1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/c2compiler.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/escape.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/rootnode.hpp"
41 #include "utilities/macros.hpp"
42 #if INCLUDE_G1GC
43 #include "gc/g1/g1ThreadLocalData.hpp"
44 #endif // INCLUDE_G1GC
45 #if INCLUDE_ZGC
46 #include "gc/z/c2/zBarrierSetC2.hpp"
47 #endif
48
49 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
50 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
51 _in_worklist(C->comp_arena()),
52 _next_pidx(0),
53 _collecting(true),
54 _verify(false),
55 _compile(C),
56 _igvn(igvn),
57 _node_map(C->comp_arena()) {
58 // Add unknown java object.
59 add_java_object(C->top(), PointsToNode::GlobalEscape);
60 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
61 // Add ConP(#NULL) and ConN(#NULL) nodes.
62 Node* oop_null = igvn->zerocon(T_OBJECT);
63 assert(oop_null->_idx < nodes_size(), "should be created already");
64 add_java_object(oop_null, PointsToNode::NoEscape);
65 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
66 if (UseCompressedOops) {
67 Node* noop_null = igvn->zerocon(T_NARROWOOP);
68 assert(noop_null->_idx < nodes_size(), "should be created already");
69 map_ideal_node(noop_null, null_obj);
70 }
71 _pcmp_neq = NULL; // Should be initialized
72 _pcmp_eq = NULL;
73 }
74
75 bool ConnectionGraph::has_candidates(Compile *C) {
76 // EA brings benefits only when the code has allocations and/or locks which
77 // are represented by ideal Macro nodes.
78 int cnt = C->macro_count();
79 for (int i = 0; i < cnt; i++) {
80 Node *n = C->macro_node(i);
81 if (n->is_Allocate())
82 return true;
83 if (n->is_Lock()) {
84 Node* obj = n->as_Lock()->obj_node()->uncast();
85 if (!(obj->is_Parm() || obj->is_Con()))
86 return true;
87 }
88 if (n->is_CallStaticJava() &&
89 n->as_CallStaticJava()->is_boxing_method()) {
90 return true;
91 }
92 }
93 return false;
94 }
95
96 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
97 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]);
98 ResourceMark rm;
99
100 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
101 // to create space for them in ConnectionGraph::_nodes[].
102 Node* oop_null = igvn->zerocon(T_OBJECT);
103 Node* noop_null = igvn->zerocon(T_NARROWOOP);
104 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
105 // Perform escape analysis
106 if (congraph->compute_escape()) {
107 // There are non escaping objects.
108 C->set_congraph(congraph);
109 }
110 // Cleanup.
111 if (oop_null->outcnt() == 0)
112 igvn->hash_delete(oop_null);
113 if (noop_null->outcnt() == 0)
114 igvn->hash_delete(noop_null);
115 }
116
117 bool ConnectionGraph::compute_escape() {
118 Compile* C = _compile;
119 PhaseGVN* igvn = _igvn;
120
121 // Worklists used by EA.
122 Unique_Node_List delayed_worklist;
123 GrowableArray<Node*> alloc_worklist;
124 GrowableArray<Node*> ptr_cmp_worklist;
125 GrowableArray<Node*> storestore_worklist;
126 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
127 GrowableArray<PointsToNode*> ptnodes_worklist;
128 GrowableArray<JavaObjectNode*> java_objects_worklist;
129 GrowableArray<JavaObjectNode*> non_escaped_worklist;
130 GrowableArray<FieldNode*> oop_fields_worklist;
131 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
132
133 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]);
134
135 // 1. Populate Connection Graph (CG) with PointsTo nodes.
136 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
137 // Initialize worklist
138 if (C->root() != NULL) {
139 ideal_nodes.push(C->root());
140 }
141 // Processed ideal nodes are unique on ideal_nodes list
142 // but several ideal nodes are mapped to the phantom_obj.
143 // To avoid duplicated entries on the following worklists
144 // add the phantom_obj only once to them.
145 ptnodes_worklist.append(phantom_obj);
146 java_objects_worklist.append(phantom_obj);
147 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
148 Node* n = ideal_nodes.at(next);
149 // Create PointsTo nodes and add them to Connection Graph. Called
150 // only once per ideal node since ideal_nodes is Unique_Node list.
151 add_node_to_connection_graph(n, &delayed_worklist);
152 PointsToNode* ptn = ptnode_adr(n->_idx);
153 if (ptn != NULL && ptn != phantom_obj) {
154 ptnodes_worklist.append(ptn);
155 if (ptn->is_JavaObject()) {
156 java_objects_worklist.append(ptn->as_JavaObject());
157 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
158 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
159 // Only allocations and java static calls results are interesting.
160 non_escaped_worklist.append(ptn->as_JavaObject());
161 }
162 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
163 oop_fields_worklist.append(ptn->as_Field());
164 }
165 }
166 if (n->is_MergeMem()) {
167 // Collect all MergeMem nodes to add memory slices for
168 // scalar replaceable objects in split_unique_types().
169 _mergemem_worklist.append(n->as_MergeMem());
170 } else if (OptimizePtrCompare && n->is_Cmp() &&
171 ((n->Opcode() == Op_CmpP && !(((CmpPNode*)n)->has_perturbed_operand() != NULL)) ||
172 n->Opcode() == Op_CmpN)) {
173 // Collect compare pointers nodes.
174 ptr_cmp_worklist.append(n);
175 } else if (n->is_MemBarStoreStore()) {
176 // Collect all MemBarStoreStore nodes so that depending on the
177 // escape status of the associated Allocate node some of them
178 // may be eliminated.
179 storestore_worklist.append(n);
180 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
181 (n->req() > MemBarNode::Precedent)) {
182 record_for_optimizer(n);
183 #ifdef ASSERT
184 } else if (n->is_AddP()) {
185 // Collect address nodes for graph verification.
186 addp_worklist.append(n);
187 #endif
188 } else if (n->is_ArrayCopy()) {
189 // Keep a list of ArrayCopy nodes so if one of its input is non
190 // escaping, we can record a unique type
191 arraycopy_worklist.append(n->as_ArrayCopy());
192 }
193 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
194 Node* m = n->fast_out(i); // Get user
195 ideal_nodes.push(m);
196 }
197 }
198 if (non_escaped_worklist.length() == 0) {
199 _collecting = false;
200 return false; // Nothing to do.
201 }
202 // Add final simple edges to graph.
203 while(delayed_worklist.size() > 0) {
204 Node* n = delayed_worklist.pop();
205 add_final_edges(n);
206 }
207 int ptnodes_length = ptnodes_worklist.length();
208
209 #ifdef ASSERT
210 if (VerifyConnectionGraph) {
211 // Verify that no new simple edges could be created and all
212 // local vars has edges.
213 _verify = true;
214 for (int next = 0; next < ptnodes_length; ++next) {
215 PointsToNode* ptn = ptnodes_worklist.at(next);
216 add_final_edges(ptn->ideal_node());
217 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
218 ptn->dump();
219 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
220 }
221 }
222 _verify = false;
223 }
224 #endif
225 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
226 // processing, calls to CI to resolve symbols (types, fields, methods)
227 // referenced in bytecode. During symbol resolution VM may throw
228 // an exception which CI cleans and converts to compilation failure.
229 if (C->failing()) return false;
230
231 // 2. Finish Graph construction by propagating references to all
232 // java objects through graph.
233 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
234 java_objects_worklist, oop_fields_worklist)) {
235 // All objects escaped or hit time or iterations limits.
236 _collecting = false;
237 return false;
238 }
239
240 // 3. Adjust scalar_replaceable state of nonescaping objects and push
241 // scalar replaceable allocations on alloc_worklist for processing
242 // in split_unique_types().
243 int non_escaped_length = non_escaped_worklist.length();
244 for (int next = 0; next < non_escaped_length; next++) {
245 JavaObjectNode* ptn = non_escaped_worklist.at(next);
246 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
247 Node* n = ptn->ideal_node();
248 if (n->is_Allocate()) {
249 n->as_Allocate()->_is_non_escaping = noescape;
250 }
251 if (n->is_CallStaticJava()) {
252 n->as_CallStaticJava()->_is_non_escaping = noescape;
253 }
254 if (noescape && ptn->scalar_replaceable()) {
255 adjust_scalar_replaceable_state(ptn);
256 if (ptn->scalar_replaceable()) {
257 alloc_worklist.append(ptn->ideal_node());
258 }
259 }
260 }
261
262 #ifdef ASSERT
263 if (VerifyConnectionGraph) {
264 // Verify that graph is complete - no new edges could be added or needed.
265 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
266 java_objects_worklist, addp_worklist);
267 }
268 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
269 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
270 null_obj->edge_count() == 0 &&
271 !null_obj->arraycopy_src() &&
272 !null_obj->arraycopy_dst(), "sanity");
273 #endif
274
275 _collecting = false;
276
277 } // TracePhase t3("connectionGraph")
278
279 // 4. Optimize ideal graph based on EA information.
280 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
281 if (has_non_escaping_obj) {
282 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
283 }
284
285 #ifndef PRODUCT
286 if (PrintEscapeAnalysis) {
287 dump(ptnodes_worklist); // Dump ConnectionGraph
288 }
289 #endif
290
291 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
292 #ifdef ASSERT
293 if (VerifyConnectionGraph) {
294 int alloc_length = alloc_worklist.length();
295 for (int next = 0; next < alloc_length; ++next) {
296 Node* n = alloc_worklist.at(next);
297 PointsToNode* ptn = ptnode_adr(n->_idx);
298 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
299 }
300 }
301 #endif
302
303 // 5. Separate memory graph for scalar replaceable allcations.
304 if (has_scalar_replaceable_candidates &&
305 C->AliasLevel() >= 3 && EliminateAllocations) {
306 // Now use the escape information to create unique types for
307 // scalar replaceable objects.
308 split_unique_types(alloc_worklist, arraycopy_worklist);
309 if (C->failing()) return false;
310 C->print_method(PHASE_AFTER_EA, 2);
311
312 #ifdef ASSERT
313 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
314 tty->print("=== No allocations eliminated for ");
315 C->method()->print_short_name();
316 if(!EliminateAllocations) {
317 tty->print(" since EliminateAllocations is off ===");
318 } else if(!has_scalar_replaceable_candidates) {
319 tty->print(" since there are no scalar replaceable candidates ===");
320 } else if(C->AliasLevel() < 3) {
321 tty->print(" since AliasLevel < 3 ===");
322 }
323 tty->cr();
324 #endif
325 }
326 return has_non_escaping_obj;
327 }
328
329 // Utility function for nodes that load an object
330 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
331 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
332 // ThreadLocal has RawPtr type.
333 const Type* t = _igvn->type(n);
334 if (t->make_ptr() != NULL) {
335 Node* adr = n->in(MemNode::Address);
336 #ifdef ASSERT
337 if (!adr->is_AddP()) {
338 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
339 } else {
340 assert((ptnode_adr(adr->_idx) == NULL ||
341 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
342 }
343 #endif
344 add_local_var_and_edge(n, PointsToNode::NoEscape,
345 adr, delayed_worklist);
346 }
347 }
348
349 // Populate Connection Graph with PointsTo nodes and create simple
350 // connection graph edges.
351 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
352 assert(!_verify, "this method should not be called for verification");
353 PhaseGVN* igvn = _igvn;
354 uint n_idx = n->_idx;
355 PointsToNode* n_ptn = ptnode_adr(n_idx);
356 if (n_ptn != NULL)
357 return; // No need to redefine PointsTo node during first iteration.
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 } else if (n->as_Call()->tf()->returns_value_type_as_fields()) {
383 bool returns_oop = false;
384 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
385 ProjNode* pn = n->fast_out(i)->as_Proj();
386 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
387 returns_oop = true;
388 }
389 }
390 if (returns_oop) {
391 add_call_node(n->as_Call());
392 }
393 }
394 }
395 return;
396 }
397 // Put this check here to process call arguments since some call nodes
398 // point to phantom_obj.
399 if (n_ptn == phantom_obj || n_ptn == null_obj)
400 return; // Skip predefined nodes.
401
402 int opcode = n->Opcode();
403 switch (opcode) {
404 case Op_AddP: {
405 Node* base = get_addp_base(n);
406 PointsToNode* ptn_base = ptnode_adr(base->_idx);
407 // Field nodes are created for all field types. They are used in
408 // adjust_scalar_replaceable_state() and split_unique_types().
409 // Note, non-oop fields will have only base edges in Connection
410 // Graph because such fields are not used for oop loads and stores.
411 int offset = address_offset(n, igvn);
412 add_field(n, PointsToNode::NoEscape, offset);
413 if (ptn_base == NULL) {
414 delayed_worklist->push(n); // Process it later.
415 } else {
416 n_ptn = ptnode_adr(n_idx);
417 add_base(n_ptn->as_Field(), ptn_base);
418 }
419 break;
420 }
421 case Op_CastX2P: {
422 map_ideal_node(n, phantom_obj);
423 break;
424 }
425 case Op_CastPP:
426 case Op_CheckCastPP:
427 case Op_EncodeP:
428 case Op_DecodeN:
429 case Op_EncodePKlass:
430 case Op_DecodeNKlass: {
431 add_local_var_and_edge(n, PointsToNode::NoEscape,
432 n->in(1), delayed_worklist);
433 break;
434 }
435 case Op_CMoveP: {
436 add_local_var(n, PointsToNode::NoEscape);
437 // Do not add edges during first iteration because some could be
438 // not defined yet.
439 delayed_worklist->push(n);
440 break;
441 }
442 case Op_ConP:
443 case Op_ConN:
444 case Op_ConNKlass: {
445 // assume all oop constants globally escape except for null
446 PointsToNode::EscapeState es;
447 const Type* t = igvn->type(n);
448 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
449 es = PointsToNode::NoEscape;
450 } else {
451 es = PointsToNode::GlobalEscape;
452 }
453 add_java_object(n, es);
454 break;
455 }
456 case Op_CreateEx: {
457 // assume that all exception objects globally escape
458 map_ideal_node(n, phantom_obj);
459 break;
460 }
461 case Op_LoadKlass:
462 case Op_LoadNKlass: {
463 // Unknown class is loaded
464 map_ideal_node(n, phantom_obj);
465 break;
466 }
467 case Op_LoadP:
468 #if INCLUDE_ZGC
469 case Op_LoadBarrierSlowReg:
470 case Op_LoadBarrierWeakSlowReg:
471 #endif
472 case Op_LoadN:
473 case Op_LoadPLocked: {
474 add_objload_to_connection_graph(n, delayed_worklist);
475 break;
476 }
477 case Op_Parm: {
478 map_ideal_node(n, phantom_obj);
479 break;
480 }
481 case Op_PartialSubtypeCheck: {
482 // Produces Null or notNull and is used in only in CmpP so
483 // phantom_obj could be used.
484 map_ideal_node(n, phantom_obj); // Result is unknown
485 break;
486 }
487 case Op_Phi: {
488 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
489 // ThreadLocal has RawPtr type.
490 const Type* t = n->as_Phi()->type();
491 if (t->make_ptr() != NULL) {
492 add_local_var(n, PointsToNode::NoEscape);
493 // Do not add edges during first iteration because some could be
494 // not defined yet.
