndSetP:
505 case Op_GetAndSetN: {
506 add_objload_to_connection_graph(n, delayed_worklist);
507 // fallthrough
508 }
509 case Op_StoreP:
510 case Op_StoreN:
511 case Op_StoreNKlass:
512 case Op_StorePConditional:
513 case Op_WeakCompareAndSwapP:
514 case Op_WeakCompareAndSwapN:
515 case Op_CompareAndSwapP:
516 case Op_CompareAndSwapN: {
517 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
518 break;
519 }
520 case Op_AryEq:
521 case Op_HasNegatives:
522 case Op_StrComp:
523 case Op_StrEquals:
524 case Op_StrIndexOf:
525 case Op_StrIndexOfChar:
526 case Op_StrInflatedCopy:
527 case Op_StrCompressedCopy:
528 case Op_EncodeISOArray: {
529 add_local_var(n, PointsToNode::ArgEscape);
530 delayed_worklist->push(n); // Process it later.
531 break;
532 }
533 case Op_ThreadLocal: {
534 add_java_object(n, PointsToNode::ArgEscape);
535 break;
536 }
537 default:
538 ; // Do nothing for nodes not related to EA.
539 }
540 return;
541 }
542
543 #ifdef ASSERT
544 #define ELSE_FAIL(name) \
545 /* Should not be called for not pointer type. */ \
546 n->dump(1); \
547 assert(false, name); \
548 break;
549 #else
550 #define ELSE_FAIL(name) \
551 break;
552 #endif
553
554 // Add final simple edges to graph.
555 void ConnectionGraph::add_final_edges(Node *n) {
556 PointsToNode* n_ptn = ptnode_adr(n->_idx);
557 #ifdef ASSERT
558 if (_verify && n_ptn->is_JavaObject())
559 return; // This method does not change graph for JavaObject.
560 #endif
561
562 if (n->is_Call()) {
563 process_call_arguments(n->as_Call());
564 return;
565 }
566 assert(n->is_Store() || n->is_LoadStore() ||
567 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
568 "node should be registered already");
569 int opcode = n->Opcode();
570 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
571 if (gc_handled) {
572 return; // Ignore node if already handled by GC.
573 }
574 switch (opcode) {
575 case Op_AddP: {
576 Node* base = get_addp_base(n);
577 PointsToNode* ptn_base = ptnode_adr(base->_idx);
578 assert(ptn_base != NULL, "field's base should be registered");
579 add_base(n_ptn->as_Field(), ptn_base);
580 break;
581 }
582 case Op_CastPP:
583 case Op_CheckCastPP:
584 case Op_EncodeP:
585 case Op_DecodeN:
586 case Op_EncodePKlass:
587 case Op_DecodeNKlass: {
588 add_local_var_and_edge(n, PointsToNode::NoEscape,
589 n->in(1), NULL);
590 break;
591 }
592 case Op_CMoveP: {
593 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
594 Node* in = n->in(i);
595 if (in == NULL)
596 continue; // ignore NULL
597 Node* uncast_in = in->uncast();
598 if (uncast_in->is_top() || uncast_in == n)
599 continue; // ignore top or inputs which go back this node
600 PointsToNode* ptn = ptnode_adr(in->_idx);
601 assert(ptn != NULL, "node should be registered");
602 add_edge(n_ptn, ptn);
603 }
604 break;
605 }
606 case Op_LoadP:
607 case Op_LoadN:
608 case Op_LoadPLocked: {
609 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
610 // ThreadLocal has RawPtr type.
611 const Type* t = _igvn->type(n);
612 if (t->make_ptr() != NULL) {
613 Node* adr = n->in(MemNode::Address);
614 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
615 break;
616 }
617 ELSE_FAIL("Op_LoadP");
618 }
619 case Op_Phi: {
620 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
621 // ThreadLocal has RawPtr type.
622 const Type* t = n->as_Phi()->type();
623 if (t->make_ptr() != NULL) {
624 for (uint i = 1; i < n->req(); i++) {
625 Node* in = n->in(i);
626 if (in == NULL)
627 continue; // ignore NULL
628 Node* uncast_in = in->uncast();
629 if (uncast_in->is_top() || uncast_in == n)
630 continue; // ignore top or inputs which go back this node
631 PointsToNode* ptn = ptnode_adr(in->_idx);
632 assert(ptn != NULL, "node should be registered");
633 add_edge(n_ptn, ptn);
634 }
635 break;
636 }
637 ELSE_FAIL("Op_Phi");
638 }
639 case Op_Proj: {
640 // we are only interested in the oop result projection from a call
641 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
642 n->in(0)->as_Call()->returns_pointer()) {
643 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
644 break;
645 }
646 ELSE_FAIL("Op_Proj");
647 }
648 case Op_Rethrow: // Exception object escapes
649 case Op_Return: {
650 if (n->req() > TypeFunc::Parms &&
651 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
652 // Treat Return value as LocalVar with GlobalEscape escape state.
653 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
654 n->in(TypeFunc::Parms), NULL);
655 break;
656 }
657 ELSE_FAIL("Op_Return");
658 }
659 case Op_StoreP:
660 case Op_StoreN:
661 case Op_StoreNKlass:
662 case Op_StorePConditional:
663 case Op_CompareAndExchangeP:
664 case Op_CompareAndExchangeN:
665 case Op_CompareAndSwapP:
666 case Op_CompareAndSwapN:
667 case Op_WeakCompareAndSwapP:
668 case Op_WeakCompareAndSwapN:
669 case Op_GetAndSetP:
670 case Op_GetAndSetN: {
671 if (add_final_edges_unsafe_access(n, opcode)) {
672 break;
673 }
674 ELSE_FAIL("Op_StoreP");
675 }
676 case Op_AryEq:
677 case Op_HasNegatives:
678 case Op_StrComp:
679 case Op_StrEquals:
680 case Op_StrIndexOf:
681 case Op_StrIndexOfChar:
682 case Op_StrInflatedCopy:
683 case Op_StrCompressedCopy:
684 case Op_EncodeISOArray: {
685 // char[]/byte[] arrays passed to string intrinsic do not escape but
686 // they are not scalar replaceable. Adjust escape state for them.
687 // Start from in(2) edge since in(1) is memory edge.
688 for (uint i = 2; i < n->req(); i++) {
689 Node* adr = n->in(i);
690 const Type* at = _igvn->type(adr);
691 if (!adr->is_top() && at->isa_ptr()) {
692 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
693 at->isa_ptr() != NULL, "expecting a pointer");
694 if (adr->is_AddP()) {
695 adr = get_addp_base(adr);
696 }
697 PointsToNode* ptn = ptnode_adr(adr->_idx);
698 assert(ptn != NULL, "node should be registered");
699 add_edge(n_ptn, ptn);
700 }
701 }
702 break;
703 }
704 default: {
705 // This method should be called only for EA specific nodes which may
706 // miss some edges when they were created.
707 #ifdef ASSERT
708 n->dump(1);
709 #endif
710 guarantee(false, "unknown node");
711 }
712 }
713 return;
714 }
715
716 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
717 Node* adr = n->in(MemNode::Address);
718 const Type* adr_type = _igvn->type(adr);
719 adr_type = adr_type->make_ptr();
720 if (adr_type == NULL) {
721 return; // skip dead nodes
722 }
723 if (adr_type->isa_oopptr()
724 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
725 && adr_type == TypeRawPtr::NOTNULL
726 && adr->in(AddPNode::Address)->is_Proj()
727 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
728 delayed_worklist->push(n); // Process it later.
729 #ifdef ASSERT
730 assert (adr->is_AddP(), "expecting an AddP");
731 if (adr_type == TypeRawPtr::NOTNULL) {
732 // Verify a raw address for a store captured by Initialize node.
733 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
734 assert(offs != Type::OffsetBot, "offset must be a constant");
735 }
736 #endif
737 } else {
738 // Ignore copy the displaced header to the BoxNode (OSR compilation).
739 if (adr->is_BoxLock()) {
740 return;
741 }
742 // Stored value escapes in unsafe access.
743 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
744 delayed_worklist->push(n); // Process unsafe access later.
745 return;
746 }
747 #ifdef ASSERT
748 n->dump(1);
749 assert(false, "not unsafe");
750 #endif
751 }
752 }
753
754 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
755 Node* adr = n->in(MemNode::Address);
756 const Type *adr_type = _igvn->type(adr);
757 adr_type = adr_type->make_ptr();
758 #ifdef ASSERT
759 if (adr_type == NULL) {
760 n->dump(1);
761 assert(adr_type != NULL, "dead node should not be on list");
762 return true;
763 }
764 #endif
765
766 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN ||
767 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) {
768 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
769 }
770
771 if (adr_type->isa_oopptr()
772 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
773 && adr_type == TypeRawPtr::NOTNULL
774 && adr->in(AddPNode::Address)->is_Proj()
775 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
776 // Point Address to Value
777 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
778 assert(adr_ptn != NULL &&
779 adr_ptn->as_Field()->is_oop(), "node should be registered");
780 Node* val = n->in(MemNode::ValueIn);
781 PointsToNode* ptn = ptnode_adr(val->_idx);
782 assert(ptn != NULL, "node should be registered");
783 add_edge(adr_ptn, ptn);
784 return true;
785 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
786 // Stored value escapes in unsafe access.
787 Node* val = n->in(MemNode::ValueIn);
788 PointsToNode* ptn = ptnode_adr(val->_idx);
789 assert(ptn != NULL, "node should be registered");
790 set_escape_state(ptn, PointsToNode::GlobalEscape);
791 // Add edge to object for unsafe access with offset.
792 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
793 assert(adr_ptn != NULL, "node should be registered");
794 if (adr_ptn->is_Field()) {
795 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
796 add_edge(adr_ptn, ptn);
797 }
798 return true;
799 }
800 return false;
801 }
802
803 void ConnectionGraph::add_call_node(CallNode* call) {
804 assert(call->returns_pointer(), "only for call which returns pointer");
805 uint call_idx = call->_idx;
806 if (call->is_Allocate()) {
807 Node* k = call->in(AllocateNode::KlassNode);
808 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
809 assert(kt != NULL, "TypeKlassPtr required.");
810 ciKlass* cik = kt->klass();
811 PointsToNode::EscapeState es = PointsToNode::NoEscape;
812 bool scalar_replaceable = true;
813 if (call->is_AllocateArray()) {
814 if (!cik->is_array_klass()) { // StressReflectiveCode
815 es = PointsToNode::GlobalEscape;
816 } else {
817 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
818 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
819 // Not scalar replaceable if the length is not constant or too big.
820 scalar_replaceable = false;
821 }
822 }
823 } else { // Allocate instance
824 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
825 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
826 !cik->is_instance_klass() || // StressReflectiveCode
827 !cik->as_instance_klass()->can_be_instantiated() ||
828 cik->as_instance_klass()->has_finalizer()) {
829 es = PointsToNode::GlobalEscape;
830 }
831 }
832 add_java_object(call, es);
833 PointsToNode* ptn = ptnode_adr(call_idx);
834 if (!scalar_replaceable && ptn->scalar_replaceable()) {
835 ptn->set_scalar_replaceable(false);
836 }
837 } else if (call->is_CallStaticJava()) {
838 // Call nodes could be different types:
839 //
840 // 1. CallDynamicJavaNode (what happened during call is unknown):
841 //
842 // - mapped to GlobalEscape JavaObject node if oop is returned;
843 //
844 // - all oop arguments are escaping globally;
845 //
846 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
847 //
848 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
849 //
850 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
851 // - mapped to NoEscape JavaObject node if non-escaping object allocated
852 // during call is returned;
853 // - mapped to ArgEscape LocalVar node pointed to object arguments
854 // which are returned and does not escape during call;
855 //
856 // - oop arguments escaping status is defined by bytecode analysis;
857 //
858 // For a static call, we know exactly what method is being called.
859 // Use bytecode estimator to record whether the call's return value escapes.
860 ciMethod* meth = call->as_CallJava()->method();
861 if (meth == NULL) {
862 const char* name = call->as_CallStaticJava()->_name;
863 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
864 // Returns a newly allocated unescaped object.
865 add_java_object(call, PointsToNode::NoEscape);
866 ptnode_adr(call_idx)->set_scalar_replaceable(false);
867 } else if (meth->is_boxing_method()) {
868 // Returns boxing object
869 PointsToNode::EscapeState es;
870 vmIntrinsics::ID intr = meth->intrinsic_id();
871 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
872 // It does not escape if object is always allocated.
873 es = PointsToNode::NoEscape;
874 } else {
875 // It escapes globally if object could be loaded from cache.
876 es = PointsToNode::GlobalEscape;
877 }
878 add_java_object(call, es);
879 } else {
880 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
881 call_analyzer->copy_dependencies(_compile->dependencies());
882 if (call_analyzer->is_return_allocated()) {
883 // Returns a newly allocated unescaped object, simply
884 // update dependency information.
885 // Mark it as NoEscape so that objects referenced by
886 // it's fields will be marked as NoEscape at least.
887 add_java_object(call, PointsToNode::NoEscape);
888 ptnode_adr(call_idx)->set_scalar_replaceable(false);
889 } else {
890 // Determine whether any arguments are returned.
