1 /*
2 * Copyright (c) 2005, 2025, 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 "ci/bcEscapeAnalyzer.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/metaspace.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/escape.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "utilities/macros.hpp"
48
49 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
50 // If ReduceAllocationMerges is enabled we might call split_through_phi during
51 // split_unique_types and that will create additional nodes that need to be
52 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
53 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
54 // the array will be reallocated.
55 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
56 _in_worklist(C->comp_arena()),
57 _next_pidx(0),
58 _collecting(true),
59 _verify(false),
60 _compile(C),
61 _igvn(igvn),
62 _invocation(invocation),
63 _build_iterations(0),
64 _build_time(0.),
65 _node_map(C->comp_arena()) {
66 // Add unknown java object.
67 add_java_object(C->top(), PointsToNode::GlobalEscape);
68 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
69 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
70 // Add ConP and ConN null oop nodes
71 Node* oop_null = igvn->zerocon(T_OBJECT);
72 assert(oop_null->_idx < nodes_size(), "should be created already");
73 add_java_object(oop_null, PointsToNode::NoEscape);
74 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
75 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
76 if (UseCompressedOops) {
77 Node* noop_null = igvn->zerocon(T_NARROWOOP);
78 assert(noop_null->_idx < nodes_size(), "should be created already");
79 map_ideal_node(noop_null, null_obj);
80 }
81 }
82
83 bool ConnectionGraph::has_candidates(Compile *C) {
84 // EA brings benefits only when the code has allocations and/or locks which
85 // are represented by ideal Macro nodes.
86 int cnt = C->macro_count();
87 for (int i = 0; i < cnt; i++) {
88 Node *n = C->macro_node(i);
89 if (n->is_Allocate()) {
90 return true;
91 }
92 if (n->is_Lock()) {
93 Node* obj = n->as_Lock()->obj_node()->uncast();
94 if (!(obj->is_Parm() || obj->is_Con())) {
95 return true;
96 }
97 }
98 if (n->is_CallStaticJava() &&
99 n->as_CallStaticJava()->is_boxing_method()) {
100 return true;
101 }
102 }
103 return false;
104 }
105
106 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
107 Compile::TracePhase tp(Phase::_t_escapeAnalysis);
108 ResourceMark rm;
109
110 // Add ConP and ConN null oop nodes before ConnectionGraph construction
111 // to create space for them in ConnectionGraph::_nodes[].
112 Node* oop_null = igvn->zerocon(T_OBJECT);
113 Node* noop_null = igvn->zerocon(T_NARROWOOP);
114 int invocation = 0;
115 if (C->congraph() != nullptr) {
116 invocation = C->congraph()->_invocation + 1;
117 }
118 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
119 NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(congraph);)
120 // Perform escape analysis
121 if (congraph->compute_escape()) {
122 // There are non escaping objects.
123 C->set_congraph(congraph);
124 }
125 NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(nullptr);)
126 // Cleanup.
127 if (oop_null->outcnt() == 0) {
128 igvn->hash_delete(oop_null);
129 }
130 if (noop_null->outcnt() == 0) {
131 igvn->hash_delete(noop_null);
132 }
133
134 C->print_method(PHASE_AFTER_EA, 2);
135 }
136
137 bool ConnectionGraph::compute_escape() {
138 Compile* C = _compile;
139 PhaseGVN* igvn = _igvn;
140
141 // Worklists used by EA.
142 Unique_Node_List delayed_worklist;
143 Unique_Node_List reducible_merges;
144 GrowableArray<Node*> alloc_worklist;
145 GrowableArray<Node*> ptr_cmp_worklist;
146 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
147 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
148 GrowableArray<PointsToNode*> ptnodes_worklist;
149 GrowableArray<JavaObjectNode*> java_objects_worklist;
150 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
151 GrowableArray<FieldNode*> oop_fields_worklist;
152 GrowableArray<SafePointNode*> sfn_worklist;
153 GrowableArray<MergeMemNode*> mergemem_worklist;
154 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
155
156 { Compile::TracePhase tp(Phase::_t_connectionGraph);
157
158 // 1. Populate Connection Graph (CG) with PointsTo nodes.
159 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
160 // Initialize worklist
161 if (C->root() != nullptr) {
162 ideal_nodes.push(C->root());
163 }
164 // Processed ideal nodes are unique on ideal_nodes list
165 // but several ideal nodes are mapped to the phantom_obj.
166 // To avoid duplicated entries on the following worklists
167 // add the phantom_obj only once to them.
168 ptnodes_worklist.append(phantom_obj);
169 java_objects_worklist.append(phantom_obj);
170 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
171 Node* n = ideal_nodes.at(next);
172 if ((n->Opcode() == Op_LoadX || n->Opcode() == Op_StoreX) &&
173 !n->in(MemNode::Address)->is_AddP() &&
174 _igvn->type(n->in(MemNode::Address))->isa_oopptr()) {
175 // Load/Store at mark work address is at offset 0 so has no AddP which confuses EA
176 Node* addp = new AddPNode(n->in(MemNode::Address), n->in(MemNode::Address), _igvn->MakeConX(0));
177 _igvn->register_new_node_with_optimizer(addp);
178 _igvn->replace_input_of(n, MemNode::Address, addp);
179 ideal_nodes.push(addp);
180 _nodes.at_put_grow(addp->_idx, nullptr, nullptr);
181 }
182 // Create PointsTo nodes and add them to Connection Graph. Called
183 // only once per ideal node since ideal_nodes is Unique_Node list.
184 add_node_to_connection_graph(n, &delayed_worklist);
185 PointsToNode* ptn = ptnode_adr(n->_idx);
186 if (ptn != nullptr && ptn != phantom_obj) {
187 ptnodes_worklist.append(ptn);
188 if (ptn->is_JavaObject()) {
189 java_objects_worklist.append(ptn->as_JavaObject());
190 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
191 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
192 // Only allocations and java static calls results are interesting.
193 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
194 }
195 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
196 oop_fields_worklist.append(ptn->as_Field());
197 }
198 }
199 // Collect some interesting nodes for further use.
200 switch (n->Opcode()) {
201 case Op_MergeMem:
202 // Collect all MergeMem nodes to add memory slices for
203 // scalar replaceable objects in split_unique_types().
204 mergemem_worklist.append(n->as_MergeMem());
205 break;
206 case Op_CmpP:
207 case Op_CmpN:
208 // Collect compare pointers nodes.
209 if (OptimizePtrCompare) {
210 ptr_cmp_worklist.append(n);
211 }
212 break;
213 case Op_MemBarStoreStore:
214 // Collect all MemBarStoreStore nodes so that depending on the
215 // escape status of the associated Allocate node some of them
216 // may be eliminated.
217 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
218 storestore_worklist.append(n->as_MemBarStoreStore());
219 }
220 // If MemBarStoreStore has a precedent edge add it to the worklist (like MemBarRelease)
221 case Op_MemBarRelease:
222 if (n->req() > MemBarNode::Precedent) {
223 record_for_optimizer(n);
224 }
225 break;
226 #ifdef ASSERT
227 case Op_AddP:
228 // Collect address nodes for graph verification.
229 addp_worklist.append(n);
230 break;
231 #endif
232 case Op_ArrayCopy:
233 // Keep a list of ArrayCopy nodes so if one of its input is non
234 // escaping, we can record a unique type
235 arraycopy_worklist.append(n->as_ArrayCopy());
236 break;
237 default:
238 // not interested now, ignore...
239 break;
240 }
241 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
242 Node* m = n->fast_out(i); // Get user
243 ideal_nodes.push(m);
244 }
245 if (n->is_SafePoint()) {
246 sfn_worklist.append(n->as_SafePoint());
247 }
248 }
249
250 #ifndef PRODUCT
251 if (_compile->directive()->TraceEscapeAnalysisOption) {
252 tty->print("+++++ Initial worklist for ");
253 _compile->method()->print_name();
254 tty->print_cr(" (ea_inv=%d)", _invocation);
255 for (int i = 0; i < ptnodes_worklist.length(); i++) {
256 PointsToNode* ptn = ptnodes_worklist.at(i);
257 ptn->dump();
258 }
259 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
260 }
261 #endif
262
263 if (non_escaped_allocs_worklist.length() == 0) {
264 _collecting = false;
265 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
266 return false; // Nothing to do.
267 }
268 // Add final simple edges to graph.
269 while(delayed_worklist.size() > 0) {
270 Node* n = delayed_worklist.pop();
271 add_final_edges(n);
272 }
273
274 #ifdef ASSERT
275 if (VerifyConnectionGraph) {
276 // Verify that no new simple edges could be created and all
277 // local vars has edges.
278 _verify = true;
279 int ptnodes_length = ptnodes_worklist.length();
280 for (int next = 0; next < ptnodes_length; ++next) {
281 PointsToNode* ptn = ptnodes_worklist.at(next);
282 add_final_edges(ptn->ideal_node());
283 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
284 ptn->dump();
285 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
286 }
287 }
288 _verify = false;
289 }
290 #endif
291 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
292 // processing, calls to CI to resolve symbols (types, fields, methods)
293 // referenced in bytecode. During symbol resolution VM may throw
294 // an exception which CI cleans and converts to compilation failure.
295 if (C->failing()) {
296 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
297 return false;
298 }
299
300 _compile->print_method(PHASE_EA_AFTER_INITIAL_CONGRAPH, 4);
301
302 // 2. Finish Graph construction by propagating references to all
303 // java objects through graph.
304 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
305 java_objects_worklist, oop_fields_worklist)) {
306 // All objects escaped or hit time or iterations limits.
307 _collecting = false;
308 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
309 return false;
310 }
311
312 _compile->print_method(PHASE_EA_AFTER_COMPLETE_CONGRAPH, 4);
313
314 // 3. Adjust scalar_replaceable state of nonescaping objects and push
315 // scalar replaceable allocations on alloc_worklist for processing
316 // in split_unique_types().
317 GrowableArray<JavaObjectNode*> jobj_worklist;
318 int non_escaped_length = non_escaped_allocs_worklist.length();
319 bool found_nsr_alloc = false;
320 for (int next = 0; next < non_escaped_length; next++) {
321 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
322 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
323 Node* n = ptn->ideal_node();
324 if (n->is_Allocate()) {
325 n->as_Allocate()->_is_non_escaping = noescape;
326 }
327 if (noescape && ptn->scalar_replaceable()) {
328 adjust_scalar_replaceable_state(ptn, reducible_merges);
329 if (ptn->scalar_replaceable()) {
330 jobj_worklist.push(ptn);
331 } else {
332 found_nsr_alloc = true;
333 }
334 }
335 _compile->print_method(PHASE_EA_ADJUST_SCALAR_REPLACEABLE_ITER, 6, n);
336 }
337
338 // Propagate NSR (Not Scalar Replaceable) state.
339 if (found_nsr_alloc) {
340 find_scalar_replaceable_allocs(jobj_worklist, reducible_merges);
341 }
342
343 // alloc_worklist will be processed in reverse push order.
344 // Therefore the reducible Phis will be processed for last and that's what we
345 // want because by then the scalarizable inputs of the merge will already have
346 // an unique instance type.
347 for (uint i = 0; i < reducible_merges.size(); i++ ) {
348 Node* n = reducible_merges.at(i);
349 alloc_worklist.append(n);
350 }
351
352 for (int next = 0; next < jobj_worklist.length(); ++next) {
353 JavaObjectNode* jobj = jobj_worklist.at(next);
354 if (jobj->scalar_replaceable()) {
355 alloc_worklist.append(jobj->ideal_node());
356 }
357 }
358
359 #ifdef ASSERT
360 if (VerifyConnectionGraph) {
361 // Verify that graph is complete - no new edges could be added or needed.
362 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
363 java_objects_worklist, addp_worklist);
364 }
365 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
366 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
367 null_obj->edge_count() == 0 &&
368 !null_obj->arraycopy_src() &&
369 !null_obj->arraycopy_dst(), "sanity");
370 #endif
371
372 _collecting = false;
373
374 _compile->print_method(PHASE_EA_AFTER_PROPAGATE_NSR, 4);
375 } // TracePhase t3("connectionGraph")
376
377 // 4. Optimize ideal graph based on EA information.
378 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
379 if (has_non_escaping_obj) {
380 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
381 }
382
383 #ifndef PRODUCT
384 if (PrintEscapeAnalysis) {
385 dump(ptnodes_worklist); // Dump ConnectionGraph
386 }
387 #endif
388
389 #ifdef ASSERT
390 if (VerifyConnectionGraph) {
391 int alloc_length = alloc_worklist.length();
392 for (int next = 0; next < alloc_length; ++next) {
393 Node* n = alloc_worklist.at(next);
394 PointsToNode* ptn = ptnode_adr(n->_idx);
395 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
396 }
397 }
398
399 if (VerifyReduceAllocationMerges) {
400 for (uint i = 0; i < reducible_merges.size(); i++ ) {
401 Node* n = reducible_merges.at(i);
402 if (!can_reduce_phi(n->as_Phi())) {
403 TraceReduceAllocationMerges = true;
404 n->dump(2);
405 n->dump(-2);
406 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
407 }
408 }
409 }
410 #endif
411
412 _compile->print_method(PHASE_EA_AFTER_GRAPH_OPTIMIZATION, 4);
413
414 // 5. Separate memory graph for scalar replaceable allcations.
415 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
416 if (has_scalar_replaceable_candidates && EliminateAllocations) {
417 assert(C->do_aliasing(), "Aliasing should be enabled");
418 // Now use the escape information to create unique types for
419 // scalar replaceable objects.
420 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
421 if (C->failing()) {
422 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
423 return false;
424 }
425
426 #ifdef ASSERT
427 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
428 tty->print("=== No allocations eliminated for ");
429 C->method()->print_short_name();
430 if (!EliminateAllocations) {
431 tty->print(" since EliminateAllocations is off ===");
432 } else if(!has_scalar_replaceable_candidates) {
433 tty->print(" since there are no scalar replaceable candidates ===");
434 }
435 tty->cr();
436 #endif
437 }
438
439 // 6. Expand flat accesses if the object does not escape. This adds nodes to
440 // the graph, so it has to be after split_unique_types. This expands atomic
441 // mismatched accesses (though encapsulated in LoadFlats and StoreFlats) into
442 // non-mismatched accesses, so it is better before reduce allocation merges.
443 if (has_non_escaping_obj) {
444 optimize_flat_accesses(sfn_worklist);
445 }
446
447 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES, 4);
448
449 // 7. Reduce allocation merges used as debug information. This is done after
450 // split_unique_types because the methods used to create SafePointScalarObject
451 // need to traverse the memory graph to find values for object fields. We also
452 // set to null the scalarized inputs of reducible Phis so that the Allocate
453 // that they point can be later scalar replaced.
454 bool delay = _igvn->delay_transform();
455 _igvn->set_delay_transform(true);
456 for (uint i = 0; i < reducible_merges.size(); i++) {
457 Node* n = reducible_merges.at(i);
458 if (n->outcnt() > 0) {
459 if (!reduce_phi_on_safepoints(n->as_Phi())) {
460 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
461 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
462 return false;
463 }
464
465 // Now we set the scalar replaceable inputs of ophi to null, which is
466 // the last piece that would prevent it from being scalar replaceable.
467 reset_scalar_replaceable_entries(n->as_Phi());
468 }
469 }
470 _igvn->set_delay_transform(delay);
471
472 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
473 // java calls if they pass ArgEscape objects as parameters.
474 if (has_non_escaping_obj &&
475 (C->env()->should_retain_local_variables() ||
476 C->env()->jvmti_can_get_owned_monitor_info() ||
477 C->env()->jvmti_can_walk_any_space() ||
478 DeoptimizeObjectsALot)) {
479 int sfn_length = sfn_worklist.length();
480 for (int next = 0; next < sfn_length; next++) {
481 SafePointNode* sfn = sfn_worklist.at(next);
482 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
483 if (sfn->is_CallJava()) {
484 CallJavaNode* call = sfn->as_CallJava();
485 call->set_arg_escape(has_arg_escape(call));
486 }
487 }
488 }
489
490 _compile->print_method(PHASE_EA_AFTER_REDUCE_PHI_ON_SAFEPOINTS, 4);
491
492 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
493 return has_non_escaping_obj;
494 }
495
496 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
497 // if at least one scalar replaceable allocation participates in the merge.
498 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
499 bool found_sr_allocate = false;
500
501 for (uint i = 1; i < ophi->req(); i++) {
502 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
503 if (ptn != nullptr && ptn->scalar_replaceable()) {
504 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
505
506 // Don't handle arrays.
507 if (alloc->Opcode() != Op_Allocate) {
508 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
509 continue;
510 }
511
512 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
513 found_sr_allocate = true;
514 } else {
515 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
516 ptn->set_scalar_replaceable(false);
517 }
518 }
519 }
520
521 NOT_PRODUCT(if (TraceReduceAllocationMerges && !found_sr_allocate) tty->print_cr("Can NOT reduce Phi %d on invocation %d. No SR Allocate as input.", ophi->_idx, _invocation);)
522 return found_sr_allocate;
523 }
524
525 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
526 // I require the 'other' input to be a constant so that I can move the Cmp
527 // around safely.
528 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
529 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
530 Node* left = cmp->in(1);
531 Node* right = cmp->in(2);
532
533 return (left == n || right == n) &&
534 (left->is_Con() || right->is_Con()) &&
535 cmp->outcnt() == 1;
536 }
537
538 // We are going to check if any of the SafePointScalarMerge entries
539 // in the SafePoint reference the Phi that we are checking.
540 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
541 JVMState *jvms = sfpt->jvms();
542
543 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
544 Node* sfpt_in = sfpt->in(i);
545 if (sfpt_in->is_SafePointScalarMerge()) {
546 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
547 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
548 if (nsr_ptr == n) {
549 return true;
550 }
551 }
552 }
553
554 return false;
555 }
556
557 // Check if we are able to untangle the merge. The following patterns are
558 // supported:
559 // - Phi -> SafePoints
560 // - Phi -> CmpP/N
561 // - Phi -> AddP -> Load
562 // - Phi -> CastPP -> SafePoints
563 // - Phi -> CastPP -> AddP -> Load
564 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
565 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
566 Node* use = n->fast_out(i);
567
568 if (use->is_SafePoint()) {
569 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
570 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
571 return false;
572 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
573 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
574 return false;
575 }
576 } else if (use->is_AddP()) {
577 Node* addp = use;
578 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
579 Node* use_use = addp->fast_out(j);
580 const Type* load_type = _igvn->type(use_use);
581
582 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
583 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. AddP user isn't a [splittable] Load(): %s", n->_idx, _invocation, use_use->Name());)
584 return false;
585 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
586 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. [Narrow] Klass Load: %s", n->_idx, _invocation, use_use->Name());)
587 return false;
588 }
589 }
590 } else if (nesting > 0) {
591 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Unsupported user %s at nesting level %d.", n->_idx, _invocation, use->Name(), nesting);)
592 return false;
593 } else if (use->is_CastPP()) {
594 const Type* cast_t = _igvn->type(use);
595 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
596 #ifndef PRODUCT
597 if (TraceReduceAllocationMerges) {
598 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
599 use->dump();
600 }
601 #endif
602 return false;
603 }
604
605 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
606 if (!is_trivial_control) {
607 // If it's not a trivial control then we check if we can reduce the
608 // CmpP/N used by the If controlling the cast.
609 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
610 Node* iff = use->in(0)->in(0);
611 // We may have an OpaqueNotNull node between If and Bool nodes. But we could also have a sub class of IfNode,
612 // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
613 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
614 if (can_reduce) {
615 Node* iff_cmp = iff->in(1)->in(1);
616 int opc = iff_cmp->Opcode();
617 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
618 }
619 if (!can_reduce) {
620 #ifndef PRODUCT
621 if (TraceReduceAllocationMerges) {
622 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
623 n->dump(5);
624 }
625 #endif
626 return false;
627 }
628 }
629 }
630
631 if (!can_reduce_check_users(use, nesting+1)) {
632 return false;
633 }
634 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
635 if (!can_reduce_cmp(n, use)) {
636 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. CmpP/N %d isn't reducible.", n->_idx, _invocation, use->_idx);)
637 return false;
638 }
639 } else {
640 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. One of the uses is: %d %s", n->_idx, _invocation, use->_idx, use->Name());)
641 return false;
642 }
643 }
644
645 return true;
646 }
647
648 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
649 // only used in some certain code shapes. Check comments in
650 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
651 // details.
652 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
653 // If there was an error attempting to reduce allocation merges for this
654 // method we might have disabled the compilation and be retrying with RAM
655 // disabled.
656 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
657 return false;
658 }
659
660 const Type* phi_t = _igvn->type(ophi);
661 if (phi_t == nullptr ||
662 phi_t->make_ptr() == nullptr ||
663 phi_t->make_ptr()->isa_aryptr() != nullptr) {
664 return false;
665 }
666
667 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
668 return false;
669 }
670
671 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
672 return true;
673 }
674
675 // This method will return a CmpP/N that we need to use on the If controlling a
676 // CastPP after it was split. This method is only called on bases that are
677 // nullable therefore we always need a controlling if for the splitted CastPP.
678 //
679 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
680 // If the CastPP currently doesn't have a control then the CmpP/N will be
681 // against the null constant, otherwise it will be against the constant input of
682 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
683 // case because we have constraints on it and because the CastPP has a control
684 // input.