495 delayed_worklist->push(n);
496 }
497 break;
498 }
499 case Op_Proj: {
500 // we are only interested in the oop result projection from a call
501 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
502 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) {
503 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
504 n->in(0)->as_Call()->tf()->returns_value_type_as_fields(), "what kind of oop return is it?");
505 add_local_var_and_edge(n, PointsToNode::NoEscape,
506 n->in(0), delayed_worklist);
507 }
508 #if INCLUDE_ZGC
509 else if (UseZGC) {
510 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) {
511 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), delayed_worklist);
512 }
513 }
514 #endif
515 break;
516 }
517 case Op_Rethrow: // Exception object escapes
518 case Op_Return: {
519 if (n->req() > TypeFunc::Parms &&
520 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
521 // Treat Return value as LocalVar with GlobalEscape escape state.
522 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
523 n->in(TypeFunc::Parms), delayed_worklist);
524 }
525 break;
526 }
527 case Op_CompareAndExchangeP:
528 case Op_CompareAndExchangeN:
529 case Op_GetAndSetP:
530 case Op_GetAndSetN: {
531 add_objload_to_connection_graph(n, delayed_worklist);
532 // fallthrough
533 }
534 case Op_StoreP:
535 case Op_StoreN:
536 case Op_StoreNKlass:
537 case Op_StorePConditional:
538 case Op_WeakCompareAndSwapP:
539 case Op_WeakCompareAndSwapN:
540 case Op_CompareAndSwapP:
541 case Op_CompareAndSwapN: {
542 Node* adr = n->in(MemNode::Address);
543 const Type *adr_type = igvn->type(adr);
544 adr_type = adr_type->make_ptr();
545 if (adr_type == NULL) {
546 break; // skip dead nodes
547 }
548 if ( adr_type->isa_oopptr()
549 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
550 && adr_type == TypeRawPtr::NOTNULL
551 && adr->in(AddPNode::Address)->is_Proj()
552 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
553 delayed_worklist->push(n); // Process it later.
554 #ifdef ASSERT
555 assert(adr->is_AddP(), "expecting an AddP");
556 if (adr_type == TypeRawPtr::NOTNULL) {
557 // Verify a raw address for a store captured by Initialize node.
558 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
559 assert(offs != Type::OffsetBot, "offset must be a constant");
560 }
561 #endif
562 } else {
563 // Ignore copy the displaced header to the BoxNode (OSR compilation).
564 if (adr->is_BoxLock())
565 break;
566 // Stored value escapes in unsafe access.
567 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
568 // Pointer stores in G1 barriers looks like unsafe access.
569 // Ignore such stores to be able scalar replace non-escaping
570 // allocations.
571 #if INCLUDE_G1GC
572 if (UseG1GC && adr->is_AddP()) {
573 Node* base = get_addp_base(adr);
574 if (base->Opcode() == Op_LoadP &&
575 base->in(MemNode::Address)->is_AddP()) {
576 adr = base->in(MemNode::Address);
577 Node* tls = get_addp_base(adr);
578 if (tls->Opcode() == Op_ThreadLocal) {
579 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
580 if (offs == in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset())) {
581 break; // G1 pre barrier previous oop value store.
582 }
583 if (offs == in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset())) {
584 break; // G1 post barrier card address store.
585 }
586 }
587 }
588 }
589 #endif
590 delayed_worklist->push(n); // Process unsafe access later.
591 break;
592 }
593 #ifdef ASSERT
594 n->dump(1);
595 assert(false, "not unsafe or G1 barrier raw StoreP");
596 #endif
597 }
598 break;
599 }
600 case Op_AryEq:
601 case Op_HasNegatives:
602 case Op_StrComp:
603 case Op_StrEquals:
604 case Op_StrIndexOf:
605 case Op_StrIndexOfChar:
606 case Op_StrInflatedCopy:
607 case Op_StrCompressedCopy:
608 case Op_EncodeISOArray: {
609 add_local_var(n, PointsToNode::ArgEscape);
610 delayed_worklist->push(n); // Process it later.
611 break;
612 }
613 case Op_ThreadLocal: {
614 add_java_object(n, PointsToNode::ArgEscape);
615 break;
616 }
617 default:
618 ; // Do nothing for nodes not related to EA.
619 }
620 return;
621 }
622
623 #ifdef ASSERT
624 #define ELSE_FAIL(name) \
625 /* Should not be called for not pointer type. */ \
626 n->dump(1); \
627 assert(false, name); \
628 break;
629 #else
630 #define ELSE_FAIL(name) \
631 break;
632 #endif
633
634 // Add final simple edges to graph.
635 void ConnectionGraph::add_final_edges(Node *n) {
636 PointsToNode* n_ptn = ptnode_adr(n->_idx);
637 #ifdef ASSERT
638 if (_verify && n_ptn->is_JavaObject())
639 return; // This method does not change graph for JavaObject.
640 #endif
641
642 if (n->is_Call()) {
643 process_call_arguments(n->as_Call());
644 return;
645 }
646 assert(n->is_Store() || n->is_LoadStore() ||
647 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
648 "node should be registered already");
649 int opcode = n->Opcode();
650 switch (opcode) {
651 case Op_AddP: {
652 Node* base = get_addp_base(n);
653 PointsToNode* ptn_base = ptnode_adr(base->_idx);
654 assert(ptn_base != NULL, "field's base should be registered");
655 add_base(n_ptn->as_Field(), ptn_base);
656 break;
657 }
658 case Op_CastPP:
659 case Op_CheckCastPP:
660 case Op_EncodeP:
661 case Op_DecodeN:
662 case Op_EncodePKlass:
663 case Op_DecodeNKlass: {
664 add_local_var_and_edge(n, PointsToNode::NoEscape,
665 n->in(1), NULL);
666 break;
667 }
668 case Op_CMoveP: {
669 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
670 Node* in = n->in(i);
671 if (in == NULL)
672 continue; // ignore NULL
673 Node* uncast_in = in->uncast();
674 if (uncast_in->is_top() || uncast_in == n)
675 continue; // ignore top or inputs which go back this node
676 PointsToNode* ptn = ptnode_adr(in->_idx);
677 assert(ptn != NULL, "node should be registered");
678 add_edge(n_ptn, ptn);
679 }
680 break;
681 }
682 case Op_LoadP:
683 #if INCLUDE_ZGC
684 case Op_LoadBarrierSlowReg:
685 case Op_LoadBarrierWeakSlowReg:
686 #endif
687 case Op_LoadN:
688 case Op_LoadPLocked: {
689 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
690 // ThreadLocal has RawPtr type.
691 const Type* t = _igvn->type(n);
692 if (t->make_ptr() != NULL) {
693 Node* adr = n->in(MemNode::Address);
694 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
695 break;
696 }
697 ELSE_FAIL("Op_LoadP");
698 }
699 case Op_Phi: {
700 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
701 // ThreadLocal has RawPtr type.
702 const Type* t = n->as_Phi()->type();
703 if (t->make_ptr() != NULL) {
704 for (uint i = 1; i < n->req(); i++) {
705 Node* in = n->in(i);
706 if (in == NULL)
707 continue; // ignore NULL
708 Node* uncast_in = in->uncast();
709 if (uncast_in->is_top() || uncast_in == n)
710 continue; // ignore top or inputs which go back this node
711 PointsToNode* ptn = ptnode_adr(in->_idx);
712 assert(ptn != NULL, "node should be registered");
713 add_edge(n_ptn, ptn);
714 }
715 break;
716 }
717 ELSE_FAIL("Op_Phi");
718 }
719 case Op_Proj: {
720 // we are only interested in the oop result projection from a call
721 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
722 (n->in(0)->as_Call()->returns_pointer()|| n->bottom_type()->isa_ptr())) {
723 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
724 n->in(0)->as_Call()->tf()->returns_value_type_as_fields(), "what kind of oop return is it?");
725 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
726 break;
727 }
728 #if INCLUDE_ZGC
729 else if (UseZGC) {
730 if (n->as_Proj()->_con == LoadBarrierNode::Oop && n->in(0)->is_LoadBarrier()) {
731 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0)->in(LoadBarrierNode::Oop), NULL);
732 break;
733 }
734 }
735 #endif
736 ELSE_FAIL("Op_Proj");
737 }
738 case Op_Rethrow: // Exception object escapes
739 case Op_Return: {
740 if (n->req() > TypeFunc::Parms &&
741 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
742 // Treat Return value as LocalVar with GlobalEscape escape state.
743 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
744 n->in(TypeFunc::Parms), NULL);
745 break;
746 }
747 ELSE_FAIL("Op_Return");
748 }
749 case Op_StoreP:
750 case Op_StoreN:
751 case Op_StoreNKlass:
752 case Op_StorePConditional:
753 case Op_CompareAndExchangeP:
754 case Op_CompareAndExchangeN:
755 case Op_CompareAndSwapP:
756 case Op_CompareAndSwapN:
757 case Op_WeakCompareAndSwapP:
758 case Op_WeakCompareAndSwapN:
759 case Op_GetAndSetP:
760 case Op_GetAndSetN: {
761 Node* adr = n->in(MemNode::Address);
762 const Type *adr_type = _igvn->type(adr);
763 adr_type = adr_type->make_ptr();
764 #ifdef ASSERT
765 if (adr_type == NULL) {
766 n->dump(1);
767 assert(adr_type != NULL, "dead node should not be on list");
768 break;
769 }
770 #endif
771 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN ||
772 opcode == Op_CompareAndExchangeN || opcode == Op_CompareAndExchangeP) {
773 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
774 }
775 if ( adr_type->isa_oopptr()
776 || ( (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
777 && adr_type == TypeRawPtr::NOTNULL
778 && adr->in(AddPNode::Address)->is_Proj()
779 && adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
780 // Point Address to Value
781 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
782 assert(adr_ptn != NULL &&
783 adr_ptn->as_Field()->is_oop(), "node should be registered");
784 Node *val = n->in(MemNode::ValueIn);
785 PointsToNode* ptn = ptnode_adr(val->_idx);
786 assert(ptn != NULL, "node should be registered");
787 add_edge(adr_ptn, ptn);
788 break;
789 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
790 // Stored value escapes in unsafe access.
791 Node *val = n->in(MemNode::ValueIn);
792 PointsToNode* ptn = ptnode_adr(val->_idx);
793 assert(ptn != NULL, "node should be registered");
794 set_escape_state(ptn, PointsToNode::GlobalEscape);
795 // Add edge to object for unsafe access with offset.
796 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
797 assert(adr_ptn != NULL, "node should be registered");
798 if (adr_ptn->is_Field()) {
799 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
800 add_edge(adr_ptn, ptn);
801 }
802 break;
803 }
804 ELSE_FAIL("Op_StoreP");
805 }
806 case Op_AryEq:
807 case Op_HasNegatives:
808 case Op_StrComp:
809 case Op_StrEquals:
810 case Op_StrIndexOf:
811 case Op_StrIndexOfChar:
812 case Op_StrInflatedCopy:
813 case Op_StrCompressedCopy:
814 case Op_EncodeISOArray: {
815 // char[]/byte[] arrays passed to string intrinsic do not escape but
816 // they are not scalar replaceable. Adjust escape state for them.
817 // Start from in(2) edge since in(1) is memory edge.
818 for (uint i = 2; i < n->req(); i++) {
819 Node* adr = n->in(i);
820 const Type* at = _igvn->type(adr);
821 if (!adr->is_top() && at->isa_ptr()) {
822 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
823 at->isa_ptr() != NULL, "expecting a pointer");
824 if (adr->is_AddP()) {
825 adr = get_addp_base(adr);
826 }
827 PointsToNode* ptn = ptnode_adr(adr->_idx);
828 assert(ptn != NULL, "node should be registered");
829 add_edge(n_ptn, ptn);
830 }
831 }
832 break;
833 }
834 default: {
835 // This method should be called only for EA specific nodes which may
836 // miss some edges when they were created.
837 #ifdef ASSERT
838 n->dump(1);
839 #endif
840 guarantee(false, "unknown node");
841 }
842 }
843 return;
844 }
845
846 void ConnectionGraph::add_call_node(CallNode* call) {
847 assert(call->returns_pointer() || call->tf()->returns_value_type_as_fields(), "only for call which returns pointer");
848 uint call_idx = call->_idx;
849 if (call->is_Allocate()) {
850 Node* k = call->in(AllocateNode::KlassNode);
851 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
852 assert(kt != NULL, "TypeKlassPtr required.");
853 ciKlass* cik = kt->klass();
854 PointsToNode::EscapeState es = PointsToNode::NoEscape;
855 bool scalar_replaceable = true;
856 if (call->is_AllocateArray()) {
857 if (!cik->is_array_klass()) { // StressReflectiveCode
858 es = PointsToNode::GlobalEscape;
859 } else {
860 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
861 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
862 // Not scalar replaceable if the length is not constant or too big.
863 scalar_replaceable = false;
864 }
865 }
866 } else { // Allocate instance
867 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
868 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
869 !cik->is_instance_klass() || // StressReflectiveCode
870 !cik->as_instance_klass()->can_be_instantiated() ||
871 cik->as_instance_klass()->has_finalizer()) {
872 es = PointsToNode::GlobalEscape;
873 }
874 }
875 add_java_object(call, es);
876 PointsToNode* ptn = ptnode_adr(call_idx);
877 if (!scalar_replaceable && ptn->scalar_replaceable()) {
878 ptn->set_scalar_replaceable(false);
879 }
880 } else if (call->is_CallStaticJava()) {
881 // Call nodes could be different types:
882 //
883 // 1. CallDynamicJavaNode (what happened during call is unknown):
884 //
885 // - mapped to GlobalEscape JavaObject node if oop is returned;
886 //
887 // - all oop arguments are escaping globally;
888 //
889 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
890 //
891 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
892 //
893 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
894 // - mapped to NoEscape JavaObject node if non-escaping object allocated
895 // during call is returned;
896 // - mapped to ArgEscape LocalVar node pointed to object arguments
897 // which are returned and does not escape during call;
898 //
899 // - oop arguments escaping status is defined by bytecode analysis;
900 //
901 // For a static call, we know exactly what method is being called.
902 // Use bytecode estimator to record whether the call's return value escapes.
903 ciMethod* meth = call->as_CallJava()->method();
904 if (meth == NULL) {
905 const char* name = call->as_CallStaticJava()->_name;
906 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
907 // Returns a newly allocated unescaped object.
908 add_java_object(call, PointsToNode::NoEscape);
909 ptnode_adr(call_idx)->set_scalar_replaceable(false);
910 } else if (meth->is_boxing_method()) {
911 // Returns boxing object
912 PointsToNode::EscapeState es;
913 vmIntrinsics::ID intr = meth->intrinsic_id();
914 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
915 // It does not escape if object is always allocated.
916 es = PointsToNode::NoEscape;
917 } else {
918 // It escapes globally if object could be loaded from cache.
919 es = PointsToNode::GlobalEscape;
920 }
921 add_java_object(call, es);
922 } else {
923 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
924 call_analyzer->copy_dependencies(_compile->dependencies());
925 if (call_analyzer->is_return_allocated()) {
926 // Returns a newly allocated unescaped object, simply
927 // update dependency information.
928 // Mark it as NoEscape so that objects referenced by
929 // it's fields will be marked as NoEscape at least.
930 add_java_object(call, PointsToNode::NoEscape);
931 ptnode_adr(call_idx)->set_scalar_replaceable(false);
932 } else {
933 // Determine whether any arguments are returned.
934 const TypeTuple* d = call->tf()->domain_cc();
935 bool ret_arg = false;
936 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
937 if (d->field_at(i)->isa_ptr() != NULL &&
938 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
939 ret_arg = true;
940 break;
941 }
942 }
943 if (ret_arg) {
944 add_local_var(call, PointsToNode::ArgEscape);
945 } else {
946 // Returns unknown object.