891 const TypeTuple* d = call->tf()->domain();
892 bool ret_arg = false;
893 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
894 if (d->field_at(i)->isa_ptr() != NULL &&
895 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
896 ret_arg = true;
897 break;
898 }
899 }
900 if (ret_arg) {
901 add_local_var(call, PointsToNode::ArgEscape);
902 } else {
903 // Returns unknown object.
904 map_ideal_node(call, phantom_obj);
905 }
906 }
907 }
908 } else {
909 // An other type of call, assume the worst case:
910 // returned value is unknown and globally escapes.
911 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
912 map_ideal_node(call, phantom_obj);
913 }
914 }
915
916 void ConnectionGraph::process_call_arguments(CallNode *call) {
917 bool is_arraycopy = false;
918 switch (call->Opcode()) {
919 #ifdef ASSERT
920 case Op_Allocate:
921 case Op_AllocateArray:
922 case Op_Lock:
923 case Op_Unlock:
924 assert(false, "should be done already");
925 break;
926 #endif
927 case Op_ArrayCopy:
928 case Op_CallLeafNoFP:
929 // Most array copies are ArrayCopy nodes at this point but there
930 // are still a few direct calls to the copy subroutines (See
931 // PhaseStringOpts::copy_string())
932 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
933 call->as_CallLeaf()->is_call_to_arraycopystub();
934 // fall through
935 case Op_CallLeaf: {
936 // Stub calls, objects do not escape but they are not scale replaceable.
937 // Adjust escape state for outgoing arguments.
938 const TypeTuple * d = call->tf()->domain();
939 bool src_has_oops = false;
940 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
941 const Type* at = d->field_at(i);
942 Node *arg = call->in(i);
943 if (arg == NULL) {
944 continue;
945 }
946 const Type *aat = _igvn->type(arg);
947 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
948 continue;
949 if (arg->is_AddP()) {
950 //
951 // The inline_native_clone() case when the arraycopy stub is called
952 // after the allocation before Initialize and CheckCastPP nodes.
953 // Or normal arraycopy for object arrays case.
954 //
955 // Set AddP's base (Allocate) as not scalar replaceable since
956 // pointer to the base (with offset) is passed as argument.
957 //
958 arg = get_addp_base(arg);
959 }
960 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
961 assert(arg_ptn != NULL, "should be registered");
962 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
963 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
964 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
965 aat->isa_ptr() != NULL, "expecting an Ptr");
966 bool arg_has_oops = aat->isa_oopptr() &&
967 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
968 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
969 if (i == TypeFunc::Parms) {
970 src_has_oops = arg_has_oops;
971 }
972 //
973 // src or dst could be j.l.Object when other is basic type array:
974 //
975 // arraycopy(char[],0,Object*,0,size);
976 // arraycopy(Object*,0,char[],0,size);
977 //
978 // Don't add edges in such cases.
979 //
980 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
981 arg_has_oops && (i > TypeFunc::Parms);
982 #ifdef ASSERT
983 if (!(is_arraycopy ||
984 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
985 (call->as_CallLeaf()->_name != NULL &&
986 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
987 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
988 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
989 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
990 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
991 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
992 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
993 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
994 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
995 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
996 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
997 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
998 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
999 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
1000 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
1001 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
1002 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
1003 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
1004 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
1005 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
1006 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
1007 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
1008 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
1009 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0)
1010 ))) {
1011 call->dump();
1012 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
1013 }
1014 #endif
1015 // Always process arraycopy's destination object since
1016 // we need to add all possible edges to references in
1017 // source object.
1018 if (arg_esc >= PointsToNode::ArgEscape &&
1019 !arg_is_arraycopy_dest) {
1020 continue;
1021 }
1022 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
1023 if (call->is_ArrayCopy()) {
1024 ArrayCopyNode* ac = call->as_ArrayCopy();
1025 if (ac->is_clonebasic() ||
1026 ac->is_arraycopy_validated() ||
1027 ac->is_copyof_validated() ||
1028 ac->is_copyofrange_validated()) {
1029 es = PointsToNode::NoEscape;
1030 }
1031 }
1032 set_escape_state(arg_ptn, es);
1033 if (arg_is_arraycopy_dest) {
1034 Node* src = call->in(TypeFunc::Parms);
1035 if (src->is_AddP()) {
1036 src = get_addp_base(src);
1037 }
1038 PointsToNode* src_ptn = ptnode_adr(src->_idx);
1039 assert(src_ptn != NULL, "should be registered");
1040 if (arg_ptn != src_ptn) {
1041 // Special arraycopy edge:
1042 // A destination object's field can't have the source object
1043 // as base since objects escape states are not related.
1044 // Only escape state of destination object's fields affects
1045 // escape state of fields in source object.
1046 add_arraycopy(call, es, src_ptn, arg_ptn);
1047 }
1048 }
1049 }
1050 }
1051 break;
1052 }
1053 case Op_CallStaticJava: {
1054 // For a static call, we know exactly what method is being called.
1055 // Use bytecode estimator to record the call's escape affects
1056 #ifdef ASSERT
1057 const char* name = call->as_CallStaticJava()->_name;
1058 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1059 #endif
1060 ciMethod* meth = call->as_CallJava()->method();
1061 if ((meth != NULL) && meth->is_boxing_method()) {
1062 break; // Boxing methods do not modify any oops.
1063 }
1064 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1065 // fall-through if not a Java method or no analyzer information
1066 if (call_analyzer != NULL) {
1067 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1068 const TypeTuple* d = call->tf()->domain();
1069 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1070 const Type* at = d->field_at(i);
1071 int k = i - TypeFunc::Parms;
1072 Node* arg = call->in(i);
1073 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1074 if (at->isa_ptr() != NULL &&
1075 call_analyzer->is_arg_returned(k)) {
1076 // The call returns arguments.
1077 if (call_ptn != NULL) { // Is call's result used?
1078 assert(call_ptn->is_LocalVar(), "node should be registered");
1079 assert(arg_ptn != NULL, "node should be registered");
1080 add_edge(call_ptn, arg_ptn);
1081 }
1082 }
1083 if (at->isa_oopptr() != NULL &&
1084 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1085 if (!call_analyzer->is_arg_stack(k)) {
1086 // The argument global escapes
1087 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1088 } else {
1089 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1090 if (!call_analyzer->is_arg_local(k)) {
1091 // The argument itself doesn't escape, but any fields might
1092 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1093 }
1094 }
1095 }
1096 }
1097 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1098 // The call returns arguments.
1099 assert(call_ptn->edge_count() > 0, "sanity");
1100 if (!call_analyzer->is_return_local()) {
1101 // Returns also unknown object.
1102 add_edge(call_ptn, phantom_obj);
1103 }
1104 }
1105 break;
1106 }
1107 }
1108 default: {
1109 // Fall-through here if not a Java method or no analyzer information
1110 // or some other type of call, assume the worst case: all arguments
1111 // globally escape.
1112 const TypeTuple* d = call->tf()->domain();
1113 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1114 const Type* at = d->field_at(i);
1115 if (at->isa_oopptr() != NULL) {
1116 Node* arg = call->in(i);
1117 if (arg->is_AddP()) {
1118 arg = get_addp_base(arg);
1119 }
1120 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1121 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1122 }
1123 }
1124 }
1125 }
1126 }
1127
1128
1129 // Finish Graph construction.
1130 bool ConnectionGraph::complete_connection_graph(
1131 GrowableArray<PointsToNode*>& ptnodes_worklist,
1132 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1133 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1134 GrowableArray<FieldNode*>& oop_fields_worklist) {
1135 // Normally only 1-3 passes needed to build Connection Graph depending
1136 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1137 // Set limit to 20 to catch situation when something did go wrong and
1138 // bailout Escape Analysis.
1139 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1140 #define CG_BUILD_ITER_LIMIT 20
1141
1142 // Propagate GlobalEscape and ArgEscape escape states and check that
1143 // we still have non-escaping objects. The method pushs on _worklist
1144 // Field nodes which reference phantom_object.
1145 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1146 return false; // Nothing to do.
1147 }
1148 // Now propagate references to all JavaObject nodes.
1149 int java_objects_length = java_objects_worklist.length();
1150 elapsedTimer time;
1151 bool timeout = false;
1152 int new_edges = 1;
1153 int iterations = 0;
1154 do {
1155 while ((new_edges > 0) &&
1156 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1157 double start_time = time.seconds();
1158 time.start();
1159 new_edges = 0;
1160 // Propagate references to phantom_object for nodes pushed on _worklist
1161 // by find_non_escaped_objects() and find_field_value().
1162 new_edges += add_java_object_edges(phantom_obj, false);
1163 for (int next = 0; next < java_objects_length; ++next) {
1164 JavaObjectNode* ptn = java_objects_worklist.at(next);
1165 new_edges += add_java_object_edges(ptn, true);
1166
1167 #define SAMPLE_SIZE 4
1168 if ((next % SAMPLE_SIZE) == 0) {
1169 // Each 4 iterations calculate how much time it will take
1170 // to complete graph construction.
1171 time.stop();
1172 // Poll for requests from shutdown mechanism to quiesce compiler
1173 // because Connection graph construction may take long time.
1174 CompileBroker::maybe_block();
1175 double stop_time = time.seconds();
1176 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1177 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1178 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1179 timeout = true;
1180 break; // Timeout
1181 }
1182 start_time = stop_time;
1183 time.start();
1184 }
1185 #undef SAMPLE_SIZE
1186
1187 }
1188 if (timeout) break;
1189 if (new_edges > 0) {
1190 // Update escape states on each iteration if graph was updated.
1191 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1192 return false; // Nothing to do.
1193 }
1194 }
1195 time.stop();
1196 if (time.seconds() >= EscapeAnalysisTimeout) {
1197 timeout = true;
1198 break;
1199 }
1200 }
1201 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1202 time.start();
1203 // Find fields which have unknown value.
1204 int fields_length = oop_fields_worklist.length();
1205 for (int next = 0; next < fields_length; next++) {
1206 FieldNode* field = oop_fields_worklist.at(next);
1207 if (field->edge_count() == 0) {
1208 new_edges += find_field_value(field);
1209 // This code may added new edges to phantom_object.
1210 // Need an other cycle to propagate references to phantom_object.
1211 }
1212 }
1213 time.stop();
1214 if (time.seconds() >= EscapeAnalysisTimeout) {
1215 timeout = true;
1216 break;
1217 }
1218 } else {
1219 new_edges = 0; // Bailout
1220 }
1221 } while (new_edges > 0);
1222
1223 // Bailout if passed limits.
1224 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1225 Compile* C = _compile;
1226 if (C->log() != NULL) {
1227 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1228 C->log()->text("%s", timeout ? "time" : "iterations");
1229 C->log()->end_elem(" limit'");
1230 }
1231 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1232 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length());
1233 // Possible infinite build_connection_graph loop,
1234 // bailout (no changes to ideal graph were made).
1235 return false;
1236 }
1237 #ifdef ASSERT
1238 if (Verbose && PrintEscapeAnalysis) {
1239 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1240 iterations, nodes_size(), ptnodes_worklist.length());
1241 }
1242 #endif
1243
1244 #undef CG_BUILD_ITER_LIMIT
1245
1246 // Find fields initialized by NULL for non-escaping Allocations.
1247 int non_escaped_length = non_escaped_worklist.length();
1248 for (int next = 0; next < non_escaped_length; next++) {
1249 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1250 PointsToNode::EscapeState es = ptn->escape_state();
1251 assert(es <= PointsToNode::ArgEscape, "sanity");
1252 if (es == PointsToNode::NoEscape) {
1253 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1254 // Adding references to NULL object does not change escape states
1255 // since it does not escape. Also no fields are added to NULL object.
1256 add_java_object_edges(null_obj, false);
1257 }
1258 }
1259 Node* n = ptn->ideal_node();
1260 if (n->is_Allocate()) {
1261 // The object allocated by this Allocate node will never be
1262 // seen by an other thread. Mark it so that when it is
1263 // expanded no MemBarStoreStore is added.
1264 InitializeNode* ini = n->as_Allocate()->initialization();
1265 if (ini != NULL)
1266 ini->set_does_not_escape();
1267 }
1268 }
1269 return true; // Finished graph construction.
1270 }
1271
1272 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1273 // and check that we still have non-escaping java objects.
1274 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1275 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1276 GrowableArray<PointsToNode*> escape_worklist;
1277 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1278 int ptnodes_length = ptnodes_worklist.length();
1279 for (int next = 0; next < ptnodes_length; ++next) {
1280 PointsToNode* ptn = ptnodes_worklist.at(next);
1281 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1282 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1283 escape_worklist.push(ptn);
1284 }
1285 }
1286 // Set escape states to referenced nodes (edges list).
1287 while (escape_worklist.length() > 0) {
1288 PointsToNode* ptn = escape_worklist.pop();
1289 PointsToNode::EscapeState es = ptn->escape_state();
1290 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1291 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1292 es >= PointsToNode::ArgEscape) {
1293 // GlobalEscape or ArgEscape state of field means it has unknown value.