685 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
686 const Type* t = base->bottom_type();
687 Node* con = nullptr;
688
689 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
690 con = _igvn->zerocon(t->basic_type());
691 } else {
692 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
693 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
694 Node* bol = curr_ctrl->in(0)->in(1);
695 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
696 Node* curr_cmp = bol->in(1);
697 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
698 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
699 }
700
701 return CmpNode::make(base, con, t->basic_type());
702 }
703
704 // This method 'specializes' the CastPP passed as parameter to the base passed
705 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
706 // means that the CastPP now will be specific for a given base instead of a Phi.
707 // An If-Then-Else-Region block is inserted to control the CastPP. The control
708 // of the CastPP is a copy of the current one (if there is one) or a check
709 // against null.
710 //
711 // Before:
712 //
713 // C1 C2 ... Cn
714 // \ | /
715 // \ | /
716 // \ | /
717 // \ | /
718 // \ | /
719 // \ | /
720 // \|/
721 // Region B1 B2 ... Bn
722 // | \ | /
723 // | \ | /
724 // | \ | /
725 // | \ | /
726 // | \ | /
727 // | \ | /
728 // ---------------> Phi
729 // |
730 // X |
731 // | |
732 // | |
733 // ------> CastPP
734 //
735 // After (only partial illustration; base = B2, current_control = C2):
736 //
737 // C2
738 // |
739 // If
740 // / \
741 // / \
742 // T F
743 // /\ /
744 // / \ /
745 // / \ /
746 // C1 CastPP Reg Cn
747 // | | |
748 // | | |
749 // | | |
750 // -------------- | ----------
751 // | | |
752 // Region
753 //
754 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
755 Node* control_successor = current_control->unique_ctrl_out();
756 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
757 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
758 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
759 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
760 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
761 Node* end_region = _igvn->transform(new RegionNode(3));
762
763 // Insert the new if-else-region block into the graph
764 end_region->set_req(1, not_eq_control);
765 end_region->set_req(2, yes_eq_control);
766 control_successor->replace_edge(current_control, end_region, _igvn);
767
768 _igvn->_worklist.push(current_control);
769 _igvn->_worklist.push(control_successor);
770
771 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr));
772 }
773
774 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
775 const Type* load_type = _igvn->type(curr_load);
776 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
777 Node* memory = curr_load->in(MemNode::Memory);
778
779 // The data_phi merging the loads needs to be nullable if
780 // we are loading pointers.
781 if (load_type->make_ptr() != nullptr) {
782 if (load_type->isa_narrowoop()) {
783 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
784 } else if (load_type->isa_ptr()) {
785 load_type = load_type->meet(TypePtr::NULL_PTR);
786 } else {
787 assert(false, "Unexpected load ptr type.");
788 }
789 }
790
791 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
792
793 for (int i = 1; i < bases_for_loads->length(); i++) {
794 Node* base = bases_for_loads->at(i);
795 Node* cmp_region = nullptr;
796 if (base != nullptr) {
797 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
798 cmp_region = base->unique_ctrl_out_or_null();
799 assert(cmp_region != nullptr, "There should be.");
800 base = base->find_out_with(Op_CastPP);
801 }
802
803 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
804 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
805 Node* load = curr_load->clone();
806 load->set_req(0, nullptr);
807 load->set_req(1, mem);
808 load->set_req(2, addr);
809
810 if (cmp_region != nullptr) { // see comment on previous if
811 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
812 intermediate_phi->set_req(1, _igvn->transform(load));
813 load = intermediate_phi;
814 }
815
816 data_phi->set_req(i, _igvn->transform(load));
817 } else {
818 // Just use the default, which is already in phi
819 }
820 }
821
822 // Takes care of updating CG and split_unique_types worklists due
823 // to cloned AddP->Load.
824 updates_after_load_split(data_phi, curr_load, alloc_worklist);
825
826 return _igvn->transform(data_phi);
827 }
828
829 // This method only reduces CastPP fields loads; SafePoints are handled
830 // separately. The idea here is basically to clone the CastPP and place copies
831 // on each input of the Phi, including non-scalar replaceable inputs.
832 // Experimentation shows that the resulting IR graph is simpler that way than if
833 // we just split the cast through scalar-replaceable inputs.
834 //
835 // The reduction process requires that CastPP's control be one of:
836 // 1) no control,
837 // 2) the same region as Ophi, or
838 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
839 //
840 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
841 // flow. If the CastPP originally had an IfTrue/False control input then we'll
842 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
843 // juse use a CmpP/N against the null constant.
844 //
845 // The If-Then-Else-Region isn't always needed. For instance, if input to
846 // splitted cast was not nullable (or if it was the null constant) then we don't
847 // need (shouldn't) use a CastPP at all.
848 //
849 // After the casts are splitted we'll split the AddP->Loads through the Phi and
850 // connect them to the just split CastPPs.
851 //
852 // Before (CastPP control is same as Phi):
853 //
854 // Region Allocate Null Call
855 // | \ | /
856 // | \ | /
857 // | \ | /
858 // | \ | /
859 // | \ | /
860 // | \ | /
861 // ------------------> Phi # Oop Phi
862 // | |
863 // | |
864 // | |
865 // | |
866 // ----------------> CastPP
867 // |
868 // AddP
869 // |
870 // Load
871 //
872 // After (Very much simplified):
873 //
874 // Call Null
875 // \ /
876 // CmpP
877 // |
878 // Bool#NE
879 // |
880 // If
881 // / \
882 // T F
883 // / \ /
884 // / R
885 // CastPP |
886 // | |
887 // AddP |
888 // | |
889 // Load |
890 // \ | 0
891 // Allocate \ | /
892 // \ \ | /
893 // AddP Phi
894 // \ /
895 // Load /
896 // \ 0 /
897 // \ | /
898 // \|/
899 // Phi # "Field" Phi
900 //
901 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node*> &alloc_worklist) {
902 Node* ophi = curr_castpp->in(1);
903 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
904
905 // Identify which base should be used for AddP->Load later when spliting the
906 // CastPP->Loads through ophi. Three kind of values may be stored in this
907 // array, depending on the nullability status of the corresponding input in
908 // ophi.
909 //
910 // - nullptr: Meaning that the base is actually the null constant and therefore
911 // we won't try to load from it.
912 //
913 // - CFG Node: Meaning that the base is a CastPP that was specialized for
914 // this input of Ophi. I.e., we added an If->Then->Else-Region
915 // that will 'activate' the CastPp only when the input is not Null.
916 //
917 // - Other Node: Meaning that the base is not nullable and therefore we'll try
918 // to load directly from it.
919 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
920
921 for (uint i = 1; i < ophi->req(); i++) {
922 Node* base = ophi->in(i);
923 const Type* base_t = _igvn->type(base);
924
925 if (base_t->maybe_null()) {
926 if (base->is_Con()) {
927 // Nothing todo as bases_for_loads[i] is already null
928 } else {
929 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
930 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
931 }
932 } else {
933 bases_for_loads.at_put(i, base);
934 }
935 }
936
937 // Now let's split the CastPP->Loads through the Phi
938 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
939 Node* use = curr_castpp->raw_out(i);
940 if (use->is_AddP()) {
941 for (int j = use->outcnt()-1; j >= 0;) {
942 Node* use_use = use->raw_out(j);
943 assert(use_use->is_Load(), "Expected this to be a Load node.");
944
945 // We can't make an unconditional load from a nullable input. The
946 // 'split_castpp_load_through_phi` method will add an
947 // 'If-Then-Else-Region` around nullable bases and only load from them
948 // when the input is not null.
949 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
950 _igvn->replace_node(use_use, phi);
951
952 --j;
953 j = MIN2(j, (int)use->outcnt()-1);
954 }
955
956 _igvn->remove_dead_node(use);
957 }
958 --i;
959 i = MIN2(i, (int)curr_castpp->outcnt()-1);
960 }
961 }
962
963 // This method split a given CmpP/N through the Phi used in one of its inputs.
964 // As a result we convert a comparison with a pointer to a comparison with an
965 // integer.
966 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
967 // while the other must be a constant.
968 // The splitting process is basically just cloning the CmpP/N above the input
969 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
970 // can prove at compile time the result of the comparison.
971 //
972 // Before:
973 //
974 // in1 in2 ... inN
975 // \ | /
976 // \ | /
977 // \ | /
978 // \ | /
979 // \ | /
980 // \ | /
981 // Phi
982 // | Other
983 // | /
984 // | /
985 // | /
986 // CmpP/N
987 //
988 // After:
989 //
990 // in1 Other in2 Other inN Other
991 // | | | | | |
992 // \ | | | | |
993 // \ / | / | /
994 // CmpP/N CmpP/N CmpP/N
995 // Bool Bool Bool
996 // \ | /
997 // \ | /
998 // \ | /
999 // \ | /
1000 // \ | /
1001 // \ | /
1002 // \ | /
1003 // \ | /
1004 // Phi
1005 // |
1006 // | Zero
1007 // | /
1008 // | /
1009 // | /
1010 // CmpI
1011 //
1012 //
1013 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
1014 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
1015 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
1016
1017 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
1018 Node* zero = _igvn->intcon(0);
1019 Node* one = _igvn->intcon(1);
1020 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
1021
1022 // This Phi will merge the result of the Cmps split through the Phi
1023 Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
1024
1025 for (uint i=1; i<ophi->req(); i++) {
1026 Node* ophi_input = ophi->in(i);
1027 Node* res_phi_input = nullptr;
1028
1029 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1030 if (tcmp->singleton()) {
1031 if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
1032 (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
1033 res_phi_input = one;
1034 } else {
1035 res_phi_input = zero;
1036 }
1037 } else {
1038 Node* ncmp = _igvn->transform(cmp->clone());
1039 ncmp->set_req(1, ophi_input);
1040 ncmp->set_req(2, other);
1041 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1042 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1043 }
1044
1045 res_phi->set_req(i, res_phi_input);
1046 }
1047
1048 // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1049 Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1050 _igvn->replace_node(cmp, new_cmp);
1051 }
1052
1053 // Push the newly created AddP on alloc_worklist and patch
1054 // the connection graph. Note that the changes in the CG below
1055 // won't affect the ES of objects since the new nodes have the
1056 // same status as the old ones.
1057 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1058 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1059 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1060 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1061
1062 if (data_phi == nullptr || !data_phi->is_Phi()) {
1063 // Make this a retry?
1064 return ;
1065 }
1066
1067 Node* previous_addp = previous_load->in(MemNode::Address);
1068 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1069 for (uint i = 1; i < data_phi->req(); i++) {
1070 Node* new_load = data_phi->in(i);
1071
1072 if (new_load->is_Phi()) {
1073 // new_load is currently the "intermediate_phi" from an specialized
1074 // CastPP.
1075 new_load = new_load->in(1);
1076 }
1077
1078 // "new_load" might actually be a constant, parameter, etc.
1079 if (new_load->is_Load()) {
1080 Node* new_addp = new_load->in(MemNode::Address);
1081 Node* base = get_addp_base(new_addp);
1082
1083 // The base might not be something that we can create an unique
1084 // type for. If that's the case we are done with that input.
1085 PointsToNode* jobj_ptn = unique_java_object(base);
1086 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1087 continue;
1088 }
1089
1090 // Push to alloc_worklist since the base has an unique_type
1091 alloc_worklist.append_if_missing(new_addp);
1092
1093 // Now let's add the node to the connection graph
1094 _nodes.at_grow(new_addp->_idx, nullptr);
1095 add_field(new_addp, fn->escape_state(), fn->offset());
1096 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1097
1098 // If the load doesn't load an object then it won't be
1099 // part of the connection graph
1100 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1101 if (curr_load_ptn != nullptr) {
1102 _nodes.at_grow(new_load->_idx, nullptr);
1103 add_local_var(new_load, curr_load_ptn->escape_state());
1104 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1105 }
1106 }
1107 }
1108 }
1109
1110 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1111 // We'll pass this to 'split_through_phi' so that it'll do the split even
1112 // though the load doesn't have an unique instance type.
1113 bool ignore_missing_instance_id = true;
1114
1115 // All AddPs are present in the connection graph
1116 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1117
1118 // Iterate over AddP looking for a Load
1119 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1120 Node* previous_load = previous_addp->raw_out(k);
1121 if (previous_load->is_Load()) {
1122 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1123
1124 // Takes care of updating CG and split_unique_types worklists due to cloned
1125 // AddP->Load.
1126 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1127
1128 _igvn->replace_node(previous_load, data_phi);
1129 }
1130 --k;
1131 k = MIN2(k, (int)previous_addp->outcnt()-1);
1132 }
1133
1134 // Remove the old AddP from the processing list because it's dead now
1135 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1136 alloc_worklist.remove_if_existing(previous_addp);
1137 }
1138
1139 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1140 // selector is:
1141 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1142 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1143 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1144 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1145 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1146 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1147 uint number_of_sr_objects = 0;
1148 for (uint i = 1; i < ophi->req(); i++) {
1149 Node* base = ophi->in(i);
1150 JavaObjectNode* ptn = unique_java_object(base);
1151
1152 if (ptn != nullptr && ptn->scalar_replaceable()) {
1153 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1154 selector->set_req(i, sr_obj_idx);
1155 number_of_sr_objects++;
1156 }
1157 }
1158
1159 return selector->as_Phi();
1160 }
1161
1162 // Returns true if the AddP node 'n' has at least one base that is a reducible
1163 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1164 // checked instead.
1165 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1166 PointsToNode* ptn = ptnode_adr(n->_idx);
1167 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1168 return false;
1169 }
1170
1171 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1172 Node* base = i.get()->ideal_node();
1173
1174 if (reducible_merges.member(base)) {
1175 return true;
1176 }
1177
1178 if (base->is_CastPP() || base->is_CheckCastPP()) {
1179 base = base->in(1);
1180 if (reducible_merges.member(base)) {
1181 return true;
1182 }
1183 }
1184 }
1185
1186 return false;
1187 }
1188
1189 // This method will call its helper method to reduce SafePoint nodes that use
1190 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1191 // "version" of Phi use the same debug information (regarding the Phi).
1192 // Therefore, I collect all safepoints and patch them all at once.
1193 //
1194 // The safepoints using the Phi node have to be processed before safepoints of
1195 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1196 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1197 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1198 // algorithm that process Phi's safepoints will think that the added Phi
1199 // reference is a regular reference.
1200 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1201 PhiNode* selector = create_selector(ophi);
1202 Unique_Node_List safepoints;
1203 Unique_Node_List casts;
1204
1205 // Just collect the users of the Phis for later processing
1206 // in the needed order.
1207 for (uint i = 0; i < ophi->outcnt(); i++) {
1208 Node* use = ophi->raw_out(i);
1209 if (use->is_SafePoint()) {
1210 safepoints.push(use);
1211 } else if (use->is_CastPP()) {
1212 casts.push(use);
1213 } else {
1214 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1215 }
1216 }
1217
1218 // Need to process safepoints using the Phi first
1219 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1220 return false;
1221 }
1222
1223 // Now process CastPP->safepoints
1224 for (uint i = 0; i < casts.size(); i++) {
1225 Node* cast = casts.at(i);
1226 Unique_Node_List cast_sfpts;
1227
1228 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1229 Node* use_use = cast->fast_out(j);
1230 if (use_use->is_SafePoint()) {
1231 cast_sfpts.push(use_use);
1232 } else {
1233 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1234 }
1235 }
1236
1237 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1238 return false;
1239 }
1240 }
1241
1242 return true;
1243 }
1244
1245 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1246 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1247 // SafePointScalarObjectNode for each scalar replaceable input. Each
1248 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1249 // check detailed description in SafePointScalarMergeNode class header.
1250 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1251 PhaseMacroExpand mexp(*_igvn);
1252 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1253 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1254
1255 Node* nsr_merge_pointer = ophi;
1256 if (cast != nullptr) {
1257 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1258 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::DependencyType::FloatingNarrowing, nullptr));
1259 }
1260
1261 for (uint spi = 0; spi < safepoints.size(); spi++) {
1262 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1263 JVMState *jvms = sfpt->jvms();
1264 uint merge_idx = (sfpt->req() - jvms->scloff());
1265 int debug_start = jvms->debug_start();
1266
1267 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1268 smerge->init_req(0, _compile->root());
1269 _igvn->register_new_node_with_optimizer(smerge);
1270
1271 // The next two inputs are:
1272 // (1) A copy of the original pointer to NSR objects.
1273 // (2) A selector, used to decide if we need to rematerialize an object
1274 // or use the pointer to a NSR object.
1275 // See more details of these fields in the declaration of SafePointScalarMergeNode
1276 sfpt->add_req(nsr_merge_pointer);
1277 sfpt->add_req(selector);
1278
1279 for (uint i = 1; i < ophi->req(); i++) {
1280 Node* base = ophi->in(i);
1281 JavaObjectNode* ptn = unique_java_object(base);
1282
1283 // If the base is not scalar replaceable we don't need to register information about
1284 // it at this time.
1285 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1286 continue;
1287 }
1288
1289 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1290 Unique_Node_List value_worklist;
1291 #ifdef ASSERT
1292 const Type* res_type = alloc->result_cast()->bottom_type();
1293 if (res_type->is_inlinetypeptr() && !Compile::current()->has_circular_inline_type()) {
1294 PhiNode* phi = ophi->as_Phi();
1295 assert(!ophi->as_Phi()->can_push_inline_types_down(_igvn), "missed earlier scalarization opportunity");
1296 }
1297 #endif
1298 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt, &value_worklist);
1299 if (sobj == nullptr) {
1300 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1301 return false;
1302 }
1303
1304 // Now make a pass over the debug information replacing any references
1305 // to the allocated object with "sobj"
1306 Node* ccpp = alloc->result_cast();
1307 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1308
1309 // Register the scalarized object as a candidate for reallocation
1310 smerge->add_req(sobj);
1311
1312 // Scalarize inline types that were added to the safepoint.
1313 // Don't allow linking a constant oop (if available) for flat array elements
1314 // because Deoptimization::reassign_flat_array_elements needs field values.
1315 const bool allow_oop = !merge_t->is_flat();
1316 for (uint j = 0; j < value_worklist.size(); ++j) {
1317 InlineTypeNode* vt = value_worklist.at(j)->as_InlineType();
1318 vt->make_scalar_in_safepoints(_igvn, allow_oop);
1319 }
1320 }
1321
1322 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1323 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1324
1325 // The call to 'replace_edges_in_range' above might have removed the
1326 // reference to ophi that we need at _merge_pointer_idx. The line below make
1327 // sure the reference is maintained.
1328 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1329 _igvn->_worklist.push(sfpt);
1330 }
1331
1332 return true;
1333 }
1334
1335 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1336 bool delay = _igvn->delay_transform();
1337 _igvn->set_delay_transform(true);
1338 _igvn->hash_delete(ophi);
1339
1340 // Copying all users first because some will be removed and others won't.
1341 // Ophi also may acquire some new users as part of Cast reduction.
1342 // CastPPs also need to be processed before CmpPs.
1343 Unique_Node_List castpps;
1344 Unique_Node_List others;
1345 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1346 Node* use = ophi->fast_out(i);
1347
1348 if (use->is_CastPP()) {
1349 castpps.push(use);
1350 } else if (use->is_AddP() || use->is_Cmp()) {
1351 others.push(use);
1352 } else {
1353 // Safepoints to be processed later; other users aren't expected here
1354 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1355 }
1356 }
1357
1358 _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1359
1360 // CastPPs need to be processed before Cmps because during the process of
1361 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1362 // by the If controlling the CastPP.
1363 for (uint i = 0; i < castpps.size(); i++) {
1364 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1365 _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1366 }
1367
1368 for (uint i = 0; i < others.size(); i++) {
1369 Node* use = others.at(i);
1370
1371 if (use->is_AddP()) {
1372 reduce_phi_on_field_access(use, alloc_worklist);
1373 _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1374 } else if(use->is_Cmp()) {
1375 reduce_phi_on_cmp(use);
1376 _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1377 }
1378 }
1379
1380 _igvn->set_delay_transform(delay);
1381 }
1382
1383 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1384 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1385 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1386 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1387 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1388
1389 for (uint i = 1; i < ophi->req(); i++) {
1390 Node* base = ophi->in(i);
1391 JavaObjectNode* ptn = unique_java_object(base);
1392
1393 if (ptn != nullptr && ptn->scalar_replaceable()) {
1394 new_phi->set_req(i, null_ptr);
1395 } else {
1396 new_phi->set_req(i, ophi->in(i));
1397 }
1398 }
1399
1400 for (int i = ophi->outcnt()-1; i >= 0;) {
1401 Node* out = ophi->raw_out(i);
1402
1403 if (out->is_ConstraintCast()) {
1404 const Type* out_t = _igvn->type(out)->make_ptr();
1405 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1406 bool change = out_new_t != out_t;
1407
1408 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1409 Node* out2 = out->raw_out(j);
1410 if (!out2->is_SafePoint()) {
1411 change = false;
1412 break;
1413 }
1414 }
1415
1416 if (change) {
1417 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::DependencyType::NonFloatingNarrowing, nullptr);
1418 _igvn->replace_node(out, new_cast);
1419 _igvn->register_new_node_with_optimizer(new_cast);
1420 }
1421 }
1422
1423 --i;
1424 i = MIN2(i, (int)ophi->outcnt()-1);
1425 }
1426
1427 _igvn->replace_node(ophi, new_phi);
1428 }
1429
1430 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1431 if (!C->do_reduce_allocation_merges()) return;
1432
1433 Unique_Node_List ideal_nodes;
1434 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1435 ideal_nodes.push(root);
1436
1437 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1438 Node* n = ideal_nodes.at(next);
1439
1440 if (n->is_SafePointScalarMerge()) {
1441 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1442
1443 // Validate inputs of merge
1444 for (uint i = 1; i < merge->req(); i++) {
1445 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1446 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1447 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1448 }
1449 }
1450
1451 // Validate users of merge
1452 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1453 Node* sfpt = merge->fast_out(i);
1454 if (sfpt->is_SafePoint()) {
1455 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1456
1457 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1458 assert(false, "SafePointScalarMerge nodes can't be nested.");
1459 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1460 }
1461 } else {
1462 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1463 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1464 }
1465 }
1466 }
1467
1468 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1469 Node* m = n->fast_out(i);
1470 ideal_nodes.push(m);
1471 }
1472 }
1473 }
1474
1475 // Returns true if there is an object in the scope of sfn that does not escape globally.
