1 /*
2 * Copyright (c) 2005, 2026, 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
1082 // If new_load is a Load but not from an AddP, it means that the load is folded into another
1083 // load. And since this load is not from a field, we cannot create a unique type for it.
1084 // For example:
1085 //
1086 // if (b) {
1087 // Holder h1 = new Holder();
1088 // Object o = ...;
1089 // h.o = o.getClass();
1090 // } else {
1091 // Holder h2 = ...;
1092 // }
1093 // Holder h = Phi(h1, h2);
1094 // Object r = h.o;
1095 //
1096 // Then, splitting r through the merge point results in:
1097 //
1098 // if (b) {
1099 // Holder h1 = new Holder();
1100 // Object o = ...;
1101 // h.o = o.getClass();
1102 // Object o1 = h.o;
1103 // } else {
1104 // Holder h2 = ...;
1105 // Object o2 = h2.o;
1106 // }
1107 // Object r = Phi(o1, o2);
1108 //
1109 // In this case, o1 is folded to o.getClass() which is a Load but not from an AddP, but from
1110 // an OopHandle that is loaded from the Klass of o.
1111 if (!new_addp->is_AddP()) {
1112 continue;
1113 }
1114 Node* base = get_addp_base(new_addp);
1115
1116 // The base might not be something that we can create an unique
1117 // type for. If that's the case we are done with that input.
1118 PointsToNode* jobj_ptn = unique_java_object(base);
1119 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1120 continue;
1121 }
1122
1123 // Push to alloc_worklist since the base has an unique_type
1124 alloc_worklist.append_if_missing(new_addp);
1125
1126 // Now let's add the node to the connection graph
1127 _nodes.at_grow(new_addp->_idx, nullptr);
1128 add_field(new_addp, fn->escape_state(), fn->offset());
1129 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1130
1131 // If the load doesn't load an object then it won't be
1132 // part of the connection graph
1133 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1134 if (curr_load_ptn != nullptr) {
1135 _nodes.at_grow(new_load->_idx, nullptr);
1136 add_local_var(new_load, curr_load_ptn->escape_state());
1137 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1138 }
1139 }
1140 }
1141 }
1142
1143 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1144 // We'll pass this to 'split_through_phi' so that it'll do the split even
1145 // though the load doesn't have an unique instance type.
1146 bool ignore_missing_instance_id = true;
1147
1148 // All AddPs are present in the connection graph
1149 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1150
1151 // Iterate over AddP looking for a Load
1152 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1153 Node* previous_load = previous_addp->raw_out(k);
1154 if (previous_load->is_Load()) {
1155 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1156
1157 // Takes care of updating CG and split_unique_types worklists due to cloned
1158 // AddP->Load.
1159 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1160
1161 _igvn->replace_node(previous_load, data_phi);
1162 }
1163 --k;
1164 k = MIN2(k, (int)previous_addp->outcnt()-1);
1165 }
1166
1167 // Remove the old AddP from the processing list because it's dead now
1168 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1169 alloc_worklist.remove_if_existing(previous_addp);
1170 }
1171
1172 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1173 // selector is:
1174 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1175 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1176 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1177 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1178 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1179 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1180 uint number_of_sr_objects = 0;
1181 for (uint i = 1; i < ophi->req(); i++) {
1182 Node* base = ophi->in(i);
1183 JavaObjectNode* ptn = unique_java_object(base);
1184
1185 if (ptn != nullptr && ptn->scalar_replaceable()) {
1186 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1187 selector->set_req(i, sr_obj_idx);
1188 number_of_sr_objects++;
1189 }
1190 }
1191
1192 return selector->as_Phi();
1193 }
1194
1195 // Returns true if the AddP node 'n' has at least one base that is a reducible
1196 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1197 // checked instead.
1198 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1199 PointsToNode* ptn = ptnode_adr(n->_idx);
1200 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1201 return false;
1202 }
1203
1204 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1205 Node* base = i.get()->ideal_node();
1206
1207 if (reducible_merges.member(base)) {
1208 return true;
1209 }
1210
1211 if (base->is_CastPP() || base->is_CheckCastPP()) {
1212 base = base->in(1);
1213 if (reducible_merges.member(base)) {
1214 return true;
1215 }
1216 }
1217 }
1218
1219 return false;
1220 }
1221
1222 // This method will call its helper method to reduce SafePoint nodes that use
1223 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1224 // "version" of Phi use the same debug information (regarding the Phi).