1294 if (add_edge(ptn, phantom_obj)) {
1295 // New edge was added
1296 add_field_uses_to_worklist(ptn->as_Field());
1297 }
1298 }
1299 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1300 PointsToNode* e = i.get();
1301 if (e->is_Arraycopy()) {
1302 assert(ptn->arraycopy_dst(), "sanity");
1303 // Propagate only fields escape state through arraycopy edge.
1304 if (e->fields_escape_state() < field_es) {
1305 set_fields_escape_state(e, field_es);
1306 escape_worklist.push(e);
1307 }
1308 } else if (es >= field_es) {
1309 // fields_escape_state is also set to 'es' if it is less than 'es'.
1310 if (e->escape_state() < es) {
1311 set_escape_state(e, es);
1312 escape_worklist.push(e);
1313 }
1314 } else {
1315 // Propagate field escape state.
1316 bool es_changed = false;
1317 if (e->fields_escape_state() < field_es) {
1318 set_fields_escape_state(e, field_es);
1319 es_changed = true;
1320 }
1321 if ((e->escape_state() < field_es) &&
1322 e->is_Field() && ptn->is_JavaObject() &&
1323 e->as_Field()->is_oop()) {
1324 // Change escape state of referenced fields.
1325 set_escape_state(e, field_es);
1326 es_changed = true;
1327 } else if (e->escape_state() < es) {
1328 set_escape_state(e, es);
1329 es_changed = true;
1330 }
1331 if (es_changed) {
1332 escape_worklist.push(e);
1333 }
1334 }
1335 }
1336 }
1337 // Remove escaped objects from non_escaped list.
1338 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1339 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1340 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1341 non_escaped_worklist.delete_at(next);
1342 }
1343 if (ptn->escape_state() == PointsToNode::NoEscape) {
1344 // Find fields in non-escaped allocations which have unknown value.
1345 find_init_values(ptn, phantom_obj, NULL);
1346 }
1347 }
1348 return (non_escaped_worklist.length() > 0);
1349 }
1350
1351 // Add all references to JavaObject node by walking over all uses.
1352 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1353 int new_edges = 0;
1354 if (populate_worklist) {
1355 // Populate _worklist by uses of jobj's uses.
1356 for (UseIterator i(jobj); i.has_next(); i.next()) {
1357 PointsToNode* use = i.get();
1358 if (use->is_Arraycopy())
1359 continue;
1360 add_uses_to_worklist(use);
1361 if (use->is_Field() && use->as_Field()->is_oop()) {
1362 // Put on worklist all field's uses (loads) and
1363 // related field nodes (same base and offset).
1364 add_field_uses_to_worklist(use->as_Field());
1365 }
1366 }
1367 }
1368 for (int l = 0; l < _worklist.length(); l++) {
1369 PointsToNode* use = _worklist.at(l);
1370 if (PointsToNode::is_base_use(use)) {
1371 // Add reference from jobj to field and from field to jobj (field's base).
1372 use = PointsToNode::get_use_node(use)->as_Field();
1373 if (add_base(use->as_Field(), jobj)) {
1374 new_edges++;
1375 }
1376 continue;
1377 }
1378 assert(!use->is_JavaObject(), "sanity");
1379 if (use->is_Arraycopy()) {
1380 if (jobj == null_obj) // NULL object does not have field edges
1381 continue;
1382 // Added edge from Arraycopy node to arraycopy's source java object
1383 if (add_edge(use, jobj)) {
1384 jobj->set_arraycopy_src();
1385 new_edges++;
1386 }
1387 // and stop here.
1388 continue;
1389 }
1390 if (!add_edge(use, jobj))
1391 continue; // No new edge added, there was such edge already.
1392 new_edges++;
1393 if (use->is_LocalVar()) {
1394 add_uses_to_worklist(use);
1395 if (use->arraycopy_dst()) {
1396 for (EdgeIterator i(use); i.has_next(); i.next()) {
1397 PointsToNode* e = i.get();
1398 if (e->is_Arraycopy()) {
1399 if (jobj == null_obj) // NULL object does not have field edges
1400 continue;
1401 // Add edge from arraycopy's destination java object to Arraycopy node.
1402 if (add_edge(jobj, e)) {
1403 new_edges++;
1404 jobj->set_arraycopy_dst();
1405 }
1406 }
1407 }
1408 }
1409 } else {
1410 // Added new edge to stored in field values.
1411 // Put on worklist all field's uses (loads) and
1412 // related field nodes (same base and offset).
1413 add_field_uses_to_worklist(use->as_Field());
1414 }
1415 }
1416 _worklist.clear();
1417 _in_worklist.Reset();
1418 return new_edges;
1419 }
1420
1421 // Put on worklist all related field nodes.
1422 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1423 assert(field->is_oop(), "sanity");
1424 int offset = field->offset();
1425 add_uses_to_worklist(field);
1426 // Loop over all bases of this field and push on worklist Field nodes
1427 // with the same offset and base (since they may reference the same field).
1428 for (BaseIterator i(field); i.has_next(); i.next()) {
1429 PointsToNode* base = i.get();
1430 add_fields_to_worklist(field, base);
1431 // Check if the base was source object of arraycopy and go over arraycopy's
1432 // destination objects since values stored to a field of source object are
1433 // accessable by uses (loads) of fields of destination objects.
1434 if (base->arraycopy_src()) {
1435 for (UseIterator j(base); j.has_next(); j.next()) {
1436 PointsToNode* arycp = j.get();
1437 if (arycp->is_Arraycopy()) {
1438 for (UseIterator k(arycp); k.has_next(); k.next()) {
1439 PointsToNode* abase = k.get();
1440 if (abase->arraycopy_dst() && abase != base) {
1441 // Look for the same arraycopy reference.
1442 add_fields_to_worklist(field, abase);
1443 }
1444 }
1445 }
1446 }
1447 }
1448 }
1449 }
1450
1451 // Put on worklist all related field nodes.
1452 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1453 int offset = field->offset();
1454 if (base->is_LocalVar()) {
1455 for (UseIterator j(base); j.has_next(); j.next()) {
1456 PointsToNode* f = j.get();
1457 if (PointsToNode::is_base_use(f)) { // Field
1458 f = PointsToNode::get_use_node(f);
1459 if (f == field || !f->as_Field()->is_oop())
1460 continue;
1461 int offs = f->as_Field()->offset();
1462 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1463 add_to_worklist(f);
1464 }
1465 }
1466 }
1467 } else {
1468 assert(base->is_JavaObject(), "sanity");
1469 if (// Skip phantom_object since it is only used to indicate that
1470 // this field's content globally escapes.
1471 (base != phantom_obj) &&
1472 // NULL object node does not have fields.
1473 (base != null_obj)) {
1474 for (EdgeIterator i(base); i.has_next(); i.next()) {
1475 PointsToNode* f = i.get();
1476 // Skip arraycopy edge since store to destination object field
1477 // does not update value in source object field.
1478 if (f->is_Arraycopy()) {
1479 assert(base->arraycopy_dst(), "sanity");
1480 continue;
1481 }
1482 if (f == field || !f->as_Field()->is_oop())
1483 continue;
1484 int offs = f->as_Field()->offset();
1485 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1486 add_to_worklist(f);
1487 }
1488 }
1489 }
1490 }
1491 }
1492
1493 // Find fields which have unknown value.
1494 int ConnectionGraph::find_field_value(FieldNode* field) {
1495 // Escaped fields should have init value already.
1496 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1497 int new_edges = 0;
1498 for (BaseIterator i(field); i.has_next(); i.next()) {
1499 PointsToNode* base = i.get();
1500 if (base->is_JavaObject()) {
1501 // Skip Allocate's fields which will be processed later.
1502 if (base->ideal_node()->is_Allocate())
1503 return 0;
1504 assert(base == null_obj, "only NULL ptr base expected here");
1505 }
1506 }
1507 if (add_edge(field, phantom_obj)) {
1508 // New edge was added
1509 new_edges++;
1510 add_field_uses_to_worklist(field);
1511 }
1512 return new_edges;
1513 }
1514
1515 // Find fields initializing values for allocations.
1516 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1517 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1518 int new_edges = 0;
1519 Node* alloc = pta->ideal_node();
1520 if (init_val == phantom_obj) {
1521 // Do nothing for Allocate nodes since its fields values are
1522 // "known" unless they are initialized by arraycopy/clone.
1523 if (alloc->is_Allocate() && !pta->arraycopy_dst())
1524 return 0;
1525 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
1526 #ifdef ASSERT
1527 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) {
1528 const char* name = alloc->as_CallStaticJava()->_name;
1529 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1530 }
1531 #endif
1532 // Non-escaped allocation returned from Java or runtime call have
1533 // unknown values in fields.
1534 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1535 PointsToNode* field = i.get();
1536 if (field->is_Field() && field->as_Field()->is_oop()) {
1537 if (add_edge(field, phantom_obj)) {
1538 // New edge was added
1539 new_edges++;
1540 add_field_uses_to_worklist(field->as_Field());
1541 }
1542 }
1543 }
1544 return new_edges;
1545 }
1546 assert(init_val == null_obj, "sanity");
1547 // Do nothing for Call nodes since its fields values are unknown.
1548 if (!alloc->is_Allocate())
1549 return 0;
1550
1551 InitializeNode* ini = alloc->as_Allocate()->initialization();
1552 bool visited_bottom_offset = false;
1553 GrowableArray<int> offsets_worklist;
1554
1555 // Check if an oop field's initializing value is recorded and add
1556 // a corresponding NULL if field's value if it is not recorded.
1557 // Connection Graph does not record a default initialization by NULL
1558 // captured by Initialize node.
1559 //
1560 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1561 PointsToNode* field = i.get(); // Field (AddP)
1562 if (!field->is_Field() || !field->as_Field()->is_oop())
1563 continue; // Not oop field
1564 int offset = field->as_Field()->offset();
1565 if (offset == Type::OffsetBot) {
1566 if (!visited_bottom_offset) {
1567 // OffsetBot is used to reference array's element,
1568 // always add reference to NULL to all Field nodes since we don't
1569 // known which element is referenced.
1570 if (add_edge(field, null_obj)) {
1571 // New edge was added
1572 new_edges++;
1573 add_field_uses_to_worklist(field->as_Field());
1574 visited_bottom_offset = true;
1575 }
1576 }
1577 } else {
1578 // Check only oop fields.
1579 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1580 if (adr_type->isa_rawptr()) {
1581 #ifdef ASSERT
1582 // Raw pointers are used for initializing stores so skip it
1583 // since it should be recorded already
1584 Node* base = get_addp_base(field->ideal_node());
1585 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1586 (base->in(0) == alloc),"unexpected pointer type");
1587 #endif
1588 continue;
1589 }
1590 if (!offsets_worklist.contains(offset)) {
1591 offsets_worklist.append(offset);
1592 Node* value = NULL;
1593 if (ini != NULL) {
1594 // StoreP::memory_type() == T_ADDRESS
1595 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1596 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1597 // Make sure initializing store has the same type as this AddP.
1598 // This AddP may reference non existing field because it is on a
1599 // dead branch of bimorphic call which is not eliminated yet.
1600 if (store != NULL && store->is_Store() &&
1601 store->as_Store()->memory_type() == ft) {
1602 value = store->in(MemNode::ValueIn);
1603 #ifdef ASSERT
1604 if (VerifyConnectionGraph) {
1605 // Verify that AddP already points to all objects the value points to.
1606 PointsToNode* val = ptnode_adr(value->_idx);
1607 assert((val != NULL), "should be processed already");
1608 PointsToNode* missed_obj = NULL;
1609 if (val->is_JavaObject()) {
1610 if (!field->points_to(val->as_JavaObject())) {
1611 missed_obj = val;
1612 }
1613 } else {
1614 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1615 tty->print_cr("----------init store has invalid value -----");
1616 store->dump();
1617 val->dump();
1618 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1619 }
1620 for (EdgeIterator j(val); j.has_next(); j.next()) {
1621 PointsToNode* obj = j.get();
1622 if (obj->is_JavaObject()) {
1623 if (!field->points_to(obj->as_JavaObject())) {
1624 missed_obj = obj;
1625 break;
1626 }
1627 }
1628 }
1629 }
1630 if (missed_obj != NULL) {
1631 tty->print_cr("----------field---------------------------------");
1632 field->dump();
1633 tty->print_cr("----------missed referernce to object-----------");
1634 missed_obj->dump();
1635 tty->print_cr("----------object referernced by init store -----");
1636 store->dump();
1637 val->dump();
1638 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1639 }
1640 }
1641 #endif
1642 } else {
1643 // There could be initializing stores which follow allocation.
1644 // For example, a volatile field store is not collected
1645 // by Initialize node.
1646 //
1647 // Need to check for dependent loads to separate such stores from
1648 // stores which follow loads. For now, add initial value NULL so
1649 // that compare pointers optimization works correctly.
1650 }
1651 }
1652 if (value == NULL) {
1653 // A field's initializing value was not recorded. Add NULL.