1476 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1477 Compile* C = _compile;
1478 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1479 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1480 DeoptimizeObjectsALot) {
1481 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1482 int num_locs = jvms->loc_size();
1483 for (int idx = 0; idx < num_locs; idx++) {
1484 Node* l = sfn->local(jvms, idx);
1485 if (not_global_escape(l)) {
1486 return true;
1487 }
1488 }
1489 }
1490 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1491 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1492 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1493 int num_mon = jvms->nof_monitors();
1494 for (int idx = 0; idx < num_mon; idx++) {
1495 Node* m = sfn->monitor_obj(jvms, idx);
1496 if (m != nullptr && not_global_escape(m)) {
1497 return true;
1498 }
1499 }
1500 }
1501 }
1502 return false;
1503 }
1504
1505 // Returns true if at least one of the arguments to the call is an object
1506 // that does not escape globally.
1507 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1508 if (call->method() != nullptr) {
1509 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1510 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1511 Node* p = call->in(idx);
1512 if (not_global_escape(p)) {
1513 return true;
1514 }
1515 }
1516 } else {
1517 const char* name = call->as_CallStaticJava()->_name;
1518 assert(name != nullptr, "no name");
1519 // no arg escapes through uncommon traps
1520 if (strcmp(name, "uncommon_trap") != 0) {
1521 // process_call_arguments() assumes that all arguments escape globally
1522 const TypeTuple* d = call->tf()->domain_sig();
1523 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1524 const Type* at = d->field_at(i);
1525 if (at->isa_oopptr() != nullptr) {
1526 return true;
1527 }
1528 }
1529 }
1530 }
1531 return false;
1532 }
1533
1534
1535
1536 // Utility function for nodes that load an object
1537 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1538 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1539 // ThreadLocal has RawPtr type.
1540 const Type* t = _igvn->type(n);
1541 if (t->make_ptr() != nullptr) {
1542 Node* adr = n->in(MemNode::Address);
1543 #ifdef ASSERT
1544 if (!adr->is_AddP()) {
1545 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1546 } else {
1547 assert((ptnode_adr(adr->_idx) == nullptr ||
1548 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1549 }
1550 #endif
1551 add_local_var_and_edge(n, PointsToNode::NoEscape,
1552 adr, delayed_worklist);
1553 }
1554 }
1555
1556 // Populate Connection Graph with PointsTo nodes and create simple
1557 // connection graph edges.
1558 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1559 assert(!_verify, "this method should not be called for verification");
1560 PhaseGVN* igvn = _igvn;
1561 uint n_idx = n->_idx;
1562 PointsToNode* n_ptn = ptnode_adr(n_idx);
1563 if (n_ptn != nullptr) {
1564 return; // No need to redefine PointsTo node during first iteration.
1565 }
1566 int opcode = n->Opcode();
1567 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1568 if (gc_handled) {
1569 return; // Ignore node if already handled by GC.
1570 }
1571
1572 if (n->is_Call()) {
1573 // Arguments to allocation and locking don't escape.
1574 if (n->is_AbstractLock()) {
1575 // Put Lock and Unlock nodes on IGVN worklist to process them during
1576 // first IGVN optimization when escape information is still available.
1577 record_for_optimizer(n);
1578 } else if (n->is_Allocate()) {
1579 add_call_node(n->as_Call());
1580 record_for_optimizer(n);
1581 } else {
1582 if (n->is_CallStaticJava()) {
1583 const char* name = n->as_CallStaticJava()->_name;
1584 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1585 return; // Skip uncommon traps
1586 }
1587 }
1588 // Don't mark as processed since call's arguments have to be processed.
1589 delayed_worklist->push(n);
1590 // Check if a call returns an object.
1591 if ((n->as_Call()->returns_pointer() &&
1592 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1593 (n->is_CallStaticJava() &&
1594 n->as_CallStaticJava()->is_boxing_method())) {
1595 add_call_node(n->as_Call());
1596 } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1597 bool returns_oop = false;
1598 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1599 ProjNode* pn = n->fast_out(i)->as_Proj();
1600 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1601 returns_oop = true;
1602 }
1603 }
1604 if (returns_oop) {
1605 add_call_node(n->as_Call());
1606 }
1607 }
1608 }
1609 return;
1610 }
1611 // Put this check here to process call arguments since some call nodes
1612 // point to phantom_obj.
1613 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1614 return; // Skip predefined nodes.
1615 }
1616 switch (opcode) {
1617 case Op_AddP: {
1618 Node* base = get_addp_base(n);
1619 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1620 // Field nodes are created for all field types. They are used in
1621 // adjust_scalar_replaceable_state() and split_unique_types().
1622 // Note, non-oop fields will have only base edges in Connection
1623 // Graph because such fields are not used for oop loads and stores.
1624 int offset = address_offset(n, igvn);
1625 add_field(n, PointsToNode::NoEscape, offset);
1626 if (ptn_base == nullptr) {
1627 delayed_worklist->push(n); // Process it later.
1628 } else {
1629 n_ptn = ptnode_adr(n_idx);
1630 add_base(n_ptn->as_Field(), ptn_base);
1631 }
1632 break;
1633 }
1634 case Op_CastX2P:
1635 case Op_CastI2N: {
1636 map_ideal_node(n, phantom_obj);
1637 break;
1638 }
1639 case Op_InlineType:
1640 case Op_CastPP:
1641 case Op_CheckCastPP:
1642 case Op_EncodeP:
1643 case Op_DecodeN:
1644 case Op_EncodePKlass:
1645 case Op_DecodeNKlass: {
1646 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1647 break;
1648 }
1649 case Op_CMoveP: {
1650 add_local_var(n, PointsToNode::NoEscape);
1651 // Do not add edges during first iteration because some could be
1652 // not defined yet.
1653 delayed_worklist->push(n);
1654 break;
1655 }
1656 case Op_ConP:
1657 case Op_ConN:
1658 case Op_ConNKlass: {
1659 // assume all oop constants globally escape except for null
1660 PointsToNode::EscapeState es;
1661 const Type* t = igvn->type(n);
1662 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1663 es = PointsToNode::NoEscape;
1664 } else {
1665 es = PointsToNode::GlobalEscape;
1666 }
1667 PointsToNode* ptn_con = add_java_object(n, es);
1668 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1669 break;
1670 }
1671 case Op_CreateEx: {
1672 // assume that all exception objects globally escape
1673 map_ideal_node(n, phantom_obj);
1674 break;
1675 }
1676 case Op_LoadKlass:
1677 case Op_LoadNKlass: {
1678 // Unknown class is loaded
1679 map_ideal_node(n, phantom_obj);
1680 break;
1681 }
1682 case Op_LoadP:
1683 case Op_LoadN: {
1684 add_objload_to_connection_graph(n, delayed_worklist);
1685 break;
1686 }
1687 case Op_Parm: {
1688 map_ideal_node(n, phantom_obj);
1689 break;
1690 }
1691 case Op_PartialSubtypeCheck: {
1692 // Produces Null or notNull and is used in only in CmpP so
1693 // phantom_obj could be used.
1694 map_ideal_node(n, phantom_obj); // Result is unknown
1695 break;
1696 }
1697 case Op_Phi: {
1698 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1699 // ThreadLocal has RawPtr type.
1700 const Type* t = n->as_Phi()->type();
1701 if (t->make_ptr() != nullptr) {
1702 add_local_var(n, PointsToNode::NoEscape);
1703 // Do not add edges during first iteration because some could be
1704 // not defined yet.
1705 delayed_worklist->push(n);
1706 }
1707 break;
1708 }
1709 case Op_LoadFlat:
1710 // Treat LoadFlat similar to an unknown call that receives nothing and produces its results
1711 map_ideal_node(n, phantom_obj);
1712 break;
1713 case Op_StoreFlat:
1714 // Treat StoreFlat similar to a call that escapes the stored flattened fields
1715 delayed_worklist->push(n);
1716 break;
1717 case Op_Proj: {
1718 // we are only interested in the oop result projection from a call
1719 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
1720 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) {
1721 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1722 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1723 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1724 } else if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_LoadFlat() && igvn->type(n)->isa_ptr()) {
1725 // Treat LoadFlat outputs similar to a call return value
1726 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1727 }
1728 break;
1729 }
1730 case Op_Rethrow: // Exception object escapes
1731 case Op_Return: {
1732 if (n->req() > TypeFunc::Parms &&
1733 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1734 // Treat Return value as LocalVar with GlobalEscape escape state.
1735 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1736 }
1737 break;
1738 }
1739 case Op_CompareAndExchangeP:
1740 case Op_CompareAndExchangeN:
1741 case Op_GetAndSetP:
1742 case Op_GetAndSetN: {
1743 add_objload_to_connection_graph(n, delayed_worklist);
1744 // fall-through
1745 }
1746 case Op_StoreP:
1747 case Op_StoreN:
1748 case Op_StoreNKlass:
1749 case Op_WeakCompareAndSwapP:
1750 case Op_WeakCompareAndSwapN:
1751 case Op_CompareAndSwapP:
1752 case Op_CompareAndSwapN: {
1753 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1754 break;
1755 }
1756 case Op_AryEq:
1757 case Op_CountPositives:
1758 case Op_StrComp:
1759 case Op_StrEquals:
1760 case Op_StrIndexOf:
1761 case Op_StrIndexOfChar:
1762 case Op_StrInflatedCopy:
1763 case Op_StrCompressedCopy:
1764 case Op_VectorizedHashCode:
1765 case Op_EncodeISOArray: {
1766 add_local_var(n, PointsToNode::ArgEscape);
1767 delayed_worklist->push(n); // Process it later.
1768 break;
1769 }
1770 case Op_ThreadLocal: {
1771 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1772 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1773 break;
1774 }
1775 case Op_Blackhole: {
1776 // All blackhole pointer arguments are globally escaping.
1777 // Only do this if there is at least one pointer argument.
1778 // Do not add edges during first iteration because some could be
1779 // not defined yet, defer to final step.
1780 for (uint i = 0; i < n->req(); i++) {
1781 Node* in = n->in(i);
1782 if (in != nullptr) {
1783 const Type* at = _igvn->type(in);
1784 if (!at->isa_ptr()) continue;
1785
1786 add_local_var(n, PointsToNode::GlobalEscape);
1787 delayed_worklist->push(n);
1788 break;
1789 }
1790 }
1791 break;
1792 }
1793 default:
1794 ; // Do nothing for nodes not related to EA.
1795 }
1796 return;
1797 }
1798
1799 // Add final simple edges to graph.
1800 void ConnectionGraph::add_final_edges(Node *n) {
1801 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1802 #ifdef ASSERT
1803 if (_verify && n_ptn->is_JavaObject())
1804 return; // This method does not change graph for JavaObject.
1805 #endif
1806
1807 if (n->is_Call()) {
1808 process_call_arguments(n->as_Call());
1809 return;
1810 }
1811 assert(n->is_Store() || n->is_LoadStore() || n->is_StoreFlat() ||
1812 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1813 "node should be registered already");
1814 int opcode = n->Opcode();
1815 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1816 if (gc_handled) {
1817 return; // Ignore node if already handled by GC.
1818 }
1819 switch (opcode) {
1820 case Op_AddP: {
1821 Node* base = get_addp_base(n);
1822 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1823 assert(ptn_base != nullptr, "field's base should be registered");
1824 add_base(n_ptn->as_Field(), ptn_base);
1825 break;
1826 }
1827 case Op_InlineType:
1828 case Op_CastPP:
1829 case Op_CheckCastPP:
1830 case Op_EncodeP:
1831 case Op_DecodeN:
1832 case Op_EncodePKlass:
1833 case Op_DecodeNKlass: {
1834 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1835 break;
1836 }
1837 case Op_CMoveP: {
1838 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1839 Node* in = n->in(i);
1840 if (in == nullptr) {
1841 continue; // ignore null
1842 }
1843 Node* uncast_in = in->uncast();
1844 if (uncast_in->is_top() || uncast_in == n) {
1845 continue; // ignore top or inputs which go back this node
1846 }
1847 PointsToNode* ptn = ptnode_adr(in->_idx);
1848 assert(ptn != nullptr, "node should be registered");
1849 add_edge(n_ptn, ptn);
1850 }
1851 break;
1852 }
1853 case Op_LoadP:
1854 case Op_LoadN: {
1855 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1856 // ThreadLocal has RawPtr type.
1857 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1858 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1859 break;
1860 }
1861 case Op_Phi: {
1862 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1863 // ThreadLocal has RawPtr type.
1864 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1865 for (uint i = 1; i < n->req(); i++) {
1866 Node* in = n->in(i);
1867 if (in == nullptr) {
1868 continue; // ignore null
1869 }
1870 Node* uncast_in = in->uncast();
1871 if (uncast_in->is_top() || uncast_in == n) {
1872 continue; // ignore top or inputs which go back this node
1873 }
1874 PointsToNode* ptn = ptnode_adr(in->_idx);
1875 assert(ptn != nullptr, "node should be registered");
1876 add_edge(n_ptn, ptn);
1877 }
1878 break;
1879 }
1880 case Op_StoreFlat: {
1881 // StoreFlat globally escapes its stored flattened fields
1882 InlineTypeNode* value = n->as_StoreFlat()->value();
1883 ciInlineKlass* vk = _igvn->type(value)->inline_klass();
1884 for (int i = 0; i < vk->nof_nonstatic_fields(); i++) {
1885 ciField* field = vk->nonstatic_field_at(i);
1886 if (field->type()->is_primitive_type()) {
1887 continue;
1888 }
1889
1890 Node* field_value = value->field_value_by_offset(field->offset_in_bytes(), true);
1891 PointsToNode* field_value_ptn = ptnode_adr(field_value->_idx);
1892 set_escape_state(field_value_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "store into a flat field"));
1893 }
1894 break;
1895 }
1896 case Op_Proj: {
1897 if (n->in(0)->is_Call()) {
1898 // we are only interested in the oop result projection from a call
1899 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1900 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1901 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1902 } else if (n->in(0)->is_LoadFlat()) {
1903 // Treat LoadFlat outputs similar to a call return value
1904 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1905 }
1906 break;
1907 }
1908 case Op_Rethrow: // Exception object escapes
1909 case Op_Return: {
1910 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1911 "Unexpected node type");
1912 // Treat Return value as LocalVar with GlobalEscape escape state.
1913 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1914 break;
1915 }
1916 case Op_CompareAndExchangeP:
1917 case Op_CompareAndExchangeN:
1918 case Op_GetAndSetP:
1919 case Op_GetAndSetN:{
1920 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1921 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1922 // fall-through
1923 }
1924 case Op_CompareAndSwapP:
1925 case Op_CompareAndSwapN:
1926 case Op_WeakCompareAndSwapP:
1927 case Op_WeakCompareAndSwapN:
1928 case Op_StoreP:
1929 case Op_StoreN:
1930 case Op_StoreNKlass:{
1931 add_final_edges_unsafe_access(n, opcode);
1932 break;
1933 }
1934 case Op_VectorizedHashCode:
1935 case Op_AryEq:
1936 case Op_CountPositives:
1937 case Op_StrComp:
1938 case Op_StrEquals:
1939 case Op_StrIndexOf:
1940 case Op_StrIndexOfChar:
1941 case Op_StrInflatedCopy:
1942 case Op_StrCompressedCopy:
1943 case Op_EncodeISOArray: {
1944 // char[]/byte[] arrays passed to string intrinsic do not escape but
1945 // they are not scalar replaceable. Adjust escape state for them.
1946 // Start from in(2) edge since in(1) is memory edge.
1947 for (uint i = 2; i < n->req(); i++) {
1948 Node* adr = n->in(i);
1949 const Type* at = _igvn->type(adr);
1950 if (!adr->is_top() && at->isa_ptr()) {
1951 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1952 at->isa_ptr() != nullptr, "expecting a pointer");
1953 if (adr->is_AddP()) {
1954 adr = get_addp_base(adr);
1955 }
1956 PointsToNode* ptn = ptnode_adr(adr->_idx);
1957 assert(ptn != nullptr, "node should be registered");
1958 add_edge(n_ptn, ptn);
1959 }
1960 }
1961 break;
1962 }
1963 case Op_Blackhole: {
1964 // All blackhole pointer arguments are globally escaping.
1965 for (uint i = 0; i < n->req(); i++) {
1966 Node* in = n->in(i);
1967 if (in != nullptr) {
1968 const Type* at = _igvn->type(in);
1969 if (!at->isa_ptr()) continue;
1970
1971 if (in->is_AddP()) {
1972 in = get_addp_base(in);
1973 }
1974
1975 PointsToNode* ptn = ptnode_adr(in->_idx);
1976 assert(ptn != nullptr, "should be defined already");
1977 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1978 add_edge(n_ptn, ptn);
1979 }
1980 }
1981 break;
1982 }
1983 default: {
1984 // This method should be called only for EA specific nodes which may
1985 // miss some edges when they were created.
1986 #ifdef ASSERT
1987 n->dump(1);
1988 #endif
1989 guarantee(false, "unknown node");
1990 }
1991 }
1992 return;
1993 }
1994
1995 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1996 Node* adr = n->in(MemNode::Address);
1997 const Type* adr_type = _igvn->type(adr);
1998 adr_type = adr_type->make_ptr();
1999 if (adr_type == nullptr) {
2000 return; // skip dead nodes
2001 }
2002 if (adr_type->isa_oopptr()
2003 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2004 && adr_type == TypeRawPtr::NOTNULL
2005 && is_captured_store_address(adr))) {
2006 delayed_worklist->push(n); // Process it later.
2007 #ifdef ASSERT
2008 assert (adr->is_AddP(), "expecting an AddP");
2009 if (adr_type == TypeRawPtr::NOTNULL) {
2010 // Verify a raw address for a store captured by Initialize node.
2011 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2012 assert(offs != Type::OffsetBot, "offset must be a constant");
2013 }
2014 #endif
2015 } else {
2016 // Ignore copy the displaced header to the BoxNode (OSR compilation).
2017 if (adr->is_BoxLock()) {
2018 return;
2019 }
2020 // Stored value escapes in unsafe access.
2021 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2022 delayed_worklist->push(n); // Process unsafe access later.
2023 return;
2024 }
2025 #ifdef ASSERT
2026 n->dump(1);
2027 assert(false, "not unsafe");
2028 #endif
2029 }
2030 }
2031
2032 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
2033 Node* adr = n->in(MemNode::Address);
2034 const Type *adr_type = _igvn->type(adr);
2035 adr_type = adr_type->make_ptr();
2036 #ifdef ASSERT
2037 if (adr_type == nullptr) {
2038 n->dump(1);
2039 assert(adr_type != nullptr, "dead node should not be on list");
2040 return true;
2041 }
2042 #endif
2043
2044 if (adr_type->isa_oopptr()
2045 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2046 && adr_type == TypeRawPtr::NOTNULL
2047 && is_captured_store_address(adr))) {
2048 // Point Address to Value
2049 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2050 assert(adr_ptn != nullptr &&
2051 adr_ptn->as_Field()->is_oop(), "node should be registered");
2052 Node* val = n->in(MemNode::ValueIn);
2053 PointsToNode* ptn = ptnode_adr(val->_idx);
2054 assert(ptn != nullptr, "node should be registered");
2055 add_edge(adr_ptn, ptn);
2056 return true;
2057 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2058 // Stored value escapes in unsafe access.
2059 Node* val = n->in(MemNode::ValueIn);
2060 PointsToNode* ptn = ptnode_adr(val->_idx);
2061 assert(ptn != nullptr, "node should be registered");
2062 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2063 // Add edge to object for unsafe access with offset.
2064 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2065 assert(adr_ptn != nullptr, "node should be registered");
2066 if (adr_ptn->is_Field()) {
2067 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2068 add_edge(adr_ptn, ptn);
2069 }
2070 return true;
2071 }
2072 #ifdef ASSERT
2073 n->dump(1);
2074 assert(false, "not unsafe");
2075 #endif
2076 return false;
2077 }
2078
2079 void ConnectionGraph::add_call_node(CallNode* call) {
2080 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2081 uint call_idx = call->_idx;
2082 if (call->is_Allocate()) {
2083 Node* k = call->in(AllocateNode::KlassNode);
2084 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2085 assert(kt != nullptr, "TypeKlassPtr required.");
2086 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2087 bool scalar_replaceable = true;
2088 NOT_PRODUCT(const char* nsr_reason = "");
2089 if (call->is_AllocateArray()) {
2090 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2091 es = PointsToNode::GlobalEscape;
2092 } else {
2093 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2094 if (length < 0) {
2095 // Not scalar replaceable if the length is not constant.
2096 scalar_replaceable = false;
2097 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2098 } else if (length > EliminateAllocationArraySizeLimit) {
2099 // Not scalar replaceable if the length is too big.
2100 scalar_replaceable = false;
2101 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2102 }
2103 }
2104 } else { // Allocate instance
2105 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2106 es = PointsToNode::GlobalEscape;
2107 } else {
2108 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2109 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2110 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2111 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2112 !ik->can_be_instantiated() ||
2113 ik->has_finalizer()) {
2114 es = PointsToNode::GlobalEscape;
2115 } else {
2116 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2117 if (nfields > EliminateAllocationFieldsLimit) {
2118 // Not scalar replaceable if there are too many fields.