1225 // Therefore, I collect all safepoints and patch them all at once.
1226 //
1227 // The safepoints using the Phi node have to be processed before safepoints of
1228 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1229 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1230 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1231 // algorithm that process Phi's safepoints will think that the added Phi
1232 // reference is a regular reference.
1233 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1234 PhiNode* selector = create_selector(ophi);
1235 Unique_Node_List safepoints;
1236 Unique_Node_List casts;
1237
1238 // Just collect the users of the Phis for later processing
1239 // in the needed order.
1240 for (uint i = 0; i < ophi->outcnt(); i++) {
1241 Node* use = ophi->raw_out(i);
1242 if (use->is_SafePoint()) {
1243 safepoints.push(use);
1244 } else if (use->is_CastPP()) {
1245 casts.push(use);
1246 } else {
1247 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1248 }
1249 }
1250
1251 // Need to process safepoints using the Phi first
1252 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1253 return false;
1254 }
1255
1256 // Now process CastPP->safepoints
1257 for (uint i = 0; i < casts.size(); i++) {
1258 Node* cast = casts.at(i);
1259 Unique_Node_List cast_sfpts;
1260
1261 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1262 Node* use_use = cast->fast_out(j);
1263 if (use_use->is_SafePoint()) {
1264 cast_sfpts.push(use_use);
1265 } else {
1266 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1267 }
1268 }
1269
1270 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1271 return false;
1272 }
1273 }
1274
1275 return true;
1276 }
1277
1278 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1279 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1280 // SafePointScalarObjectNode for each scalar replaceable input. Each
1281 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1282 // check detailed description in SafePointScalarMergeNode class header.
1283 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1284 PhaseMacroExpand mexp(*_igvn);
1285 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1286 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1287
1288 Node* nsr_merge_pointer = ophi;
1289 if (cast != nullptr) {
1290 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1291 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::DependencyType::FloatingNarrowing, nullptr));
1292 }
1293
1294 for (uint spi = 0; spi < safepoints.size(); spi++) {
1295 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1296 JVMState *jvms = sfpt->jvms();
1297 uint merge_idx = (sfpt->req() - jvms->scloff());
1298 int debug_start = jvms->debug_start();
1299
1300 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1301 smerge->init_req(0, _compile->root());
1302 _igvn->register_new_node_with_optimizer(smerge);
1303
1304 // The next two inputs are:
1305 // (1) A copy of the original pointer to NSR objects.
1306 // (2) A selector, used to decide if we need to rematerialize an object
1307 // or use the pointer to a NSR object.
1308 // See more details of these fields in the declaration of SafePointScalarMergeNode
1309 sfpt->add_req(nsr_merge_pointer);
1310 sfpt->add_req(selector);
1311
1312 for (uint i = 1; i < ophi->req(); i++) {
1313 Node* base = ophi->in(i);
1314 JavaObjectNode* ptn = unique_java_object(base);
1315
1316 // If the base is not scalar replaceable we don't need to register information about
1317 // it at this time.
1318 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1319 continue;
1320 }
1321
1322 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1323 Unique_Node_List value_worklist;
1324 #ifdef ASSERT
1325 const Type* res_type = alloc->result_cast()->bottom_type();
1326 if (res_type->is_inlinetypeptr() && !Compile::current()->has_circular_inline_type()) {
1327 PhiNode* phi = ophi->as_Phi();
1328 assert(!ophi->as_Phi()->can_push_inline_types_down(_igvn), "missed earlier scalarization opportunity");
1329 }
1330 #endif
1331 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt, &value_worklist);
1332 if (sobj == nullptr) {
1333 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1334 return false;
1335 }
1336
1337 // Now make a pass over the debug information replacing any references
1338 // to the allocated object with "sobj"
1339 Node* ccpp = alloc->result_cast();
1340 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1341
1342 // Register the scalarized object as a candidate for reallocation
1343 smerge->add_req(sobj);
1344
1345 // Scalarize inline types that were added to the safepoint.