1654 if (add_edge(field, null_obj)) {
1655 // New edge was added
1656 new_edges++;
1657 add_field_uses_to_worklist(field->as_Field());
1658 }
1659 }
1660 }
1661 }
1662 }
1663 return new_edges;
1664 }
1665
1666 // Adjust scalar_replaceable state after Connection Graph is built.
1667 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1668 // Search for non-escaping objects which are not scalar replaceable
1669 // and mark them to propagate the state to referenced objects.
1670
1671 // 1. An object is not scalar replaceable if the field into which it is
1672 // stored has unknown offset (stored into unknown element of an array).
1673 //
1674 for (UseIterator i(jobj); i.has_next(); i.next()) {
1675 PointsToNode* use = i.get();
1676 if (use->is_Arraycopy()) {
1677 continue;
1678 }
1679 if (use->is_Field()) {
1680 FieldNode* field = use->as_Field();
1681 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
1682 if (field->offset() == Type::OffsetBot) {
1683 jobj->set_scalar_replaceable(false);
1684 return;
1685 }
1686 // 2. An object is not scalar replaceable if the field into which it is
1687 // stored has multiple bases one of which is null.
1688 if (field->base_count() > 1) {
1689 for (BaseIterator i(field); i.has_next(); i.next()) {
1690 PointsToNode* base = i.get();
1691 if (base == null_obj) {
1692 jobj->set_scalar_replaceable(false);
1693 return;
1694 }
1695 }
1696 }
1697 }
1698 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1699 // 3. An object is not scalar replaceable if it is merged with other objects.
1700 for (EdgeIterator j(use); j.has_next(); j.next()) {
1701 PointsToNode* ptn = j.get();
1702 if (ptn->is_JavaObject() && ptn != jobj) {
1703 // Mark all objects.
1704 jobj->set_scalar_replaceable(false);
1705 ptn->set_scalar_replaceable(false);
1706 }
1707 }
1708 if (!jobj->scalar_replaceable()) {
1709 return;
1710 }
1711 }
1712
1713 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1714 if (j.get()->is_Arraycopy()) {
1715 continue;
1716 }
1717
1718 // Non-escaping object node should point only to field nodes.
1719 FieldNode* field = j.get()->as_Field();
1720 int offset = field->as_Field()->offset();
1721
1722 // 4. An object is not scalar replaceable if it has a field with unknown
1723 // offset (array's element is accessed in loop).
1724 if (offset == Type::OffsetBot) {
1725 jobj->set_scalar_replaceable(false);
1726 return;
1727 }
1728 // 5. Currently an object is not scalar replaceable if a LoadStore node
1729 // access its field since the field value is unknown after it.
1730 //
1731 Node* n = field->ideal_node();
1732
1733 // Test for an unsafe access that was parsed as maybe off heap
1734 // (with a CheckCastPP to raw memory).
1735 assert(n->is_AddP(), "expect an address computation");
1736 if (n->in(AddPNode::Base)->is_top() &&
1737 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
1738 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
1739 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
1740 jobj->set_scalar_replaceable(false);
1741 return;
1742 }
1743
1744 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1745 Node* u = n->fast_out(i);
1746 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
1747 jobj->set_scalar_replaceable(false);
1748 return;
1749 }
1750 }
1751
1752 // 6. Or the address may point to more then one object. This may produce
1753 // the false positive result (set not scalar replaceable)
1754 // since the flow-insensitive escape analysis can't separate
1755 // the case when stores overwrite the field's value from the case
1756 // when stores happened on different control branches.
1757 //
1758 // Note: it will disable scalar replacement in some cases:
1759 //
1760 // Point p[] = new Point[1];
1761 // p[0] = new Point(); // Will be not scalar replaced
1762 //
1763 // but it will save us from incorrect optimizations in next cases:
1764 //
1765 // Point p[] = new Point[1];
1766 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1767 //
1768 if (field->base_count() > 1) {
1769 for (BaseIterator i(field); i.has_next(); i.next()) {
1770 PointsToNode* base = i.get();
1771 // Don't take into account LocalVar nodes which
1772 // may point to only one object which should be also
1773 // this field's base by now.
1774 if (base->is_JavaObject() && base != jobj) {
1775 // Mark all bases.
1776 jobj->set_scalar_replaceable(false);
1777 base->set_scalar_replaceable(false);
1778 }
1779 }
1780 }
1781 }
1782 }
1783
1784 #ifdef ASSERT
1785 void ConnectionGraph::verify_connection_graph(
1786 GrowableArray<PointsToNode*>& ptnodes_worklist,
1787 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1788 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1789 GrowableArray<Node*>& addp_worklist) {
1790 // Verify that graph is complete - no new edges could be added.
1791 int java_objects_length = java_objects_worklist.length();
1792 int non_escaped_length = non_escaped_worklist.length();
1793 int new_edges = 0;
1794 for (int next = 0; next < java_objects_length; ++next) {
1795 JavaObjectNode* ptn = java_objects_worklist.at(next);
1796 new_edges += add_java_object_edges(ptn, true);
1797 }
1798 assert(new_edges == 0, "graph was not complete");
1799 // Verify that escape state is final.
1800 int length = non_escaped_worklist.length();
1801 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1802 assert((non_escaped_length == non_escaped_worklist.length()) &&
1803 (non_escaped_length == length) &&
1804 (_worklist.length() == 0), "escape state was not final");
1805
1806 // Verify fields information.
1807 int addp_length = addp_worklist.length();
1808 for (int next = 0; next < addp_length; ++next ) {
1809 Node* n = addp_worklist.at(next);
1810 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1811 if (field->is_oop()) {
1812 // Verify that field has all bases
1813 Node* base = get_addp_base(n);
1814 PointsToNode* ptn = ptnode_adr(base->_idx);
1815 if (ptn->is_JavaObject()) {
1816 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1817 } else {
1818 assert(ptn->is_LocalVar(), "sanity");
1819 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1820 PointsToNode* e = i.get();
1821 if (e->is_JavaObject()) {
1822 assert(field->has_base(e->as_JavaObject()), "sanity");
1823 }
1824 }
1825 }
1826 // Verify that all fields have initializing values.
1827 if (field->edge_count() == 0) {
1828 tty->print_cr("----------field does not have references----------");
1829 field->dump();
1830 for (BaseIterator i(field); i.has_next(); i.next()) {
1831 PointsToNode* base = i.get();
1832 tty->print_cr("----------field has next base---------------------");
1833 base->dump();
1834 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1835 tty->print_cr("----------base has fields-------------------------");
1836 for (EdgeIterator j(base); j.has_next(); j.next()) {
1837 j.get()->dump();
1838 }
1839 tty->print_cr("----------base has references---------------------");
1840 for (UseIterator j(base); j.has_next(); j.next()) {
1841 j.get()->dump();
1842 }
1843 }
1844 }
1845 for (UseIterator i(field); i.has_next(); i.next()) {
1846 i.get()->dump();
1847 }
1848 assert(field->edge_count() > 0, "sanity");
1849 }
1850 }
1851 }
1852 }
1853 #endif
1854
1855 // Optimize ideal graph.
1856 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1857 GrowableArray<Node*>& storestore_worklist) {
1858 Compile* C = _compile;
1859 PhaseIterGVN* igvn = _igvn;
1860 if (EliminateLocks) {
1861 // Mark locks before changing ideal graph.
1862 int cnt = C->macro_count();
1863 for( int i=0; i < cnt; i++ ) {
1864 Node *n = C->macro_node(i);
1865 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1866 AbstractLockNode* alock = n->as_AbstractLock();
1867 if (!alock->is_non_esc_obj()) {
1868 if (not_global_escape(alock->obj_node())) {
1869 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1870 // The lock could be marked eliminated by lock coarsening
1871 // code during first IGVN before EA. Replace coarsened flag
1872 // to eliminate all associated locks/unlocks.
1873 #ifdef ASSERT
1874 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1875 #endif
1876 alock->set_non_esc_obj();
1877 }
1878 }
1879 }
1880 }
1881 }
1882
1883 if (OptimizePtrCompare) {
1884 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1885 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1886 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1887 // Optimize objects compare.
1888 while (ptr_cmp_worklist.length() != 0) {
1889 Node *n = ptr_cmp_worklist.pop();
1890 Node *res = optimize_ptr_compare(n);
1891 if (res != NULL) {
1892 #ifndef PRODUCT
1893 if (PrintOptimizePtrCompare) {
1894 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"));
1895 if (Verbose) {
1896 n->dump(1);
1897 }
1898 }
1899 #endif
1900 igvn->replace_node(n, res);
1901 }
1902 }
1903 // cleanup
1904 if (_pcmp_neq->outcnt() == 0)
1905 igvn->hash_delete(_pcmp_neq);
1906 if (_pcmp_eq->outcnt() == 0)
1907 igvn->hash_delete(_pcmp_eq);
1908 }
1909
1910 // For MemBarStoreStore nodes added in library_call.cpp, check
1911 // escape status of associated AllocateNode and optimize out
1912 // MemBarStoreStore node if the allocated object never escapes.
1913 while (storestore_worklist.length() != 0) {
1914 Node *n = storestore_worklist.pop();
1915 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1916 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1917 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1918 if (not_global_escape(alloc)) {
1919 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1920 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1921 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1922 igvn->register_new_node_with_optimizer(mb);
1923 igvn->replace_node(storestore, mb);
1924 }
1925 }
1926 }
1927
1928 // Optimize objects compare.
1929 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1930 assert(OptimizePtrCompare, "sanity");
1931 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1932 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1933 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1934 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1935 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1936 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1937
1938 // Check simple cases first.
1939 if (jobj1 != NULL) {
1940 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1941 if (jobj1 == jobj2) {
1942 // Comparing the same not escaping object.
1943 return _pcmp_eq;
1944 }
1945 Node* obj = jobj1->ideal_node();
1946 // Comparing not escaping allocation.
1947 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1948 !ptn2->points_to(jobj1)) {
1949 return _pcmp_neq; // This includes nullness check.
1950 }
1951 }
1952 }
1953 if (jobj2 != NULL) {
1954 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1955 Node* obj = jobj2->ideal_node();
1956 // Comparing not escaping allocation.
1957 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1958 !ptn1->points_to(jobj2)) {
1959 return _pcmp_neq; // This includes nullness check.
1960 }
1961 }
1962 }
1963 if (jobj1 != NULL && jobj1 != phantom_obj &&
1964 jobj2 != NULL && jobj2 != phantom_obj &&
1965 jobj1->ideal_node()->is_Con() &&
1966 jobj2->ideal_node()->is_Con()) {
1967 // Klass or String constants compare. Need to be careful with
1968 // compressed pointers - compare types of ConN and ConP instead of nodes.
1969 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1970 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1971 if (t1->make_ptr() == t2->make_ptr()) {
1972 return _pcmp_eq;
1973 } else {
1974 return _pcmp_neq;
1975 }
1976 }
1977 if (ptn1->meet(ptn2)) {
1978 return NULL; // Sets are not disjoint
1979 }
1980
1981 // Sets are disjoint.
1982 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1983 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1984 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1985 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1986 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
1987 (set2_has_unknown_ptr && set1_has_null_ptr)) {
1988 // Check nullness of unknown object.
1989 return NULL;
1990 }
1991
1992 // Disjointness by itself is not sufficient since
1993 // alias analysis is not complete for escaped objects.
1994 // Disjoint sets are definitely unrelated only when
1995 // at least one set has only not escaping allocations.
1996 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1997 if (ptn1->non_escaping_allocation()) {
1998 return _pcmp_neq;
1999 }
2000 }
2001 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2002 if (ptn2->non_escaping_allocation()) {
2003 return _pcmp_neq;
2004 }
2005 }
2006 return NULL;
2007 }
2008
2009 // Connection Graph constuction functions.
2010
2011 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
2012 PointsToNode* ptadr = _nodes.at(n->_idx);
2013 if (ptadr != NULL) {
2014 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
2015 return;
2016 }
2017 Compile* C = _compile;
2018 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
2019 _nodes.at_put(n->_idx, ptadr);
2020 }
2021
2022 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
2023 PointsToNode* ptadr = _nodes.at(n->_idx);
2024 if (ptadr != NULL) {
2025 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
2026 return;
2027 }
2028 Compile* C = _compile;
2029 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
2030 _nodes.at_put(n->_idx, ptadr);
2031 }
2032
2033 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
2034 PointsToNode* ptadr = _nodes.at(n->_idx);
2035 if (ptadr != NULL) {
2036 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
2037 return;
2038 }
2039 bool unsafe = false;
2040 bool is_oop = is_oop_field(n, offset, &unsafe);
2041 if (unsafe) {
2042 es = PointsToNode::GlobalEscape;
2043 }
2044 Compile* C = _compile;
2045 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
2046 _nodes.at_put(n->_idx, field);
2047 }
2048
2049 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
2050 PointsToNode* src, PointsToNode* dst) {
2051 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
2052 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
2053 PointsToNode* ptadr = _nodes.at(n->_idx);
2054 if (ptadr != NULL) {
2055 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
2056 return;
2057 }
2058 Compile* C = _compile;
2059 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
2060 _nodes.at_put(n->_idx, ptadr);
2061 // Add edge from arraycopy node to source object.