2119 scalar_replaceable = false;
2120 NOT_PRODUCT(nsr_reason = "has too many fields");
2121 }
2122 }
2123 }
2124 }
2125 add_java_object(call, es);
2126 PointsToNode* ptn = ptnode_adr(call_idx);
2127 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2128 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2129 }
2130 } else if (call->is_CallStaticJava()) {
2131 // Call nodes could be different types:
2132 //
2133 // 1. CallDynamicJavaNode (what happened during call is unknown):
2134 //
2135 // - mapped to GlobalEscape JavaObject node if oop is returned;
2136 //
2137 // - all oop arguments are escaping globally;
2138 //
2139 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2140 //
2141 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2142 //
2143 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2144 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2145 // during call is returned;
2146 // - mapped to ArgEscape LocalVar node pointed to object arguments
2147 // which are returned and does not escape during call;
2148 //
2149 // - oop arguments escaping status is defined by bytecode analysis;
2150 //
2151 // For a static call, we know exactly what method is being called.
2152 // Use bytecode estimator to record whether the call's return value escapes.
2153 ciMethod* meth = call->as_CallJava()->method();
2154 if (meth == nullptr) {
2155 const char* name = call->as_CallStaticJava()->_name;
2156 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0 ||
2157 strncmp(name, "C2 Runtime load_unknown_inline", 30) == 0 ||
2158 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "TODO: add failed case check");
2159 // Returns a newly allocated non-escaped object.
2160 add_java_object(call, PointsToNode::NoEscape);
2161 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2162 } else if (meth->is_boxing_method()) {
2163 // Returns boxing object
2164 PointsToNode::EscapeState es;
2165 vmIntrinsics::ID intr = meth->intrinsic_id();
2166 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2167 // It does not escape if object is always allocated.
2168 es = PointsToNode::NoEscape;
2169 } else {
2170 // It escapes globally if object could be loaded from cache.
2171 es = PointsToNode::GlobalEscape;
2172 }
2173 add_java_object(call, es);
2174 if (es == PointsToNode::GlobalEscape) {
2175 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2176 }
2177 } else {
2178 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2179 call_analyzer->copy_dependencies(_compile->dependencies());
2180 if (call_analyzer->is_return_allocated()) {
2181 // Returns a newly allocated non-escaped object, simply
2182 // update dependency information.
2183 // Mark it as NoEscape so that objects referenced by
2184 // it's fields will be marked as NoEscape at least.
2185 add_java_object(call, PointsToNode::NoEscape);
2186 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2187 } else {
2188 // Determine whether any arguments are returned.
2189 const TypeTuple* d = call->tf()->domain_cc();
2190 bool ret_arg = false;
2191 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2192 if (d->field_at(i)->isa_ptr() != nullptr &&
2193 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2194 ret_arg = true;
2195 break;
2196 }
2197 }
2198 if (ret_arg) {
2199 add_local_var(call, PointsToNode::ArgEscape);
2200 } else {
2201 // Returns unknown object.
2202 map_ideal_node(call, phantom_obj);
2203 }
2204 }
2205 }
2206 } else {
2207 // An other type of call, assume the worst case:
2208 // returned value is unknown and globally escapes.
2209 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2210 map_ideal_node(call, phantom_obj);
2211 }
2212 }
2213
2214 void ConnectionGraph::process_call_arguments(CallNode *call) {
2215 bool is_arraycopy = false;
2216 switch (call->Opcode()) {
2217 #ifdef ASSERT
2218 case Op_Allocate:
2219 case Op_AllocateArray:
2220 case Op_Lock:
2221 case Op_Unlock:
2222 assert(false, "should be done already");
2223 break;
2224 #endif
2225 case Op_ArrayCopy:
2226 case Op_CallLeafNoFP:
2227 // Most array copies are ArrayCopy nodes at this point but there
2228 // are still a few direct calls to the copy subroutines (See
2229 // PhaseStringOpts::copy_string())
2230 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2231 call->as_CallLeaf()->is_call_to_arraycopystub();
2232 // fall through
2233 case Op_CallLeafVector:
2234 case Op_CallLeaf: {
2235 // Stub calls, objects do not escape but they are not scale replaceable.
2236 // Adjust escape state for outgoing arguments.
2237 const TypeTuple * d = call->tf()->domain_sig();
2238 bool src_has_oops = false;
2239 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2240 const Type* at = d->field_at(i);
2241 Node *arg = call->in(i);
2242 if (arg == nullptr) {
2243 continue;
2244 }
2245 const Type *aat = _igvn->type(arg);
2246 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2247 continue;
2248 }
2249 if (arg->is_AddP()) {
2250 //
2251 // The inline_native_clone() case when the arraycopy stub is called
2252 // after the allocation before Initialize and CheckCastPP nodes.
2253 // Or normal arraycopy for object arrays case.
2254 //
2255 // Set AddP's base (Allocate) as not scalar replaceable since
2256 // pointer to the base (with offset) is passed as argument.
2257 //
2258 arg = get_addp_base(arg);
2259 }
2260 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2261 assert(arg_ptn != nullptr, "should be registered");
2262 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2263 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2264 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2265 aat->isa_ptr() != nullptr, "expecting an Ptr");
2266 bool arg_has_oops = aat->isa_oopptr() &&
2267 (aat->isa_instptr() ||
2268 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2269 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2270 aat->isa_aryptr()->is_flat() &&
2271 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2272 if (i == TypeFunc::Parms) {
2273 src_has_oops = arg_has_oops;
2274 }
2275 //
2276 // src or dst could be j.l.Object when other is basic type array:
2277 //
2278 // arraycopy(char[],0,Object*,0,size);
2279 // arraycopy(Object*,0,char[],0,size);
2280 //
2281 // Don't add edges in such cases.
2282 //
2283 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2284 arg_has_oops && (i > TypeFunc::Parms);
2285 #ifdef ASSERT
2286 if (!(is_arraycopy ||
2287 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2288 (call->as_CallLeaf()->_name != nullptr &&
2289 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2290 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2291 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2292 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2293 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2294 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2295 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2296 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2297 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2298 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2299 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2300 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2301 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2302 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2303 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2304 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2305 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2306 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2307 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2308 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2309 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2310 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2311 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2312 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2313 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2314 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2315 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2316 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2317 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2318 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2319 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2320 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2321 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2322 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2323 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2324 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2325 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2326 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2327 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2328 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2329 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2330 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2331 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2332 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2333 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2334 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2335 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2336 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2337 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2338 strcmp(call->as_CallLeaf()->_name, "store_inline_type_fields_to_buf") == 0 ||
2339 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2340 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2341 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2342 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2343 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2344 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2345 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2346 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2347 ))) {
2348 call->dump();
2349 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2350 }
2351 #endif
2352 // Always process arraycopy's destination object since
2353 // we need to add all possible edges to references in
2354 // source object.
2355 if (arg_esc >= PointsToNode::ArgEscape &&
2356 !arg_is_arraycopy_dest) {
2357 continue;
2358 }
2359 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2360 if (call->is_ArrayCopy()) {
2361 ArrayCopyNode* ac = call->as_ArrayCopy();
2362 if (ac->is_clonebasic() ||
2363 ac->is_arraycopy_validated() ||
2364 ac->is_copyof_validated() ||
2365 ac->is_copyofrange_validated()) {
2366 es = PointsToNode::NoEscape;
2367 }
2368 }
2369 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2370 if (arg_is_arraycopy_dest) {
2371 Node* src = call->in(TypeFunc::Parms);
2372 if (src->is_AddP()) {
2373 src = get_addp_base(src);
2374 }
2375 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2376 assert(src_ptn != nullptr, "should be registered");
2377 // Special arraycopy edge:
2378 // Only escape state of destination object's fields affects
2379 // escape state of fields in source object.
2380 add_arraycopy(call, es, src_ptn, arg_ptn);
2381 }
2382 }
2383 }
2384 break;
2385 }
2386 case Op_CallStaticJava: {
2387 // For a static call, we know exactly what method is being called.
2388 // Use bytecode estimator to record the call's escape affects
2389 #ifdef ASSERT
2390 const char* name = call->as_CallStaticJava()->_name;
2391 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2392 #endif
2393 ciMethod* meth = call->as_CallJava()->method();
2394 if ((meth != nullptr) && meth->is_boxing_method()) {
2395 break; // Boxing methods do not modify any oops.
2396 }
2397 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2398 // fall-through if not a Java method or no analyzer information
2399 if (call_analyzer != nullptr) {
2400 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2401 const TypeTuple* d = call->tf()->domain_cc();
2402 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2403 const Type* at = d->field_at(i);
2404 int k = i - TypeFunc::Parms;
2405 Node* arg = call->in(i);
2406 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2407 if (at->isa_ptr() != nullptr &&
2408 call_analyzer->is_arg_returned(k)) {
2409 // The call returns arguments.
2410 if (call_ptn != nullptr) { // Is call's result used?
2411 assert(call_ptn->is_LocalVar(), "node should be registered");
2412 assert(arg_ptn != nullptr, "node should be registered");
2413 add_edge(call_ptn, arg_ptn);
2414 }
2415 }
2416 if (at->isa_oopptr() != nullptr &&
2417 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2418 if (!call_analyzer->is_arg_stack(k)) {
2419 // The argument global escapes
2420 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2421 } else {
2422 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2423 if (!call_analyzer->is_arg_local(k)) {
2424 // The argument itself doesn't escape, but any fields might
2425 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2426 }
2427 }
2428 }
2429 }
2430 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2431 // The call returns arguments.
2432 assert(call_ptn->edge_count() > 0, "sanity");
2433 if (!call_analyzer->is_return_local()) {
2434 // Returns also unknown object.
2435 add_edge(call_ptn, phantom_obj);
2436 }
2437 }
2438 break;
2439 }
2440 }
2441 default: {
2442 // Fall-through here if not a Java method or no analyzer information
2443 // or some other type of call, assume the worst case: all arguments
2444 // globally escape.
2445 const TypeTuple* d = call->tf()->domain_cc();
2446 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2447 const Type* at = d->field_at(i);
2448 if (at->isa_oopptr() != nullptr) {
2449 Node* arg = call->in(i);
2450 if (arg->is_AddP()) {
2451 arg = get_addp_base(arg);
2452 }
2453 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2454 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2455 }
2456 }
2457 }
2458 }
2459 }
2460
2461
2462 // Finish Graph construction.
2463 bool ConnectionGraph::complete_connection_graph(
2464 GrowableArray<PointsToNode*>& ptnodes_worklist,
2465 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2466 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2467 GrowableArray<FieldNode*>& oop_fields_worklist) {
2468 // Normally only 1-3 passes needed to build Connection Graph depending
2469 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2470 // Set limit to 20 to catch situation when something did go wrong and
2471 // bailout Escape Analysis.
2472 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2473 #define GRAPH_BUILD_ITER_LIMIT 20
2474
2475 // Propagate GlobalEscape and ArgEscape escape states and check that
2476 // we still have non-escaping objects. The method pushs on _worklist
2477 // Field nodes which reference phantom_object.
2478 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2479 return false; // Nothing to do.
2480 }
2481 // Now propagate references to all JavaObject nodes.
2482 int java_objects_length = java_objects_worklist.length();
2483 elapsedTimer build_time;
2484 build_time.start();
2485 elapsedTimer time;
2486 bool timeout = false;
2487 int new_edges = 1;
2488 int iterations = 0;
2489 do {
2490 while ((new_edges > 0) &&
2491 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2492 double start_time = time.seconds();
2493 time.start();
2494 new_edges = 0;
2495 // Propagate references to phantom_object for nodes pushed on _worklist
2496 // by find_non_escaped_objects() and find_field_value().
2497 new_edges += add_java_object_edges(phantom_obj, false);
2498 for (int next = 0; next < java_objects_length; ++next) {
2499 JavaObjectNode* ptn = java_objects_worklist.at(next);
2500 new_edges += add_java_object_edges(ptn, true);
2501
2502 #define SAMPLE_SIZE 4
2503 if ((next % SAMPLE_SIZE) == 0) {
2504 // Each 4 iterations calculate how much time it will take
2505 // to complete graph construction.
2506 time.stop();
2507 // Poll for requests from shutdown mechanism to quiesce compiler
2508 // because Connection graph construction may take long time.
2509 CompileBroker::maybe_block();
2510 double stop_time = time.seconds();
2511 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2512 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2513 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2514 timeout = true;
2515 break; // Timeout
2516 }
2517 start_time = stop_time;
2518 time.start();
2519 }
2520 #undef SAMPLE_SIZE
2521
2522 }
2523 if (timeout) break;
2524 if (new_edges > 0) {
2525 // Update escape states on each iteration if graph was updated.
2526 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2527 return false; // Nothing to do.
2528 }
2529 }
2530 time.stop();
2531 if (time.seconds() >= EscapeAnalysisTimeout) {
2532 timeout = true;
2533 break;
2534 }
2535 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2536 }
2537 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2538 time.start();
2539 // Find fields which have unknown value.
2540 int fields_length = oop_fields_worklist.length();
2541 for (int next = 0; next < fields_length; next++) {
2542 FieldNode* field = oop_fields_worklist.at(next);
2543 if (field->edge_count() == 0) {
2544 new_edges += find_field_value(field);
2545 // This code may added new edges to phantom_object.
2546 // Need an other cycle to propagate references to phantom_object.
2547 }
2548 }
2549 time.stop();
2550 if (time.seconds() >= EscapeAnalysisTimeout) {
2551 timeout = true;
2552 break;
2553 }
2554 } else {
2555 new_edges = 0; // Bailout
2556 }
2557 } while (new_edges > 0);
2558
2559 build_time.stop();
2560 _build_time = build_time.seconds();
2561 _build_iterations = iterations;
2562
2563 // Bailout if passed limits.
2564 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2565 Compile* C = _compile;
2566 if (C->log() != nullptr) {
2567 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2568 C->log()->text("%s", timeout ? "time" : "iterations");
2569 C->log()->end_elem(" limit'");
2570 }
2571 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2572 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2573 // Possible infinite build_connection_graph loop,
2574 // bailout (no changes to ideal graph were made).
2575 return false;
2576 }
2577
2578 #undef GRAPH_BUILD_ITER_LIMIT
2579
2580 // Find fields initialized by null for non-escaping Allocations.
2581 int non_escaped_length = non_escaped_allocs_worklist.length();
2582 for (int next = 0; next < non_escaped_length; next++) {
2583 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2584 PointsToNode::EscapeState es = ptn->escape_state();
2585 assert(es <= PointsToNode::ArgEscape, "sanity");
2586 if (es == PointsToNode::NoEscape) {
2587 if (find_init_values_null(ptn, _igvn) > 0) {
2588 // Adding references to null object does not change escape states
2589 // since it does not escape. Also no fields are added to null object.
2590 add_java_object_edges(null_obj, false);
2591 }
2592 }
2593 Node* n = ptn->ideal_node();
2594 if (n->is_Allocate()) {
2595 // The object allocated by this Allocate node will never be
2596 // seen by an other thread. Mark it so that when it is
2597 // expanded no MemBarStoreStore is added.
2598 InitializeNode* ini = n->as_Allocate()->initialization();
2599 if (ini != nullptr)
2600 ini->set_does_not_escape();
2601 }
2602 }
2603 return true; // Finished graph construction.
2604 }
2605
2606 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2607 // and check that we still have non-escaping java objects.
2608 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2609 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2610 bool print_method) {
2611 GrowableArray<PointsToNode*> escape_worklist;
2612 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2613 int ptnodes_length = ptnodes_worklist.length();
2614 for (int next = 0; next < ptnodes_length; ++next) {
2615 PointsToNode* ptn = ptnodes_worklist.at(next);
2616 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2617 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2618 escape_worklist.push(ptn);
2619 }
2620 }
2621 // Set escape states to referenced nodes (edges list).
2622 while (escape_worklist.length() > 0) {
2623 PointsToNode* ptn = escape_worklist.pop();
2624 PointsToNode::EscapeState es = ptn->escape_state();
2625 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2626 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2627 es >= PointsToNode::ArgEscape) {
2628 // GlobalEscape or ArgEscape state of field means it has unknown value.
2629 if (add_edge(ptn, phantom_obj)) {
2630 // New edge was added
2631 add_field_uses_to_worklist(ptn->as_Field());
2632 }
2633 }
2634 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2635 PointsToNode* e = i.get();
2636 if (e->is_Arraycopy()) {
2637 assert(ptn->arraycopy_dst(), "sanity");
2638 // Propagate only fields escape state through arraycopy edge.
2639 if (e->fields_escape_state() < field_es) {
2640 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2641 escape_worklist.push(e);
2642 }
2643 } else if (es >= field_es) {
2644 // fields_escape_state is also set to 'es' if it is less than 'es'.
2645 if (e->escape_state() < es) {
2646 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2647 escape_worklist.push(e);
2648 }
2649 } else {
2650 // Propagate field escape state.
2651 bool es_changed = false;
2652 if (e->fields_escape_state() < field_es) {
2653 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2654 es_changed = true;
2655 }
2656 if ((e->escape_state() < field_es) &&
2657 e->is_Field() && ptn->is_JavaObject() &&
2658 e->as_Field()->is_oop()) {
2659 // Change escape state of referenced fields.
2660 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2661 es_changed = true;
2662 } else if (e->escape_state() < es) {
2663 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2664 es_changed = true;
2665 }
2666 if (es_changed) {
2667 escape_worklist.push(e);
2668 }
2669 }
2670 if (print_method) {
2671 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2672 }
2673 }
2674 }
2675 // Remove escaped objects from non_escaped list.
2676 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2677 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2678 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2679 non_escaped_allocs_worklist.delete_at(next);
2680 }
2681 if (ptn->escape_state() == PointsToNode::NoEscape) {
2682 // Find fields in non-escaped allocations which have unknown value.
2683 find_init_values_phantom(ptn);
2684 }
2685 }
2686 return (non_escaped_allocs_worklist.length() > 0);
2687 }
2688
2689 // Add all references to JavaObject node by walking over all uses.
2690 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2691 int new_edges = 0;
2692 if (populate_worklist) {
2693 // Populate _worklist by uses of jobj's uses.
2694 for (UseIterator i(jobj); i.has_next(); i.next()) {
2695 PointsToNode* use = i.get();
2696 if (use->is_Arraycopy()) {
2697 continue;
2698 }
2699 add_uses_to_worklist(use);
2700 if (use->is_Field() && use->as_Field()->is_oop()) {
2701 // Put on worklist all field's uses (loads) and
2702 // related field nodes (same base and offset).
2703 add_field_uses_to_worklist(use->as_Field());
2704 }
2705 }
2706 }
2707 for (int l = 0; l < _worklist.length(); l++) {
2708 PointsToNode* use = _worklist.at(l);
2709 if (PointsToNode::is_base_use(use)) {
2710 // Add reference from jobj to field and from field to jobj (field's base).
2711 use = PointsToNode::get_use_node(use)->as_Field();
2712 if (add_base(use->as_Field(), jobj)) {
2713 new_edges++;
2714 }
2715 continue;
2716 }
2717 assert(!use->is_JavaObject(), "sanity");
2718 if (use->is_Arraycopy()) {
2719 if (jobj == null_obj) { // null object does not have field edges
2720 continue;
2721 }
2722 // Added edge from Arraycopy node to arraycopy's source java object
2723 if (add_edge(use, jobj)) {
2724 jobj->set_arraycopy_src();
2725 new_edges++;
2726 }
2727 // and stop here.
2728 continue;
2729 }
2730 if (!add_edge(use, jobj)) {
2731 continue; // No new edge added, there was such edge already.
2732 }
2733 new_edges++;
2734 if (use->is_LocalVar()) {
2735 add_uses_to_worklist(use);
2736 if (use->arraycopy_dst()) {
2737 for (EdgeIterator i(use); i.has_next(); i.next()) {
2738 PointsToNode* e = i.get();
2739 if (e->is_Arraycopy()) {
2740 if (jobj == null_obj) { // null object does not have field edges
2741 continue;
2742 }
2743 // Add edge from arraycopy's destination java object to Arraycopy node.
2744 if (add_edge(jobj, e)) {
2745 new_edges++;
2746 jobj->set_arraycopy_dst();
2747 }
2748 }
2749 }
2750 }
2751 } else {
2752 // Added new edge to stored in field values.
2753 // Put on worklist all field's uses (loads) and
2754 // related field nodes (same base and offset).
2755 add_field_uses_to_worklist(use->as_Field());
2756 }
2757 }
2758 _worklist.clear();
2759 _in_worklist.reset();
2760 return new_edges;
2761 }
2762
2763 // Put on worklist all related field nodes.
2764 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2765 assert(field->is_oop(), "sanity");
2766 int offset = field->offset();
2767 add_uses_to_worklist(field);
2768 // Loop over all bases of this field and push on worklist Field nodes
2769 // with the same offset and base (since they may reference the same field).
2770 for (BaseIterator i(field); i.has_next(); i.next()) {
2771 PointsToNode* base = i.get();
2772 add_fields_to_worklist(field, base);
2773 // Check if the base was source object of arraycopy and go over arraycopy's
2774 // destination objects since values stored to a field of source object are
2775 // accessible by uses (loads) of fields of destination objects.
2776 if (base->arraycopy_src()) {
2777 for (UseIterator j(base); j.has_next(); j.next()) {
2778 PointsToNode* arycp = j.get();
2779 if (arycp->is_Arraycopy()) {
2780 for (UseIterator k(arycp); k.has_next(); k.next()) {
2781 PointsToNode* abase = k.get();
2782 if (abase->arraycopy_dst() && abase != base) {
2783 // Look for the same arraycopy reference.