1346 // Don't allow linking a constant oop (if available) for flat array elements
1347 // because Deoptimization::reassign_flat_array_elements needs field values.
1348 const bool allow_oop = !merge_t->is_flat();
1349 for (uint j = 0; j < value_worklist.size(); ++j) {
1350 InlineTypeNode* vt = value_worklist.at(j)->as_InlineType();
1351 vt->make_scalar_in_safepoints(_igvn, allow_oop);
1352 }
1353 }
1354
1355 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1356 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1357
1358 // The call to 'replace_edges_in_range' above might have removed the
1359 // reference to ophi that we need at _merge_pointer_idx. The line below make
1360 // sure the reference is maintained.
1361 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1362 _igvn->_worklist.push(sfpt);
1363 }
1364
1365 return true;
1366 }
1367
1368 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1369 bool delay = _igvn->delay_transform();
1370 _igvn->set_delay_transform(true);
1371 _igvn->hash_delete(ophi);
1372
1373 // Copying all users first because some will be removed and others won't.
1374 // Ophi also may acquire some new users as part of Cast reduction.
1375 // CastPPs also need to be processed before CmpPs.
1376 Unique_Node_List castpps;
1377 Unique_Node_List others;
1378 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1379 Node* use = ophi->fast_out(i);
1380
1381 if (use->is_CastPP()) {
1382 castpps.push(use);
1383 } else if (use->is_AddP() || use->is_Cmp()) {
1384 others.push(use);
1385 } else {
1386 // Safepoints to be processed later; other users aren't expected here
1387 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1388 }
1389 }
1390
1391 _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1392
1393 // CastPPs need to be processed before Cmps because during the process of
1394 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1395 // by the If controlling the CastPP.
1396 for (uint i = 0; i < castpps.size(); i++) {
1397 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1398 _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1399 }
1400
1401 for (uint i = 0; i < others.size(); i++) {
1402 Node* use = others.at(i);
1403
1404 if (use->is_AddP()) {
1405 reduce_phi_on_field_access(use, alloc_worklist);
1406 _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1407 } else if(use->is_Cmp()) {
1408 reduce_phi_on_cmp(use);
1409 _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1410 }
1411 }
1412
1413 _igvn->set_delay_transform(delay);
1414 }
1415
1416 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1417 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1418 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1419 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1420 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1421
1422 for (uint i = 1; i < ophi->req(); i++) {
1423 Node* base = ophi->in(i);
1424 JavaObjectNode* ptn = unique_java_object(base);
1425
1426 if (ptn != nullptr && ptn->scalar_replaceable()) {
1427 new_phi->set_req(i, null_ptr);
1428 } else {
1429 new_phi->set_req(i, ophi->in(i));
1430 }
1431 }
1432
1433 for (int i = ophi->outcnt()-1; i >= 0;) {
1434 Node* out = ophi->raw_out(i);
1435
1436 if (out->is_ConstraintCast()) {
1437 const Type* out_t = _igvn->type(out)->make_ptr();
1438 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1439 bool change = out_new_t != out_t;
1440
1441 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1442 Node* out2 = out->raw_out(j);
1443 if (!out2->is_SafePoint()) {
1444 change = false;
1445 break;
1446 }
1447 }
1448
1449 if (change) {
1450 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::DependencyType::NonFloatingNarrowing, nullptr);
1451 _igvn->replace_node(out, new_cast);
1452 _igvn->register_new_node_with_optimizer(new_cast);
1453 }
1454 }
1455
1456 --i;
1457 i = MIN2(i, (int)ophi->outcnt()-1);
1458 }
1459
1460 _igvn->replace_node(ophi, new_phi);
1461 }
1462
1463 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1464 if (!C->do_reduce_allocation_merges()) return;
1465
1466 Unique_Node_List ideal_nodes;
1467 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1468 ideal_nodes.push(root);
1469
1470 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1471 Node* n = ideal_nodes.at(next);
1472
1473 if (n->is_SafePointScalarMerge()) {
1474 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1475
1476 // Validate inputs of merge
1477 for (uint i = 1; i < merge->req(); i++) {
1478 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1479 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1480 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1481 }
1482 }
1483
1484 // Validate users of merge
1485 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1486 Node* sfpt = merge->fast_out(i);
1487 if (sfpt->is_SafePoint()) {
1488 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1489
1490 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1491 assert(false, "SafePointScalarMerge nodes can't be nested.");
1492 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1493 }
1494 } else {
1495 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1496 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1497 }
1498 }
1499 }
1500
1501 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1502 Node* m = n->fast_out(i);
1503 ideal_nodes.push(m);
1504 }
1505 }
1506 }
1507
1508 // Returns true if there is an object in the scope of sfn that does not escape globally.