2062 (void)add_edge(ptadr, src);
2063 src->set_arraycopy_src();
2064 // Add edge from destination object to arraycopy node.
2065 (void)add_edge(dst, ptadr);
2066 dst->set_arraycopy_dst();
2067 }
2068
2069 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
2070 const Type* adr_type = n->as_AddP()->bottom_type();
2071 BasicType bt = T_INT;
2072 if (offset == Type::OffsetBot) {
2073 // Check only oop fields.
2074 if (!adr_type->isa_aryptr() ||
2075 (adr_type->isa_aryptr()->klass() == NULL) ||
2076 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2077 // OffsetBot is used to reference array's element. Ignore first AddP.
2078 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2079 bt = T_OBJECT;
2080 }
2081 }
2082 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2083 if (adr_type->isa_instptr()) {
2084 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
2085 if (field != NULL) {
2086 bt = field->layout_type();
2087 } else {
2088 // Check for unsafe oop field access
2089 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2090 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2091 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
2092 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
2093 bt = T_OBJECT;
2094 (*unsafe) = true;
2095 }
2096 }
2097 } else if (adr_type->isa_aryptr()) {
2098 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2099 // Ignore array length load.
2100 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2101 // Ignore first AddP.
2102 } else {
2103 const Type* elemtype = adr_type->isa_aryptr()->elem();
2104 bt = elemtype->array_element_basic_type();
2105 }
2106 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2107 // Allocation initialization, ThreadLocal field access, unsafe access
2108 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2109 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2110 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
2111 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
2112 bt = T_OBJECT;
2113 }
2114 }
2115 }
2116 // Note: T_NARROWOOP is not classed as a real reference type
2117 return (is_reference_type(bt) || bt == T_NARROWOOP);
2118 }
2119
2120 // Returns unique pointed java object or NULL.
2121 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2122 assert(!_collecting, "should not call when contructed graph");
2123 // If the node was created after the escape computation we can't answer.
2124 uint idx = n->_idx;
2125 if (idx >= nodes_size()) {
2126 return NULL;
2127 }
2128 PointsToNode* ptn = ptnode_adr(idx);
2129 if (ptn == NULL) {
2130 return NULL;
2131 }
2132 if (ptn->is_JavaObject()) {
2133 return ptn->as_JavaObject();
2134 }
2135 assert(ptn->is_LocalVar(), "sanity");
2136 // Check all java objects it points to.
2137 JavaObjectNode* jobj = NULL;
2138 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2139 PointsToNode* e = i.get();
2140 if (e->is_JavaObject()) {
2141 if (jobj == NULL) {
2142 jobj = e->as_JavaObject();
2143 } else if (jobj != e) {
2144 return NULL;
2145 }
2146 }
2147 }
2148 return jobj;
2149 }
2150
2151 // Return true if this node points only to non-escaping allocations.
2152 bool PointsToNode::non_escaping_allocation() {
2153 if (is_JavaObject()) {
2154 Node* n = ideal_node();
2155 if (n->is_Allocate() || n->is_CallStaticJava()) {
2156 return (escape_state() == PointsToNode::NoEscape);
2157 } else {
2158 return false;
2159 }
2160 }
2161 assert(is_LocalVar(), "sanity");
2162 // Check all java objects it points to.
2163 for (EdgeIterator i(this); i.has_next(); i.next()) {
2164 PointsToNode* e = i.get();
2165 if (e->is_JavaObject()) {
2166 Node* n = e->ideal_node();
2167 if ((e->escape_state() != PointsToNode::NoEscape) ||
2168 !(n->is_Allocate() || n->is_CallStaticJava())) {
2169 return false;
2170 }
2171 }
2172 }
2173 return true;
2174 }
2175
2176 // Return true if we know the node does not escape globally.
2177 bool ConnectionGraph::not_global_escape(Node *n) {
2178 assert(!_collecting, "should not call during graph construction");
2179 // If the node was created after the escape computation we can't answer.
2180 uint idx = n->_idx;
2181 if (idx >= nodes_size()) {
2182 return false;
2183 }
2184 PointsToNode* ptn = ptnode_adr(idx);
2185 if (ptn == NULL) {
2186 return false; // not in congraph (e.g. ConI)
2187 }
2188 PointsToNode::EscapeState es = ptn->escape_state();
2189 // If we have already computed a value, return it.
2190 if (es >= PointsToNode::GlobalEscape)
2191 return false;
2192 if (ptn->is_JavaObject()) {
2193 return true; // (es < PointsToNode::GlobalEscape);
2194 }
2195 assert(ptn->is_LocalVar(), "sanity");
2196 // Check all java objects it points to.
2197 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2198 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2199 return false;
2200 }
2201 return true;
2202 }
2203
2204
2205 // Helper functions
2206
2207 // Return true if this node points to specified node or nodes it points to.
2208 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2209 if (is_JavaObject()) {
2210 return (this == ptn);
2211 }
2212 assert(is_LocalVar() || is_Field(), "sanity");
2213 for (EdgeIterator i(this); i.has_next(); i.next()) {
2214 if (i.get() == ptn)
2215 return true;
2216 }
2217 return false;
2218 }
2219
2220 // Return true if one node points to an other.
2221 bool PointsToNode::meet(PointsToNode* ptn) {
2222 if (this == ptn) {
2223 return true;
2224 } else if (ptn->is_JavaObject()) {
2225 return this->points_to(ptn->as_JavaObject());
2226 } else if (this->is_JavaObject()) {
2227 return ptn->points_to(this->as_JavaObject());
2228 }
2229 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2230 int ptn_count = ptn->edge_count();
2231 for (EdgeIterator i(this); i.has_next(); i.next()) {
2232 PointsToNode* this_e = i.get();
2233 for (int j = 0; j < ptn_count; j++) {
2234 if (this_e == ptn->edge(j))
2235 return true;
2236 }
2237 }
2238 return false;
2239 }
2240
2241 #ifdef ASSERT
2242 // Return true if bases point to this java object.
2243 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2244 for (BaseIterator i(this); i.has_next(); i.next()) {
2245 if (i.get() == jobj)
2246 return true;
2247 }
2248 return false;
2249 }
2250 #endif
2251
2252 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2253 const Type *adr_type = phase->type(adr);
2254 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2255 adr->in(AddPNode::Address)->is_Proj() &&
2256 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2257 // We are computing a raw address for a store captured by an Initialize
2258 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2259 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2260 assert(offs != Type::OffsetBot ||
2261 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2262 "offset must be a constant or it is initialization of array");
2263 return offs;
2264 }
2265 const TypePtr *t_ptr = adr_type->isa_ptr();
2266 assert(t_ptr != NULL, "must be a pointer type");
2267 return t_ptr->offset();
2268 }
2269
2270 Node* ConnectionGraph::get_addp_base(Node *addp) {
2271 assert(addp->is_AddP(), "must be AddP");
2272 //
2273 // AddP cases for Base and Address inputs:
2274 // case #1. Direct object's field reference:
2275 // Allocate
2276 // |
2277 // Proj #5 ( oop result )
2278 // |
2279 // CheckCastPP (cast to instance type)
2280 // | |
2281 // AddP ( base == address )
2282 //
2283 // case #2. Indirect object's field reference:
2284 // Phi
2285 // |
2286 // CastPP (cast to instance type)
2287 // | |
2288 // AddP ( base == address )
2289 //
2290 // case #3. Raw object's field reference for Initialize node:
2291 // Allocate
2292 // |
2293 // Proj #5 ( oop result )
2294 // top |
2295 // \ |
2296 // AddP ( base == top )
2297 //
2298 // case #4. Array's element reference:
2299 // {CheckCastPP | CastPP}
2300 // | | |
2301 // | AddP ( array's element offset )
2302 // | |
2303 // AddP ( array's offset )
2304 //
2305 // case #5. Raw object's field reference for arraycopy stub call:
2306 // The inline_native_clone() case when the arraycopy stub is called
2307 // after the allocation before Initialize and CheckCastPP nodes.
2308 // Allocate
2309 // |
2310 // Proj #5 ( oop result )
2311 // | |
2312 // AddP ( base == address )
2313 //
2314 // case #6. Constant Pool, ThreadLocal, CastX2P or
2315 // Raw object's field reference:
2316 // {ConP, ThreadLocal, CastX2P, raw Load}
2317 // top |
2318 // \ |
2319 // AddP ( base == top )
2320 //
2321 // case #7. Klass's field reference.
2322 // LoadKlass
2323 // | |
2324 // AddP ( base == address )
2325 //
2326 // case #8. narrow Klass's field reference.
2327 // LoadNKlass
2328 // |
2329 // DecodeN
2330 // | |
2331 // AddP ( base == address )
2332 //
2333 // case #9. Mixed unsafe access
2334 // {instance}
2335 // |
2336 // CheckCastPP (raw)
2337 // top |
2338 // \ |
2339 // AddP ( base == top )
2340 //
2341 Node *base = addp->in(AddPNode::Base);
2342 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
2343 base = addp->in(AddPNode::Address);
2344 while (base->is_AddP()) {
2345 // Case #6 (unsafe access) may have several chained AddP nodes.
2346 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2347 base = base->in(AddPNode::Address);
2348 }
2349 if (base->Opcode() == Op_CheckCastPP &&
2350 base->bottom_type()->isa_rawptr() &&
2351 _igvn->type(base->in(1))->isa_oopptr()) {
2352 base = base->in(1); // Case #9
2353 } else {
2354 Node* uncast_base = base->uncast();
2355 int opcode = uncast_base->Opcode();
2356 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2357 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2358 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2359 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2360 }
2361 }
2362 return base;
2363 }
2364
2365 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2366 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2367 Node* addp2 = addp->raw_out(0);
2368 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2369 addp2->in(AddPNode::Base) == n &&
2370 addp2->in(AddPNode::Address) == addp) {
2371 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2372 //
2373 // Find array's offset to push it on worklist first and
2374 // as result process an array's element offset first (pushed second)
2375 // to avoid CastPP for the array's offset.
2376 // Otherwise the inserted CastPP (LocalVar) will point to what
2377 // the AddP (Field) points to. Which would be wrong since
2378 // the algorithm expects the CastPP has the same point as
2379 // as AddP's base CheckCastPP (LocalVar).
2380 //
2381 // ArrayAllocation
2382 // |
2383 // CheckCastPP
2384 // |
2385 // memProj (from ArrayAllocation CheckCastPP)
2386 // | ||
2387 // | || Int (element index)
2388 // | || | ConI (log(element size))
2389 // | || | /
2390 // | || LShift
2391 // | || /
2392 // | AddP (array's element offset)
2393 // | |
2394 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2395 // | / /
2396 // AddP (array's offset)
2397 // |
2398 // Load/Store (memory operation on array's element)
2399 //
2400 return addp2;
2401 }
2402 return NULL;
2403 }
2404
2405 //
2406 // Adjust the type and inputs of an AddP which computes the
2407 // address of a field of an instance
2408 //
2409 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2410 PhaseGVN* igvn = _igvn;
2411 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2412 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2413 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2414 if (t == NULL) {
2415 // We are computing a raw address for a store captured by an Initialize
2416 // compute an appropriate address type (cases #3 and #5).
2417 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2418 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2419 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2420 assert(offs != Type::OffsetBot, "offset must be a constant");
2421 t = base_t->add_offset(offs)->is_oopptr();
2422 }
2423 int inst_id = base_t->instance_id();
2424 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2425 "old type must be non-instance or match new type");
2426
2427 // The type 't' could be subclass of 'base_t'.
2428 // As result t->offset() could be large then base_t's size and it will
2429 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2430 // constructor verifies correctness of the offset.
2431 //
2432 // It could happened on subclass's branch (from the type profiling
2433 // inlining) which was not eliminated during parsing since the exactness
2434 // of the allocation type was not propagated to the subclass type check.
2435 //
2436 // Or the type 't' could be not related to 'base_t' at all.
2437 // It could happened when CHA type is different from MDO type on a dead path
2438 // (for example, from instanceof check) which is not collapsed during parsing.
2439 //
2440 // Do nothing for such AddP node and don't process its users since
2441 // this code branch will go away.
2442 //
2443 if (!t->is_known_instance() &&
2444 !base_t->klass()->is_subtype_of(t->klass())) {
2445 return false; // bail out
2446 }
2447 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2448 // Do NOT remove the next line: ensure a new alias index is allocated
2449 // for the instance type. Note: C++ will not remove it since the call
2450 // has side effect.
2451 int alias_idx = _compile->get_alias_index(tinst);
2452 igvn->set_type(addp, tinst);
2453 // record the allocation in the node map
2454 set_map(addp, get_map(base->_idx));
2455 // Set addp's Base and Address to 'base'.
2456 Node *abase = addp->in(AddPNode::Base);
2457 Node *adr = addp->in(AddPNode::Address);
2458 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2459 adr->in(0)->_idx == (uint)inst_id) {
2460 // Skip AddP cases #3 and #5.