947 map_ideal_node(call, phantom_obj);
948 }
949 }
950 }
951 } else {
952 // An other type of call, assume the worst case:
953 // returned value is unknown and globally escapes.
954 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
955 map_ideal_node(call, phantom_obj);
956 }
957 }
958
959 void ConnectionGraph::process_call_arguments(CallNode *call) {
960 bool is_arraycopy = false;
961 switch (call->Opcode()) {
962 #ifdef ASSERT
963 case Op_Allocate:
964 case Op_AllocateArray:
965 case Op_Lock:
966 case Op_Unlock:
967 assert(false, "should be done already");
968 break;
969 #endif
970 case Op_ArrayCopy:
971 case Op_CallLeafNoFP:
972 // Most array copies are ArrayCopy nodes at this point but there
973 // are still a few direct calls to the copy subroutines (See
974 // PhaseStringOpts::copy_string())
975 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
976 call->as_CallLeaf()->is_call_to_arraycopystub();
977 // fall through
978 case Op_CallLeaf: {
979 // Stub calls, objects do not escape but they are not scale replaceable.
980 // Adjust escape state for outgoing arguments.
981 const TypeTuple * d = call->tf()->domain_sig();
982 bool src_has_oops = false;
983 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
984 const Type* at = d->field_at(i);
985 Node *arg = call->in(i);
986 if (arg == NULL) {
987 continue;
988 }
989 const Type *aat = _igvn->type(arg);
990 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
991 continue;
992 if (arg->is_AddP()) {
993 //
994 // The inline_native_clone() case when the arraycopy stub is called
995 // after the allocation before Initialize and CheckCastPP nodes.
996 // Or normal arraycopy for object arrays case.
997 //
998 // Set AddP's base (Allocate) as not scalar replaceable since
999 // pointer to the base (with offset) is passed as argument.
1000 //
1001 arg = get_addp_base(arg);
1002 }
1003 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1004 assert(arg_ptn != NULL, "should be registered");
1005 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
1006 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
1007 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1008 aat->isa_ptr() != NULL, "expecting an Ptr");
1009 bool arg_has_oops = aat->isa_oopptr() &&
1010 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
1011 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
1012 if (i == TypeFunc::Parms) {
1013 src_has_oops = arg_has_oops;
1014 }
1015 //
1016 // src or dst could be j.l.Object when other is basic type array:
1017 //
1018 // arraycopy(char[],0,Object*,0,size);
1019 // arraycopy(Object*,0,char[],0,size);
1020 //
1021 // Don't add edges in such cases.
1022 //
1023 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
1024 arg_has_oops && (i > TypeFunc::Parms);
1025 #ifdef ASSERT
1026 if (!(is_arraycopy ||
1027 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
1028 (call->as_CallLeaf()->_name != NULL &&
1029 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
1030 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
1031 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
1032 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
1033 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
1034 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
1035 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
1036 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
1037 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
1038 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
1039 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
1040 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
1041 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
1042 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
1043 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
1044 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
1045 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
1046 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
1047 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
1048 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
1049 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
1050 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
1051 strcmp(call->as_CallLeaf()->_name, "load_unknown_value") == 0 ||
1052 strcmp(call->as_CallLeaf()->_name, "store_unknown_value") == 0)
1053 ))) {
1054 call->dump();
1055 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
1056 }
1057 #endif
1058 // Always process arraycopy's destination object since
1059 // we need to add all possible edges to references in
1060 // source object.
1061 if (arg_esc >= PointsToNode::ArgEscape &&
1062 !arg_is_arraycopy_dest) {
1063 continue;
1064 }
1065 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
1066 if (call->is_ArrayCopy()) {
1067 ArrayCopyNode* ac = call->as_ArrayCopy();
1068 if (ac->is_clonebasic() ||
1069 ac->is_arraycopy_validated() ||
1070 ac->is_copyof_validated() ||
1071 ac->is_copyofrange_validated()) {
1072 es = PointsToNode::NoEscape;
1073 }
1074 }
1075 set_escape_state(arg_ptn, es);
1076 if (arg_is_arraycopy_dest) {
1077 Node* src = call->in(TypeFunc::Parms);
1078 if (src->is_AddP()) {
1079 src = get_addp_base(src);
1080 }
1081 PointsToNode* src_ptn = ptnode_adr(src->_idx);
1082 assert(src_ptn != NULL, "should be registered");
1083 if (arg_ptn != src_ptn) {
1084 // Special arraycopy edge:
1085 // A destination object's field can't have the source object
1086 // as base since objects escape states are not related.
1087 // Only escape state of destination object's fields affects
1088 // escape state of fields in source object.
1089 add_arraycopy(call, es, src_ptn, arg_ptn);
1090 }
1091 }
1092 }
1093 }
1094 break;
1095 }
1096 case Op_CallStaticJava: {
1097 // For a static call, we know exactly what method is being called.
1098 // Use bytecode estimator to record the call's escape affects
1099 #ifdef ASSERT
1100 const char* name = call->as_CallStaticJava()->_name;
1101 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
1102 #endif
1103 ciMethod* meth = call->as_CallJava()->method();
1104 if ((meth != NULL) && meth->is_boxing_method()) {
1105 break; // Boxing methods do not modify any oops.
1106 }
1107 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1108 // fall-through if not a Java method or no analyzer information
1109 if (call_analyzer != NULL) {
1110 PointsToNode* call_ptn = ptnode_adr(call->_idx);
1111 const TypeTuple* d = call->tf()->domain_cc();
1112 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1113 const Type* at = d->field_at(i);
1114 int k = i - TypeFunc::Parms;
1115 Node* arg = call->in(i);
1116 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
1117 if (at->isa_ptr() != NULL &&
1118 call_analyzer->is_arg_returned(k)) {
1119 // The call returns arguments.
1120 if (call_ptn != NULL) { // Is call's result used?
1121 assert(call_ptn->is_LocalVar(), "node should be registered");
1122 assert(arg_ptn != NULL, "node should be registered");
1123 add_edge(call_ptn, arg_ptn);
1124 }
1125 }
1126 if (at->isa_oopptr() != NULL &&
1127 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1128 if (!call_analyzer->is_arg_stack(k)) {
1129 // The argument global escapes
1130 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1131 } else {
1132 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1133 if (!call_analyzer->is_arg_local(k)) {
1134 // The argument itself doesn't escape, but any fields might
1135 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1136 }
1137 }
1138 }
1139 }
1140 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1141 // The call returns arguments.
1142 assert(call_ptn->edge_count() > 0, "sanity");
1143 if (!call_analyzer->is_return_local()) {
1144 // Returns also unknown object.
1145 add_edge(call_ptn, phantom_obj);
1146 }
1147 }
1148 break;
1149 }
1150 }
1151 default: {
1152 // Fall-through here if not a Java method or no analyzer information
1153 // or some other type of call, assume the worst case: all arguments
1154 // globally escape.
1155 const TypeTuple* d = call->tf()->domain_cc();
1156 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1157 const Type* at = d->field_at(i);
1158 if (at->isa_oopptr() != NULL) {
1159 Node* arg = call->in(i);
1160 if (arg->is_AddP()) {
1161 arg = get_addp_base(arg);
1162 }
1163 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1164 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1165 }
1166 }
1167 }
1168 }
1169 }
1170
1171
1172 // Finish Graph construction.
1173 bool ConnectionGraph::complete_connection_graph(
1174 GrowableArray<PointsToNode*>& ptnodes_worklist,
1175 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1176 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1177 GrowableArray<FieldNode*>& oop_fields_worklist) {
1178 // Normally only 1-3 passes needed to build Connection Graph depending
1179 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1180 // Set limit to 20 to catch situation when something did go wrong and
1181 // bailout Escape Analysis.
1182 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
1183 #define CG_BUILD_ITER_LIMIT 20
1184
1185 // Propagate GlobalEscape and ArgEscape escape states and check that
1186 // we still have non-escaping objects. The method pushs on _worklist
1187 // Field nodes which reference phantom_object.
1188 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1189 return false; // Nothing to do.
1190 }
1191 // Now propagate references to all JavaObject nodes.
1192 int java_objects_length = java_objects_worklist.length();
1193 elapsedTimer time;
1194 bool timeout = false;
1195 int new_edges = 1;
1196 int iterations = 0;
1197 do {
1198 while ((new_edges > 0) &&
1199 (iterations++ < CG_BUILD_ITER_LIMIT)) {
1200 double start_time = time.seconds();
1201 time.start();
1202 new_edges = 0;
1203 // Propagate references to phantom_object for nodes pushed on _worklist
1204 // by find_non_escaped_objects() and find_field_value().
1205 new_edges += add_java_object_edges(phantom_obj, false);
1206 for (int next = 0; next < java_objects_length; ++next) {
1207 JavaObjectNode* ptn = java_objects_worklist.at(next);
1208 new_edges += add_java_object_edges(ptn, true);
1209
1210 #define SAMPLE_SIZE 4
1211 if ((next % SAMPLE_SIZE) == 0) {
1212 // Each 4 iterations calculate how much time it will take
1213 // to complete graph construction.
1214 time.stop();
1215 // Poll for requests from shutdown mechanism to quiesce compiler
1216 // because Connection graph construction may take long time.
1217 CompileBroker::maybe_block();
1218 double stop_time = time.seconds();
1219 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
1220 double time_until_end = time_per_iter * (double)(java_objects_length - next);
1221 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
1222 timeout = true;
1223 break; // Timeout
1224 }
1225 start_time = stop_time;
1226 time.start();
1227 }
1228 #undef SAMPLE_SIZE
1229
1230 }
1231 if (timeout) break;
1232 if (new_edges > 0) {
1233 // Update escape states on each iteration if graph was updated.
1234 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1235 return false; // Nothing to do.
1236 }
1237 }
1238 time.stop();
1239 if (time.seconds() >= EscapeAnalysisTimeout) {
1240 timeout = true;
1241 break;
1242 }
1243 }
1244 if ((iterations < CG_BUILD_ITER_LIMIT) && !timeout) {
1245 time.start();
1246 // Find fields which have unknown value.
1247 int fields_length = oop_fields_worklist.length();
1248 for (int next = 0; next < fields_length; next++) {
1249 FieldNode* field = oop_fields_worklist.at(next);
1250 if (field->edge_count() == 0) {
1251 new_edges += find_field_value(field);
1252 // This code may added new edges to phantom_object.
1253 // Need an other cycle to propagate references to phantom_object.
1254 }
1255 }
1256 time.stop();
1257 if (time.seconds() >= EscapeAnalysisTimeout) {
1258 timeout = true;
1259 break;
1260 }
1261 } else {
1262 new_edges = 0; // Bailout
1263 }
1264 } while (new_edges > 0);
1265
1266 // Bailout if passed limits.
1267 if ((iterations >= CG_BUILD_ITER_LIMIT) || timeout) {
1268 Compile* C = _compile;
1269 if (C->log() != NULL) {
1270 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1271 C->log()->text("%s", timeout ? "time" : "iterations");
1272 C->log()->end_elem(" limit'");
1273 }
1274 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1275 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length());
1276 // Possible infinite build_connection_graph loop,
1277 // bailout (no changes to ideal graph were made).
1278 return false;
1279 }
1280 #ifdef ASSERT
1281 if (Verbose && PrintEscapeAnalysis) {
1282 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1283 iterations, nodes_size(), ptnodes_worklist.length());
1284 }
1285 #endif
1286
1287 #undef CG_BUILD_ITER_LIMIT
1288
1289 // Find fields initialized by NULL for non-escaping Allocations.
1290 int non_escaped_length = non_escaped_worklist.length();
1291 for (int next = 0; next < non_escaped_length; next++) {
1292 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1293 PointsToNode::EscapeState es = ptn->escape_state();
1294 assert(es <= PointsToNode::ArgEscape, "sanity");
1295 if (es == PointsToNode::NoEscape) {
1296 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1297 // Adding references to NULL object does not change escape states
1298 // since it does not escape. Also no fields are added to NULL object.
1299 add_java_object_edges(null_obj, false);
1300 }
1301 }
1302 Node* n = ptn->ideal_node();
1303 if (n->is_Allocate()) {
1304 // The object allocated by this Allocate node will never be
1305 // seen by an other thread. Mark it so that when it is
1306 // expanded no MemBarStoreStore is added.
1307 InitializeNode* ini = n->as_Allocate()->initialization();
1308 if (ini != NULL)
1309 ini->set_does_not_escape();
1310 }
1311 }
1312 return true; // Finished graph construction.
1313 }
1314
1315 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1316 // and check that we still have non-escaping java objects.
1317 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1318 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1319 GrowableArray<PointsToNode*> escape_worklist;
1320 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1321 int ptnodes_length = ptnodes_worklist.length();
1322 for (int next = 0; next < ptnodes_length; ++next) {
1323 PointsToNode* ptn = ptnodes_worklist.at(next);
1324 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1325 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1326 escape_worklist.push(ptn);
1327 }
1328 }
1329 // Set escape states to referenced nodes (edges list).
1330 while (escape_worklist.length() > 0) {
1331 PointsToNode* ptn = escape_worklist.pop();
1332 PointsToNode::EscapeState es = ptn->escape_state();
1333 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1334 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1335 es >= PointsToNode::ArgEscape) {
1336 // GlobalEscape or ArgEscape state of field means it has unknown value.
1337 if (add_edge(ptn, phantom_obj)) {
1338 // New edge was added
1339 add_field_uses_to_worklist(ptn->as_Field());
1340 }
1341 }
1342 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1343 PointsToNode* e = i.get();
1344 if (e->is_Arraycopy()) {
1345 assert(ptn->arraycopy_dst(), "sanity");
1346 // Propagate only fields escape state through arraycopy edge.
1347 if (e->fields_escape_state() < field_es) {
1348 set_fields_escape_state(e, field_es);
1349 escape_worklist.push(e);
1350 }
1351 } else if (es >= field_es) {
1352 // fields_escape_state is also set to 'es' if it is less than 'es'.
1353 if (e->escape_state() < es) {
1354 set_escape_state(e, es);
1355 escape_worklist.push(e);
1356 }
1357 } else {
1358 // Propagate field escape state.
1359 bool es_changed = false;
1360 if (e->fields_escape_state() < field_es) {
1361 set_fields_escape_state(e, field_es);
1362 es_changed = true;
1363 }
1364 if ((e->escape_state() < field_es) &&
1365 e->is_Field() && ptn->is_JavaObject() &&
1366 e->as_Field()->is_oop()) {
1367 // Change escape state of referenced fields.
1368 set_escape_state(e, field_es);
1369 es_changed = true;
1370 } else if (e->escape_state() < es) {
1371 set_escape_state(e, es);
1372 es_changed = true;
1373 }
1374 if (es_changed) {
1375 escape_worklist.push(e);
1376 }
1377 }
1378 }
1379 }
1380 // Remove escaped objects from non_escaped list.
1381 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1382 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1383 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1384 non_escaped_worklist.delete_at(next);
1385 }
1386 if (ptn->escape_state() == PointsToNode::NoEscape) {
1387 // Find fields in non-escaped allocations which have unknown value.
1388 find_init_values(ptn, phantom_obj, NULL);
1389 }
1390 }
1391 return (non_escaped_worklist.length() > 0);
1392 }
1393
1394 // Add all references to JavaObject node by walking over all uses.
1395 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1396 int new_edges = 0;
1397 if (populate_worklist) {
1398 // Populate _worklist by uses of jobj's uses.