1509 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1510 Compile* C = _compile;
1511 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1512 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1513 DeoptimizeObjectsALot) {
1514 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1515 int num_locs = jvms->loc_size();
1516 for (int idx = 0; idx < num_locs; idx++) {
1517 Node* l = sfn->local(jvms, idx);
1518 if (not_global_escape(l)) {
1519 return true;
1520 }
1521 }
1522 }
1523 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1524 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1525 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1526 int num_mon = jvms->nof_monitors();
1527 for (int idx = 0; idx < num_mon; idx++) {
1528 Node* m = sfn->monitor_obj(jvms, idx);
1529 if (m != nullptr && not_global_escape(m)) {
1530 return true;
1531 }
1532 }
1533 }
1534 }
1535 return false;
1536 }
1537
1538 // Returns true if at least one of the arguments to the call is an object
1539 // that does not escape globally.
1540 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1541 if (call->method() != nullptr) {
1542 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1543 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1544 Node* p = call->in(idx);
1545 if (not_global_escape(p)) {
1546 return true;
1547 }
1548 }
1549 } else {
1550 const char* name = call->as_CallStaticJava()->_name;
1551 assert(name != nullptr, "no name");
1552 // no arg escapes through uncommon traps
1553 if (strcmp(name, "uncommon_trap") != 0) {
1554 // process_call_arguments() assumes that all arguments escape globally
1555 const TypeTuple* d = call->tf()->domain_sig();
1556 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1557 const Type* at = d->field_at(i);
1558 if (at->isa_oopptr() != nullptr) {
1559 return true;
1560 }
1561 }
1562 }
1563 }
1564 return false;
1565 }
1566
1567
1568
1569 // Utility function for nodes that load an object
1570 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1571 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1572 // ThreadLocal has RawPtr type.
1573 const Type* t = _igvn->type(n);
1574 if (t->make_ptr() != nullptr) {
1575 Node* adr = n->in(MemNode::Address);
1576 #ifdef ASSERT
1577 if (!adr->is_AddP()) {
1578 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1579 } else {
1580 assert((ptnode_adr(adr->_idx) == nullptr ||
1581 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1582 }
1583 #endif
1584 add_local_var_and_edge(n, PointsToNode::NoEscape,
1585 adr, delayed_worklist);
1586 }
1587 }
1588
1589 // Populate Connection Graph with PointsTo nodes and create simple
1590 // connection graph edges.
1591 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1592 assert(!_verify, "this method should not be called for verification");
1593 PhaseGVN* igvn = _igvn;
1594 uint n_idx = n->_idx;
1595 PointsToNode* n_ptn = ptnode_adr(n_idx);
1596 if (n_ptn != nullptr) {
1597 return; // No need to redefine PointsTo node during first iteration.
1598 }
1599 int opcode = n->Opcode();
1600 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1601 if (gc_handled) {
1602 return; // Ignore node if already handled by GC.
1603 }
1604
1605 if (n->is_Call()) {
1606 // Arguments to allocation and locking don't escape.
1607 if (n->is_AbstractLock()) {
1608 // Put Lock and Unlock nodes on IGVN worklist to process them during
1609 // first IGVN optimization when escape information is still available.
1610 record_for_optimizer(n);
1611 } else if (n->is_Allocate()) {
1612 add_call_node(n->as_Call());
1613 record_for_optimizer(n);
1614 } else {
1615 if (n->is_CallStaticJava()) {
1616 const char* name = n->as_CallStaticJava()->_name;
1617 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1618 return; // Skip uncommon traps
1619 }
1620 }
1621 // Don't mark as processed since call's arguments have to be processed.
1622 delayed_worklist->push(n);
1623 // Check if a call returns an object.