2461 } else {
2462 assert(!abase->is_top(), "sanity"); // AddP case #3
2463 if (abase != base) {
2464 igvn->hash_delete(addp);
2465 addp->set_req(AddPNode::Base, base);
2466 if (abase == adr) {
2467 addp->set_req(AddPNode::Address, base);
2468 } else {
2469 // AddP case #4 (adr is array's element offset AddP node)
2470 #ifdef ASSERT
2471 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2472 assert(adr->is_AddP() && atype != NULL &&
2473 atype->instance_id() == inst_id, "array's element offset should be processed first");
2474 #endif
2475 }
2476 igvn->hash_insert(addp);
2477 }
2478 }
2479 // Put on IGVN worklist since at least addp's type was changed above.
2480 record_for_optimizer(addp);
2481 return true;
2482 }
2483
2484 //
2485 // Create a new version of orig_phi if necessary. Returns either the newly
2486 // created phi or an existing phi. Sets create_new to indicate whether a new
2487 // phi was created. Cache the last newly created phi in the node map.
2488 //
2489 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2490 Compile *C = _compile;
2491 PhaseGVN* igvn = _igvn;
2492 new_created = false;
2493 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2494 // nothing to do if orig_phi is bottom memory or matches alias_idx
2495 if (phi_alias_idx == alias_idx) {
2496 return orig_phi;
2497 }
2498 // Have we recently created a Phi for this alias index?
2499 PhiNode *result = get_map_phi(orig_phi->_idx);
2500 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2501 return result;
2502 }
2503 // Previous check may fail when the same wide memory Phi was split into Phis
2504 // for different memory slices. Search all Phis for this region.
2505 if (result != NULL) {
2506 Node* region = orig_phi->in(0);
2507 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2508 Node* phi = region->fast_out(i);
2509 if (phi->is_Phi() &&
2510 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2511 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2512 return phi->as_Phi();
2513 }
2514 }
2515 }
2516 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2517 if (C->do_escape_analysis() == true && !C->failing()) {
2518 // Retry compilation without escape analysis.
2519 // If this is the first failure, the sentinel string will "stick"
2520 // to the Compile object, and the C2Compiler will see it and retry.
2521 C->record_failure(C2Compiler::retry_no_escape_analysis());
2522 }
2523 return NULL;
2524 }
2525 orig_phi_worklist.append_if_missing(orig_phi);
2526 const TypePtr *atype = C->get_adr_type(alias_idx);
2527 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2528 C->copy_node_notes_to(result, orig_phi);
2529 igvn->set_type(result, result->bottom_type());
2530 record_for_optimizer(result);
2531 set_map(orig_phi, result);
2532 new_created = true;
2533 return result;
2534 }
2535
2536 //
2537 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2538 // specified alias index.
2539 //
2540 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2541 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2542 Compile *C = _compile;
2543 PhaseGVN* igvn = _igvn;
2544 bool new_phi_created;
2545 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2546 if (!new_phi_created) {
2547 return result;
2548 }
2549 GrowableArray<PhiNode *> phi_list;
2550 GrowableArray<uint> cur_input;
2551 PhiNode *phi = orig_phi;
2552 uint idx = 1;
2553 bool finished = false;
2554 while(!finished) {
2555 while (idx < phi->req()) {
2556 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2557 if (mem != NULL && mem->is_Phi()) {
2558 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2559 if (new_phi_created) {
2560 // found an phi for which we created a new split, push current one on worklist and begin
2561 // processing new one
2562 phi_list.push(phi);
2563 cur_input.push(idx);
2564 phi = mem->as_Phi();
2565 result = newphi;
2566 idx = 1;
2567 continue;
2568 } else {
2569 mem = newphi;
2570 }
2571 }
2572 if (C->failing()) {
2573 return NULL;
2574 }
2575 result->set_req(idx++, mem);
2576 }
2577 #ifdef ASSERT
2578 // verify that the new Phi has an input for each input of the original
2579 assert( phi->req() == result->req(), "must have same number of inputs.");
2580 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2581 #endif
2582 // Check if all new phi's inputs have specified alias index.
2583 // Otherwise use old phi.
2584 for (uint i = 1; i < phi->req(); i++) {
2585 Node* in = result->in(i);
2586 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2587 }
2588 // we have finished processing a Phi, see if there are any more to do
2589 finished = (phi_list.length() == 0 );
2590 if (!finished) {
2591 phi = phi_list.pop();
2592 idx = cur_input.pop();
2593 PhiNode *prev_result = get_map_phi(phi->_idx);
2594 prev_result->set_req(idx++, result);
2595 result = prev_result;
2596 }
2597 }
2598 return result;
2599 }
2600
2601 //
2602 // The next methods are derived from methods in MemNode.
2603 //
2604 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2605 Node *mem = mmem;
2606 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2607 // means an array I have not precisely typed yet. Do not do any
2608 // alias stuff with it any time soon.
2609 if (toop->base() != Type::AnyPtr &&
2610 !(toop->klass() != NULL &&
2611 toop->klass()->is_java_lang_Object() &&
2612 toop->offset() == Type::OffsetBot)) {
2613 mem = mmem->memory_at(alias_idx);
2614 // Update input if it is progress over what we have now
2615 }
2616 return mem;
2617 }
2618
2619 //
2620 // Move memory users to their memory slices.
2621 //
2622 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2623 Compile* C = _compile;
2624 PhaseGVN* igvn = _igvn;
2625 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2626 assert(tp != NULL, "ptr type");
2627 int alias_idx = C->get_alias_index(tp);
2628 int general_idx = C->get_general_index(alias_idx);
2629
2630 // Move users first
2631 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2632 Node* use = n->fast_out(i);
2633 if (use->is_MergeMem()) {
2634 MergeMemNode* mmem = use->as_MergeMem();
2635 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2636 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2637 continue; // Nothing to do
2638 }
2639 // Replace previous general reference to mem node.
2640 uint orig_uniq = C->unique();
2641 Node* m = find_inst_mem(n, general_idx, orig_phis);
2642 assert(orig_uniq == C->unique(), "no new nodes");
2643 mmem->set_memory_at(general_idx, m);
2644 --imax;
2645 --i;
2646 } else if (use->is_MemBar()) {
2647 assert(!use->is_Initialize(), "initializing stores should not be moved");
2648 if (use->req() > MemBarNode::Precedent &&
2649 use->in(MemBarNode::Precedent) == n) {
2650 // Don't move related membars.
2651 record_for_optimizer(use);
2652 continue;
2653 }
2654 tp = use->as_MemBar()->adr_type()->isa_ptr();
2655 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) ||
2656 alias_idx == general_idx) {
2657 continue; // Nothing to do
2658 }
2659 // Move to general memory slice.
2660 uint orig_uniq = C->unique();
2661 Node* m = find_inst_mem(n, general_idx, orig_phis);
2662 assert(orig_uniq == C->unique(), "no new nodes");
2663 igvn->hash_delete(use);
2664 imax -= use->replace_edge(n, m);
2665 igvn->hash_insert(use);
2666 record_for_optimizer(use);
2667 --i;
2668 #ifdef ASSERT
2669 } else if (use->is_Mem()) {
2670 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2671 // Don't move related cardmark.
2672 continue;
2673 }
2674 // Memory nodes should have new memory input.
2675 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2676 assert(tp != NULL, "ptr type");
2677 int idx = C->get_alias_index(tp);
2678 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2679 "Following memory nodes should have new memory input or be on the same memory slice");
2680 } else if (use->is_Phi()) {
2681 // Phi nodes should be split and moved already.
2682 tp = use->as_Phi()->adr_type()->isa_ptr();
2683 assert(tp != NULL, "ptr type");
2684 int idx = C->get_alias_index(tp);
2685 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2686 } else {
2687 use->dump();
2688 assert(false, "should not be here");
2689 #endif
2690 }
2691 }
2692 }
2693
2694 //
2695 // Search memory chain of "mem" to find a MemNode whose address
2696 // is the specified alias index.
2697 //
2698 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2699 if (orig_mem == NULL)
2700 return orig_mem;
2701 Compile* C = _compile;
2702 PhaseGVN* igvn = _igvn;
2703 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2704 bool is_instance = (toop != NULL) && toop->is_known_instance();
2705 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
2706 Node *prev = NULL;
2707 Node *result = orig_mem;
2708 while (prev != result) {
2709 prev = result;
2710 if (result == start_mem)
2711 break; // hit one of our sentinels
2712 if (result->is_Mem()) {
2713 const Type *at = igvn->type(result->in(MemNode::Address));
2714 if (at == Type::TOP)
2715 break; // Dead
2716 assert (at->isa_ptr() != NULL, "pointer type required.");
2717 int idx = C->get_alias_index(at->is_ptr());
2718 if (idx == alias_idx)
2719 break; // Found
2720 if (!is_instance && (at->isa_oopptr() == NULL ||
2721 !at->is_oopptr()->is_known_instance())) {
2722 break; // Do not skip store to general memory slice.
2723 }
2724 result = result->in(MemNode::Memory);
2725 }
2726 if (!is_instance)
2727 continue; // don't search further for non-instance types
2728 // skip over a call which does not affect this memory slice
2729 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2730 Node *proj_in = result->in(0);
2731 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2732 break; // hit one of our sentinels
2733 } else if (proj_in->is_Call()) {
2734 // ArrayCopy node processed here as well
2735 CallNode *call = proj_in->as_Call();
2736 if (!call->may_modify(toop, igvn)) {
2737 result = call->in(TypeFunc::Memory);
2738 }
2739 } else if (proj_in->is_Initialize()) {
2740 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2741 // Stop if this is the initialization for the object instance which
2742 // which contains this memory slice, otherwise skip over it.
2743 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2744 result = proj_in->in(TypeFunc::Memory);
2745 }
2746 } else if (proj_in->is_MemBar()) {
2747 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() &&
2748 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() &&
2749 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) {
2750 // clone
2751 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy();
2752 if (ac->may_modify(toop, igvn)) {
2753 break;
2754 }
2755 }
2756 result = proj_in->in(TypeFunc::Memory);
2757 }
2758 } else if (result->is_MergeMem()) {
2759 MergeMemNode *mmem = result->as_MergeMem();
2760 result = step_through_mergemem(mmem, alias_idx, toop);
2761 if (result == mmem->base_memory()) {
2762 // Didn't find instance memory, search through general slice recursively.
2763 result = mmem->memory_at(C->get_general_index(alias_idx));
2764 result = find_inst_mem(result, alias_idx, orig_phis);
2765 if (C->failing()) {
2766 return NULL;
2767 }
2768 mmem->set_memory_at(alias_idx, result);
2769 }
2770 } else if (result->is_Phi() &&
2771 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2772 Node *un = result->as_Phi()->unique_input(igvn);
2773 if (un != NULL) {
2774 orig_phis.append_if_missing(result->as_Phi());
2775 result = un;
2776 } else {
2777 break;
2778 }
2779 } else if (result->is_ClearArray()) {
2780 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2781 // Can not bypass initialization of the instance
2782 // we are looking for.
2783 break;
2784 }
2785 // Otherwise skip it (the call updated 'result' value).
2786 } else if (result->Opcode() == Op_SCMemProj) {
2787 Node* mem = result->in(0);
2788 Node* adr = NULL;
2789 if (mem->is_LoadStore()) {
2790 adr = mem->in(MemNode::Address);
2791 } else {
2792 assert(mem->Opcode() == Op_EncodeISOArray ||
2793 mem->Opcode() == Op_StrCompressedCopy, "sanity");
2794 adr = mem->in(3); // Memory edge corresponds to destination array
2795 }
2796 const Type *at = igvn->type(adr);
2797 if (at != Type::TOP) {
2798 assert(at->isa_ptr() != NULL, "pointer type required.");
2799 int idx = C->get_alias_index(at->is_ptr());
2800 if (idx == alias_idx) {
2801 // Assert in debug mode
2802 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
2803 break; // In product mode return SCMemProj node
2804 }
2805 }
2806 result = mem->in(MemNode::Memory);
2807 } else if (result->Opcode() == Op_StrInflatedCopy) {
2808 Node* adr = result->in(3); // Memory edge corresponds to destination array
2809 const Type *at = igvn->type(adr);
2810 if (at != Type::TOP) {
2811 assert(at->isa_ptr() != NULL, "pointer type required.");
2812 int idx = C->get_alias_index(at->is_ptr());
2813 if (idx == alias_idx) {
2814 // Assert in debug mode
2815 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
2816 break; // In product mode return SCMemProj node
2817 }
2818 }
2819 result = result->in(MemNode::Memory);
2820 }
2821 }
2822 if (result->is_Phi()) {
2823 PhiNode *mphi = result->as_Phi();
2824 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2825 const TypePtr *t = mphi->adr_type();
2826 if (!is_instance) {
2827 // Push all non-instance Phis on the orig_phis worklist to update inputs
2828 // during Phase 4 if needed.
2829 orig_phis.append_if_missing(mphi);
2830 } else if (C->get_alias_index(t) != alias_idx) {
2831 // Create a new Phi with the specified alias index type.
2832 result = split_memory_phi(mphi, alias_idx, orig_phis);
2833 }
2834 }
2835 // the result is either MemNode, PhiNode, InitializeNode.