1399 for (UseIterator i(jobj); i.has_next(); i.next()) {
1400 PointsToNode* use = i.get();
1401 if (use->is_Arraycopy())
1402 continue;
1403 add_uses_to_worklist(use);
1404 if (use->is_Field() && use->as_Field()->is_oop()) {
1405 // Put on worklist all field's uses (loads) and
1406 // related field nodes (same base and offset).
1407 add_field_uses_to_worklist(use->as_Field());
1408 }
1409 }
1410 }
1411 for (int l = 0; l < _worklist.length(); l++) {
1412 PointsToNode* use = _worklist.at(l);
1413 if (PointsToNode::is_base_use(use)) {
1414 // Add reference from jobj to field and from field to jobj (field's base).
1415 use = PointsToNode::get_use_node(use)->as_Field();
1416 if (add_base(use->as_Field(), jobj)) {
1417 new_edges++;
1418 }
1419 continue;
1420 }
1421 assert(!use->is_JavaObject(), "sanity");
1422 if (use->is_Arraycopy()) {
1423 if (jobj == null_obj) // NULL object does not have field edges
1424 continue;
1425 // Added edge from Arraycopy node to arraycopy's source java object
1426 if (add_edge(use, jobj)) {
1427 jobj->set_arraycopy_src();
1428 new_edges++;
1429 }
1430 // and stop here.
1431 continue;
1432 }
1433 if (!add_edge(use, jobj))
1434 continue; // No new edge added, there was such edge already.
1435 new_edges++;
1436 if (use->is_LocalVar()) {
1437 add_uses_to_worklist(use);
1438 if (use->arraycopy_dst()) {
1439 for (EdgeIterator i(use); i.has_next(); i.next()) {
1440 PointsToNode* e = i.get();
1441 if (e->is_Arraycopy()) {
1442 if (jobj == null_obj) // NULL object does not have field edges
1443 continue;
1444 // Add edge from arraycopy's destination java object to Arraycopy node.
1445 if (add_edge(jobj, e)) {
1446 new_edges++;
1447 jobj->set_arraycopy_dst();
1448 }
1449 }
1450 }
1451 }
1452 } else {
1453 // Added new edge to stored in field values.
1454 // Put on worklist all field's uses (loads) and
1455 // related field nodes (same base and offset).
1456 add_field_uses_to_worklist(use->as_Field());
1457 }
1458 }
1459 _worklist.clear();
1460 _in_worklist.Reset();
1461 return new_edges;
1462 }
1463
1464 // Put on worklist all related field nodes.
1465 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1466 assert(field->is_oop(), "sanity");
1467 int offset = field->offset();
1468 add_uses_to_worklist(field);
1469 // Loop over all bases of this field and push on worklist Field nodes
1470 // with the same offset and base (since they may reference the same field).
1471 for (BaseIterator i(field); i.has_next(); i.next()) {
1472 PointsToNode* base = i.get();
1473 add_fields_to_worklist(field, base);
1474 // Check if the base was source object of arraycopy and go over arraycopy's
1475 // destination objects since values stored to a field of source object are
1476 // accessable by uses (loads) of fields of destination objects.
1477 if (base->arraycopy_src()) {
1478 for (UseIterator j(base); j.has_next(); j.next()) {
1479 PointsToNode* arycp = j.get();
1480 if (arycp->is_Arraycopy()) {
1481 for (UseIterator k(arycp); k.has_next(); k.next()) {
1482 PointsToNode* abase = k.get();
1483 if (abase->arraycopy_dst() && abase != base) {
1484 // Look for the same arraycopy reference.
1485 add_fields_to_worklist(field, abase);
1486 }
1487 }
1488 }
1489 }
1490 }
1491 }
1492 }
1493
1494 // Put on worklist all related field nodes.
1495 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1496 int offset = field->offset();
1497 if (base->is_LocalVar()) {
1498 for (UseIterator j(base); j.has_next(); j.next()) {
1499 PointsToNode* f = j.get();
1500 if (PointsToNode::is_base_use(f)) { // Field
1501 f = PointsToNode::get_use_node(f);
1502 if (f == field || !f->as_Field()->is_oop())
1503 continue;
1504 int offs = f->as_Field()->offset();
1505 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1506 add_to_worklist(f);
1507 }
1508 }
1509 }
1510 } else {
1511 assert(base->is_JavaObject(), "sanity");
1512 if (// Skip phantom_object since it is only used to indicate that
1513 // this field's content globally escapes.
1514 (base != phantom_obj) &&
1515 // NULL object node does not have fields.
1516 (base != null_obj)) {
1517 for (EdgeIterator i(base); i.has_next(); i.next()) {
1518 PointsToNode* f = i.get();
1519 // Skip arraycopy edge since store to destination object field
1520 // does not update value in source object field.
1521 if (f->is_Arraycopy()) {
1522 assert(base->arraycopy_dst(), "sanity");
1523 continue;
1524 }
1525 if (f == field || !f->as_Field()->is_oop())
1526 continue;
1527 int offs = f->as_Field()->offset();
1528 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1529 add_to_worklist(f);
1530 }
1531 }
1532 }
1533 }
1534 }
1535
1536 // Find fields which have unknown value.
1537 int ConnectionGraph::find_field_value(FieldNode* field) {
1538 // Escaped fields should have init value already.
1539 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1540 int new_edges = 0;
1541 for (BaseIterator i(field); i.has_next(); i.next()) {
1542 PointsToNode* base = i.get();
1543 if (base->is_JavaObject()) {
1544 // Skip Allocate's fields which will be processed later.
1545 if (base->ideal_node()->is_Allocate())
1546 return 0;
1547 assert(base == null_obj, "only NULL ptr base expected here");
1548 }
1549 }
1550 if (add_edge(field, phantom_obj)) {
1551 // New edge was added
1552 new_edges++;
1553 add_field_uses_to_worklist(field);
1554 }
1555 return new_edges;
1556 }
1557
1558 // Find fields initializing values for allocations.
1559 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1560 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1561 int new_edges = 0;
1562 Node* alloc = pta->ideal_node();
1563 if (init_val == phantom_obj) {
1564 // Do nothing for Allocate nodes since its fields values are
1565 // "known" unless they are initialized by arraycopy/clone.
1566 if (alloc->is_Allocate() && !pta->arraycopy_dst())
1567 return 0;
1568 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
1569 #ifdef ASSERT
1570 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == NULL) {
1571 const char* name = alloc->as_CallStaticJava()->_name;
1572 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1573 }
1574 #endif
1575 // Non-escaped allocation returned from Java or runtime call have
1576 // unknown values in fields.
1577 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1578 PointsToNode* field = i.get();
1579 if (field->is_Field() && field->as_Field()->is_oop()) {
1580 if (add_edge(field, phantom_obj)) {
1581 // New edge was added
1582 new_edges++;
1583 add_field_uses_to_worklist(field->as_Field());
1584 }
1585 }
1586 }
1587 return new_edges;
1588 }
1589 assert(init_val == null_obj, "sanity");
1590 // Do nothing for Call nodes since its fields values are unknown.
1591 if (!alloc->is_Allocate())
1592 return 0;
1593
1594 InitializeNode* ini = alloc->as_Allocate()->initialization();
1595 bool visited_bottom_offset = false;
1596 GrowableArray<int> offsets_worklist;
1597
1598 // Check if an oop field's initializing value is recorded and add
1599 // a corresponding NULL if field's value if it is not recorded.
1600 // Connection Graph does not record a default initialization by NULL
1601 // captured by Initialize node.
1602 //
1603 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1604 PointsToNode* field = i.get(); // Field (AddP)
1605 if (!field->is_Field() || !field->as_Field()->is_oop())
1606 continue; // Not oop field
1607 int offset = field->as_Field()->offset();
1608 if (offset == Type::OffsetBot) {
1609 if (!visited_bottom_offset) {
1610 // OffsetBot is used to reference array's element,
1611 // always add reference to NULL to all Field nodes since we don't
1612 // known which element is referenced.
1613 if (add_edge(field, null_obj)) {
1614 // New edge was added
1615 new_edges++;
1616 add_field_uses_to_worklist(field->as_Field());
1617 visited_bottom_offset = true;
1618 }
1619 }
1620 } else {
1621 // Check only oop fields.
1622 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1623 if (adr_type->isa_rawptr()) {
1624 #ifdef ASSERT
1625 // Raw pointers are used for initializing stores so skip it
1626 // since it should be recorded already
1627 Node* base = get_addp_base(field->ideal_node());
1628 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1629 (base->in(0) == alloc),"unexpected pointer type");
1630 #endif
1631 continue;
1632 }
1633 if (!offsets_worklist.contains(offset)) {
1634 offsets_worklist.append(offset);
1635 Node* value = NULL;
1636 if (ini != NULL) {
1637 // StoreP::memory_type() == T_ADDRESS
1638 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1639 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1640 // Make sure initializing store has the same type as this AddP.
1641 // This AddP may reference non existing field because it is on a
1642 // dead branch of bimorphic call which is not eliminated yet.
1643 if (store != NULL && store->is_Store() &&
1644 store->as_Store()->memory_type() == ft) {
1645 value = store->in(MemNode::ValueIn);
1646 #ifdef ASSERT
1647 if (VerifyConnectionGraph) {
1648 // Verify that AddP already points to all objects the value points to.
1649 PointsToNode* val = ptnode_adr(value->_idx);
1650 assert((val != NULL), "should be processed already");
1651 PointsToNode* missed_obj = NULL;
1652 if (val->is_JavaObject()) {
1653 if (!field->points_to(val->as_JavaObject())) {
1654 missed_obj = val;
1655 }
1656 } else {
1657 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1658 tty->print_cr("----------init store has invalid value -----");
1659 store->dump();
1660 val->dump();
1661 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1662 }
1663 for (EdgeIterator j(val); j.has_next(); j.next()) {
1664 PointsToNode* obj = j.get();
1665 if (obj->is_JavaObject()) {
1666 if (!field->points_to(obj->as_JavaObject())) {
1667 missed_obj = obj;
1668 break;
1669 }
1670 }
1671 }
1672 }
1673 if (missed_obj != NULL) {
1674 tty->print_cr("----------field---------------------------------");
1675 field->dump();
1676 tty->print_cr("----------missed reference to object------------");
1677 missed_obj->dump();
1678 tty->print_cr("----------object referenced by init store-------");
1679 store->dump();
1680 val->dump();
1681 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1682 }
1683 }
1684 #endif
1685 } else {
1686 // There could be initializing stores which follow allocation.
1687 // For example, a volatile field store is not collected
1688 // by Initialize node.
1689 //
1690 // Need to check for dependent loads to separate such stores from
1691 // stores which follow loads. For now, add initial value NULL so
1692 // that compare pointers optimization works correctly.
1693 }
1694 }
1695 if (value == NULL) {
1696 // A field's initializing value was not recorded. Add NULL.
1697 if (add_edge(field, null_obj)) {
1698 // New edge was added
1699 new_edges++;
1700 add_field_uses_to_worklist(field->as_Field());
1701 }
1702 }
1703 }
1704 }
1705 }
1706 return new_edges;
1707 }
1708
1709 // Adjust scalar_replaceable state after Connection Graph is built.
1710 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1711 // Search for non-escaping objects which are not scalar replaceable
1712 // and mark them to propagate the state to referenced objects.
1713
1714 // 1. An object is not scalar replaceable if the field into which it is
1715 // stored has unknown offset (stored into unknown element of an array).
1716 //
1717 for (UseIterator i(jobj); i.has_next(); i.next()) {
1718 PointsToNode* use = i.get();
1719 if (use->is_Arraycopy()) {
1720 continue;
1721 }
1722 if (use->is_Field()) {
1723 FieldNode* field = use->as_Field();
1724 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
1725 if (field->offset() == Type::OffsetBot) {
1726 jobj->set_scalar_replaceable(false);
1727 return;
1728 }
1729 // 2. An object is not scalar replaceable if the field into which it is
1730 // stored has multiple bases one of which is null.
1731 if (field->base_count() > 1) {
1732 for (BaseIterator i(field); i.has_next(); i.next()) {
1733 PointsToNode* base = i.get();
1734 if (base == null_obj) {
1735 jobj->set_scalar_replaceable(false);
1736 return;
1737 }
1738 }
1739 }
1740 }
1741 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1742 // 3. An object is not scalar replaceable if it is merged with other objects.
1743 for (EdgeIterator j(use); j.has_next(); j.next()) {
1744 PointsToNode* ptn = j.get();
1745 if (ptn->is_JavaObject() && ptn != jobj) {
1746 // Mark all objects.
1747 jobj->set_scalar_replaceable(false);
1748 ptn->set_scalar_replaceable(false);
1749 }
1750 }
1751 if (!jobj->scalar_replaceable()) {
1752 return;
1753 }
1754 }
1755
1756 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1757 if (j.get()->is_Arraycopy()) {
1758 continue;
1759 }
1760
1761 // Non-escaping object node should point only to field nodes.
1762 FieldNode* field = j.get()->as_Field();
1763 int offset = field->as_Field()->offset();
1764
1765 // 4. An object is not scalar replaceable if it has a field with unknown
1766 // offset (array's element is accessed in loop).
1767 if (offset == Type::OffsetBot) {
1768 jobj->set_scalar_replaceable(false);
1769 return;
1770 }
1771 // 5. Currently an object is not scalar replaceable if a LoadStore node
1772 // access its field since the field value is unknown after it.
1773 //
1774 Node* n = field->ideal_node();
1775 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1776 if (n->fast_out(i)->is_LoadStore()) {
1777 jobj->set_scalar_replaceable(false);
1778 return;
1779 }
1780 }
1781
1782 // 6. Or the address may point to more then one object. This may produce
1783 // the false positive result (set not scalar replaceable)
1784 // since the flow-insensitive escape analysis can't separate
1785 // the case when stores overwrite the field's value from the case
1786 // when stores happened on different control branches.
1787 //
1788 // Note: it will disable scalar replacement in some cases:
1789 //
1790 // Point p[] = new Point[1];
1791 // p[0] = new Point(); // Will be not scalar replaced
1792 //
1793 // but it will save us from incorrect optimizations in next cases:
1794 //
1795 // Point p[] = new Point[1];
1796 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1797 //
1798 if (field->base_count() > 1) {
1799 for (BaseIterator i(field); i.has_next(); i.next()) {
1800 PointsToNode* base = i.get();
1801 // Don't take into account LocalVar nodes which
1802 // may point to only one object which should be also
1803 // this field's base by now.
1804 if (base->is_JavaObject() && base != jobj) {
1805 // Mark all bases.
1806 jobj->set_scalar_replaceable(false);
1807 base->set_scalar_replaceable(false);
1808 }
1809 }
1810 }
1811 }
1812 }
1813
1814 #ifdef ASSERT
1815 void ConnectionGraph::verify_connection_graph(
1816 GrowableArray<PointsToNode*>& ptnodes_worklist,
1817 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1818 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1819 GrowableArray<Node*>& addp_worklist) {
1820 // Verify that graph is complete - no new edges could be added.
1821 int java_objects_length = java_objects_worklist.length();
1822 int non_escaped_length = non_escaped_worklist.length();
1823 int new_edges = 0;
1824 for (int next = 0; next < java_objects_length; ++next) {
1825 JavaObjectNode* ptn = java_objects_worklist.at(next);
1826 new_edges += add_java_object_edges(ptn, true);
1827 }
1828 assert(new_edges == 0, "graph was not complete");
1829 // Verify that escape state is final.