1624 if ((n->as_Call()->returns_pointer() &&
1625 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1626 (n->is_CallStaticJava() &&
1627 n->as_CallStaticJava()->is_boxing_method())) {
1628 add_call_node(n->as_Call());
1629 } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1630 bool returns_oop = false;
1631 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1632 ProjNode* pn = n->fast_out(i)->as_Proj();
1633 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1634 returns_oop = true;
1635 }
1636 }
1637 if (returns_oop) {
1638 add_call_node(n->as_Call());
1639 }
1640 }
1641 }
1642 return;
1643 }
1644 // Put this check here to process call arguments since some call nodes
1645 // point to phantom_obj.
1646 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1647 return; // Skip predefined nodes.
1648 }
1649 switch (opcode) {
1650 case Op_AddP: {
1651 Node* base = get_addp_base(n);
1652 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1653 // Field nodes are created for all field types. They are used in
1654 // adjust_scalar_replaceable_state() and split_unique_types().
1655 // Note, non-oop fields will have only base edges in Connection
1656 // Graph because such fields are not used for oop loads and stores.
1657 int offset = address_offset(n, igvn);
1658 add_field(n, PointsToNode::NoEscape, offset);
1659 if (ptn_base == nullptr) {
1660 delayed_worklist->push(n); // Process it later.
1661 } else {
1662 n_ptn = ptnode_adr(n_idx);
1663 add_base(n_ptn->as_Field(), ptn_base);
1664 }
1665 break;
1666 }
1667 case Op_CastX2P:
1668 case Op_CastI2N: {
1669 map_ideal_node(n, phantom_obj);
1670 break;
1671 }
1672 case Op_InlineType:
1673 case Op_CastPP:
1674 case Op_CheckCastPP:
1675 case Op_EncodeP:
1676 case Op_DecodeN:
1677 case Op_EncodePKlass:
1678 case Op_DecodeNKlass: {
1679 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1680 break;
1681 }
1682 case Op_CMoveP: {
1683 add_local_var(n, PointsToNode::NoEscape);
1684 // Do not add edges during first iteration because some could be
1685 // not defined yet.
1686 delayed_worklist->push(n);
1687 break;
1688 }
1689 case Op_ConP:
1690 case Op_ConN:
1691 case Op_ConNKlass: {
1692 // assume all oop constants globally escape except for null
1693 PointsToNode::EscapeState es;
1694 const Type* t = igvn->type(n);
1695 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1696 es = PointsToNode::NoEscape;
1697 } else {
1698 es = PointsToNode::GlobalEscape;
1699 }
1700 PointsToNode* ptn_con = add_java_object(n, es);
1701 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1702 break;
1703 }
1704 case Op_CreateEx: {
1705 // assume that all exception objects globally escape
1706 map_ideal_node(n, phantom_obj);
1707 break;
1708 }
1709 case Op_LoadKlass:
1710 case Op_LoadNKlass: {
1711 // Unknown class is loaded
1712 map_ideal_node(n, phantom_obj);
1713 break;
1714 }
1715 case Op_LoadP:
1716 case Op_LoadN: {
1717 add_objload_to_connection_graph(n, delayed_worklist);
1718 break;
1719 }
1720 case Op_Parm: {
1721 map_ideal_node(n, phantom_obj);
1722 break;
1723 }
1724 case Op_PartialSubtypeCheck: {
1725 // Produces Null or notNull and is used in only in CmpP so
1726 // phantom_obj could be used.
1727 map_ideal_node(n, phantom_obj); // Result is unknown
1728 break;
1729 }
1730 case Op_Phi: {
1731 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1732 // ThreadLocal has RawPtr type.
1733 const Type* t = n->as_Phi()->type();
1734 if (t->make_ptr() != nullptr) {
1735 add_local_var(n, PointsToNode::NoEscape);
1736 // Do not add edges during first iteration because some could be
1737 // not defined yet.