2836 return result;
2837 }
2838
2839 //
2840 // Convert the types of unescaped object to instance types where possible,
2841 // propagate the new type information through the graph, and update memory
2842 // edges and MergeMem inputs to reflect the new type.
2843 //
2844 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2845 // The processing is done in 4 phases:
2846 //
2847 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2848 // types for the CheckCastPP for allocations where possible.
2849 // Propagate the new types through users as follows:
2850 // casts and Phi: push users on alloc_worklist
2851 // AddP: cast Base and Address inputs to the instance type
2852 // push any AddP users on alloc_worklist and push any memnode
2853 // users onto memnode_worklist.
2854 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2855 // search the Memory chain for a store with the appropriate type
2856 // address type. If a Phi is found, create a new version with
2857 // the appropriate memory slices from each of the Phi inputs.
2858 // For stores, process the users as follows:
2859 // MemNode: push on memnode_worklist
2860 // MergeMem: push on mergemem_worklist
2861 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2862 // moving the first node encountered of each instance type to the
2863 // the input corresponding to its alias index.
2864 // appropriate memory slice.
2865 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2866 //
2867 // In the following example, the CheckCastPP nodes are the cast of allocation
2868 // results and the allocation of node 29 is unescaped and eligible to be an
2869 // instance type.
2870 //
2871 // We start with:
2872 //
2873 // 7 Parm #memory
2874 // 10 ConI "12"
2875 // 19 CheckCastPP "Foo"
2876 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2877 // 29 CheckCastPP "Foo"
2878 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2879 //
2880 // 40 StoreP 25 7 20 ... alias_index=4
2881 // 50 StoreP 35 40 30 ... alias_index=4
2882 // 60 StoreP 45 50 20 ... alias_index=4
2883 // 70 LoadP _ 60 30 ... alias_index=4
2884 // 80 Phi 75 50 60 Memory alias_index=4
2885 // 90 LoadP _ 80 30 ... alias_index=4
2886 // 100 LoadP _ 80 20 ... alias_index=4
2887 //
2888 //
2889 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2890 // and creating a new alias index for node 30. This gives:
2891 //
2892 // 7 Parm #memory
2893 // 10 ConI "12"
2894 // 19 CheckCastPP "Foo"
2895 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2896 // 29 CheckCastPP "Foo" iid=24
2897 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2898 //
2899 // 40 StoreP 25 7 20 ... alias_index=4
2900 // 50 StoreP 35 40 30 ... alias_index=6
2901 // 60 StoreP 45 50 20 ... alias_index=4
2902 // 70 LoadP _ 60 30 ... alias_index=6
2903 // 80 Phi 75 50 60 Memory alias_index=4
2904 // 90 LoadP _ 80 30 ... alias_index=6
2905 // 100 LoadP _ 80 20 ... alias_index=4
2906 //
2907 // In phase 2, new memory inputs are computed for the loads and stores,
2908 // And a new version of the phi is created. In phase 4, the inputs to
2909 // node 80 are updated and then the memory nodes are updated with the
2910 // values computed in phase 2. This results in:
2911 //
2912 // 7 Parm #memory
2913 // 10 ConI "12"
2914 // 19 CheckCastPP "Foo"
2915 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2916 // 29 CheckCastPP "Foo" iid=24
2917 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2918 //
2919 // 40 StoreP 25 7 20 ... alias_index=4
2920 // 50 StoreP 35 7 30 ... alias_index=6
2921 // 60 StoreP 45 40 20 ... alias_index=4
2922 // 70 LoadP _ 50 30 ... alias_index=6
2923 // 80 Phi 75 40 60 Memory alias_index=4
2924 // 120 Phi 75 50 50 Memory alias_index=6
2925 // 90 LoadP _ 120 30 ... alias_index=6
2926 // 100 LoadP _ 80 20 ... alias_index=4
2927 //
2928 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) {
2929 GrowableArray<Node *> memnode_worklist;
2930 GrowableArray<PhiNode *> orig_phis;
2931 PhaseIterGVN *igvn = _igvn;
2932 uint new_index_start = (uint) _compile->num_alias_types();
2933 Arena* arena = Thread::current()->resource_area();
2934 VectorSet visited(arena);
2935 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2936 uint unique_old = _compile->unique();
2937
2938 // Phase 1: Process possible allocations from alloc_worklist.
2939 // Create instance types for the CheckCastPP for allocations where possible.
2940 //
2941 // (Note: don't forget to change the order of the second AddP node on
2942 // the alloc_worklist if the order of the worklist processing is changed,
2943 // see the comment in find_second_addp().)
2944 //
2945 while (alloc_worklist.length() != 0) {
2946 Node *n = alloc_worklist.pop();
2947 uint ni = n->_idx;
2948 if (n->is_Call()) {
2949 CallNode *alloc = n->as_Call();
2950 // copy escape information to call node
2951 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2952 PointsToNode::EscapeState es = ptn->escape_state();
2953 // We have an allocation or call which returns a Java object,
2954 // see if it is unescaped.
2955 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2956 continue;
2957 // Find CheckCastPP for the allocate or for the return value of a call
2958 n = alloc->result_cast();
2959 if (n == NULL) { // No uses except Initialize node
2960 if (alloc->is_Allocate()) {
2961 // Set the scalar_replaceable flag for allocation
2962 // so it could be eliminated if it has no uses.
2963 alloc->as_Allocate()->_is_scalar_replaceable = true;
2964 }
2965 if (alloc->is_CallStaticJava()) {
2966 // Set the scalar_replaceable flag for boxing method
2967 // so it could be eliminated if it has no uses.
2968 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2969 }
2970 continue;
2971 }
2972 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2973 assert(!alloc->is_Allocate(), "allocation should have unique type");
2974 continue;
2975 }
2976
2977 // The inline code for Object.clone() casts the allocation result to
2978 // java.lang.Object and then to the actual type of the allocated
2979 // object. Detect this case and use the second cast.
2980 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2981 // the allocation result is cast to java.lang.Object and then
2982 // to the actual Array type.
2983 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2984 && (alloc->is_AllocateArray() ||
2985 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2986 Node *cast2 = NULL;
2987 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2988 Node *use = n->fast_out(i);
2989 if (use->is_CheckCastPP()) {
2990 cast2 = use;
2991 break;
2992 }
2993 }
2994 if (cast2 != NULL) {
2995 n = cast2;
2996 } else {
2997 // Non-scalar replaceable if the allocation type is unknown statically
2998 // (reflection allocation), the object can't be restored during
2999 // deoptimization without precise type.
3000 continue;
3001 }
3002 }
3003
3004 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
3005 if (t == NULL)
3006 continue; // not a TypeOopPtr
3007 if (!t->klass_is_exact())
3008 continue; // not an unique type
3009
3010 if (alloc->is_Allocate()) {
3011 // Set the scalar_replaceable flag for allocation
3012 // so it could be eliminated.
3013 alloc->as_Allocate()->_is_scalar_replaceable = true;
3014 }
3015 if (alloc->is_CallStaticJava()) {
3016 // Set the scalar_replaceable flag for boxing method
3017 // so it could be eliminated.
3018 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3019 }
3020 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
3021 // in order for an object to be scalar-replaceable, it must be:
3022 // - a direct allocation (not a call returning an object)
3023 // - non-escaping
3024 // - eligible to be a unique type
3025 // - not determined to be ineligible by escape analysis
3026 set_map(alloc, n);
3027 set_map(n, alloc);
3028 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
3029 igvn->hash_delete(n);
3030 igvn->set_type(n, tinst);
3031 n->raise_bottom_type(tinst);
3032 igvn->hash_insert(n);
3033 record_for_optimizer(n);
3034 // Allocate an alias index for the header fields. Accesses to
3035 // the header emitted during macro expansion wouldn't have
3036 // correct memory state otherwise.
3037 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
3038 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
3039 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
3040
3041 // First, put on the worklist all Field edges from Connection Graph
3042 // which is more accurate than putting immediate users from Ideal Graph.
3043 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
3044 PointsToNode* tgt = e.get();
3045 if (tgt->is_Arraycopy()) {
3046 continue;
3047 }
3048 Node* use = tgt->ideal_node();
3049 assert(tgt->is_Field() && use->is_AddP(),
3050 "only AddP nodes are Field edges in CG");
3051 if (use->outcnt() > 0) { // Don't process dead nodes
3052 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
3053 if (addp2 != NULL) {
3054 assert(alloc->is_AllocateArray(),"array allocation was expected");
3055 alloc_worklist.append_if_missing(addp2);
3056 }
3057 alloc_worklist.append_if_missing(use);
3058 }
3059 }
3060
3061 // An allocation may have an Initialize which has raw stores. Scan
3062 // the users of the raw allocation result and push AddP users
3063 // on alloc_worklist.
3064 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
3065 assert (raw_result != NULL, "must have an allocation result");
3066 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
3067 Node *use = raw_result->fast_out(i);
3068 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
3069 Node* addp2 = find_second_addp(use, raw_result);
3070 if (addp2 != NULL) {
3071 assert(alloc->is_AllocateArray(),"array allocation was expected");
3072 alloc_worklist.append_if_missing(addp2);
3073 }
3074 alloc_worklist.append_if_missing(use);
3075 } else if (use->is_MemBar()) {
3076 memnode_worklist.append_if_missing(use);
3077 }
3078 }
3079 }
3080 } else if (n->is_AddP()) {
3081 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
3082 if (jobj == NULL || jobj == phantom_obj) {
3083 #ifdef ASSERT
3084 ptnode_adr(get_addp_base(n)->_idx)->dump();
3085 ptnode_adr(n->_idx)->dump();
3086 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3087 #endif
3088 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3089 return;
3090 }
3091 Node *base = get_map(jobj->idx()); // CheckCastPP node
3092 if (!split_AddP(n, base)) continue; // wrong type from dead path
3093 } else if (n->is_Phi() ||
3094 n->is_CheckCastPP() ||
3095 n->is_EncodeP() ||
3096 n->is_DecodeN() ||
3097 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
3098 if (visited.test_set(n->_idx)) {
3099 assert(n->is_Phi(), "loops only through Phi's");
3100 continue; // already processed
3101 }
3102 JavaObjectNode* jobj = unique_java_object(n);
3103 if (jobj == NULL || jobj == phantom_obj) {
3104 #ifdef ASSERT
3105 ptnode_adr(n->_idx)->dump();
3106 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3107 #endif
3108 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3109 return;
3110 } else {
3111 Node *val = get_map(jobj->idx()); // CheckCastPP node
3112 TypeNode *tn = n->as_Type();
3113 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
3114 assert(tinst != NULL && tinst->is_known_instance() &&
3115 tinst->instance_id() == jobj->idx() , "instance type expected.");
3116
3117 const Type *tn_type = igvn->type(tn);
3118 const TypeOopPtr *tn_t;
3119 if (tn_type->isa_narrowoop()) {
3120 tn_t = tn_type->make_ptr()->isa_oopptr();
3121 } else {
3122 tn_t = tn_type->isa_oopptr();
3123 }
3124 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
3125 if (tn_type->isa_narrowoop()) {
3126 tn_type = tinst->make_narrowoop();
3127 } else {
3128 tn_type = tinst;
3129 }
3130 igvn->hash_delete(tn);
3131 igvn->set_type(tn, tn_type);
3132 tn->set_type(tn_type);
3133 igvn->hash_insert(tn);
3134 record_for_optimizer(n);
3135 } else {
3136 assert(tn_type == TypePtr::NULL_PTR ||
3137 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3138 "unexpected type");
3139 continue; // Skip dead path with different type
3140 }
3141 }
3142 } else {
3143 debug_only(n->dump();)
3144 assert(false, "EA: unexpected node");
3145 continue;
3146 }
3147 // push allocation's users on appropriate worklist
3148 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3149 Node *use = n->fast_out(i);
3150 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3151 // Load/store to instance's field
3152 memnode_worklist.append_if_missing(use);
3153 } else if (use->is_MemBar()) {
3154 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3155 memnode_worklist.append_if_missing(use);
3156 }
3157 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3158 Node* addp2 = find_second_addp(use, n);
3159 if (addp2 != NULL) {
3160 alloc_worklist.append_if_missing(addp2);
3161 }
3162 alloc_worklist.append_if_missing(use);
3163 } else if (use->is_Phi() ||
3164 use->is_CheckCastPP() ||
3165 use->is_EncodeNarrowPtr() ||
3166 use->is_DecodeNarrowPtr() ||
3167 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3168 alloc_worklist.append_if_missing(use);
3169 #ifdef ASSERT
3170 } else if (use->is_Mem()) {
3171 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3172 } else if (use->is_MergeMem()) {
3173 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3174 } else if (use->is_SafePoint()) {
3175 // Look for MergeMem nodes for calls which reference unique allocation
3176 // (through CheckCastPP nodes) even for debug info.