1830 int length = non_escaped_worklist.length();
1831 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1832 assert((non_escaped_length == non_escaped_worklist.length()) &&
1833 (non_escaped_length == length) &&
1834 (_worklist.length() == 0), "escape state was not final");
1835
1836 // Verify fields information.
1837 int addp_length = addp_worklist.length();
1838 for (int next = 0; next < addp_length; ++next ) {
1839 Node* n = addp_worklist.at(next);
1840 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1841 if (field->is_oop()) {
1842 // Verify that field has all bases
1843 Node* base = get_addp_base(n);
1844 PointsToNode* ptn = ptnode_adr(base->_idx);
1845 if (ptn->is_JavaObject()) {
1846 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1847 } else {
1848 assert(ptn->is_LocalVar(), "sanity");
1849 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1850 PointsToNode* e = i.get();
1851 if (e->is_JavaObject()) {
1852 assert(field->has_base(e->as_JavaObject()), "sanity");
1853 }
1854 }
1855 }
1856 // Verify that all fields have initializing values.
1857 if (field->edge_count() == 0) {
1858 tty->print_cr("----------field does not have references----------");
1859 field->dump();
1860 for (BaseIterator i(field); i.has_next(); i.next()) {
1861 PointsToNode* base = i.get();
1862 tty->print_cr("----------field has next base---------------------");
1863 base->dump();
1864 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1865 tty->print_cr("----------base has fields-------------------------");
1866 for (EdgeIterator j(base); j.has_next(); j.next()) {
1867 j.get()->dump();
1868 }
1869 tty->print_cr("----------base has references---------------------");
1870 for (UseIterator j(base); j.has_next(); j.next()) {
1871 j.get()->dump();
1872 }
1873 }
1874 }
1875 for (UseIterator i(field); i.has_next(); i.next()) {
1876 i.get()->dump();
1877 }
1878 assert(field->edge_count() > 0, "sanity");
1879 }
1880 }
1881 }
1882 }
1883 #endif
1884
1885 // Optimize ideal graph.
1886 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1887 GrowableArray<Node*>& storestore_worklist) {
1888 Compile* C = _compile;
1889 PhaseIterGVN* igvn = _igvn;
1890 if (EliminateLocks) {
1891 // Mark locks before changing ideal graph.
1892 int cnt = C->macro_count();
1893 for( int i=0; i < cnt; i++ ) {
1894 Node *n = C->macro_node(i);
1895 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1896 AbstractLockNode* alock = n->as_AbstractLock();
1897 if (!alock->is_non_esc_obj()) {
1898 if (not_global_escape(alock->obj_node())) {
1899 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1900 // The lock could be marked eliminated by lock coarsening
1901 // code during first IGVN before EA. Replace coarsened flag
1902 // to eliminate all associated locks/unlocks.
1903 #ifdef ASSERT
1904 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
1905 #endif
1906 alock->set_non_esc_obj();
1907 }
1908 }
1909 }
1910 }
1911 }
1912
1913 if (OptimizePtrCompare) {
1914 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1915 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1916 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1917 // Optimize objects compare.
1918 while (ptr_cmp_worklist.length() != 0) {
1919 Node *n = ptr_cmp_worklist.pop();
1920 Node *res = optimize_ptr_compare(n);
1921 if (res != NULL) {
1922 #ifndef PRODUCT
1923 if (PrintOptimizePtrCompare) {
1924 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"));
1925 if (Verbose) {
1926 n->dump(1);
1927 }
1928 }
1929 #endif
1930 igvn->replace_node(n, res);
1931 }
1932 }
1933 // cleanup
1934 if (_pcmp_neq->outcnt() == 0)
1935 igvn->hash_delete(_pcmp_neq);
1936 if (_pcmp_eq->outcnt() == 0)
1937 igvn->hash_delete(_pcmp_eq);
1938 }
1939
1940 // For MemBarStoreStore nodes added in library_call.cpp, check
1941 // escape status of associated AllocateNode and optimize out
1942 // MemBarStoreStore node if the allocated object never escapes.
1943 while (storestore_worklist.length() != 0) {
1944 Node *n = storestore_worklist.pop();
1945 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1946 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1947 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1948 if (not_global_escape(alloc)) {
1949 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1950 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1951 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1952 igvn->register_new_node_with_optimizer(mb);
1953 igvn->replace_node(storestore, mb);
1954 }
1955 }
1956 }
1957
1958 // Optimize objects compare.
1959 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1960 assert(OptimizePtrCompare, "sanity");
1961 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1962 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1963 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1964 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1965 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1966 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1967
1968 // Check simple cases first.
1969 if (jobj1 != NULL) {
1970 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1971 if (jobj1 == jobj2) {
1972 // Comparing the same not escaping object.
1973 return _pcmp_eq;
1974 }
1975 Node* obj = jobj1->ideal_node();
1976 // Comparing not escaping allocation.
1977 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1978 !ptn2->points_to(jobj1)) {
1979 return _pcmp_neq; // This includes nullness check.
1980 }
1981 }
1982 }
1983 if (jobj2 != NULL) {
1984 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1985 Node* obj = jobj2->ideal_node();
1986 // Comparing not escaping allocation.
1987 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1988 !ptn1->points_to(jobj2)) {
1989 return _pcmp_neq; // This includes nullness check.
1990 }
1991 }
1992 }
1993 if (jobj1 != NULL && jobj1 != phantom_obj &&
1994 jobj2 != NULL && jobj2 != phantom_obj &&
1995 jobj1->ideal_node()->is_Con() &&
1996 jobj2->ideal_node()->is_Con()) {
1997 // Klass or String constants compare. Need to be careful with
1998 // compressed pointers - compare types of ConN and ConP instead of nodes.
1999 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
2000 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
2001 if (t1->make_ptr() == t2->make_ptr()) {
2002 return _pcmp_eq;
2003 } else {
2004 return _pcmp_neq;
2005 }
2006 }
2007 if (ptn1->meet(ptn2)) {
2008 return NULL; // Sets are not disjoint
2009 }
2010
2011 // Sets are disjoint.
2012 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
2013 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
2014 bool set1_has_null_ptr = ptn1->points_to(null_obj);
2015 bool set2_has_null_ptr = ptn2->points_to(null_obj);
2016 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
2017 (set2_has_unknown_ptr && set1_has_null_ptr)) {
2018 // Check nullness of unknown object.
2019 return NULL;
2020 }
2021
2022 // Disjointness by itself is not sufficient since
2023 // alias analysis is not complete for escaped objects.
2024 // Disjoint sets are definitely unrelated only when
2025 // at least one set has only not escaping allocations.
2026 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
2027 if (ptn1->non_escaping_allocation()) {
2028 return _pcmp_neq;
2029 }
2030 }
2031 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2032 if (ptn2->non_escaping_allocation()) {
2033 return _pcmp_neq;
2034 }
2035 }
2036 return NULL;
2037 }
2038
2039 // Connection Graph constuction functions.
2040
2041 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
2042 PointsToNode* ptadr = _nodes.at(n->_idx);
2043 if (ptadr != NULL) {
2044 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
2045 return;
2046 }
2047 Compile* C = _compile;
2048 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
2049 _nodes.at_put(n->_idx, ptadr);
2050 }
2051
2052 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
2053 PointsToNode* ptadr = _nodes.at(n->_idx);
2054 if (ptadr != NULL) {
2055 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
2056 return;
2057 }
2058 Compile* C = _compile;
2059 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
2060 _nodes.at_put(n->_idx, ptadr);
2061 }
2062
2063 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
2064 PointsToNode* ptadr = _nodes.at(n->_idx);
2065 if (ptadr != NULL) {
2066 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
2067 return;
2068 }
2069 bool unsafe = false;
2070 bool is_oop = is_oop_field(n, offset, &unsafe);
2071 if (unsafe) {
2072 es = PointsToNode::GlobalEscape;
2073 }
2074 Compile* C = _compile;
2075 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
2076 _nodes.at_put(n->_idx, field);
2077 }
2078
2079 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
2080 PointsToNode* src, PointsToNode* dst) {
2081 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
2082 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
2083 PointsToNode* ptadr = _nodes.at(n->_idx);
2084 if (ptadr != NULL) {
2085 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
2086 return;
2087 }
2088 Compile* C = _compile;
2089 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
2090 _nodes.at_put(n->_idx, ptadr);
2091 // Add edge from arraycopy node to source object.
2092 (void)add_edge(ptadr, src);
2093 src->set_arraycopy_src();
2094 // Add edge from destination object to arraycopy node.
2095 (void)add_edge(dst, ptadr);
2096 dst->set_arraycopy_dst();
2097 }
2098
2099 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
2100 const Type* adr_type = n->as_AddP()->bottom_type();
2101 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
2102 BasicType bt = T_INT;
2103 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
2104 // Check only oop fields.
2105 if (!adr_type->isa_aryptr() ||
2106 (adr_type->isa_aryptr()->klass() == NULL) ||
2107 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
2108 // OffsetBot is used to reference array's element. Ignore first AddP.
2109 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
2110 bt = T_OBJECT;
2111 }
2112 }
2113 } else if (offset != oopDesc::klass_offset_in_bytes()) {
2114 if (adr_type->isa_instptr()) {
2115 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
2116 if (field != NULL) {
2117 bt = field->layout_type();
2118 } else {
2119 // Check for unsafe oop field access
2120 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2121 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2122 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2123 bt = T_OBJECT;
2124 (*unsafe) = true;
2125 }
2126 }
2127 } else if (adr_type->isa_aryptr()) {
2128 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2129 // Ignore array length load.
2130 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
2131 // Ignore first AddP.
2132 } else {
2133 const Type* elemtype = adr_type->isa_aryptr()->elem();
2134 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) {
2135 ciValueKlass* vk = elemtype->is_valuetype()->value_klass();
2136 field_offset += vk->first_field_offset();
2137 bt = vk->get_field_by_offset(field_offset, false)->layout_type();
2138 } else {
2139 bt = elemtype->array_element_basic_type();
2140 }
2141 }
2142 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
2143 // Allocation initialization, ThreadLocal field access, unsafe access
2144 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
2145 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
2146 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN)) {
2147 bt = T_OBJECT;
2148 }
2149 }
2150 }
2151 // TODO enable when using T_VALUETYPEPTR
2152 //assert(bt != T_VALUETYPE, "should not have valuetype here");
2153 return (bt == T_OBJECT || bt == T_VALUETYPE || bt == T_VALUETYPEPTR || bt == T_NARROWOOP || bt == T_ARRAY);
2154 }
2155
2156 // Returns unique pointed java object or NULL.
2157 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2158 assert(!_collecting, "should not call when constructed graph");
2159 // If the node was created after the escape computation we can't answer.
2160 uint idx = n->_idx;
2161 if (idx >= nodes_size()) {
2162 return NULL;
2163 }
2164 PointsToNode* ptn = ptnode_adr(idx);
2165 if (ptn->is_JavaObject()) {
2166 return ptn->as_JavaObject();
2167 }
2168 assert(ptn->is_LocalVar(), "sanity");
2169 // Check all java objects it points to.
2170 JavaObjectNode* jobj = NULL;
2171 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2172 PointsToNode* e = i.get();
2173 if (e->is_JavaObject()) {
2174 if (jobj == NULL) {
2175 jobj = e->as_JavaObject();
2176 } else if (jobj != e) {
2177 return NULL;
2178 }
2179 }
2180 }
2181 return jobj;
2182 }
2183
2184 // Return true if this node points only to non-escaping allocations.
2185 bool PointsToNode::non_escaping_allocation() {
2186 if (is_JavaObject()) {
2187 Node* n = ideal_node();
2188 if (n->is_Allocate() || n->is_CallStaticJava()) {
2189 return (escape_state() == PointsToNode::NoEscape);
2190 } else {
2191 return false;
2192 }
2193 }
2194 assert(is_LocalVar(), "sanity");
2195 // Check all java objects it points to.
2196 for (EdgeIterator i(this); i.has_next(); i.next()) {
2197 PointsToNode* e = i.get();
2198 if (e->is_JavaObject()) {
2199 Node* n = e->ideal_node();
2200 if ((e->escape_state() != PointsToNode::NoEscape) ||
2201 !(n->is_Allocate() || n->is_CallStaticJava())) {
2202 return false;
2203 }
2204 }
2205 }
2206 return true;
2207 }
2208
2209 // Return true if we know the node does not escape globally.
2210 bool ConnectionGraph::not_global_escape(Node *n) {
2211 assert(!_collecting, "should not call during graph construction");
2212 // If the node was created after the escape computation we can't answer.
2213 uint idx = n->_idx;
2214 if (idx >= nodes_size()) {
2215 return false;
2216 }
2217 PointsToNode* ptn = ptnode_adr(idx);
2218 PointsToNode::EscapeState es = ptn->escape_state();
2219 // If we have already computed a value, return it.
2220 if (es >= PointsToNode::GlobalEscape)
2221 return false;
2222 if (ptn->is_JavaObject()) {
2223 return true; // (es < PointsToNode::GlobalEscape);
2224 }
2225 assert(ptn->is_LocalVar(), "sanity");
2226 // Check all java objects it points to.
2227 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2228 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2229 return false;
2230 }
2231 return true;
2232 }
2233
2234
2235 // Helper functions
2236
2237 // Return true if this node points to specified node or nodes it points to.
2238 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2239 if (is_JavaObject()) {
2240 return (this == ptn);
2241 }
2242 assert(is_LocalVar() || is_Field(), "sanity");
2243 for (EdgeIterator i(this); i.has_next(); i.next()) {
2244 if (i.get() == ptn)
2245 return true;
2246 }
2247 return false;
2248 }
2249
2250 // Return true if one node points to an other.
2251 bool PointsToNode::meet(PointsToNode* ptn) {
2252 if (this == ptn) {
2253 return true;
2254 } else if (ptn->is_JavaObject()) {
2255 return this->points_to(ptn->as_JavaObject());
2256 } else if (this->is_JavaObject()) {
2257 return ptn->points_to(this->as_JavaObject());
2258 }
2259 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2260 int ptn_count = ptn->edge_count();
2261 for (EdgeIterator i(this); i.has_next(); i.next()) {
2262 PointsToNode* this_e = i.get();
2263 for (int j = 0; j < ptn_count; j++) {
2264 if (this_e == ptn->edge(j))
2265 return true;
2266 }
2267 }
2268 return false;
2269 }
2270
2271 #ifdef ASSERT
2272 // Return true if bases point to this java object.
2273 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2274 for (BaseIterator i(this); i.has_next(); i.next()) {
2275 if (i.get() == jobj)
2276 return true;
2277 }
2278 return false;
2279 }
2280 #endif
2281
2282 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2283 const Type *adr_type = phase->type(adr);
2284 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2285 adr->in(AddPNode::Address)->is_Proj() &&
2286 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2287 // We are computing a raw address for a store captured by an Initialize
2288 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2289 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2290 assert(offs != Type::OffsetBot ||
2291 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2292 "offset must be a constant or it is initialization of array");
2293 return offs;
2294 }
2295 return adr_type->is_ptr()->flattened_offset();
2296 }
2297
2298 Node* ConnectionGraph::get_addp_base(Node *addp) {
2299 assert(addp->is_AddP(), "must be AddP");
2300 //
2301 // AddP cases for Base and Address inputs:
2302 // case #1. Direct object's field reference:
2303 // Allocate
2304 // |
2305 // Proj #5 ( oop result )
2306 // |
2307 // CheckCastPP (cast to instance type)
2308 // | |
2309 // AddP ( base == address )
2310 //
2311 // case #2. Indirect object's field reference:
2312 // Phi
2313 // |
2314 // CastPP (cast to instance type)
2315 // | |
2316 // AddP ( base == address )
2317 //
2318 // case #3. Raw object's field reference for Initialize node:
2319 // Allocate
2320 // |
2321 // Proj #5 ( oop result )
2322 // top |
2323 // \ |
2324 // AddP ( base == top )
2325 //
2326 // case #4. Array's element reference:
2327 // {CheckCastPP | CastPP}
2328 // | | |
2329 // | AddP ( array's element offset )
2330 // | |
2331 // AddP ( array's offset )
2332 //
2333 // case #5. Raw object's field reference for arraycopy stub call:
2334 // The inline_native_clone() case when the arraycopy stub is called
2335 // after the allocation before Initialize and CheckCastPP nodes.