2784 add_fields_to_worklist(field, abase);
2785 }
2786 }
2787 }
2788 }
2789 }
2790 }
2791 }
2792
2793 // Put on worklist all related field nodes.
2794 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2795 int offset = field->offset();
2796 if (base->is_LocalVar()) {
2797 for (UseIterator j(base); j.has_next(); j.next()) {
2798 PointsToNode* f = j.get();
2799 if (PointsToNode::is_base_use(f)) { // Field
2800 f = PointsToNode::get_use_node(f);
2801 if (f == field || !f->as_Field()->is_oop()) {
2802 continue;
2803 }
2804 int offs = f->as_Field()->offset();
2805 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2806 add_to_worklist(f);
2807 }
2808 }
2809 }
2810 } else {
2811 assert(base->is_JavaObject(), "sanity");
2812 if (// Skip phantom_object since it is only used to indicate that
2813 // this field's content globally escapes.
2814 (base != phantom_obj) &&
2815 // null object node does not have fields.
2816 (base != null_obj)) {
2817 for (EdgeIterator i(base); i.has_next(); i.next()) {
2818 PointsToNode* f = i.get();
2819 // Skip arraycopy edge since store to destination object field
2820 // does not update value in source object field.
2821 if (f->is_Arraycopy()) {
2822 assert(base->arraycopy_dst(), "sanity");
2823 continue;
2824 }
2825 if (f == field || !f->as_Field()->is_oop()) {
2826 continue;
2827 }
2828 int offs = f->as_Field()->offset();
2829 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2830 add_to_worklist(f);
2831 }
2832 }
2833 }
2834 }
2835 }
2836
2837 // Find fields which have unknown value.
2838 int ConnectionGraph::find_field_value(FieldNode* field) {
2839 // Escaped fields should have init value already.
2840 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2841 int new_edges = 0;
2842 for (BaseIterator i(field); i.has_next(); i.next()) {
2843 PointsToNode* base = i.get();
2844 if (base->is_JavaObject()) {
2845 // Skip Allocate's fields which will be processed later.
2846 if (base->ideal_node()->is_Allocate()) {
2847 return 0;
2848 }
2849 assert(base == null_obj, "only null ptr base expected here");
2850 }
2851 }
2852 if (add_edge(field, phantom_obj)) {
2853 // New edge was added
2854 new_edges++;
2855 add_field_uses_to_worklist(field);
2856 }
2857 return new_edges;
2858 }
2859
2860 // Find fields initializing values for allocations.
2861 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2862 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2863 PointsToNode* init_val = phantom_obj;
2864 Node* alloc = pta->ideal_node();
2865
2866 // Do nothing for Allocate nodes since its fields values are
2867 // "known" unless they are initialized by arraycopy/clone.
2868 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2869 if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2870 // Null-free inline type arrays are initialized with an init value instead of null
2871 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
2872 assert(init_val != nullptr, "init value should be registered");
2873 } else {
2874 return 0;
2875 }
2876 }
2877 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
2878 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
2879 #ifdef ASSERT
2880 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
2881 const char* name = alloc->as_CallStaticJava()->_name;
2882 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0 ||
2883 strncmp(name, "C2 Runtime load_unknown_inline", 30) == 0 ||
2884 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "sanity");
2885 }
2886 #endif
2887 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2888 int new_edges = 0;
2889 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2890 PointsToNode* field = i.get();
2891 if (field->is_Field() && field->as_Field()->is_oop()) {
2892 if (add_edge(field, init_val)) {
2893 // New edge was added
2894 new_edges++;
2895 add_field_uses_to_worklist(field->as_Field());
2896 }
2897 }
2898 }
2899 return new_edges;
2900 }
2901
2902 // Find fields initializing values for allocations.
2903 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2904 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2905 Node* alloc = pta->ideal_node();
2906 // Do nothing for Call nodes since its fields values are unknown.
2907 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2908 return 0;
2909 }
2910 InitializeNode* ini = alloc->as_Allocate()->initialization();
2911 bool visited_bottom_offset = false;
2912 GrowableArray<int> offsets_worklist;
2913 int new_edges = 0;
2914
2915 // Check if an oop field's initializing value is recorded and add
2916 // a corresponding null if field's value if it is not recorded.
2917 // Connection Graph does not record a default initialization by null
2918 // captured by Initialize node.
2919 //
2920 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2921 PointsToNode* field = i.get(); // Field (AddP)
2922 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2923 continue; // Not oop field
2924 }
2925 int offset = field->as_Field()->offset();
2926 if (offset == Type::OffsetBot) {
2927 if (!visited_bottom_offset) {
2928 // OffsetBot is used to reference array's element,
2929 // always add reference to null to all Field nodes since we don't
2930 // known which element is referenced.
2931 if (add_edge(field, null_obj)) {
2932 // New edge was added
2933 new_edges++;
2934 add_field_uses_to_worklist(field->as_Field());
2935 visited_bottom_offset = true;
2936 }
2937 }
2938 } else {
2939 // Check only oop fields.
2940 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2941 if (adr_type->isa_rawptr()) {
2942 #ifdef ASSERT
2943 // Raw pointers are used for initializing stores so skip it
2944 // since it should be recorded already
2945 Node* base = get_addp_base(field->ideal_node());
2946 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2947 #endif
2948 continue;
2949 }
2950 if (!offsets_worklist.contains(offset)) {
2951 offsets_worklist.append(offset);
2952 Node* value = nullptr;
2953 if (ini != nullptr) {
2954 // StoreP::value_basic_type() == T_ADDRESS
2955 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2956 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2957 // Make sure initializing store has the same type as this AddP.
2958 // This AddP may reference non existing field because it is on a
2959 // dead branch of bimorphic call which is not eliminated yet.
2960 if (store != nullptr && store->is_Store() &&
2961 store->as_Store()->value_basic_type() == ft) {
2962 value = store->in(MemNode::ValueIn);
2963 #ifdef ASSERT
2964 if (VerifyConnectionGraph) {
2965 // Verify that AddP already points to all objects the value points to.
2966 PointsToNode* val = ptnode_adr(value->_idx);
2967 assert((val != nullptr), "should be processed already");
2968 PointsToNode* missed_obj = nullptr;
2969 if (val->is_JavaObject()) {
2970 if (!field->points_to(val->as_JavaObject())) {
2971 missed_obj = val;
2972 }
2973 } else {
2974 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2975 tty->print_cr("----------init store has invalid value -----");
2976 store->dump();
2977 val->dump();
2978 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2979 }
2980 for (EdgeIterator j(val); j.has_next(); j.next()) {
2981 PointsToNode* obj = j.get();
2982 if (obj->is_JavaObject()) {
2983 if (!field->points_to(obj->as_JavaObject())) {
2984 missed_obj = obj;
2985 break;
2986 }
2987 }
2988 }
2989 }
2990 if (missed_obj != nullptr) {
2991 tty->print_cr("----------field---------------------------------");
2992 field->dump();
2993 tty->print_cr("----------missed reference to object------------");
2994 missed_obj->dump();
2995 tty->print_cr("----------object referenced by init store-------");
2996 store->dump();
2997 val->dump();
2998 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2999 }
3000 }
3001 #endif
3002 } else {
3003 // There could be initializing stores which follow allocation.
3004 // For example, a volatile field store is not collected
3005 // by Initialize node.
3006 //
3007 // Need to check for dependent loads to separate such stores from
3008 // stores which follow loads. For now, add initial value null so
3009 // that compare pointers optimization works correctly.
3010 }
3011 }
3012 if (value == nullptr) {
3013 // A field's initializing value was not recorded. Add null.
3014 if (add_edge(field, null_obj)) {
3015 // New edge was added
3016 new_edges++;
3017 add_field_uses_to_worklist(field->as_Field());
3018 }
3019 }
3020 }
3021 }
3022 }
3023 return new_edges;
3024 }
3025
3026 // Adjust scalar_replaceable state after Connection Graph is built.
3027 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
3028 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
3029 // returns true. If one of the constraints in this method set 'jobj' to NSR
3030 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
3031 // input, 'adjust_scalar_replaceable_state' will eventually be called with
3032 // that other object and the Phi will become a reducible Phi.
3033 // There could be multiple merges involving the same jobj.
3034 Unique_Node_List candidates;
3035
3036 // Search for non-escaping objects which are not scalar replaceable
3037 // and mark them to propagate the state to referenced objects.
3038
3039 for (UseIterator i(jobj); i.has_next(); i.next()) {
3040 PointsToNode* use = i.get();
3041 if (use->is_Arraycopy()) {
3042 continue;
3043 }
3044 if (use->is_Field()) {
3045 FieldNode* field = use->as_Field();
3046 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3047 // 1. An object is not scalar replaceable if the field into which it is
3048 // stored has unknown offset (stored into unknown element of an array).
3049 if (field->offset() == Type::OffsetBot) {
3050 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
3051 return;
3052 }
3053 for (BaseIterator i(field); i.has_next(); i.next()) {
3054 PointsToNode* base = i.get();
3055 // 2. An object is not scalar replaceable if the field into which it is
3056 // stored has multiple bases one of which is null.
3057 if ((base == null_obj) && (field->base_count() > 1)) {
3058 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3059 return;
3060 }
3061 // 2.5. An object is not scalar replaceable if the field into which it is
3062 // stored has NSR base.
3063 if (!base->scalar_replaceable()) {
3064 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3065 return;
3066 }
3067 }
3068 }
3069 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3070 // 3. An object is not scalar replaceable if it is merged with other objects
3071 // and we can't remove the merge
3072 for (EdgeIterator j(use); j.has_next(); j.next()) {
3073 PointsToNode* ptn = j.get();
3074 if (ptn->is_JavaObject() && ptn != jobj) {
3075 Node* use_n = use->ideal_node();
3076
3077 // These other local vars may point to multiple objects through a Phi
3078 // In this case we skip them and see if we can reduce the Phi.
3079 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3080 use_n = use_n->in(1);
3081 }
3082
3083 // If it's already a candidate or confirmed reducible merge we can skip verification
3084 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3085 continue;
3086 }
3087
3088 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3089 candidates.push(use_n);
3090 } else {
3091 // Mark all objects as NSR if we can't remove the merge
3092 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3093 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3094 }
3095 }
3096 }
3097 if (!jobj->scalar_replaceable()) {
3098 return;
3099 }
3100 }
3101
3102 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3103 if (j.get()->is_Arraycopy()) {
3104 continue;
3105 }
3106
3107 // Non-escaping object node should point only to field nodes.
3108 FieldNode* field = j.get()->as_Field();
3109 int offset = field->as_Field()->offset();
3110
3111 // 4. An object is not scalar replaceable if it has a field with unknown
3112 // offset (array's element is accessed in loop).
3113 if (offset == Type::OffsetBot) {
3114 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3115 return;
3116 }
3117 // 5. Currently an object is not scalar replaceable if a LoadStore node
3118 // access its field since the field value is unknown after it.
3119 //
3120 Node* n = field->ideal_node();
3121
3122 // Test for an unsafe access that was parsed as maybe off heap
3123 // (with a CheckCastPP to raw memory).
3124 assert(n->is_AddP(), "expect an address computation");
3125 if (n->in(AddPNode::Base)->is_top() &&
3126 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3127 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3128 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3129 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3130 return;
3131 }
3132
3133 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3134 Node* u = n->fast_out(i);
3135 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3136 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3137 return;
3138 }
3139 }
3140
3141 // 6. Or the address may point to more then one object. This may produce
3142 // the false positive result (set not scalar replaceable)
3143 // since the flow-insensitive escape analysis can't separate
3144 // the case when stores overwrite the field's value from the case
3145 // when stores happened on different control branches.
3146 //
3147 // Note: it will disable scalar replacement in some cases:
3148 //
3149 // Point p[] = new Point[1];
3150 // p[0] = new Point(); // Will be not scalar replaced
3151 //
3152 // but it will save us from incorrect optimizations in next cases:
3153 //
3154 // Point p[] = new Point[1];
3155 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3156 //
3157 if (field->base_count() > 1 && candidates.size() == 0) {
3158 if (has_non_reducible_merge(field, reducible_merges)) {
3159 for (BaseIterator i(field); i.has_next(); i.next()) {
3160 PointsToNode* base = i.get();
3161 // Don't take into account LocalVar nodes which
3162 // may point to only one object which should be also
3163 // this field's base by now.
3164 if (base->is_JavaObject() && base != jobj) {
3165 // Mark all bases.
3166 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3167 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3168 }
3169 }
3170
3171 if (!jobj->scalar_replaceable()) {
3172 return;
3173 }
3174 }
3175 }
3176 }
3177
3178 // The candidate is truly a reducible merge only if none of the other
3179 // constraints ruled it as NSR. There could be multiple merges involving the
3180 // same jobj.
3181 assert(jobj->scalar_replaceable(), "sanity");
3182 for (uint i = 0; i < candidates.size(); i++ ) {
3183 Node* candidate = candidates.at(i);
3184 reducible_merges.push(candidate);
3185 }
3186 }
3187
3188 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3189 for (BaseIterator i(field); i.has_next(); i.next()) {
3190 Node* base = i.get()->ideal_node();
3191 if (base->is_Phi() && !reducible_merges.member(base)) {
3192 return true;
3193 }
3194 }
3195 return false;
3196 }
3197
3198 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3199 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3200
3201 // Look for 'phis' that refer to 'jobj' as the last
3202 // remaining scalar replaceable input.
3203 uint reducible_merges_cnt = reducible_merges.size();
3204 for (uint i = 0; i < reducible_merges_cnt; i++) {
3205 Node* phi = reducible_merges.at(i);
3206
3207 // This 'Phi' will be a 'good' if it still points to
3208 // at least one scalar replaceable object. Note that 'obj'
3209 // was/should be marked as NSR before calling this function.
3210 bool good_phi = false;
3211
3212 for (uint j = 1; j < phi->req(); j++) {
3213 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3214 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3215 good_phi = true;
3216 break;
3217 }
3218 }
3219
3220 if (!good_phi) {
3221 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3222 reducible_merges.remove(i);
3223
3224 // Decrement the index because the 'remove' call above actually
3225 // moves the last entry of the list to position 'i'.
3226 i--;
3227
3228 reducible_merges_cnt--;
3229 }
3230 }
3231 }
3232
3233 // Propagate NSR (Not scalar replaceable) state.
3234 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3235 int jobj_length = jobj_worklist.length();
3236 bool found_nsr_alloc = true;
3237 while (found_nsr_alloc) {
3238 found_nsr_alloc = false;
3239 for (int next = 0; next < jobj_length; ++next) {
3240 JavaObjectNode* jobj = jobj_worklist.at(next);
3241 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3242 PointsToNode* use = i.get();
3243 if (use->is_Field()) {
3244 FieldNode* field = use->as_Field();
3245 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3246 assert(field->offset() != Type::OffsetBot, "sanity");
3247 for (BaseIterator i(field); i.has_next(); i.next()) {
3248 PointsToNode* base = i.get();
3249 // An object is not scalar replaceable if the field into which
3250 // it is stored has NSR base.
3251 if ((base != null_obj) && !base->scalar_replaceable()) {
3252 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3253 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3254 // because there is no point in reducing a Phi that won't improve the number of SR
3255 // objects.
3256 revisit_reducible_phi_status(jobj, reducible_merges);
3257 found_nsr_alloc = true;
3258 break;
3259 }
3260 }
3261 } else if (use->is_LocalVar()) {
3262 Node* phi = use->ideal_node();
3263 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3264 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3265 reducible_merges.yank(phi);
3266 found_nsr_alloc = true;
3267 break;
3268 }
3269 }
3270 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3271 }
3272 }
3273 }
3274 }
3275
3276 #ifdef ASSERT
3277 void ConnectionGraph::verify_connection_graph(
3278 GrowableArray<PointsToNode*>& ptnodes_worklist,
3279 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3280 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3281 GrowableArray<Node*>& addp_worklist) {
3282 // Verify that graph is complete - no new edges could be added.
3283 int java_objects_length = java_objects_worklist.length();
3284 int non_escaped_length = non_escaped_allocs_worklist.length();
3285 int new_edges = 0;
3286 for (int next = 0; next < java_objects_length; ++next) {
3287 JavaObjectNode* ptn = java_objects_worklist.at(next);
3288 new_edges += add_java_object_edges(ptn, true);
3289 }
3290 assert(new_edges == 0, "graph was not complete");
3291 // Verify that escape state is final.
3292 int length = non_escaped_allocs_worklist.length();
3293 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3294 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3295 (non_escaped_length == length) &&
3296 (_worklist.length() == 0), "escape state was not final");
3297
3298 // Verify fields information.
3299 int addp_length = addp_worklist.length();
3300 for (int next = 0; next < addp_length; ++next ) {
3301 Node* n = addp_worklist.at(next);
3302 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3303 if (field->is_oop()) {
3304 // Verify that field has all bases
3305 Node* base = get_addp_base(n);
3306 PointsToNode* ptn = ptnode_adr(base->_idx);
3307 if (ptn->is_JavaObject()) {
3308 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3309 } else {
3310 assert(ptn->is_LocalVar(), "sanity");
3311 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3312 PointsToNode* e = i.get();
3313 if (e->is_JavaObject()) {
3314 assert(field->has_base(e->as_JavaObject()), "sanity");
3315 }
3316 }
3317 }
3318 // Verify that all fields have initializing values.
3319 if (field->edge_count() == 0) {
3320 tty->print_cr("----------field does not have references----------");
3321 field->dump();
3322 for (BaseIterator i(field); i.has_next(); i.next()) {
3323 PointsToNode* base = i.get();
3324 tty->print_cr("----------field has next base---------------------");
3325 base->dump();
3326 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3327 tty->print_cr("----------base has fields-------------------------");
3328 for (EdgeIterator j(base); j.has_next(); j.next()) {
3329 j.get()->dump();
3330 }
3331 tty->print_cr("----------base has references---------------------");
3332 for (UseIterator j(base); j.has_next(); j.next()) {
3333 j.get()->dump();
3334 }
3335 }
3336 }
3337 for (UseIterator i(field); i.has_next(); i.next()) {
3338 i.get()->dump();
3339 }
3340 assert(field->edge_count() > 0, "sanity");
3341 }
3342 }
3343 }
3344 }
3345 #endif
3346
3347 // Optimize ideal graph.
3348 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3349 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3350 Compile* C = _compile;
3351 PhaseIterGVN* igvn = _igvn;
3352 if (EliminateLocks) {
3353 // Mark locks before changing ideal graph.
3354 int cnt = C->macro_count();
3355 for (int i = 0; i < cnt; i++) {
3356 Node *n = C->macro_node(i);
3357 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3358 AbstractLockNode* alock = n->as_AbstractLock();
3359 if (!alock->is_non_esc_obj()) {
3360 const Type* obj_type = igvn->type(alock->obj_node());
3361 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3362 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3363 // The lock could be marked eliminated by lock coarsening
3364 // code during first IGVN before EA. Replace coarsened flag
3365 // to eliminate all associated locks/unlocks.
3366 #ifdef ASSERT
3367 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3368 #endif
3369 alock->set_non_esc_obj();
3370 }
3371 }
3372 }
3373 }
3374 }
3375
3376 if (OptimizePtrCompare) {
3377 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3378 Node *n = ptr_cmp_worklist.at(i);
3379 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3380 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3381 if (tcmp->singleton()) {
3382 Node* cmp = igvn->makecon(tcmp);
3383 #ifndef PRODUCT
3384 if (PrintOptimizePtrCompare) {
3385 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, (tcmp == TypeInt::CC_EQ ? "EQ" : "NotEQ"));
3386 if (Verbose) {
3387 n->dump(1);
3388 }
3389 }
3390 #endif
3391 igvn->replace_node(n, cmp);
3392 }
3393 }
3394 }
3395
3396 // For MemBarStoreStore nodes added in library_call.cpp, check
3397 // escape status of associated AllocateNode and optimize out
3398 // MemBarStoreStore node if the allocated object never escapes.
3399 for (int i = 0; i < storestore_worklist.length(); i++) {
3400 Node* storestore = storestore_worklist.at(i);
3401 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3402 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3403 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3404 // Non-escaping inline type buffer allocations don't require a membar
3405 storestore->as_MemBar()->remove(_igvn);
3406 } else {
3407 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3408 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3409 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3410 igvn->register_new_node_with_optimizer(mb);
3411 igvn->replace_node(storestore, mb);
3412 }
3413 }
3414 }
3415 }
3416
3417 // Atomic flat accesses on non-escaping objects can be optimized to non-atomic accesses
3418 void ConnectionGraph::optimize_flat_accesses(GrowableArray<SafePointNode*>& sfn_worklist) {
3419 PhaseIterGVN& igvn = *_igvn;
3420 bool delay = igvn.delay_transform();
3421 igvn.set_delay_transform(true);
3422 igvn.C->for_each_flat_access([&](Node* n) {
3423 Node* base = n->is_LoadFlat() ? n->as_LoadFlat()->base() : n->as_StoreFlat()->base();
3424 if (!not_global_escape(base)) {
3425 return;
3426 }
3427
3428 bool expanded;
3429 if (n->is_LoadFlat()) {
3430 expanded = n->as_LoadFlat()->expand_non_atomic(igvn);
3431 } else {
3432 expanded = n->as_StoreFlat()->expand_non_atomic(igvn);
3433 }
3434 if (expanded) {
3435 sfn_worklist.remove(n->as_SafePoint());
3436 igvn.C->remove_flat_access(n);
3437 }
3438 });
3439 igvn.set_delay_transform(delay);
3440 }
3441
3442 // Optimize objects compare.