3177 Node* m = use->in(TypeFunc::Memory);
3178 if (m->is_MergeMem()) {
3179 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3180 }
3181 } else if (use->Opcode() == Op_EncodeISOArray) {
3182 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3183 // EncodeISOArray overwrites destination array
3184 memnode_worklist.append_if_missing(use);
3185 }
3186 } else {
3187 uint op = use->Opcode();
3188 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
3189 (use->in(MemNode::Memory) == n)) {
3190 // They overwrite memory edge corresponding to destination array,
3191 memnode_worklist.append_if_missing(use);
3192 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
3193 op == Op_CastP2X || op == Op_StoreCM ||
3194 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3195 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3196 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
3197 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
3198 n->dump();
3199 use->dump();
3200 assert(false, "EA: missing allocation reference path");
3201 }
3202 #endif
3203 }
3204 }
3205
3206 }
3207
3208 // Go over all ArrayCopy nodes and if one of the inputs has a unique
3209 // type, record it in the ArrayCopy node so we know what memory this
3210 // node uses/modified.
3211 for (int next = 0; next < arraycopy_worklist.length(); next++) {
3212 ArrayCopyNode* ac = arraycopy_worklist.at(next);
3213 Node* dest = ac->in(ArrayCopyNode::Dest);
3214 if (dest->is_AddP()) {
3215 dest = get_addp_base(dest);
3216 }
3217 JavaObjectNode* jobj = unique_java_object(dest);
3218 if (jobj != NULL) {
3219 Node *base = get_map(jobj->idx());
3220 if (base != NULL) {
3221 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3222 ac->_dest_type = base_t;
3223 }
3224 }
3225 Node* src = ac->in(ArrayCopyNode::Src);
3226 if (src->is_AddP()) {
3227 src = get_addp_base(src);
3228 }
3229 jobj = unique_java_object(src);
3230 if (jobj != NULL) {
3231 Node* base = get_map(jobj->idx());
3232 if (base != NULL) {
3233 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3234 ac->_src_type = base_t;
3235 }
3236 }
3237 }
3238
3239 // New alias types were created in split_AddP().
3240 uint new_index_end = (uint) _compile->num_alias_types();
3241 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3242
3243 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3244 // compute new values for Memory inputs (the Memory inputs are not
3245 // actually updated until phase 4.)
3246 if (memnode_worklist.length() == 0)
3247 return; // nothing to do
3248 while (memnode_worklist.length() != 0) {
3249 Node *n = memnode_worklist.pop();
3250 if (visited.test_set(n->_idx))
3251 continue;
3252 if (n->is_Phi() || n->is_ClearArray()) {
3253 // we don't need to do anything, but the users must be pushed
3254 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3255 // we don't need to do anything, but the users must be pushed
3256 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
3257 if (n == NULL)
3258 continue;
3259 } else if (n->Opcode() == Op_StrCompressedCopy ||
3260 n->Opcode() == Op_EncodeISOArray) {
3261 // get the memory projection
3262 n = n->find_out_with(Op_SCMemProj);
3263 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3264 } else {
3265 assert(n->is_Mem(), "memory node required.");
3266 Node *addr = n->in(MemNode::Address);
3267 const Type *addr_t = igvn->type(addr);
3268 if (addr_t == Type::TOP)
3269 continue;
3270 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3271 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3272 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3273 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3274 if (_compile->failing()) {
3275 return;
3276 }
3277 if (mem != n->in(MemNode::Memory)) {
3278 // We delay the memory edge update since we need old one in
3279 // MergeMem code below when instances memory slices are separated.
3280 set_map(n, mem);
3281 }
3282 if (n->is_Load()) {
3283 continue; // don't push users
3284 } else if (n->is_LoadStore()) {
3285 // get the memory projection
3286 n = n->find_out_with(Op_SCMemProj);
3287 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3288 }
3289 }
3290 // push user on appropriate worklist
3291 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3292 Node *use = n->fast_out(i);
3293 if (use->is_Phi() || use->is_ClearArray()) {
3294 memnode_worklist.append_if_missing(use);
3295 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3296 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3297 continue;
3298 memnode_worklist.append_if_missing(use);
3299 } else if (use->is_MemBar()) {
3300 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3301 memnode_worklist.append_if_missing(use);
3302 }
3303 #ifdef ASSERT
3304 } else if(use->is_Mem()) {
3305 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3306 } else if (use->is_MergeMem()) {
3307 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3308 } else if (use->Opcode() == Op_EncodeISOArray) {
3309 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3310 // EncodeISOArray overwrites destination array
3311 memnode_worklist.append_if_missing(use);
3312 }
3313 } else {
3314 uint op = use->Opcode();
3315 if ((use->in(MemNode::Memory) == n) &&
3316 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
3317 // They overwrite memory edge corresponding to destination array,
3318 memnode_worklist.append_if_missing(use);
3319 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3320 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3321 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3322 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
3323 n->dump();
3324 use->dump();
3325 assert(false, "EA: missing memory path");
3326 }
3327 #endif
3328 }
3329 }
3330 }
3331
3332 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3333 // Walk each memory slice moving the first node encountered of each
3334 // instance type to the the input corresponding to its alias index.
3335 uint length = _mergemem_worklist.length();
3336 for( uint next = 0; next < length; ++next ) {
3337 MergeMemNode* nmm = _mergemem_worklist.at(next);
3338 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3339 // Note: we don't want to use MergeMemStream here because we only want to
3340 // scan inputs which exist at the start, not ones we add during processing.
3341 // Note 2: MergeMem may already contains instance memory slices added
3342 // during find_inst_mem() call when memory nodes were processed above.
3343 igvn->hash_delete(nmm);
3344 uint nslices = MIN2(nmm->req(), new_index_start);
3345 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3346 Node* mem = nmm->in(i);
3347 Node* cur = NULL;
3348 if (mem == NULL || mem->is_top())
3349 continue;
3350 // First, update mergemem by moving memory nodes to corresponding slices
3351 // if their type became more precise since this mergemem was created.
3352 while (mem->is_Mem()) {
3353 const Type *at = igvn->type(mem->in(MemNode::Address));
3354 if (at != Type::TOP) {
3355 assert (at->isa_ptr() != NULL, "pointer type required.");
3356 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3357 if (idx == i) {
3358 if (cur == NULL)
3359 cur = mem;
3360 } else {
3361 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3362 nmm->set_memory_at(idx, mem);
3363 }
3364 }
3365 }
3366 mem = mem->in(MemNode::Memory);
3367 }
3368 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3369 // Find any instance of the current type if we haven't encountered
3370 // already a memory slice of the instance along the memory chain.
3371 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3372 if((uint)_compile->get_general_index(ni) == i) {
3373 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3374 if (nmm->is_empty_memory(m)) {
3375 Node* result = find_inst_mem(mem, ni, orig_phis);
3376 if (_compile->failing()) {
3377 return;
3378 }
3379 nmm->set_memory_at(ni, result);
3380 }
3381 }
3382 }
3383 }
3384 // Find the rest of instances values
3385 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3386 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3387 Node* result = step_through_mergemem(nmm, ni, tinst);
3388 if (result == nmm->base_memory()) {
3389 // Didn't find instance memory, search through general slice recursively.
3390 result = nmm->memory_at(_compile->get_general_index(ni));
3391 result = find_inst_mem(result, ni, orig_phis);
3392 if (_compile->failing()) {
3393 return;
3394 }
3395 nmm->set_memory_at(ni, result);
3396 }
3397 }
3398 igvn->hash_insert(nmm);
3399 record_for_optimizer(nmm);
3400 }
3401
3402 // Phase 4: Update the inputs of non-instance memory Phis and
3403 // the Memory input of memnodes
3404 // First update the inputs of any non-instance Phi's from
3405 // which we split out an instance Phi. Note we don't have
3406 // to recursively process Phi's encounted on the input memory
3407 // chains as is done in split_memory_phi() since they will
3408 // also be processed here.
3409 for (int j = 0; j < orig_phis.length(); j++) {
3410 PhiNode *phi = orig_phis.at(j);
3411 int alias_idx = _compile->get_alias_index(phi->adr_type());
3412 igvn->hash_delete(phi);
3413 for (uint i = 1; i < phi->req(); i++) {
3414 Node *mem = phi->in(i);
3415 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3416 if (_compile->failing()) {
3417 return;
3418 }
3419 if (mem != new_mem) {
3420 phi->set_req(i, new_mem);
3421 }
3422 }
3423 igvn->hash_insert(phi);
3424 record_for_optimizer(phi);
3425 }
3426
3427 // Update the memory inputs of MemNodes with the value we computed
3428 // in Phase 2 and move stores memory users to corresponding memory slices.
3429 // Disable memory split verification code until the fix for 6984348.
3430 // Currently it produces false negative results since it does not cover all cases.
3431 #if 0 // ifdef ASSERT
3432 visited.Reset();
3433 Node_Stack old_mems(arena, _compile->unique() >> 2);
3434 #endif
3435 for (uint i = 0; i < ideal_nodes.size(); i++) {
3436 Node* n = ideal_nodes.at(i);
3437 Node* nmem = get_map(n->_idx);
3438 assert(nmem != NULL, "sanity");
3439 if (n->is_Mem()) {
3440 #if 0 // ifdef ASSERT
3441 Node* old_mem = n->in(MemNode::Memory);
3442 if (!visited.test_set(old_mem->_idx)) {
3443 old_mems.push(old_mem, old_mem->outcnt());
3444 }
3445 #endif
3446 assert(n->in(MemNode::Memory) != nmem, "sanity");
3447 if (!n->is_Load()) {
3448 // Move memory users of a store first.
3449 move_inst_mem(n, orig_phis);
3450 }
3451 // Now update memory input
3452 igvn->hash_delete(n);
3453 n->set_req(MemNode::Memory, nmem);
3454 igvn->hash_insert(n);
3455 record_for_optimizer(n);
3456 } else {
3457 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3458 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3459 }
3460 }
3461 #if 0 // ifdef ASSERT
3462 // Verify that memory was split correctly
3463 while (old_mems.is_nonempty()) {
3464 Node* old_mem = old_mems.node();
3465 uint old_cnt = old_mems.index();
3466 old_mems.pop();
3467 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3468 }
3469 #endif
3470 }
3471
3472 #ifndef PRODUCT
3473 static const char *node_type_names[] = {
3474 "UnknownType",
3475 "JavaObject",
3476 "LocalVar",
3477 "Field",
3478 "Arraycopy"
3479 };
3480
3481 static const char *esc_names[] = {
3482 "UnknownEscape",
3483 "NoEscape",
3484 "ArgEscape",
3485 "GlobalEscape"
3486 };
3487
3488 void PointsToNode::dump(bool print_state) const {
3489 NodeType nt = node_type();
3490 tty->print("%s ", node_type_names[(int) nt]);
3491 if (print_state) {
3492 EscapeState es = escape_state();
3493 EscapeState fields_es = fields_escape_state();
3494 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3495 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3496 tty->print("NSR ");
3497 }
3498 if (is_Field()) {
3499 FieldNode* f = (FieldNode*)this;
3500 if (f->is_oop())
3501 tty->print("oop ");
3502 if (f->offset() > 0)
3503 tty->print("+%d ", f->offset());
3504 tty->print("(");
3505 for (BaseIterator i(f); i.has_next(); i.next()) {
3506 PointsToNode* b = i.get();
3507 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3508 }
3509 tty->print(" )");
3510 }
3511 tty->print("[");
3512 for (EdgeIterator i(this); i.has_next(); i.next()) {
3513 PointsToNode* e = i.get();
3514 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3515 }
3516 tty->print(" [");
3517 for (UseIterator i(this); i.has_next(); i.next()) {
3518 PointsToNode* u = i.get();
3519 bool is_base = false;
3520 if (PointsToNode::is_base_use(u)) {
3521 is_base = true;
3522 u = PointsToNode::get_use_node(u)->as_Field();
3523 }
3524 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3525 }
3526 tty->print(" ]] ");
3527 if (_node == NULL)
3528 tty->print_cr("<null>");
3529 else
3530 _node->dump();
3531 }
3532
3533 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3534 bool first = true;
3535 int ptnodes_length = ptnodes_worklist.length();
3536 for (int i = 0; i < ptnodes_length; i++) {
3537 PointsToNode *ptn = ptnodes_worklist.at(i);
3538 if (ptn == NULL || !ptn->is_JavaObject())
3539 continue;
3540 PointsToNode::EscapeState es = ptn->escape_state();
3541 if ((es != PointsToNode::NoEscape) && !Verbose) {
3542 continue;
3543 }
3544 Node* n = ptn->ideal_node();
3545 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3546 n->as_CallStaticJava()->is_boxing_method())) {
3547 if (first) {
3548 tty->cr();
3549 tty->print("======== Connection graph for ");
3550 _compile->method()->print_short_name();
3551 tty->cr();
3552 first = false;
3553 }
3554 ptn->dump();
3555 // Print all locals and fields which reference this allocation
3556 for (UseIterator j(ptn); j.has_next(); j.next()) {
3557 PointsToNode* use = j.get();
3558 if (use->is_LocalVar()) {
3559 use->dump(Verbose);
3560 } else if (Verbose) {
3561 use->dump();
3562 }
3563 }
3564 tty->cr();
3565 }
3566 }
3567 }
3568 #endif
3569
3570 void ConnectionGraph::record_for_optimizer(Node *n) {
3571 _igvn->_worklist.push(n);
3572 _igvn->add_users_to_worklist(n);
3573 }