2336 // Allocate
2337 // |
2338 // Proj #5 ( oop result )
2339 // | |
2340 // AddP ( base == address )
2341 //
2342 // case #6. Constant Pool, ThreadLocal, CastX2P or
2343 // Raw object's field reference:
2344 // {ConP, ThreadLocal, CastX2P, raw Load}
2345 // top |
2346 // \ |
2347 // AddP ( base == top )
2348 //
2349 // case #7. Klass's field reference.
2350 // LoadKlass
2351 // | |
2352 // AddP ( base == address )
2353 //
2354 // case #8. narrow Klass's field reference.
2355 // LoadNKlass
2356 // |
2357 // DecodeN
2358 // | |
2359 // AddP ( base == address )
2360 //
2361 // case #9. Mixed unsafe access
2362 // {instance}
2363 // |
2364 // CheckCastPP (raw)
2365 // top |
2366 // \ |
2367 // AddP ( base == top )
2368 //
2369 Node *base = addp->in(AddPNode::Base);
2370 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
2371 base = addp->in(AddPNode::Address);
2372 while (base->is_AddP()) {
2373 // Case #6 (unsafe access) may have several chained AddP nodes.
2374 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2375 base = base->in(AddPNode::Address);
2376 }
2377 if (base->Opcode() == Op_CheckCastPP &&
2378 base->bottom_type()->isa_rawptr() &&
2379 _igvn->type(base->in(1))->isa_oopptr()) {
2380 base = base->in(1); // Case #9
2381 } else {
2382 Node* uncast_base = base->uncast();
2383 int opcode = uncast_base->Opcode();
2384 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2385 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2386 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2387 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2388 }
2389 }
2390 return base;
2391 }
2392
2393 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2394 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2395 Node* addp2 = addp->raw_out(0);
2396 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2397 addp2->in(AddPNode::Base) == n &&
2398 addp2->in(AddPNode::Address) == addp) {
2399 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2400 //
2401 // Find array's offset to push it on worklist first and
2402 // as result process an array's element offset first (pushed second)
2403 // to avoid CastPP for the array's offset.
2404 // Otherwise the inserted CastPP (LocalVar) will point to what
2405 // the AddP (Field) points to. Which would be wrong since
2406 // the algorithm expects the CastPP has the same point as
2407 // as AddP's base CheckCastPP (LocalVar).
2408 //
2409 // ArrayAllocation
2410 // |
2411 // CheckCastPP
2412 // |
2413 // memProj (from ArrayAllocation CheckCastPP)
2414 // | ||
2415 // | || Int (element index)
2416 // | || | ConI (log(element size))
2417 // | || | /
2418 // | || LShift
2419 // | || /
2420 // | AddP (array's element offset)
2421 // | |
2422 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2423 // | / /
2424 // AddP (array's offset)
2425 // |
2426 // Load/Store (memory operation on array's element)
2427 //
2428 return addp2;
2429 }
2430 return NULL;
2431 }
2432
2433 //
2434 // Adjust the type and inputs of an AddP which computes the
2435 // address of a field of an instance
2436 //
2437 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2438 PhaseGVN* igvn = _igvn;
2439 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2440 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2441 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2442 if (t == NULL) {
2443 // We are computing a raw address for a store captured by an Initialize
2444 // compute an appropriate address type (cases #3 and #5).
2445 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2446 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2447 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2448 assert(offs != Type::OffsetBot, "offset must be a constant");
2449 if (base_t->isa_aryptr() != NULL) {
2450 // In the case of a flattened value type array, each field has its
2451 // own slice so we need to extract the field being accessed from
2452 // the address computation
2453 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
2454 } else {
2455 t = base_t->add_offset(offs)->is_oopptr();
2456 }
2457 }
2458 int inst_id = base_t->instance_id();
2459 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2460 "old type must be non-instance or match new type");
2461
2462 // The type 't' could be subclass of 'base_t'.
2463 // As result t->offset() could be large then base_t's size and it will
2464 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2465 // constructor verifies correctness of the offset.
2466 //
2467 // It could happened on subclass's branch (from the type profiling
2468 // inlining) which was not eliminated during parsing since the exactness
2469 // of the allocation type was not propagated to the subclass type check.
2470 //
2471 // Or the type 't' could be not related to 'base_t' at all.
2472 // It could happen when CHA type is different from MDO type on a dead path
2473 // (for example, from instanceof check) which is not collapsed during parsing.
2474 //
2475 // Do nothing for such AddP node and don't process its users since
2476 // this code branch will go away.
2477 //
2478 if (!t->is_known_instance() &&
2479 !base_t->klass()->is_subtype_of(t->klass())) {
2480 return false; // bail out
2481 }
2482 const TypePtr* tinst = base_t->add_offset(t->offset());
2483 if (tinst->isa_aryptr() && t->isa_aryptr()) {
2484 // In the case of a flattened value type array, each field has its
2485 // own slice so we need to keep track of the field being accessed.
2486 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
2487 }
2488
2489 // Do NOT remove the next line: ensure a new alias index is allocated
2490 // for the instance type. Note: C++ will not remove it since the call
2491 // has side effect.
2492 int alias_idx = _compile->get_alias_index(tinst);
2493 igvn->set_type(addp, tinst);
2494 // record the allocation in the node map
2495 set_map(addp, get_map(base->_idx));
2496 // Set addp's Base and Address to 'base'.
2497 Node *abase = addp->in(AddPNode::Base);
2498 Node *adr = addp->in(AddPNode::Address);
2499 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2500 adr->in(0)->_idx == (uint)inst_id) {
2501 // Skip AddP cases #3 and #5.
2502 } else {
2503 assert(!abase->is_top(), "sanity"); // AddP case #3
2504 if (abase != base) {
2505 igvn->hash_delete(addp);
2506 addp->set_req(AddPNode::Base, base);
2507 if (abase == adr) {
2508 addp->set_req(AddPNode::Address, base);
2509 } else {
2510 // AddP case #4 (adr is array's element offset AddP node)
2511 #ifdef ASSERT
2512 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2513 assert(adr->is_AddP() && atype != NULL &&
2514 atype->instance_id() == inst_id, "array's element offset should be processed first");
2515 #endif
2516 }
2517 igvn->hash_insert(addp);
2518 }
2519 }
2520 // Put on IGVN worklist since at least addp's type was changed above.
2521 record_for_optimizer(addp);
2522 return true;
2523 }
2524
2525 //
2526 // Create a new version of orig_phi if necessary. Returns either the newly
2527 // created phi or an existing phi. Sets create_new to indicate whether a new
2528 // phi was created. Cache the last newly created phi in the node map.
2529 //
2530 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2531 Compile *C = _compile;
2532 PhaseGVN* igvn = _igvn;
2533 new_created = false;
2534 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2535 // nothing to do if orig_phi is bottom memory or matches alias_idx
2536 if (phi_alias_idx == alias_idx) {
2537 return orig_phi;
2538 }
2539 // Have we recently created a Phi for this alias index?
2540 PhiNode *result = get_map_phi(orig_phi->_idx);
2541 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2542 return result;
2543 }
2544 // Previous check may fail when the same wide memory Phi was split into Phis
2545 // for different memory slices. Search all Phis for this region.
2546 if (result != NULL) {
2547 Node* region = orig_phi->in(0);
2548 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2549 Node* phi = region->fast_out(i);
2550 if (phi->is_Phi() &&
2551 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2552 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2553 return phi->as_Phi();
2554 }
2555 }
2556 }
2557 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
2558 if (C->do_escape_analysis() == true && !C->failing()) {
2559 // Retry compilation without escape analysis.
2560 // If this is the first failure, the sentinel string will "stick"
2561 // to the Compile object, and the C2Compiler will see it and retry.
2562 C->record_failure(C2Compiler::retry_no_escape_analysis());
2563 }
2564 return NULL;
2565 }
2566 orig_phi_worklist.append_if_missing(orig_phi);
2567 const TypePtr *atype = C->get_adr_type(alias_idx);
2568 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2569 C->copy_node_notes_to(result, orig_phi);
2570 igvn->set_type(result, result->bottom_type());
2571 record_for_optimizer(result);
2572 set_map(orig_phi, result);
2573 new_created = true;
2574 return result;
2575 }
2576
2577 //
2578 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2579 // specified alias index.
2580 //
2581 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2582 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2583 Compile *C = _compile;
2584 PhaseGVN* igvn = _igvn;
2585 bool new_phi_created;
2586 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2587 if (!new_phi_created) {
2588 return result;
2589 }
2590 GrowableArray<PhiNode *> phi_list;
2591 GrowableArray<uint> cur_input;
2592 PhiNode *phi = orig_phi;
2593 uint idx = 1;
2594 bool finished = false;
2595 while(!finished) {
2596 while (idx < phi->req()) {
2597 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2598 if (mem != NULL && mem->is_Phi()) {
2599 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2600 if (new_phi_created) {
2601 // found an phi for which we created a new split, push current one on worklist and begin
2602 // processing new one
2603 phi_list.push(phi);
2604 cur_input.push(idx);
2605 phi = mem->as_Phi();
2606 result = newphi;
2607 idx = 1;
2608 continue;
2609 } else {
2610 mem = newphi;
2611 }
2612 }
2613 if (C->failing()) {
2614 return NULL;
2615 }
2616 result->set_req(idx++, mem);
2617 }
2618 #ifdef ASSERT
2619 // verify that the new Phi has an input for each input of the original
2620 assert( phi->req() == result->req(), "must have same number of inputs.");
2621 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2622 #endif
2623 // Check if all new phi's inputs have specified alias index.
2624 // Otherwise use old phi.
2625 for (uint i = 1; i < phi->req(); i++) {
2626 Node* in = result->in(i);
2627 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2628 }
2629 // we have finished processing a Phi, see if there are any more to do
2630 finished = (phi_list.length() == 0 );
2631 if (!finished) {
2632 phi = phi_list.pop();
2633 idx = cur_input.pop();
2634 PhiNode *prev_result = get_map_phi(phi->_idx);
2635 prev_result->set_req(idx++, result);
2636 result = prev_result;
2637 }
2638 }
2639 return result;
2640 }
2641
2642 //
2643 // The next methods are derived from methods in MemNode.
2644 //
2645 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2646 Node *mem = mmem;
2647 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2648 // means an array I have not precisely typed yet. Do not do any
2649 // alias stuff with it any time soon.
2650 if (toop->base() != Type::AnyPtr &&
2651 !(toop->klass() != NULL &&
2652 toop->klass()->is_java_lang_Object() &&
2653 toop->offset() == Type::OffsetBot)) {
2654 mem = mmem->memory_at(alias_idx);
2655 // Update input if it is progress over what we have now
2656 }
2657 return mem;
2658 }
2659
2660 //
2661 // Move memory users to their memory slices.
2662 //
2663 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2664 Compile* C = _compile;
2665 PhaseGVN* igvn = _igvn;
2666 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2667 assert(tp != NULL, "ptr type");
2668 int alias_idx = C->get_alias_index(tp);
2669 int general_idx = C->get_general_index(alias_idx);
2670
2671 // Move users first
2672 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2673 Node* use = n->fast_out(i);
2674 if (use->is_MergeMem()) {
2675 MergeMemNode* mmem = use->as_MergeMem();
2676 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2677 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2678 continue; // Nothing to do
2679 }
2680 // Replace previous general reference to mem node.
2681 uint orig_uniq = C->unique();
2682 Node* m = find_inst_mem(n, general_idx, orig_phis);
2683 assert(orig_uniq == C->unique(), "no new nodes");
2684 mmem->set_memory_at(general_idx, m);
2685 --imax;
2686 --i;
2687 } else if (use->is_MemBar()) {
2688 assert(!use->is_Initialize(), "initializing stores should not be moved");
2689 if (use->req() > MemBarNode::Precedent &&
2690 use->in(MemBarNode::Precedent) == n) {
2691 // Don't move related membars.
2692 record_for_optimizer(use);
2693 continue;
2694 }
2695 tp = use->as_MemBar()->adr_type()->isa_ptr();
2696 if ((tp != NULL && C->get_alias_index(tp) == alias_idx) ||
2697 alias_idx == general_idx) {
2698 continue; // Nothing to do
2699 }
2700 // Move to general memory slice.
2701 uint orig_uniq = C->unique();
2702 Node* m = find_inst_mem(n, general_idx, orig_phis);
2703 assert(orig_uniq == C->unique(), "no new nodes");
2704 igvn->hash_delete(use);
2705 imax -= use->replace_edge(n, m);
2706 igvn->hash_insert(use);
2707 record_for_optimizer(use);
2708 --i;
2709 #ifdef ASSERT
2710 } else if (use->is_Mem()) {
2711 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2712 // Don't move related cardmark.
2713 continue;
2714 }
2715 // Memory nodes should have new memory input.
2716 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2717 assert(tp != NULL, "ptr type");
2718 int idx = C->get_alias_index(tp);
2719 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2720 "Following memory nodes should have new memory input or be on the same memory slice");
2721 } else if (use->is_Phi()) {
2722 // Phi nodes should be split and moved already.
2723 tp = use->as_Phi()->adr_type()->isa_ptr();
2724 assert(tp != NULL, "ptr type");
2725 int idx = C->get_alias_index(tp);
2726 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2727 } else {
2728 use->dump();
2729 assert(false, "should not be here");
2730 #endif
2731 }
2732 }
2733 }
2734
2735 //
2736 // Search memory chain of "mem" to find a MemNode whose address
2737 // is the specified alias index.
2738 //
2739 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2740 if (orig_mem == NULL)
2741 return orig_mem;
2742 Compile* C = _compile;
2743 PhaseGVN* igvn = _igvn;
2744 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2745 bool is_instance = (toop != NULL) && toop->is_known_instance();
2746 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
2747 Node *prev = NULL;
2748 Node *result = orig_mem;
2749 while (prev != result) {
2750 prev = result;
2751 if (result == start_mem)
2752 break; // hit one of our sentinels
2753 if (result->is_Mem()) {
2754 const Type *at = igvn->type(result->in(MemNode::Address));
2755 if (at == Type::TOP)
2756 break; // Dead
2757 assert (at->isa_ptr() != NULL, "pointer type required.");
2758 int idx = C->get_alias_index(at->is_ptr());
2759 if (idx == alias_idx)
2760 break; // Found
2761 if (!is_instance && (at->isa_oopptr() == NULL ||
2762 !at->is_oopptr()->is_known_instance())) {
2763 break; // Do not skip store to general memory slice.
2764 }
2765 result = result->in(MemNode::Memory);
2766 }
2767 if (!is_instance)
2768 continue; // don't search further for non-instance types
2769 // skip over a call which does not affect this memory slice
2770 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2771 Node *proj_in = result->in(0);
2772 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2773 break; // hit one of our sentinels
2774 } else if (proj_in->is_Call()) {
2775 // ArrayCopy node processed here as well
2776 CallNode *call = proj_in->as_Call();
2777 if (!call->may_modify(toop, igvn)) {
2778 result = call->in(TypeFunc::Memory);
2779 }
2780 } else if (proj_in->is_Initialize()) {
2781 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2782 // Stop if this is the initialization for the object instance which
2783 // which contains this memory slice, otherwise skip over it.