3443 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3444 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3445 if (!OptimizePtrCompare) {
3446 return UNKNOWN;
3447 }
3448 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3449 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3450
3451 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3452 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3453 JavaObjectNode* jobj1 = unique_java_object(left);
3454 JavaObjectNode* jobj2 = unique_java_object(right);
3455
3456 // The use of this method during allocation merge reduction may cause 'left'
3457 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3458 // that doesn't reference an unique java object.
3459 if (ptn1 == nullptr || ptn2 == nullptr ||
3460 jobj1 == nullptr || jobj2 == nullptr) {
3461 return UNKNOWN;
3462 }
3463
3464 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3465 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3466
3467 // Check simple cases first.
3468 if (jobj1 != nullptr) {
3469 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3470 if (jobj1 == jobj2) {
3471 // Comparing the same not escaping object.
3472 return EQ;
3473 }
3474 Node* obj = jobj1->ideal_node();
3475 // Comparing not escaping allocation.
3476 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3477 !ptn2->points_to(jobj1)) {
3478 return NE; // This includes nullness check.
3479 }
3480 }
3481 }
3482 if (jobj2 != nullptr) {
3483 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3484 Node* obj = jobj2->ideal_node();
3485 // Comparing not escaping allocation.
3486 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3487 !ptn1->points_to(jobj2)) {
3488 return NE; // This includes nullness check.
3489 }
3490 }
3491 }
3492 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3493 jobj2 != nullptr && jobj2 != phantom_obj &&
3494 jobj1->ideal_node()->is_Con() &&
3495 jobj2->ideal_node()->is_Con()) {
3496 // Klass or String constants compare. Need to be careful with
3497 // compressed pointers - compare types of ConN and ConP instead of nodes.
3498 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3499 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3500 if (t1->make_ptr() == t2->make_ptr()) {
3501 return EQ;
3502 } else {
3503 return NE;
3504 }
3505 }
3506 if (ptn1->meet(ptn2)) {
3507 return UNKNOWN; // Sets are not disjoint
3508 }
3509
3510 // Sets are disjoint.
3511 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3512 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3513 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3514 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3515 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3516 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3517 // Check nullness of unknown object.
3518 return UNKNOWN;
3519 }
3520
3521 // Disjointness by itself is not sufficient since
3522 // alias analysis is not complete for escaped objects.
3523 // Disjoint sets are definitely unrelated only when
3524 // at least one set has only not escaping allocations.
3525 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3526 if (ptn1->non_escaping_allocation()) {
3527 return NE;
3528 }
3529 }
3530 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3531 if (ptn2->non_escaping_allocation()) {
3532 return NE;
3533 }
3534 }
3535 return UNKNOWN;
3536 }
3537
3538 // Connection Graph construction functions.
3539
3540 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3541 PointsToNode* ptadr = _nodes.at(n->_idx);
3542 if (ptadr != nullptr) {
3543 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3544 return;
3545 }
3546 Compile* C = _compile;
3547 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3548 map_ideal_node(n, ptadr);
3549 }
3550
3551 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3552 PointsToNode* ptadr = _nodes.at(n->_idx);
3553 if (ptadr != nullptr) {
3554 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3555 return ptadr;
3556 }
3557 Compile* C = _compile;
3558 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3559 map_ideal_node(n, ptadr);
3560 return ptadr;
3561 }
3562
3563 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3564 PointsToNode* ptadr = _nodes.at(n->_idx);
3565 if (ptadr != nullptr) {
3566 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3567 return;
3568 }
3569 bool unsafe = false;
3570 bool is_oop = is_oop_field(n, offset, &unsafe);
3571 if (unsafe) {
3572 es = PointsToNode::GlobalEscape;
3573 }
3574 Compile* C = _compile;
3575 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3576 map_ideal_node(n, field);
3577 }
3578
3579 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3580 PointsToNode* src, PointsToNode* dst) {
3581 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3582 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3583 PointsToNode* ptadr = _nodes.at(n->_idx);
3584 if (ptadr != nullptr) {
3585 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3586 return;
3587 }
3588 Compile* C = _compile;
3589 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3590 map_ideal_node(n, ptadr);
3591 // Add edge from arraycopy node to source object.
3592 (void)add_edge(ptadr, src);
3593 src->set_arraycopy_src();
3594 // Add edge from destination object to arraycopy node.
3595 (void)add_edge(dst, ptadr);
3596 dst->set_arraycopy_dst();
3597 }
3598
3599 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3600 const Type* adr_type = n->as_AddP()->bottom_type();
3601 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3602 BasicType bt = T_INT;
3603 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3604 // Check only oop fields.
3605 if (!adr_type->isa_aryptr() ||
3606 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3607 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3608 // OffsetBot is used to reference array's element. Ignore first AddP.
3609 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3610 bt = T_OBJECT;
3611 }
3612 }
3613 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3614 if (adr_type->isa_instptr()) {
3615 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3616 if (field != nullptr) {
3617 bt = field->layout_type();
3618 } else {
3619 // Check for unsafe oop field access
3620 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3621 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3622 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3623 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3624 bt = T_OBJECT;
3625 (*unsafe) = true;
3626 }
3627 }
3628 } else if (adr_type->isa_aryptr()) {
3629 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3630 // Ignore array length load.
3631 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3632 // Ignore first AddP.
3633 } else {
3634 const Type* elemtype = adr_type->is_aryptr()->elem();
3635 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3636 ciInlineKlass* vk = elemtype->inline_klass();
3637 field_offset += vk->payload_offset();
3638 ciField* field = vk->get_field_by_offset(field_offset, false);
3639 if (field != nullptr) {
3640 bt = field->layout_type();
3641 } else {
3642 assert(field_offset == vk->payload_offset() + vk->null_marker_offset_in_payload(), "no field or null marker of %s at offset %d", vk->name()->as_utf8(), field_offset);
3643 bt = T_BOOLEAN;
3644 }
3645 } else {
3646 bt = elemtype->array_element_basic_type();
3647 }
3648 }
3649 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3650 // Allocation initialization, ThreadLocal field access, unsafe access
3651 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3652 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3653 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3654 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3655 bt = T_OBJECT;
3656 }
3657 }
3658 }
3659 // Note: T_NARROWOOP is not classed as a real reference type
3660 return (is_reference_type(bt) || bt == T_NARROWOOP);
3661 }
3662
3663 // Returns unique pointed java object or null.
3664 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3665 // If the node was created after the escape computation we can't answer.
3666 uint idx = n->_idx;
3667 if (idx >= nodes_size()) {
3668 return nullptr;
3669 }
3670 PointsToNode* ptn = ptnode_adr(idx);
3671 if (ptn == nullptr) {
3672 return nullptr;
3673 }
3674 if (ptn->is_JavaObject()) {
3675 return ptn->as_JavaObject();
3676 }
3677 assert(ptn->is_LocalVar(), "sanity");
3678 // Check all java objects it points to.
3679 JavaObjectNode* jobj = nullptr;
3680 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3681 PointsToNode* e = i.get();
3682 if (e->is_JavaObject()) {
3683 if (jobj == nullptr) {
3684 jobj = e->as_JavaObject();
3685 } else if (jobj != e) {
3686 return nullptr;
3687 }
3688 }
3689 }
3690 return jobj;
3691 }
3692
3693 // Return true if this node points only to non-escaping allocations.
3694 bool PointsToNode::non_escaping_allocation() {
3695 if (is_JavaObject()) {
3696 Node* n = ideal_node();
3697 if (n->is_Allocate() || n->is_CallStaticJava()) {
3698 return (escape_state() == PointsToNode::NoEscape);
3699 } else {
3700 return false;
3701 }
3702 }
3703 assert(is_LocalVar(), "sanity");
3704 // Check all java objects it points to.
3705 for (EdgeIterator i(this); i.has_next(); i.next()) {
3706 PointsToNode* e = i.get();
3707 if (e->is_JavaObject()) {
3708 Node* n = e->ideal_node();
3709 if ((e->escape_state() != PointsToNode::NoEscape) ||
3710 !(n->is_Allocate() || n->is_CallStaticJava())) {
3711 return false;
3712 }
3713 }
3714 }
3715 return true;
3716 }
3717
3718 // Return true if we know the node does not escape globally.
3719 bool ConnectionGraph::not_global_escape(Node *n) {
3720 assert(!_collecting, "should not call during graph construction");
3721 // If the node was created after the escape computation we can't answer.
3722 uint idx = n->_idx;
3723 if (idx >= nodes_size()) {
3724 return false;
3725 }
3726 PointsToNode* ptn = ptnode_adr(idx);
3727 if (ptn == nullptr) {
3728 return false; // not in congraph (e.g. ConI)
3729 }
3730 PointsToNode::EscapeState es = ptn->escape_state();
3731 // If we have already computed a value, return it.
3732 if (es >= PointsToNode::GlobalEscape) {
3733 return false;
3734 }
3735 if (ptn->is_JavaObject()) {
3736 return true; // (es < PointsToNode::GlobalEscape);
3737 }
3738 assert(ptn->is_LocalVar(), "sanity");
3739 // Check all java objects it points to.
3740 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3741 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3742 return false;
3743 }
3744 }
3745 return true;
3746 }
3747
3748 // Return true if locked object does not escape globally
3749 // and locked code region (identified by BoxLockNode) is balanced:
3750 // all compiled code paths have corresponding Lock/Unlock pairs.
3751 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3752 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3753 if (EliminateNestedLocks) {
3754 // We can mark whole locking region as Local only when only
3755 // one object is used for locking.
3756 alock->box_node()->as_BoxLock()->set_local();
3757 }
3758 return true;
3759 }
3760 return false;
3761 }
3762
3763 // Helper functions
3764
3765 // Return true if this node points to specified node or nodes it points to.
3766 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3767 if (is_JavaObject()) {
3768 return (this == ptn);
3769 }
3770 assert(is_LocalVar() || is_Field(), "sanity");
3771 for (EdgeIterator i(this); i.has_next(); i.next()) {
3772 if (i.get() == ptn) {
3773 return true;
3774 }
3775 }
3776 return false;
3777 }
3778
3779 // Return true if one node points to an other.
3780 bool PointsToNode::meet(PointsToNode* ptn) {
3781 if (this == ptn) {
3782 return true;
3783 } else if (ptn->is_JavaObject()) {
3784 return this->points_to(ptn->as_JavaObject());
3785 } else if (this->is_JavaObject()) {
3786 return ptn->points_to(this->as_JavaObject());
3787 }
3788 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3789 int ptn_count = ptn->edge_count();
3790 for (EdgeIterator i(this); i.has_next(); i.next()) {
3791 PointsToNode* this_e = i.get();
3792 for (int j = 0; j < ptn_count; j++) {
3793 if (this_e == ptn->edge(j)) {
3794 return true;
3795 }
3796 }
3797 }
3798 return false;
3799 }
3800
3801 #ifdef ASSERT
3802 // Return true if bases point to this java object.
3803 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3804 for (BaseIterator i(this); i.has_next(); i.next()) {
3805 if (i.get() == jobj) {
3806 return true;
3807 }
3808 }
3809 return false;
3810 }
3811 #endif
3812
3813 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3814 // Handle simple case first.
3815 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3816 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3817 return true;
3818 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3819 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3820 Node* addp_use = addp->fast_out(i);
3821 if (addp_use->is_Store()) {
3822 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3823 if (addp_use->fast_out(j)->is_Initialize()) {
3824 return true;
3825 }
3826 }
3827 }
3828 }
3829 }
3830 return false;
3831 }
3832
3833 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3834 const Type *adr_type = phase->type(adr);
3835 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3836 // We are computing a raw address for a store captured by an Initialize
3837 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3838 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3839 assert(offs != Type::OffsetBot ||
3840 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3841 "offset must be a constant or it is initialization of array");
3842 return offs;
3843 }
3844 return adr_type->is_ptr()->flat_offset();
3845 }
3846
3847 Node* ConnectionGraph::get_addp_base(Node *addp) {
3848 assert(addp->is_AddP(), "must be AddP");
3849 //
3850 // AddP cases for Base and Address inputs:
3851 // case #1. Direct object's field reference:
3852 // Allocate
3853 // |
3854 // Proj #5 ( oop result )
3855 // |
3856 // CheckCastPP (cast to instance type)
3857 // | |
3858 // AddP ( base == address )
3859 //
3860 // case #2. Indirect object's field reference:
3861 // Phi
3862 // |
3863 // CastPP (cast to instance type)
3864 // | |
3865 // AddP ( base == address )
3866 //
3867 // case #3. Raw object's field reference for Initialize node:
3868 // Allocate
3869 // |
3870 // Proj #5 ( oop result )
3871 // top |
3872 // \ |
3873 // AddP ( base == top )
3874 //
3875 // case #4. Array's element reference:
3876 // {CheckCastPP | CastPP}
3877 // | | |
3878 // | AddP ( array's element offset )
3879 // | |
3880 // AddP ( array's offset )
3881 //
3882 // case #5. Raw object's field reference for arraycopy stub call:
3883 // The inline_native_clone() case when the arraycopy stub is called
3884 // after the allocation before Initialize and CheckCastPP nodes.
3885 // Allocate
3886 // |
3887 // Proj #5 ( oop result )
3888 // | |
3889 // AddP ( base == address )
3890 //
3891 // case #6. Constant Pool, ThreadLocal, CastX2P or
3892 // Raw object's field reference:
3893 // {ConP, ThreadLocal, CastX2P, raw Load}
3894 // top |
3895 // \ |
3896 // AddP ( base == top )
3897 //
3898 // case #7. Klass's field reference.
3899 // LoadKlass
3900 // | |
3901 // AddP ( base == address )
3902 //
3903 // case #8. narrow Klass's field reference.
3904 // LoadNKlass
3905 // |
3906 // DecodeN
3907 // | |
3908 // AddP ( base == address )
3909 //
3910 // case #9. Mixed unsafe access
3911 // {instance}
3912 // |
3913 // CheckCastPP (raw)
3914 // top |
3915 // \ |
3916 // AddP ( base == top )
3917 //
3918 Node *base = addp->in(AddPNode::Base);
3919 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3920 base = addp->in(AddPNode::Address);
3921 while (base->is_AddP()) {
3922 // Case #6 (unsafe access) may have several chained AddP nodes.
3923 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3924 base = base->in(AddPNode::Address);
3925 }
3926 if (base->Opcode() == Op_CheckCastPP &&
3927 base->bottom_type()->isa_rawptr() &&
3928 _igvn->type(base->in(1))->isa_oopptr()) {
3929 base = base->in(1); // Case #9
3930 } else {
3931 Node* uncast_base = base->uncast();
3932 int opcode = uncast_base->Opcode();
3933 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3934 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3935 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3936 is_captured_store_address(addp), "sanity");
3937 }
3938 }
3939 return base;
3940 }
3941
3942 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3943 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3944 Node* addp2 = addp->raw_out(0);
3945 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3946 addp2->in(AddPNode::Base) == n &&
3947 addp2->in(AddPNode::Address) == addp) {
3948 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3949 //
3950 // Find array's offset to push it on worklist first and
3951 // as result process an array's element offset first (pushed second)
3952 // to avoid CastPP for the array's offset.
3953 // Otherwise the inserted CastPP (LocalVar) will point to what
3954 // the AddP (Field) points to. Which would be wrong since
3955 // the algorithm expects the CastPP has the same point as
3956 // as AddP's base CheckCastPP (LocalVar).
3957 //
3958 // ArrayAllocation
3959 // |
3960 // CheckCastPP
3961 // |
3962 // memProj (from ArrayAllocation CheckCastPP)
3963 // | ||
3964 // | || Int (element index)
3965 // | || | ConI (log(element size))
3966 // | || | /
3967 // | || LShift
3968 // | || /
3969 // | AddP (array's element offset)
3970 // | |
3971 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3972 // | / /
3973 // AddP (array's offset)
3974 // |
3975 // Load/Store (memory operation on array's element)
3976 //
3977 return addp2;
3978 }
3979 return nullptr;
3980 }
3981
3982 //
3983 // Adjust the type and inputs of an AddP which computes the
3984 // address of a field of an instance
3985 //
3986 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3987 PhaseGVN* igvn = _igvn;
3988 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3989 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3990 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3991 if (t == nullptr) {
3992 // We are computing a raw address for a store captured by an Initialize
3993 // compute an appropriate address type (cases #3 and #5).
3994 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3995 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3996 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3997 assert(offs != Type::OffsetBot, "offset must be a constant");
3998 if (base_t->isa_aryptr() != nullptr) {
3999 // In the case of a flat inline type array, each field has its
4000 // own slice so we need to extract the field being accessed from
4001 // the address computation
4002 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
4003 } else {
4004 t = base_t->add_offset(offs)->is_oopptr();
4005 }
4006 }
4007 int inst_id = base_t->instance_id();
4008 assert(!t->is_known_instance() || t->instance_id() == inst_id,
4009 "old type must be non-instance or match new type");
4010
4011 // The type 't' could be subclass of 'base_t'.
4012 // As result t->offset() could be large then base_t's size and it will
4013 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
4014 // constructor verifies correctness of the offset.
4015 //
4016 // It could happened on subclass's branch (from the type profiling
4017 // inlining) which was not eliminated during parsing since the exactness
4018 // of the allocation type was not propagated to the subclass type check.
4019 //
4020 // Or the type 't' could be not related to 'base_t' at all.
4021 // It could happen when CHA type is different from MDO type on a dead path
4022 // (for example, from instanceof check) which is not collapsed during parsing.
4023 //
4024 // Do nothing for such AddP node and don't process its users since
4025 // this code branch will go away.
4026 //
4027 if (!t->is_known_instance() &&
4028 !base_t->maybe_java_subtype_of(t)) {
4029 return false; // bail out
4030 }
4031 const TypePtr* tinst = base_t->add_offset(t->offset());
4032 if (tinst->isa_aryptr() && t->isa_aryptr()) {
4033 // In the case of a flat inline type array, each field has its
4034 // own slice so we need to keep track of the field being accessed.
4035 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
4036 // Keep array properties (not flat/null-free)
4037 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
4038 if (tinst == nullptr) {
4039 return false; // Skip dead path with inconsistent properties
4040 }
4041 }
4042
4043 // Do NOT remove the next line: ensure a new alias index is allocated
4044 // for the instance type. Note: C++ will not remove it since the call
4045 // has side effect.
4046 int alias_idx = _compile->get_alias_index(tinst);
4047 igvn->set_type(addp, tinst);
4048 // record the allocation in the node map
4049 set_map(addp, get_map(base->_idx));
4050 // Set addp's Base and Address to 'base'.
4051 Node *abase = addp->in(AddPNode::Base);
4052 Node *adr = addp->in(AddPNode::Address);
4053 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
4054 adr->in(0)->_idx == (uint)inst_id) {
4055 // Skip AddP cases #3 and #5.
4056 } else {
4057 assert(!abase->is_top(), "sanity"); // AddP case #3
4058 if (abase != base) {
4059 igvn->hash_delete(addp);
4060 addp->set_req(AddPNode::Base, base);
4061 if (abase == adr) {
4062 addp->set_req(AddPNode::Address, base);
4063 } else {
4064 // AddP case #4 (adr is array's element offset AddP node)
4065 #ifdef ASSERT
4066 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
4067 assert(adr->is_AddP() && atype != nullptr &&
4068 atype->instance_id() == inst_id, "array's element offset should be processed first");
4069 #endif
4070 }
4071 igvn->hash_insert(addp);
4072 }
4073 }
4074 // Put on IGVN worklist since at least addp's type was changed above.
4075 record_for_optimizer(addp);
4076 return true;
4077 }
4078
4079 //
4080 // Create a new version of orig_phi if necessary. Returns either the newly
4081 // created phi or an existing phi. Sets create_new to indicate whether a new
4082 // phi was created. Cache the last newly created phi in the node map.
4083 //
4084 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
4085 Compile *C = _compile;
4086 PhaseGVN* igvn = _igvn;
4087 new_created = false;
4088 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
4089 // nothing to do if orig_phi is bottom memory or matches alias_idx
4090 if (phi_alias_idx == alias_idx) {
4091 return orig_phi;
4092 }
4093 // Have we recently created a Phi for this alias index?
4094 PhiNode *result = get_map_phi(orig_phi->_idx);
4095 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
4096 return result;
4097 }
4098 // Previous check may fail when the same wide memory Phi was split into Phis
4099 // for different memory slices. Search all Phis for this region.
4100 if (result != nullptr) {
4101 Node* region = orig_phi->in(0);
4102 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4103 Node* phi = region->fast_out(i);
4104 if (phi->is_Phi() &&
4105 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4106 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4107 return phi->as_Phi();
4108 }
4109 }
4110 }
4111 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4112 if (C->do_escape_analysis() == true && !C->failing()) {
4113 // Retry compilation without escape analysis.
4114 // If this is the first failure, the sentinel string will "stick"
4115 // to the Compile object, and the C2Compiler will see it and retry.
4116 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4117 }
4118 return nullptr;
4119 }
4120 orig_phi_worklist.append_if_missing(orig_phi);
4121 const TypePtr *atype = C->get_adr_type(alias_idx);
4122 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4123 C->copy_node_notes_to(result, orig_phi);
4124 igvn->set_type(result, result->bottom_type());
4125 record_for_optimizer(result);
4126 set_map(orig_phi, result);
4127 new_created = true;
4128 return result;
4129 }
4130
4131 //
4132 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4133 // specified alias index.