2784 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2785 result = proj_in->in(TypeFunc::Memory);
2786 }
2787 } else if (proj_in->is_MemBar()) {
2788 if (proj_in->in(TypeFunc::Memory)->is_MergeMem() &&
2789 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->is_Proj() &&
2790 proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->is_ArrayCopy()) {
2791 // clone
2792 ArrayCopyNode* ac = proj_in->in(TypeFunc::Memory)->as_MergeMem()->in(Compile::AliasIdxRaw)->in(0)->as_ArrayCopy();
2793 if (ac->may_modify(toop, igvn)) {
2794 break;
2795 }
2796 }
2797 result = proj_in->in(TypeFunc::Memory);
2798 }
2799 } else if (result->is_MergeMem()) {
2800 MergeMemNode *mmem = result->as_MergeMem();
2801 result = step_through_mergemem(mmem, alias_idx, toop);
2802 if (result == mmem->base_memory()) {
2803 // Didn't find instance memory, search through general slice recursively.
2804 result = mmem->memory_at(C->get_general_index(alias_idx));
2805 result = find_inst_mem(result, alias_idx, orig_phis);
2806 if (C->failing()) {
2807 return NULL;
2808 }
2809 mmem->set_memory_at(alias_idx, result);
2810 }
2811 } else if (result->is_Phi() &&
2812 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2813 Node *un = result->as_Phi()->unique_input(igvn);
2814 if (un != NULL) {
2815 orig_phis.append_if_missing(result->as_Phi());
2816 result = un;
2817 } else {
2818 break;
2819 }
2820 } else if (result->is_ClearArray()) {
2821 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2822 // Can not bypass initialization of the instance
2823 // we are looking for.
2824 break;
2825 }
2826 // Otherwise skip it (the call updated 'result' value).
2827 } else if (result->Opcode() == Op_SCMemProj) {
2828 Node* mem = result->in(0);
2829 Node* adr = NULL;
2830 if (mem->is_LoadStore()) {
2831 adr = mem->in(MemNode::Address);
2832 } else {
2833 assert(mem->Opcode() == Op_EncodeISOArray ||
2834 mem->Opcode() == Op_StrCompressedCopy, "sanity");
2835 adr = mem->in(3); // Memory edge corresponds to destination array
2836 }
2837 const Type *at = igvn->type(adr);
2838 if (at != Type::TOP) {
2839 assert(at->isa_ptr() != NULL, "pointer type required.");
2840 int idx = C->get_alias_index(at->is_ptr());
2841 if (idx == alias_idx) {
2842 // Assert in debug mode
2843 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
2844 break; // In product mode return SCMemProj node
2845 }
2846 }
2847 result = mem->in(MemNode::Memory);
2848 } else if (result->Opcode() == Op_StrInflatedCopy) {
2849 Node* adr = result->in(3); // Memory edge corresponds to destination array
2850 const Type *at = igvn->type(adr);
2851 if (at != Type::TOP) {
2852 assert(at->isa_ptr() != NULL, "pointer type required.");
2853 int idx = C->get_alias_index(at->is_ptr());
2854 if (idx == alias_idx) {
2855 // Assert in debug mode
2856 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
2857 break; // In product mode return SCMemProj node
2858 }
2859 }
2860 result = result->in(MemNode::Memory);
2861 }
2862 }
2863 if (result->is_Phi()) {
2864 PhiNode *mphi = result->as_Phi();
2865 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2866 const TypePtr *t = mphi->adr_type();
2867 if (!is_instance) {
2868 // Push all non-instance Phis on the orig_phis worklist to update inputs
2869 // during Phase 4 if needed.
2870 orig_phis.append_if_missing(mphi);
2871 } else if (C->get_alias_index(t) != alias_idx) {
2872 // Create a new Phi with the specified alias index type.
2873 result = split_memory_phi(mphi, alias_idx, orig_phis);
2874 }
2875 }
2876 // the result is either MemNode, PhiNode, InitializeNode.
2877 return result;
2878 }
2879
2880 //
2881 // Convert the types of unescaped object to instance types where possible,
2882 // propagate the new type information through the graph, and update memory
2883 // edges and MergeMem inputs to reflect the new type.
2884 //
2885 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2886 // The processing is done in 4 phases:
2887 //
2888 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2889 // types for the CheckCastPP for allocations where possible.
2890 // Propagate the new types through users as follows:
2891 // casts and Phi: push users on alloc_worklist
2892 // AddP: cast Base and Address inputs to the instance type
2893 // push any AddP users on alloc_worklist and push any memnode
2894 // users onto memnode_worklist.
2895 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2896 // search the Memory chain for a store with the appropriate type
2897 // address type. If a Phi is found, create a new version with
2898 // the appropriate memory slices from each of the Phi inputs.
2899 // For stores, process the users as follows:
2900 // MemNode: push on memnode_worklist
2901 // MergeMem: push on mergemem_worklist
2902 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2903 // moving the first node encountered of each instance type to the
2904 // the input corresponding to its alias index.
2905 // appropriate memory slice.
2906 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2907 //
2908 // In the following example, the CheckCastPP nodes are the cast of allocation
2909 // results and the allocation of node 29 is unescaped and eligible to be an
2910 // instance type.
2911 //
2912 // We start with:
2913 //
2914 // 7 Parm #memory
2915 // 10 ConI "12"
2916 // 19 CheckCastPP "Foo"
2917 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2918 // 29 CheckCastPP "Foo"
2919 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2920 //
2921 // 40 StoreP 25 7 20 ... alias_index=4
2922 // 50 StoreP 35 40 30 ... alias_index=4
2923 // 60 StoreP 45 50 20 ... alias_index=4
2924 // 70 LoadP _ 60 30 ... alias_index=4
2925 // 80 Phi 75 50 60 Memory alias_index=4
2926 // 90 LoadP _ 80 30 ... alias_index=4
2927 // 100 LoadP _ 80 20 ... alias_index=4
2928 //
2929 //
2930 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2931 // and creating a new alias index for node 30. This gives:
2932 //
2933 // 7 Parm #memory
2934 // 10 ConI "12"
2935 // 19 CheckCastPP "Foo"
2936 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2937 // 29 CheckCastPP "Foo" iid=24
2938 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2939 //
2940 // 40 StoreP 25 7 20 ... alias_index=4
2941 // 50 StoreP 35 40 30 ... alias_index=6
2942 // 60 StoreP 45 50 20 ... alias_index=4
2943 // 70 LoadP _ 60 30 ... alias_index=6
2944 // 80 Phi 75 50 60 Memory alias_index=4
2945 // 90 LoadP _ 80 30 ... alias_index=6
2946 // 100 LoadP _ 80 20 ... alias_index=4
2947 //
2948 // In phase 2, new memory inputs are computed for the loads and stores,
2949 // And a new version of the phi is created. In phase 4, the inputs to
2950 // node 80 are updated and then the memory nodes are updated with the
2951 // values computed in phase 2. This results in:
2952 //
2953 // 7 Parm #memory
2954 // 10 ConI "12"
2955 // 19 CheckCastPP "Foo"
2956 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2957 // 29 CheckCastPP "Foo" iid=24
2958 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2959 //
2960 // 40 StoreP 25 7 20 ... alias_index=4
2961 // 50 StoreP 35 7 30 ... alias_index=6
2962 // 60 StoreP 45 40 20 ... alias_index=4
2963 // 70 LoadP _ 50 30 ... alias_index=6
2964 // 80 Phi 75 40 60 Memory alias_index=4
2965 // 120 Phi 75 50 50 Memory alias_index=6
2966 // 90 LoadP _ 120 30 ... alias_index=6
2967 // 100 LoadP _ 80 20 ... alias_index=4
2968 //
2969 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist, GrowableArray<ArrayCopyNode*> &arraycopy_worklist) {
2970 GrowableArray<Node *> memnode_worklist;
2971 GrowableArray<PhiNode *> orig_phis;
2972 PhaseIterGVN *igvn = _igvn;
2973 uint new_index_start = (uint) _compile->num_alias_types();
2974 Arena* arena = Thread::current()->resource_area();
2975 VectorSet visited(arena);
2976 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2977 uint unique_old = _compile->unique();
2978
2979 // Phase 1: Process possible allocations from alloc_worklist.
2980 // Create instance types for the CheckCastPP for allocations where possible.
2981 //
2982 // (Note: don't forget to change the order of the second AddP node on
2983 // the alloc_worklist if the order of the worklist processing is changed,
2984 // see the comment in find_second_addp().)
2985 //
2986 while (alloc_worklist.length() != 0) {
2987 Node *n = alloc_worklist.pop();
2988 uint ni = n->_idx;
2989 if (n->is_Call()) {
2990 CallNode *alloc = n->as_Call();
2991 // copy escape information to call node
2992 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2993 PointsToNode::EscapeState es = ptn->escape_state();
2994 // We have an allocation or call which returns a Java object,
2995 // see if it is unescaped.
2996 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2997 continue;
2998 // Find CheckCastPP for the allocate or for the return value of a call
2999 n = alloc->result_cast();
3000 if (n == NULL) { // No uses except Initialize node
3001 if (alloc->is_Allocate()) {
3002 // Set the scalar_replaceable flag for allocation
3003 // so it could be eliminated if it has no uses.
3004 alloc->as_Allocate()->_is_scalar_replaceable = true;
3005 }
3006 if (alloc->is_CallStaticJava()) {
3007 // Set the scalar_replaceable flag for boxing method
3008 // so it could be eliminated if it has no uses.
3009 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3010 }
3011 continue;
3012 }
3013 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
3014 assert(!alloc->is_Allocate(), "allocation should have unique type");
3015 continue;
3016 }
3017
3018 // The inline code for Object.clone() casts the allocation result to
3019 // java.lang.Object and then to the actual type of the allocated
3020 // object. Detect this case and use the second cast.
3021 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
3022 // the allocation result is cast to java.lang.Object and then
3023 // to the actual Array type.
3024 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
3025 && (alloc->is_AllocateArray() ||
3026 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
3027 Node *cast2 = NULL;
3028 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3029 Node *use = n->fast_out(i);
3030 if (use->is_CheckCastPP()) {
3031 cast2 = use;
3032 break;
3033 }
3034 }
3035 if (cast2 != NULL) {
3036 n = cast2;
3037 } else {
3038 // Non-scalar replaceable if the allocation type is unknown statically
3039 // (reflection allocation), the object can't be restored during
3040 // deoptimization without precise type.
3041 continue;
3042 }
3043 }
3044
3045 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
3046 if (t == NULL)
3047 continue; // not a TypeOopPtr
3048 if (!t->klass_is_exact())
3049 continue; // not an unique type
3050
3051 if (alloc->is_Allocate()) {
3052 // Set the scalar_replaceable flag for allocation
3053 // so it could be eliminated.
3054 alloc->as_Allocate()->_is_scalar_replaceable = true;
3055 }
3056 if (alloc->is_CallStaticJava()) {
3057 // Set the scalar_replaceable flag for boxing method
3058 // so it could be eliminated.
3059 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
3060 }
3061 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
3062 // in order for an object to be scalar-replaceable, it must be:
3063 // - a direct allocation (not a call returning an object)
3064 // - non-escaping
3065 // - eligible to be a unique type
3066 // - not determined to be ineligible by escape analysis
3067 set_map(alloc, n);
3068 set_map(n, alloc);
3069 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
3070 igvn->hash_delete(n);
3071 igvn->set_type(n, tinst);
3072 n->raise_bottom_type(tinst);
3073 igvn->hash_insert(n);
3074 record_for_optimizer(n);
3075 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
3076
3077 // First, put on the worklist all Field edges from Connection Graph
3078 // which is more accurate than putting immediate users from Ideal Graph.
3079 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
3080 PointsToNode* tgt = e.get();
3081 if (tgt->is_Arraycopy()) {
3082 continue;
3083 }
3084 Node* use = tgt->ideal_node();
3085 assert(tgt->is_Field() && use->is_AddP(),
3086 "only AddP nodes are Field edges in CG");
3087 if (use->outcnt() > 0) { // Don't process dead nodes
3088 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
3089 if (addp2 != NULL) {
3090 assert(alloc->is_AllocateArray(),"array allocation was expected");
3091 alloc_worklist.append_if_missing(addp2);
3092 }
3093 alloc_worklist.append_if_missing(use);
3094 }
3095 }
3096
3097 // An allocation may have an Initialize which has raw stores. Scan
3098 // the users of the raw allocation result and push AddP users
3099 // on alloc_worklist.
3100 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
3101 assert (raw_result != NULL, "must have an allocation result");
3102 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
3103 Node *use = raw_result->fast_out(i);
3104 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
3105 Node* addp2 = find_second_addp(use, raw_result);
3106 if (addp2 != NULL) {
3107 assert(alloc->is_AllocateArray(),"array allocation was expected");
3108 alloc_worklist.append_if_missing(addp2);
3109 }
3110 alloc_worklist.append_if_missing(use);
3111 } else if (use->is_MemBar()) {
3112 memnode_worklist.append_if_missing(use);
3113 }
3114 }
3115 }
3116 } else if (n->is_AddP()) {
3117 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
3118 if (jobj == NULL || jobj == phantom_obj) {
3119 #ifdef ASSERT
3120 ptnode_adr(get_addp_base(n)->_idx)->dump();
3121 ptnode_adr(n->_idx)->dump();
3122 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3123 #endif
3124 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3125 return;
3126 }
3127 Node *base = get_map(jobj->idx()); // CheckCastPP node
3128 if (!split_AddP(n, base)) continue; // wrong type from dead path
3129 } else if (n->is_Phi() ||
3130 n->is_CheckCastPP() ||
3131 n->is_EncodeP() ||
3132 n->is_DecodeN() ||
3133 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
3134 if (visited.test_set(n->_idx)) {
3135 assert(n->is_Phi(), "loops only through Phi's");
3136 continue; // already processed
3137 }
3138 JavaObjectNode* jobj = unique_java_object(n);
3139 if (jobj == NULL || jobj == phantom_obj) {
3140 #ifdef ASSERT
3141 ptnode_adr(n->_idx)->dump();
3142 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
3143 #endif
3144 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
3145 return;
3146 } else {
3147 Node *val = get_map(jobj->idx()); // CheckCastPP node
3148 TypeNode *tn = n->as_Type();
3149 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
3150 assert(tinst != NULL && tinst->is_known_instance() &&
3151 tinst->instance_id() == jobj->idx() , "instance type expected.");
3152
3153 const Type *tn_type = igvn->type(tn);
3154 const TypeOopPtr *tn_t;
3155 if (tn_type->isa_narrowoop()) {
3156 tn_t = tn_type->make_ptr()->isa_oopptr();
3157 } else {
3158 tn_t = tn_type->isa_oopptr();
3159 }
3160 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
3161 if (tn_type->isa_narrowoop()) {
3162 tn_type = tinst->make_narrowoop();
3163 } else {
3164 tn_type = tinst;
3165 }
3166 igvn->hash_delete(tn);
3167 igvn->set_type(tn, tn_type);
3168 tn->set_type(tn_type);
3169 igvn->hash_insert(tn);
3170 record_for_optimizer(n);
3171 } else {
3172 assert(tn_type == TypePtr::NULL_PTR ||
3173 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
3174 "unexpected type");
3175 continue; // Skip dead path with different type
3176 }
3177 }
3178 } else {
3179 debug_only(n->dump();)
3180 assert(false, "EA: unexpected node");
3181 continue;
3182 }
3183 // push allocation's users on appropriate worklist
3184 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3185 Node *use = n->fast_out(i);
3186 if(use->is_Mem() && use->in(MemNode::Address) == n) {
3187 // Load/store to instance's field
3188 memnode_worklist.append_if_missing(use);
3189 } else if (use->is_MemBar()) {
3190 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3191 memnode_worklist.append_if_missing(use);
3192 }
3193 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
3194 Node* addp2 = find_second_addp(use, n);
3195 if (addp2 != NULL) {
3196 alloc_worklist.append_if_missing(addp2);
3197 }
3198 alloc_worklist.append_if_missing(use);
3199 } else if (use->is_Phi() ||
3200 use->is_CheckCastPP() ||
3201 use->is_EncodeNarrowPtr() ||
3202 use->is_DecodeNarrowPtr() ||
3203 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
3204 alloc_worklist.append_if_missing(use);
3205 #ifdef ASSERT
3206 } else if (use->is_Mem()) {
3207 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
3208 } else if (use->is_MergeMem()) {
3209 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3210 } else if (use->is_SafePoint()) {
3211 // Look for MergeMem nodes for calls which reference unique allocation
3212 // (through CheckCastPP nodes) even for debug info.
3213 Node* m = use->in(TypeFunc::Memory);
3214 if (m->is_MergeMem()) {
3215 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3216 }
3217 } else if (use->Opcode() == Op_EncodeISOArray) {
3218 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3219 // EncodeISOArray overwrites destination array
3220 memnode_worklist.append_if_missing(use);
3221 }
3222 } else if (use->Opcode() == Op_Return) {
3223 assert(_compile->tf()->returns_value_type_as_fields(), "must return a value type");
3224 // Get ValueKlass by removing the tag bit from the metadata pointer
3225 Node* klass = use->in(TypeFunc::Parms);
3226 intptr_t ptr = igvn->type(klass)->isa_rawptr()->get_con();
3227 clear_nth_bit(ptr, 0);
3228 assert(Metaspace::contains((void*)ptr), "should be klass");
3229 assert(((ValueKlass*)ptr)->contains_oops(), "returned value type must contain a reference field");
3230 } else {
3231 uint op = use->Opcode();
3232 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
3233 (use->in(MemNode::Memory) == n)) {
3234 // They overwrite memory edge corresponding to destination array,
3235 memnode_worklist.append_if_missing(use);
3236 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
3237 op == Op_CastP2X || op == Op_StoreCM ||
3238 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3239 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3240 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
3241 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3242 op == Op_ValueType)) {
3243 n->dump();
3244 use->dump();
3245 assert(false, "EA: missing allocation reference path");
3246 }
3247 #endif
3248 }
3249 }
3250
3251 }
3252
3253 // Go over all ArrayCopy nodes and if one of the inputs has a unique
3254 // type, record it in the ArrayCopy node so we know what memory this
3255 // node uses/modified.
3256 for (int next = 0; next < arraycopy_worklist.length(); next++) {
3257 ArrayCopyNode* ac = arraycopy_worklist.at(next);
3258 Node* dest = ac->in(ArrayCopyNode::Dest);
3259 if (dest->is_AddP()) {
3260 dest = get_addp_base(dest);
3261 }
3262 JavaObjectNode* jobj = unique_java_object(dest);
3263 if (jobj != NULL) {
3264 Node *base = get_map(jobj->idx());
3265 if (base != NULL) {
3266 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3267 ac->_dest_type = base_t;
3268 }
3269 }
3270 Node* src = ac->in(ArrayCopyNode::Src);
3271 if (src->is_AddP()) {
3272 src = get_addp_base(src);
3273 }
3274 jobj = unique_java_object(src);
3275 if (jobj != NULL) {
3276 Node* base = get_map(jobj->idx());
3277 if (base != NULL) {
3278 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
3279 ac->_src_type = base_t;
3280 }
3281 }
3282 }
3283
3284 // New alias types were created in split_AddP().
3285 uint new_index_end = (uint) _compile->num_alias_types();
3286 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3287
3288 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3289 // compute new values for Memory inputs (the Memory inputs are not
3290 // actually updated until phase 4.)
3291 if (memnode_worklist.length() == 0)
3292 return; // nothing to do
3293 while (memnode_worklist.length() != 0) {
3294 Node *n = memnode_worklist.pop();
3295 if (visited.test_set(n->_idx))
3296 continue;
3297 if (n->is_Phi() || n->is_ClearArray()) {
3298 // we don't need to do anything, but the users must be pushed
3299 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3300 // we don't need to do anything, but the users must be pushed
3301 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
3302 if (n == NULL)
3303 continue;
3304 } else if (n->Opcode() == Op_StrCompressedCopy ||
3305 n->Opcode() == Op_EncodeISOArray) {
3306 // get the memory projection
3307 n = n->find_out_with(Op_SCMemProj);
3308 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3309 } else {
3310 assert(n->is_Mem(), "memory node required.");
3311 Node *addr = n->in(MemNode::Address);
3312 const Type *addr_t = igvn->type(addr);
3313 if (addr_t == Type::TOP)
3314 continue;
3315 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3316 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3317 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3318 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3319 if (_compile->failing()) {
3320 return;
3321 }
3322 if (mem != n->in(MemNode::Memory)) {
3323 // We delay the memory edge update since we need old one in
3324 // MergeMem code below when instances memory slices are separated.
3325 set_map(n, mem);
3326 }
3327 if (n->is_Load()) {
3328 continue; // don't push users
3329 } else if (n->is_LoadStore()) {
3330 // get the memory projection
3331 n = n->find_out_with(Op_SCMemProj);
3332 assert(n != NULL && n->Opcode() == Op_SCMemProj, "memory projection required");
3333 }
3334 }
3335 // push user on appropriate worklist
3336 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3337 Node *use = n->fast_out(i);
3338 if (use->is_Phi() || use->is_ClearArray()) {
3339 memnode_worklist.append_if_missing(use);
3340 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3341 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3342 continue;
3343 memnode_worklist.append_if_missing(use);
3344 } else if (use->is_MemBar()) {
3345 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3346 memnode_worklist.append_if_missing(use);
3347 }
3348 #ifdef ASSERT
3349 } else if(use->is_Mem()) {
3350 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3351 } else if (use->is_MergeMem()) {
3352 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3353 } else if (use->Opcode() == Op_EncodeISOArray) {
3354 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3355 // EncodeISOArray overwrites destination array
3356 memnode_worklist.append_if_missing(use);
3357 }
3358 } else {
3359 uint op = use->Opcode();
3360 if ((use->in(MemNode::Memory) == n) &&
3361 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
3362 // They overwrite memory edge corresponding to destination array,
3363 memnode_worklist.append_if_missing(use);
3364 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
3365 op == Op_AryEq || op == Op_StrComp || op == Op_HasNegatives ||
3366 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
3367 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
3368 n->dump();
3369 use->dump();
3370 assert(false, "EA: missing memory path");
3371 }
3372 #endif
3373 }
3374 }
3375 }
3376
3377 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3378 // Walk each memory slice moving the first node encountered of each
3379 // instance type to the input corresponding to its alias index.
3380 uint length = _mergemem_worklist.length();
3381 for( uint next = 0; next < length; ++next ) {
3382 MergeMemNode* nmm = _mergemem_worklist.at(next);
3383 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3384 // Note: we don't want to use MergeMemStream here because we only want to
3385 // scan inputs which exist at the start, not ones we add during processing.
3386 // Note 2: MergeMem may already contains instance memory slices added
3387 // during find_inst_mem() call when memory nodes were processed above.
3388 igvn->hash_delete(nmm);
3389 uint nslices = MIN2(nmm->req(), new_index_start);
3390 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3391 Node* mem = nmm->in(i);
3392 Node* cur = NULL;
3393 if (mem == NULL || mem->is_top())
3394 continue;
3395 // First, update mergemem by moving memory nodes to corresponding slices
3396 // if their type became more precise since this mergemem was created.
3397 while (mem->is_Mem()) {
3398 const Type *at = igvn->type(mem->in(MemNode::Address));
3399 if (at != Type::TOP) {
3400 assert (at->isa_ptr() != NULL, "pointer type required.");
3401 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3402 if (idx == i) {
3403 if (cur == NULL)
3404 cur = mem;
3405 } else {
3406 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3407 nmm->set_memory_at(idx, mem);
3408 }
3409 }
3410 }
3411 mem = mem->in(MemNode::Memory);
3412 }
3413 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3414 // Find any instance of the current type if we haven't encountered
3415 // already a memory slice of the instance along the memory chain.
3416 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3417 if((uint)_compile->get_general_index(ni) == i) {
3418 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3419 if (nmm->is_empty_memory(m)) {
3420 Node* result = find_inst_mem(mem, ni, orig_phis);
3421 if (_compile->failing()) {
3422 return;
3423 }
3424 nmm->set_memory_at(ni, result);
3425 }
3426 }
3427 }
3428 }
3429 // Find the rest of instances values
3430 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3431 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3432 Node* result = step_through_mergemem(nmm, ni, tinst);
3433 if (result == nmm->base_memory()) {
3434 // Didn't find instance memory, search through general slice recursively.
3435 result = nmm->memory_at(_compile->get_general_index(ni));
3436 result = find_inst_mem(result, ni, orig_phis);
3437 if (_compile->failing()) {
3438 return;
3439 }
3440 nmm->set_memory_at(ni, result);
3441 }
3442 }
3443 igvn->hash_insert(nmm);
3444 record_for_optimizer(nmm);
3445 }
3446
3447 // Phase 4: Update the inputs of non-instance memory Phis and
3448 // the Memory input of memnodes
3449 // First update the inputs of any non-instance Phi's from
3450 // which we split out an instance Phi. Note we don't have
3451 // to recursively process Phi's encountered on the input memory
3452 // chains as is done in split_memory_phi() since they will
3453 // also be processed here.
3454 for (int j = 0; j < orig_phis.length(); j++) {
3455 PhiNode *phi = orig_phis.at(j);
3456 int alias_idx = _compile->get_alias_index(phi->adr_type());
3457 igvn->hash_delete(phi);
3458 for (uint i = 1; i < phi->req(); i++) {
3459 Node *mem = phi->in(i);
3460 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3461 if (_compile->failing()) {
3462 return;
3463 }
3464 if (mem != new_mem) {
3465 phi->set_req(i, new_mem);
3466 }
3467 }
3468 igvn->hash_insert(phi);
3469 record_for_optimizer(phi);
3470 }
3471
3472 // Update the memory inputs of MemNodes with the value we computed
3473 // in Phase 2 and move stores memory users to corresponding memory slices.
3474 // Disable memory split verification code until the fix for 6984348.
3475 // Currently it produces false negative results since it does not cover all cases.
3476 #if 0 // ifdef ASSERT
3477 visited.Reset();
3478 Node_Stack old_mems(arena, _compile->unique() >> 2);
3479 #endif
3480 for (uint i = 0; i < ideal_nodes.size(); i++) {
3481 Node* n = ideal_nodes.at(i);
3482 Node* nmem = get_map(n->_idx);
3483 assert(nmem != NULL, "sanity");
3484 if (n->is_Mem()) {
3485 #if 0 // ifdef ASSERT
3486 Node* old_mem = n->in(MemNode::Memory);
3487 if (!visited.test_set(old_mem->_idx)) {
3488 old_mems.push(old_mem, old_mem->outcnt());
3489 }
3490 #endif
3491 assert(n->in(MemNode::Memory) != nmem, "sanity");
3492 if (!n->is_Load()) {
3493 // Move memory users of a store first.
3494 move_inst_mem(n, orig_phis);
3495 }
3496 // Now update memory input
3497 igvn->hash_delete(n);
3498 n->set_req(MemNode::Memory, nmem);
3499 igvn->hash_insert(n);
3500 record_for_optimizer(n);
3501 } else {
3502 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3503 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3504 }
3505 }
3506 #if 0 // ifdef ASSERT
3507 // Verify that memory was split correctly
3508 while (old_mems.is_nonempty()) {
3509 Node* old_mem = old_mems.node();
3510 uint old_cnt = old_mems.index();
3511 old_mems.pop();
3512 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3513 }
3514 #endif
3515 }
3516
3517 #ifndef PRODUCT
3518 static const char *node_type_names[] = {
3519 "UnknownType",
3520 "JavaObject",
3521 "LocalVar",
3522 "Field",
3523 "Arraycopy"
3524 };
3525
3526 static const char *esc_names[] = {
3527 "UnknownEscape",
3528 "NoEscape",
3529 "ArgEscape",
3530 "GlobalEscape"
3531 };
3532
3533 void PointsToNode::dump(bool print_state) const {
3534 NodeType nt = node_type();
3535 tty->print("%s ", node_type_names[(int) nt]);
3536 if (print_state) {
3537 EscapeState es = escape_state();
3538 EscapeState fields_es = fields_escape_state();
3539 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3540 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3541 tty->print("NSR ");
3542 }
3543 if (is_Field()) {
3544 FieldNode* f = (FieldNode*)this;
3545 if (f->is_oop())
3546 tty->print("oop ");
3547 if (f->offset() > 0)
3548 tty->print("+%d ", f->offset());
3549 tty->print("(");
3550 for (BaseIterator i(f); i.has_next(); i.next()) {
3551 PointsToNode* b = i.get();
3552 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3553 }
3554 tty->print(" )");
3555 }
3556 tty->print("[");
3557 for (EdgeIterator i(this); i.has_next(); i.next()) {
3558 PointsToNode* e = i.get();
3559 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3560 }
3561 tty->print(" [");
3562 for (UseIterator i(this); i.has_next(); i.next()) {
3563 PointsToNode* u = i.get();
3564 bool is_base = false;
3565 if (PointsToNode::is_base_use(u)) {
3566 is_base = true;
3567 u = PointsToNode::get_use_node(u)->as_Field();
3568 }
3569 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3570 }
3571 tty->print(" ]] ");
3572 if (_node == NULL)
3573 tty->print_cr("<null>");
3574 else
3575 _node->dump();
3576 }
3577
3578 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3579 bool first = true;
3580 int ptnodes_length = ptnodes_worklist.length();
3581 for (int i = 0; i < ptnodes_length; i++) {
3582 PointsToNode *ptn = ptnodes_worklist.at(i);
3583 if (ptn == NULL || !ptn->is_JavaObject())
3584 continue;
3585 PointsToNode::EscapeState es = ptn->escape_state();
3586 if ((es != PointsToNode::NoEscape) && !Verbose) {
3587 continue;
3588 }
3589 Node* n = ptn->ideal_node();
3590 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3591 n->as_CallStaticJava()->is_boxing_method())) {
3592 if (first) {
3593 tty->cr();
3594 tty->print("======== Connection graph for ");
3595 _compile->method()->print_short_name();
3596 tty->cr();
3597 first = false;
3598 }
3599 ptn->dump();
3600 // Print all locals and fields which reference this allocation
3601 for (UseIterator j(ptn); j.has_next(); j.next()) {
3602 PointsToNode* use = j.get();
3603 if (use->is_LocalVar()) {
3604 use->dump(Verbose);
3605 } else if (Verbose) {
3606 use->dump();
3607 }
3608 }
3609 tty->cr();
3610 }
3611 }
3612 }
3613 #endif