4134 //
4135 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4136 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4137 Compile *C = _compile;
4138 PhaseGVN* igvn = _igvn;
4139 bool new_phi_created;
4140 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4141 if (!new_phi_created) {
4142 return result;
4143 }
4144 GrowableArray<PhiNode *> phi_list;
4145 GrowableArray<uint> cur_input;
4146 PhiNode *phi = orig_phi;
4147 uint idx = 1;
4148 bool finished = false;
4149 while(!finished) {
4150 while (idx < phi->req()) {
4151 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4152 if (mem != nullptr && mem->is_Phi()) {
4153 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4154 if (new_phi_created) {
4155 // found an phi for which we created a new split, push current one on worklist and begin
4156 // processing new one
4157 phi_list.push(phi);
4158 cur_input.push(idx);
4159 phi = mem->as_Phi();
4160 result = newphi;
4161 idx = 1;
4162 continue;
4163 } else {
4164 mem = newphi;
4165 }
4166 }
4167 if (C->failing()) {
4168 return nullptr;
4169 }
4170 result->set_req(idx++, mem);
4171 }
4172 #ifdef ASSERT
4173 // verify that the new Phi has an input for each input of the original
4174 assert( phi->req() == result->req(), "must have same number of inputs.");
4175 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4176 #endif
4177 // Check if all new phi's inputs have specified alias index.
4178 // Otherwise use old phi.
4179 for (uint i = 1; i < phi->req(); i++) {
4180 Node* in = result->in(i);
4181 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4182 }
4183 // we have finished processing a Phi, see if there are any more to do
4184 finished = (phi_list.length() == 0 );
4185 if (!finished) {
4186 phi = phi_list.pop();
4187 idx = cur_input.pop();
4188 PhiNode *prev_result = get_map_phi(phi->_idx);
4189 prev_result->set_req(idx++, result);
4190 result = prev_result;
4191 }
4192 }
4193 return result;
4194 }
4195
4196 //
4197 // The next methods are derived from methods in MemNode.
4198 //
4199 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4200 Node *mem = mmem;
4201 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4202 // means an array I have not precisely typed yet. Do not do any
4203 // alias stuff with it any time soon.
4204 if (toop->base() != Type::AnyPtr &&
4205 !(toop->isa_instptr() &&
4206 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4207 toop->offset() == Type::OffsetBot)) {
4208 mem = mmem->memory_at(alias_idx);
4209 // Update input if it is progress over what we have now
4210 }
4211 return mem;
4212 }
4213
4214 //
4215 // Move memory users to their memory slices.
4216 //
4217 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4218 Compile* C = _compile;
4219 PhaseGVN* igvn = _igvn;
4220 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4221 assert(tp != nullptr, "ptr type");
4222 int alias_idx = C->get_alias_index(tp);
4223 int general_idx = C->get_general_index(alias_idx);
4224
4225 // Move users first
4226 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4227 Node* use = n->fast_out(i);
4228 if (use->is_MergeMem()) {
4229 MergeMemNode* mmem = use->as_MergeMem();
4230 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4231 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4232 continue; // Nothing to do
4233 }
4234 // Replace previous general reference to mem node.
4235 uint orig_uniq = C->unique();
4236 Node* m = find_inst_mem(n, general_idx, orig_phis);
4237 assert(orig_uniq == C->unique(), "no new nodes");
4238 mmem->set_memory_at(general_idx, m);
4239 --imax;
4240 --i;
4241 } else if (use->is_MemBar()) {
4242 assert(!use->is_Initialize(), "initializing stores should not be moved");
4243 if (use->req() > MemBarNode::Precedent &&
4244 use->in(MemBarNode::Precedent) == n) {
4245 // Don't move related membars.
4246 record_for_optimizer(use);
4247 continue;
4248 }
4249 tp = use->as_MemBar()->adr_type()->isa_ptr();
4250 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4251 alias_idx == general_idx) {
4252 continue; // Nothing to do
4253 }
4254 // Move to general memory slice.
4255 uint orig_uniq = C->unique();
4256 Node* m = find_inst_mem(n, general_idx, orig_phis);
4257 assert(orig_uniq == C->unique(), "no new nodes");
4258 igvn->hash_delete(use);
4259 imax -= use->replace_edge(n, m, igvn);
4260 igvn->hash_insert(use);
4261 record_for_optimizer(use);
4262 --i;
4263 #ifdef ASSERT
4264 } else if (use->is_Mem()) {
4265 // Memory nodes should have new memory input.
4266 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4267 assert(tp != nullptr, "ptr type");
4268 int idx = C->get_alias_index(tp);
4269 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4270 "Following memory nodes should have new memory input or be on the same memory slice");
4271 } else if (use->is_Phi()) {
4272 // Phi nodes should be split and moved already.
4273 tp = use->as_Phi()->adr_type()->isa_ptr();
4274 assert(tp != nullptr, "ptr type");
4275 int idx = C->get_alias_index(tp);
4276 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4277 } else {
4278 use->dump();
4279 assert(false, "should not be here");
4280 #endif
4281 }
4282 }
4283 }
4284
4285 //
4286 // Search memory chain of "mem" to find a MemNode whose address
4287 // is the specified alias index.
4288 //
4289 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4290 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4291 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4292 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4293 return nullptr;
4294 }
4295 if (orig_mem == nullptr) {
4296 return orig_mem;
4297 }
4298 Compile* C = _compile;
4299 PhaseGVN* igvn = _igvn;
4300 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4301 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4302 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4303 Node *prev = nullptr;
4304 Node *result = orig_mem;
4305 while (prev != result) {
4306 prev = result;
4307 if (result == start_mem) {
4308 break; // hit one of our sentinels
4309 }
4310 if (result->is_Mem()) {
4311 const Type *at = igvn->type(result->in(MemNode::Address));
4312 if (at == Type::TOP) {
4313 break; // Dead
4314 }
4315 assert (at->isa_ptr() != nullptr, "pointer type required.");
4316 int idx = C->get_alias_index(at->is_ptr());
4317 if (idx == alias_idx) {
4318 break; // Found
4319 }
4320 if (!is_instance && (at->isa_oopptr() == nullptr ||
4321 !at->is_oopptr()->is_known_instance())) {
4322 break; // Do not skip store to general memory slice.
4323 }
4324 result = result->in(MemNode::Memory);
4325 }
4326 if (!is_instance) {
4327 continue; // don't search further for non-instance types
4328 }
4329 // skip over a call which does not affect this memory slice
4330 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4331 Node *proj_in = result->in(0);
4332 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4333 break; // hit one of our sentinels
4334 } else if (proj_in->is_Call()) {
4335 // ArrayCopy node processed here as well
4336 CallNode *call = proj_in->as_Call();
4337 if (!call->may_modify(toop, igvn)) {
4338 result = call->in(TypeFunc::Memory);
4339 }
4340 } else if (proj_in->is_Initialize()) {
4341 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4342 // Stop if this is the initialization for the object instance which
4343 // which contains this memory slice, otherwise skip over it.
4344 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4345 result = proj_in->in(TypeFunc::Memory);
4346 #if 0 // TODO: Fix 8372259
4347 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4348 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4349 result = get_map(result->_idx);
4350 assert(result != nullptr, "new projection should have been allocated");
4351 break;
4352 }
4353 #else
4354 }
4355 #endif
4356 } else if (proj_in->is_MemBar()) {
4357 // Check if there is an array copy for a clone
4358 // Step over GC barrier when ReduceInitialCardMarks is disabled
4359 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4360 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4361
4362 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4363 // Stop if it is a clone
4364 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4365 if (ac->may_modify(toop, igvn)) {
4366 break;
4367 }
4368 }
4369 result = proj_in->in(TypeFunc::Memory);
4370 }
4371 } else if (result->is_MergeMem()) {
4372 MergeMemNode *mmem = result->as_MergeMem();
4373 result = step_through_mergemem(mmem, alias_idx, toop);
4374 if (result == mmem->base_memory()) {
4375 // Didn't find instance memory, search through general slice recursively.
4376 result = mmem->memory_at(C->get_general_index(alias_idx));
4377 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4378 if (C->failing()) {
4379 return nullptr;
4380 }
4381 mmem->set_memory_at(alias_idx, result);
4382 }
4383 } else if (result->is_Phi() &&
4384 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4385 Node *un = result->as_Phi()->unique_input(igvn);
4386 if (un != nullptr) {
4387 orig_phis.append_if_missing(result->as_Phi());
4388 result = un;
4389 } else {
4390 break;
4391 }
4392 } else if (result->is_ClearArray()) {
4393 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4394 // Can not bypass initialization of the instance
4395 // we are looking for.
4396 break;
4397 }
4398 // Otherwise skip it (the call updated 'result' value).
4399 } else if (result->Opcode() == Op_SCMemProj) {
4400 Node* mem = result->in(0);
4401 Node* adr = nullptr;
4402 if (mem->is_LoadStore()) {
4403 adr = mem->in(MemNode::Address);
4404 } else {
4405 assert(mem->Opcode() == Op_EncodeISOArray ||
4406 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4407 adr = mem->in(3); // Memory edge corresponds to destination array
4408 }
4409 const Type *at = igvn->type(adr);
4410 if (at != Type::TOP) {
4411 assert(at->isa_ptr() != nullptr, "pointer type required.");
4412 int idx = C->get_alias_index(at->is_ptr());
4413 if (idx == alias_idx) {
4414 // Assert in debug mode
4415 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4416 break; // In product mode return SCMemProj node
4417 }
4418 }
4419 result = mem->in(MemNode::Memory);
4420 } else if (result->Opcode() == Op_StrInflatedCopy) {
4421 Node* adr = result->in(3); // Memory edge corresponds to destination array
4422 const Type *at = igvn->type(adr);
4423 if (at != Type::TOP) {
4424 assert(at->isa_ptr() != nullptr, "pointer type required.");
4425 int idx = C->get_alias_index(at->is_ptr());
4426 if (idx == alias_idx) {
4427 // Assert in debug mode
4428 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4429 break; // In product mode return SCMemProj node
4430 }
4431 }
4432 result = result->in(MemNode::Memory);
4433 }
4434 }
4435 if (result->is_Phi()) {
4436 PhiNode *mphi = result->as_Phi();
4437 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4438 const TypePtr *t = mphi->adr_type();
4439 if (!is_instance) {
4440 // Push all non-instance Phis on the orig_phis worklist to update inputs
4441 // during Phase 4 if needed.
4442 orig_phis.append_if_missing(mphi);
4443 } else if (C->get_alias_index(t) != alias_idx) {
4444 // Create a new Phi with the specified alias index type.
4445 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4446 }
4447 }
4448 // the result is either MemNode, PhiNode, InitializeNode.
4449 return result;
4450 }
4451
4452 //
4453 // Convert the types of non-escaped object to instance types where possible,
4454 // propagate the new type information through the graph, and update memory
4455 // edges and MergeMem inputs to reflect the new type.
4456 //
4457 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4458 // The processing is done in 4 phases:
4459 //
4460 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4461 // types for the CheckCastPP for allocations where possible.
4462 // Propagate the new types through users as follows:
4463 // casts and Phi: push users on alloc_worklist
4464 // AddP: cast Base and Address inputs to the instance type
4465 // push any AddP users on alloc_worklist and push any memnode
4466 // users onto memnode_worklist.
4467 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4468 // search the Memory chain for a store with the appropriate type
4469 // address type. If a Phi is found, create a new version with
4470 // the appropriate memory slices from each of the Phi inputs.
4471 // For stores, process the users as follows:
4472 // MemNode: push on memnode_worklist
4473 // MergeMem: push on mergemem_worklist
4474 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4475 // moving the first node encountered of each instance type to the
4476 // the input corresponding to its alias index.
4477 // appropriate memory slice.
4478 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4479 //
4480 // In the following example, the CheckCastPP nodes are the cast of allocation
4481 // results and the allocation of node 29 is non-escaped and eligible to be an
4482 // instance type.
4483 //
4484 // We start with:
4485 //
4486 // 7 Parm #memory
4487 // 10 ConI "12"
4488 // 19 CheckCastPP "Foo"
4489 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4490 // 29 CheckCastPP "Foo"
4491 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4492 //
4493 // 40 StoreP 25 7 20 ... alias_index=4
4494 // 50 StoreP 35 40 30 ... alias_index=4
4495 // 60 StoreP 45 50 20 ... alias_index=4
4496 // 70 LoadP _ 60 30 ... alias_index=4
4497 // 80 Phi 75 50 60 Memory alias_index=4
4498 // 90 LoadP _ 80 30 ... alias_index=4
4499 // 100 LoadP _ 80 20 ... alias_index=4
4500 //
4501 //
4502 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4503 // and creating a new alias index for node 30. This gives:
4504 //
4505 // 7 Parm #memory
4506 // 10 ConI "12"
4507 // 19 CheckCastPP "Foo"
4508 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4509 // 29 CheckCastPP "Foo" iid=24
4510 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4511 //
4512 // 40 StoreP 25 7 20 ... alias_index=4
4513 // 50 StoreP 35 40 30 ... alias_index=6
4514 // 60 StoreP 45 50 20 ... alias_index=4
4515 // 70 LoadP _ 60 30 ... alias_index=6
4516 // 80 Phi 75 50 60 Memory alias_index=4
4517 // 90 LoadP _ 80 30 ... alias_index=6
4518 // 100 LoadP _ 80 20 ... alias_index=4
4519 //
4520 // In phase 2, new memory inputs are computed for the loads and stores,
4521 // And a new version of the phi is created. In phase 4, the inputs to
4522 // node 80 are updated and then the memory nodes are updated with the
4523 // values computed in phase 2. This results in:
4524 //
4525 // 7 Parm #memory
4526 // 10 ConI "12"
4527 // 19 CheckCastPP "Foo"
4528 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4529 // 29 CheckCastPP "Foo" iid=24
4530 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4531 //
4532 // 40 StoreP 25 7 20 ... alias_index=4
4533 // 50 StoreP 35 7 30 ... alias_index=6
4534 // 60 StoreP 45 40 20 ... alias_index=4
4535 // 70 LoadP _ 50 30 ... alias_index=6
4536 // 80 Phi 75 40 60 Memory alias_index=4
4537 // 120 Phi 75 50 50 Memory alias_index=6
4538 // 90 LoadP _ 120 30 ... alias_index=6
4539 // 100 LoadP _ 80 20 ... alias_index=4
4540 //
4541 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4542 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4543 GrowableArray<MergeMemNode*> &mergemem_worklist,
4544 Unique_Node_List &reducible_merges) {
4545 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4546 GrowableArray<Node *> memnode_worklist;
4547 GrowableArray<PhiNode *> orig_phis;
4548 PhaseIterGVN *igvn = _igvn;
4549 uint new_index_start = (uint) _compile->num_alias_types();
4550 VectorSet visited;
4551 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4552 uint unique_old = _compile->unique();
4553
4554 // Phase 1: Process possible allocations from alloc_worklist.
4555 // Create instance types for the CheckCastPP for allocations where possible.
4556 //
4557 // (Note: don't forget to change the order of the second AddP node on
4558 // the alloc_worklist if the order of the worklist processing is changed,
4559 // see the comment in find_second_addp().)
4560 //
4561 while (alloc_worklist.length() != 0) {
4562 Node *n = alloc_worklist.pop();
4563 uint ni = n->_idx;
4564 if (n->is_Call()) {
4565 CallNode *alloc = n->as_Call();
4566 // copy escape information to call node
4567 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4568 PointsToNode::EscapeState es = ptn->escape_state();
4569 // We have an allocation or call which returns a Java object,
4570 // see if it is non-escaped.
4571 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4572 continue;
4573 }
4574 // Find CheckCastPP for the allocate or for the return value of a call
4575 n = alloc->result_cast();
4576 if (n == nullptr) { // No uses except Initialize node
4577 if (alloc->is_Allocate()) {
4578 // Set the scalar_replaceable flag for allocation
4579 // so it could be eliminated if it has no uses.
4580 alloc->as_Allocate()->_is_scalar_replaceable = true;
4581 }
4582 continue;
4583 }
4584 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4585 // we could reach here for allocate case if one init is associated with many allocs.
4586 if (alloc->is_Allocate()) {
4587 alloc->as_Allocate()->_is_scalar_replaceable = false;
4588 }
4589 continue;
4590 }
4591
4592 // The inline code for Object.clone() casts the allocation result to
4593 // java.lang.Object and then to the actual type of the allocated
4594 // object. Detect this case and use the second cast.
4595 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4596 // the allocation result is cast to java.lang.Object and then
4597 // to the actual Array type.
4598 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4599 && (alloc->is_AllocateArray() ||
4600 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4601 Node *cast2 = nullptr;
4602 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4603 Node *use = n->fast_out(i);
4604 if (use->is_CheckCastPP()) {
4605 cast2 = use;
4606 break;
4607 }
4608 }
4609 if (cast2 != nullptr) {
4610 n = cast2;
4611 } else {
4612 // Non-scalar replaceable if the allocation type is unknown statically
4613 // (reflection allocation), the object can't be restored during
4614 // deoptimization without precise type.
4615 continue;
4616 }
4617 }
4618
4619 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4620 if (t == nullptr) {
4621 continue; // not a TypeOopPtr
4622 }
4623 if (!t->klass_is_exact()) {
4624 continue; // not an unique type
4625 }
4626 if (alloc->is_Allocate()) {
4627 // Set the scalar_replaceable flag for allocation
4628 // so it could be eliminated.
4629 alloc->as_Allocate()->_is_scalar_replaceable = true;
4630 }
4631 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4632 // in order for an object to be scalar-replaceable, it must be:
4633 // - a direct allocation (not a call returning an object)
4634 // - non-escaping
4635 // - eligible to be a unique type
4636 // - not determined to be ineligible by escape analysis
4637 set_map(alloc, n);
4638 set_map(n, alloc);
4639 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4640 igvn->hash_delete(n);
4641 igvn->set_type(n, tinst);
4642 n->raise_bottom_type(tinst);
4643 igvn->hash_insert(n);
4644 record_for_optimizer(n);
4645 // Allocate an alias index for the header fields. Accesses to
4646 // the header emitted during macro expansion wouldn't have
4647 // correct memory state otherwise.
4648 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4649 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4650 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4651 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4652 InitializeNode* init = alloc->as_Allocate()->initialization();
4653 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4654 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4655 const TypePtr* adr_type = proj->adr_type();
4656 const TypePtr* new_adr_type = tinst->with_offset(adr_type->offset());
4657 if (adr_type->isa_aryptr()) {
4658 // In the case of a flat inline type array, each field has its own slice so we need a
4659 // NarrowMemProj for each field of the flat array elements
4660 new_adr_type = new_adr_type->is_aryptr()->with_field_offset(adr_type->is_aryptr()->field_offset().get());
4661 }
4662 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4663 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4664 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4665 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4666 igvn->set_type(new_proj, new_proj->bottom_type());
4667 record_for_optimizer(new_proj);
4668 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4669 }
4670 };
4671 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4672
4673 // First, put on the worklist all Field edges from Connection Graph
4674 // which is more accurate than putting immediate users from Ideal Graph.
4675 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4676 PointsToNode* tgt = e.get();
4677 if (tgt->is_Arraycopy()) {
4678 continue;
4679 }
4680 Node* use = tgt->ideal_node();
4681 assert(tgt->is_Field() && use->is_AddP(),
4682 "only AddP nodes are Field edges in CG");
4683 if (use->outcnt() > 0) { // Don't process dead nodes
4684 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4685 if (addp2 != nullptr) {
4686 assert(alloc->is_AllocateArray(),"array allocation was expected");
4687 alloc_worklist.append_if_missing(addp2);
4688 }
4689 alloc_worklist.append_if_missing(use);
4690 }
4691 }
4692
4693 // An allocation may have an Initialize which has raw stores. Scan
4694 // the users of the raw allocation result and push AddP users
4695 // on alloc_worklist.
4696 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4697 assert (raw_result != nullptr, "must have an allocation result");
4698 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4699 Node *use = raw_result->fast_out(i);
4700 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4701 Node* addp2 = find_second_addp(use, raw_result);
4702 if (addp2 != nullptr) {
4703 assert(alloc->is_AllocateArray(),"array allocation was expected");
4704 alloc_worklist.append_if_missing(addp2);
4705 }
4706 alloc_worklist.append_if_missing(use);
4707 } else if (use->is_MemBar()) {
4708 memnode_worklist.append_if_missing(use);
4709 }
4710 }
4711 }
4712 } else if (n->is_AddP()) {
4713 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4714 // This AddP will go away when we reduce the Phi
4715 continue;
4716 }
4717 Node* addp_base = get_addp_base(n);
4718 JavaObjectNode* jobj = unique_java_object(addp_base);
4719 if (jobj == nullptr || jobj == phantom_obj) {
4720 #ifdef ASSERT
4721 ptnode_adr(get_addp_base(n)->_idx)->dump();
4722 ptnode_adr(n->_idx)->dump();
4723 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4724 #endif
4725 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4726 return;
4727 }
4728 Node *base = get_map(jobj->idx()); // CheckCastPP node
4729 if (!split_AddP(n, base)) continue; // wrong type from dead path
4730 } else if (n->is_Phi() ||
4731 n->is_CheckCastPP() ||
4732 n->is_EncodeP() ||
4733 n->is_DecodeN() ||
4734 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4735 if (visited.test_set(n->_idx)) {
4736 assert(n->is_Phi(), "loops only through Phi's");
4737 continue; // already processed
4738 }
4739 // Reducible Phi's will be removed from the graph after split_unique_types
4740 // finishes. For now we just try to split out the SR inputs of the merge.
4741 Node* parent = n->in(1);
4742 if (reducible_merges.member(n)) {
4743 reduce_phi(n->as_Phi(), alloc_worklist);
4744 #ifdef ASSERT
4745 if (VerifyReduceAllocationMerges) {
4746 reduced_merges.push(n);
4747 }
4748 #endif
4749 continue;
4750 } else if (reducible_merges.member(parent)) {
4751 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4752 // part of reduce_merge.
4753 continue;
4754 }
4755 JavaObjectNode* jobj = unique_java_object(n);
4756 if (jobj == nullptr || jobj == phantom_obj) {
4757 #ifdef ASSERT
4758 ptnode_adr(n->_idx)->dump();
4759 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4760 #endif
4761 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4762 return;
4763 } else {
4764 Node *val = get_map(jobj->idx()); // CheckCastPP node
4765 TypeNode *tn = n->as_Type();
4766 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4767 assert(tinst != nullptr && tinst->is_known_instance() &&
4768 tinst->instance_id() == jobj->idx() , "instance type expected.");
4769
4770 const Type *tn_type = igvn->type(tn);
4771 const TypeOopPtr *tn_t;
4772 if (tn_type->isa_narrowoop()) {
4773 tn_t = tn_type->make_ptr()->isa_oopptr();
4774 } else {
4775 tn_t = tn_type->isa_oopptr();
4776 }
4777 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4778 if (tn_t->isa_aryptr()) {
4779 // Keep array properties (not flat/null-free)
4780 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
4781 if (tinst == nullptr) {
4782 continue; // Skip dead path with inconsistent properties
4783 }
4784 }
4785 if (tn_type->isa_narrowoop()) {
4786 tn_type = tinst->make_narrowoop();
4787 } else {
4788 tn_type = tinst;
4789 }
4790 igvn->hash_delete(tn);
4791 igvn->set_type(tn, tn_type);
4792 tn->set_type(tn_type);
4793 igvn->hash_insert(tn);
4794 record_for_optimizer(n);
4795 } else {
4796 assert(tn_type == TypePtr::NULL_PTR ||
4797 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4798 "unexpected type");
4799 continue; // Skip dead path with different type
4800 }
4801 }
4802 } else {
4803 DEBUG_ONLY(n->dump();)
4804 assert(false, "EA: unexpected node");
4805 continue;
4806 }
4807 // push allocation's users on appropriate worklist
4808 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4809 Node *use = n->fast_out(i);
4810 if (use->is_Mem() && use->in(MemNode::Address) == n) {
4811 // Load/store to instance's field
4812 memnode_worklist.append_if_missing(use);
4813 } else if (use->is_MemBar()) {
4814 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4815 memnode_worklist.append_if_missing(use);
4816 }
4817 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4818 Node* addp2 = find_second_addp(use, n);
4819 if (addp2 != nullptr) {
4820 alloc_worklist.append_if_missing(addp2);
4821 }
4822 alloc_worklist.append_if_missing(use);
4823 } else if (use->is_Phi() ||
4824 use->is_CheckCastPP() ||
4825 use->is_EncodeNarrowPtr() ||
4826 use->is_DecodeNarrowPtr() ||
4827 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4828 alloc_worklist.append_if_missing(use);
4829 #ifdef ASSERT
4830 } else if (use->is_Mem()) {
4831 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4832 } else if (use->is_MergeMem()) {
4833 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4834 } else if (use->is_SafePoint()) {
4835 // Look for MergeMem nodes for calls which reference unique allocation
4836 // (through CheckCastPP nodes) even for debug info.
4837 Node* m = use->in(TypeFunc::Memory);
4838 if (m->is_MergeMem()) {
4839 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4840 }
4841 } else if (use->Opcode() == Op_EncodeISOArray) {
4842 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4843 // EncodeISOArray overwrites destination array
4844 memnode_worklist.append_if_missing(use);
4845 }
4846 } else if (use->Opcode() == Op_Return) {
4847 // Allocation is referenced by field of returned inline type
4848 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
4849 } else {
4850 uint op = use->Opcode();
4851 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4852 (use->in(MemNode::Memory) == n)) {
4853 // They overwrite memory edge corresponding to destination array,
4854 memnode_worklist.append_if_missing(use);
4855 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4856 op == Op_CastP2X ||
4857 op == Op_FastLock || op == Op_AryEq ||
4858 op == Op_StrComp || op == Op_CountPositives ||
4859 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4860 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4861 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4862 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
4863 op == Op_ReinterpretS2HF ||
4864 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4865 n->dump();
4866 use->dump();
4867 assert(false, "EA: missing allocation reference path");
4868 }
4869 #endif
4870 }
4871 }
4872
4873 }
4874
4875 #ifdef ASSERT
4876 if (VerifyReduceAllocationMerges) {
4877 for (uint i = 0; i < reducible_merges.size(); i++) {
4878 Node* phi = reducible_merges.at(i);
4879
4880 if (!reduced_merges.member(phi)) {
4881 phi->dump(2);
4882 phi->dump(-2);
4883 assert(false, "This reducible merge wasn't reduced.");
4884 }
4885
4886 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4887 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4888 Node* use = phi->fast_out(j);
4889 if (!use->is_SafePoint() && !use->is_CastPP()) {
4890 phi->dump(2);
4891 phi->dump(-2);
4892 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4893 }
4894 }
4895 }
4896 }
4897 #endif
4898
4899 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4900 // type, record it in the ArrayCopy node so we know what memory this
4901 // node uses/modified.
4902 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4903 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4904 Node* dest = ac->in(ArrayCopyNode::Dest);
4905 if (dest->is_AddP()) {
4906 dest = get_addp_base(dest);
4907 }
4908 JavaObjectNode* jobj = unique_java_object(dest);
4909 if (jobj != nullptr) {
4910 Node *base = get_map(jobj->idx());
4911 if (base != nullptr) {
4912 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4913 ac->_dest_type = base_t;
4914 }
4915 }
4916 Node* src = ac->in(ArrayCopyNode::Src);
4917 if (src->is_AddP()) {
4918 src = get_addp_base(src);
4919 }
4920 jobj = unique_java_object(src);
4921 if (jobj != nullptr) {
4922 Node* base = get_map(jobj->idx());
4923 if (base != nullptr) {
4924 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4925 ac->_src_type = base_t;
4926 }
4927 }
4928 }
4929
4930 // New alias types were created in split_AddP().
4931 uint new_index_end = (uint) _compile->num_alias_types();
4932
4933 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
4934
4935 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4936 // compute new values for Memory inputs (the Memory inputs are not
4937 // actually updated until phase 4.)
4938 if (memnode_worklist.length() == 0)
4939 return; // nothing to do
4940 while (memnode_worklist.length() != 0) {
4941 Node *n = memnode_worklist.pop();
4942 if (visited.test_set(n->_idx)) {
4943 continue;
4944 }
4945 if (n->is_Phi() || n->is_ClearArray()) {
4946 // we don't need to do anything, but the users must be pushed
4947 } else if (n->is_MemBar()) { // MemBar nodes
4948 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
4949 // we don't need to do anything, but the users must be pushed
4950 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4951 if (n == nullptr) {
4952 continue;
4953 }
4954 }
4955 } else if (n->is_CallLeaf()) {
4956 // Runtime calls with narrow memory input (no MergeMem node)
4957 // get the memory projection
4958 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4959 if (n == nullptr) {
4960 continue;
4961 }
4962 } else if (n->Opcode() == Op_StrInflatedCopy) {
4963 // Check direct uses of StrInflatedCopy.
4964 // It is memory type Node - no special SCMemProj node.
4965 } else if (n->Opcode() == Op_StrCompressedCopy ||
4966 n->Opcode() == Op_EncodeISOArray) {
4967 // get the memory projection
4968 n = n->find_out_with(Op_SCMemProj);
4969 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4970 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
4971 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
4972 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
4973 } else if (n->is_Proj()) {
4974 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
4975 } else {
4976 #ifdef ASSERT
4977 if (!n->is_Mem()) {
4978 n->dump();
4979 }
4980 assert(n->is_Mem(), "memory node required.");
4981 #endif
4982 Node *addr = n->in(MemNode::Address);
4983 const Type *addr_t = igvn->type(addr);
4984 if (addr_t == Type::TOP) {
4985 continue;
4986 }
4987 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4988 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4989 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4990 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4991 if (_compile->failing()) {
4992 return;
4993 }
4994 if (mem != n->in(MemNode::Memory)) {
4995 // We delay the memory edge update since we need old one in
4996 // MergeMem code below when instances memory slices are separated.
4997 set_map(n, mem);
4998 }
4999 if (n->is_Load()) {
5000 continue; // don't push users
5001 } else if (n->is_LoadStore()) {
5002 // get the memory projection
5003 n = n->find_out_with(Op_SCMemProj);
5004 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5005 }
5006 }
5007 // push user on appropriate worklist
5008 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5009 Node *use = n->fast_out(i);
5010 if (use->is_Phi() || use->is_ClearArray()) {
5011 memnode_worklist.append_if_missing(use);
5012 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
5013 memnode_worklist.append_if_missing(use);
5014 } else if (use->is_MemBar() || use->is_CallLeaf()) {
5015 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5016 memnode_worklist.append_if_missing(use);
5017 }
5018 } else if (use->is_Proj()) {
5019 assert(n->is_Initialize(), "We only push projections of Initialize");
5020 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
5021 memnode_worklist.append_if_missing(use);
5022 }
5023 #ifdef ASSERT
5024 } else if (use->is_Mem()) {
5025 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
5026 } else if (use->is_MergeMem()) {
5027 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5028 } else if (use->Opcode() == Op_EncodeISOArray) {
5029 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5030 // EncodeISOArray overwrites destination array
5031 memnode_worklist.append_if_missing(use);
5032 }
5033 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
5034 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5035 // store_unknown_inline overwrites destination array
5036 memnode_worklist.append_if_missing(use);
5037 } else {
5038 uint op = use->Opcode();
5039 if ((use->in(MemNode::Memory) == n) &&
5040 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
5041 // They overwrite memory edge corresponding to destination array,
5042 memnode_worklist.append_if_missing(use);
5043 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
5044 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
5045 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
5046 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
5047 n->dump();
5048 use->dump();
5049 assert(false, "EA: missing memory path");
5050 }
5051 #endif
5052 }
5053 }
5054 }
5055
5056 // Phase 3: Process MergeMem nodes from mergemem_worklist.
5057 // Walk each memory slice moving the first node encountered of each
5058 // instance type to the input corresponding to its alias index.
5059 uint length = mergemem_worklist.length();
5060 for( uint next = 0; next < length; ++next ) {
5061 MergeMemNode* nmm = mergemem_worklist.at(next);
5062 assert(!visited.test_set(nmm->_idx), "should not be visited before");
5063 // Note: we don't want to use MergeMemStream here because we only want to
5064 // scan inputs which exist at the start, not ones we add during processing.
5065 // Note 2: MergeMem may already contains instance memory slices added
5066 // during find_inst_mem() call when memory nodes were processed above.
5067 igvn->hash_delete(nmm);
5068 uint nslices = MIN2(nmm->req(), new_index_start);
5069 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
5070 Node* mem = nmm->in(i);
5071 Node* cur = nullptr;
5072 if (mem == nullptr || mem->is_top()) {
5073 continue;
5074 }
5075 // First, update mergemem by moving memory nodes to corresponding slices
5076 // if their type became more precise since this mergemem was created.
5077 while (mem->is_Mem()) {
5078 const Type* at = igvn->type(mem->in(MemNode::Address));
5079 if (at != Type::TOP) {
5080 assert (at->isa_ptr() != nullptr, "pointer type required.");
5081 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
5082 if (idx == i) {
5083 if (cur == nullptr) {
5084 cur = mem;
5085 }
5086 } else {
5087 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
5088 nmm->set_memory_at(idx, mem);
5089 }
5090 }
5091 }
5092 mem = mem->in(MemNode::Memory);
5093 }
5094 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
5095 // Find any instance of the current type if we haven't encountered
5096 // already a memory slice of the instance along the memory chain.
5097 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5098 if((uint)_compile->get_general_index(ni) == i) {
5099 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
5100 if (nmm->is_empty_memory(m)) {
5101 Node* result = find_inst_mem(mem, ni, orig_phis);
5102 if (_compile->failing()) {
5103 return;
5104 }
5105 nmm->set_memory_at(ni, result);
5106 }
5107 }
5108 }
5109 }
5110 // Find the rest of instances values
5111 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5112 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
5113 Node* result = step_through_mergemem(nmm, ni, tinst);
5114 if (result == nmm->base_memory()) {
5115 // Didn't find instance memory, search through general slice recursively.
5116 result = nmm->memory_at(_compile->get_general_index(ni));
5117 result = find_inst_mem(result, ni, orig_phis);
5118 if (_compile->failing()) {
5119 return;
5120 }
5121 nmm->set_memory_at(ni, result);
5122 }
5123 }
5124
5125 // If we have crossed the 3/4 point of max node limit it's too risky
5126 // to continue with EA/SR because we might hit the max node limit.
5127 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
5128 if (_compile->do_reduce_allocation_merges()) {
5129 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
5130 } else if (_invocation > 0) {
5131 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
5132 } else {
5133 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
5134 }
5135 return;
5136 }
5137
5138 igvn->hash_insert(nmm);
5139 record_for_optimizer(nmm);
5140 }
5141
5142 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5143
5144 // Phase 4: Update the inputs of non-instance memory Phis and
5145 // the Memory input of memnodes
5146 // First update the inputs of any non-instance Phi's from
5147 // which we split out an instance Phi. Note we don't have
5148 // to recursively process Phi's encountered on the input memory
5149 // chains as is done in split_memory_phi() since they will
5150 // also be processed here.
5151 for (int j = 0; j < orig_phis.length(); j++) {
5152 PhiNode *phi = orig_phis.at(j);
5153 int alias_idx = _compile->get_alias_index(phi->adr_type());
5154 igvn->hash_delete(phi);
5155 for (uint i = 1; i < phi->req(); i++) {
5156 Node *mem = phi->in(i);
5157 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5158 if (_compile->failing()) {
5159 return;
5160 }
5161 if (mem != new_mem) {
5162 phi->set_req(i, new_mem);
5163 }
5164 }
5165 igvn->hash_insert(phi);
5166 record_for_optimizer(phi);
5167 }
5168
5169 // Update the memory inputs of MemNodes with the value we computed
5170 // in Phase 2 and move stores memory users to corresponding memory slices.
5171 // Disable memory split verification code until the fix for 6984348.
5172 // Currently it produces false negative results since it does not cover all cases.
5173 #if 0 // ifdef ASSERT
5174 visited.Reset();
5175 Node_Stack old_mems(arena, _compile->unique() >> 2);
5176 #endif
5177 for (uint i = 0; i < ideal_nodes.size(); i++) {
5178 Node* n = ideal_nodes.at(i);
5179 Node* nmem = get_map(n->_idx);
5180 assert(nmem != nullptr, "sanity");
5181 if (n->is_Mem()) {
5182 #if 0 // ifdef ASSERT
5183 Node* old_mem = n->in(MemNode::Memory);
5184 if (!visited.test_set(old_mem->_idx)) {
5185 old_mems.push(old_mem, old_mem->outcnt());
5186 }
5187 #endif
5188 assert(n->in(MemNode::Memory) != nmem, "sanity");
5189 if (!n->is_Load()) {
5190 // Move memory users of a store first.
5191 move_inst_mem(n, orig_phis);
5192 }
5193 // Now update memory input
5194 igvn->hash_delete(n);
5195 n->set_req(MemNode::Memory, nmem);
5196 igvn->hash_insert(n);
5197 record_for_optimizer(n);
5198 } else {
5199 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5200 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5201 }
5202 }
5203 #if 0 // ifdef ASSERT
5204 // Verify that memory was split correctly
5205 while (old_mems.is_nonempty()) {
5206 Node* old_mem = old_mems.node();
5207 uint old_cnt = old_mems.index();
5208 old_mems.pop();
5209 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5210 }
5211 #endif
5212 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5213 }
5214
5215 #ifndef PRODUCT
5216 int ConnectionGraph::_no_escape_counter = 0;
5217 int ConnectionGraph::_arg_escape_counter = 0;
5218 int ConnectionGraph::_global_escape_counter = 0;
5219
5220 static const char *node_type_names[] = {
5221 "UnknownType",
5222 "JavaObject",
5223 "LocalVar",
5224 "Field",
5225 "Arraycopy"
5226 };
5227
5228 static const char *esc_names[] = {
5229 "UnknownEscape",
5230 "NoEscape",
5231 "ArgEscape",
5232 "GlobalEscape"
5233 };
5234
5235 const char* PointsToNode::esc_name() const {
5236 return esc_names[(int)escape_state()];
5237 }
5238
5239 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5240 NodeType nt = node_type();
5241 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5242 if (print_state) {
5243 EscapeState es = escape_state();
5244 EscapeState fields_es = fields_escape_state();
5245 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5246 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5247 out->print("NSR ");
5248 }
5249 }
5250 }
5251
5252 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5253 dump_header(print_state, out);
5254 if (is_Field()) {
5255 FieldNode* f = (FieldNode*)this;
5256 if (f->is_oop()) {
5257 out->print("oop ");
5258 }
5259 if (f->offset() > 0) {
5260 out->print("+%d ", f->offset());
5261 }
5262 out->print("(");
5263 for (BaseIterator i(f); i.has_next(); i.next()) {
5264 PointsToNode* b = i.get();
5265 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5266 }
5267 out->print(" )");
5268 }
5269 out->print("[");
5270 for (EdgeIterator i(this); i.has_next(); i.next()) {
5271 PointsToNode* e = i.get();
5272 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5273 }
5274 out->print(" [");
5275 for (UseIterator i(this); i.has_next(); i.next()) {
5276 PointsToNode* u = i.get();
5277 bool is_base = false;
5278 if (PointsToNode::is_base_use(u)) {
5279 is_base = true;
5280 u = PointsToNode::get_use_node(u)->as_Field();
5281 }
5282 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5283 }
5284 out->print(" ]] ");
5285 if (_node == nullptr) {
5286 out->print("<null>%s", newline ? "\n" : "");
5287 } else {
5288 _node->dump(newline ? "\n" : "", false, out);
5289 }
5290 }
5291
5292 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5293 bool first = true;
5294 int ptnodes_length = ptnodes_worklist.length();
5295 for (int i = 0; i < ptnodes_length; i++) {
5296 PointsToNode *ptn = ptnodes_worklist.at(i);
5297 if (ptn == nullptr || !ptn->is_JavaObject()) {
5298 continue;
5299 }
5300 PointsToNode::EscapeState es = ptn->escape_state();
5301 if ((es != PointsToNode::NoEscape) && !Verbose) {
5302 continue;
5303 }
5304 Node* n = ptn->ideal_node();
5305 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5306 n->as_CallStaticJava()->is_boxing_method())) {
5307 if (first) {
5308 tty->cr();
5309 tty->print("======== Connection graph for ");
5310 _compile->method()->print_short_name();
5311 tty->cr();
5312 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5313 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5314 tty->cr();
5315 first = false;
5316 }
5317 ptn->dump();
5318 // Print all locals and fields which reference this allocation
5319 for (UseIterator j(ptn); j.has_next(); j.next()) {
5320 PointsToNode* use = j.get();
5321 if (use->is_LocalVar()) {
5322 use->dump(Verbose);
5323 } else if (Verbose) {
5324 use->dump();
5325 }
5326 }
5327 tty->cr();
5328 }
5329 }
5330 }
5331
5332 void ConnectionGraph::print_statistics() {
5333 tty->print_cr("No escape = %d, Arg escape = %d, Global escape = %d", AtomicAccess::load(&_no_escape_counter), AtomicAccess::load(&_arg_escape_counter), AtomicAccess::load(&_global_escape_counter));
5334 }
5335
5336 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5337 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5338 return;
5339 }
5340 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5341 JavaObjectNode* ptn = java_objects_worklist.at(next);
5342 if (ptn->ideal_node()->is_Allocate()) {
5343 if (ptn->escape_state() == PointsToNode::NoEscape) {
5344 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5345 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5346 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5347 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5348 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5349 } else {
5350 assert(false, "Unexpected Escape State");
5351 }
5352 }
5353 }
5354 }
5355
5356 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5357 if (_compile->directive()->TraceEscapeAnalysisOption) {
5358 assert(ptn != nullptr, "should not be null");
5359 assert(reason != nullptr, "should not be null");
5360 ptn->dump_header(true);
5361 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5362 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5363 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5364 }
5365 }
5366
5367 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5368 if (_compile->directive()->TraceEscapeAnalysisOption) {
5369 stringStream ss;
5370 ss.print("propagated from: ");
5371 from->dump(true, &ss, false);
5372 return ss.as_string();
5373 } else {
5374 return nullptr;
5375 }
5376 }
5377
5378 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5379 if (_compile->directive()->TraceEscapeAnalysisOption) {
5380 stringStream ss;
5381 ss.print("escapes as arg to:");
5382 call->dump("", false, &ss);
5383 return ss.as_string();
5384 } else {
5385 return nullptr;
5386 }
5387 }
5388
5389 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5390 if (_compile->directive()->TraceEscapeAnalysisOption) {
5391 stringStream ss;
5392 ss.print("is merged with other object: ");
5393 other->dump_header(true, &ss);
5394 return ss.as_string();
5395 } else {
5396 return nullptr;
5397 }
5398 }
5399
5400 #endif
5401
5402 void ConnectionGraph::record_for_optimizer(Node *n) {
5403 _igvn->_worklist.push(n);
5404 _igvn->add_users_to_worklist(n);
5405 }