1738 delayed_worklist->push(n);
1739 }
1740 break;
1741 }
1742 case Op_LoadFlat:
1743 // Treat LoadFlat similar to an unknown call that receives nothing and produces its results
1744 map_ideal_node(n, phantom_obj);
1745 break;
1746 case Op_StoreFlat:
1747 // Treat StoreFlat similar to a call that escapes the stored flattened fields
1748 delayed_worklist->push(n);
1749 break;
1750 case Op_Proj: {
1751 // we are only interested in the oop result projection from a call
1752 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
1753 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) {
1754 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1755 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1756 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1757 } else if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_LoadFlat() && igvn->type(n)->isa_ptr()) {
1758 // Treat LoadFlat outputs similar to a call return value
1759 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1760 }
1761 break;
1762 }
1763 case Op_Rethrow: // Exception object escapes
1764 case Op_Return: {
1765 if (n->req() > TypeFunc::Parms &&
1766 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1767 // Treat Return value as LocalVar with GlobalEscape escape state.
1768 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1769 }
1770 break;
1771 }
1772 case Op_CompareAndExchangeP:
1773 case Op_CompareAndExchangeN:
1774 case Op_GetAndSetP:
1775 case Op_GetAndSetN: {
1776 add_objload_to_connection_graph(n, delayed_worklist);
1777 // fall-through
1778 }
1779 case Op_StoreP:
1780 case Op_StoreN:
1781 case Op_StoreNKlass:
1782 case Op_WeakCompareAndSwapP:
1783 case Op_WeakCompareAndSwapN:
1784 case Op_CompareAndSwapP:
1785 case Op_CompareAndSwapN: {
1786 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1787 break;
1788 }
1789 case Op_AryEq:
1790 case Op_CountPositives:
1791 case Op_StrComp:
1792 case Op_StrEquals:
1793 case Op_StrIndexOf:
1794 case Op_StrIndexOfChar:
1795 case Op_StrInflatedCopy:
1796 case Op_StrCompressedCopy:
1797 case Op_VectorizedHashCode:
1798 case Op_EncodeISOArray: {
1799 add_local_var(n, PointsToNode::ArgEscape);
1800 delayed_worklist->push(n); // Process it later.
1801 break;
1802 }
1803 case Op_ThreadLocal: {
1804 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1805 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1806 break;
1807 }
1808 case Op_Blackhole: {
1809 // All blackhole pointer arguments are globally escaping.
1810 // Only do this if there is at least one pointer argument.
1811 // Do not add edges during first iteration because some could be
1812 // not defined yet, defer to final step.
1813 for (uint i = 0; i < n->req(); i++) {
1814 Node* in = n->in(i);
1815 if (in != nullptr) {
1816 const Type* at = _igvn->type(in);
1817 if (!at->isa_ptr()) continue;
1818
1819 add_local_var(n, PointsToNode::GlobalEscape);
1820 delayed_worklist->push(n);
1821 break;
1822 }
1823 }
1824 break;
1825 }
1826 default:
1827 ; // Do nothing for nodes not related to EA.
1828 }
1829 return;
1830 }
1831
1832 // Add final simple edges to graph.
1833 void ConnectionGraph::add_final_edges(Node *n) {
1834 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1835 #ifdef ASSERT
1836 if (_verify && n_ptn->is_JavaObject())
1837 return; // This method does not change graph for JavaObject.
1838 #endif
1839
1840 if (n->is_Call()) {
1841 process_call_arguments(n->as_Call());
1842 return;
1843 }
1844 assert(n->is_Store() || n->is_LoadStore() || n->is_StoreFlat() ||
1845 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1846 "node should be registered already");
1847 int opcode = n->Opcode();
1848 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1849 if (gc_handled) {
1850 return; // Ignore node if already handled by GC.
1851 }
1852 switch (opcode) {
1853 case Op_AddP: {
1854 Node* base = get_addp_base(n);
1855 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1856 assert(ptn_base != nullptr, "field's base should be registered");
1857 add_base(n_ptn->as_Field(), ptn_base);
1858 break;
1859 }
1860 case Op_InlineType:
1861 case Op_CastPP:
1862 case Op_CheckCastPP:
1863 case Op_EncodeP:
1864 case Op_DecodeN:
1865 case Op_EncodePKlass:
1866 case Op_DecodeNKlass: {
1867 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1868 break;
1869 }
1870 case Op_CMoveP: {
1871 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1872 Node* in = n->in(i);
1873 if (in == nullptr) {
1874 continue; // ignore null
1875 }
1876 Node* uncast_in = in->uncast();
1877 if (uncast_in->is_top() || uncast_in == n) {
1878 continue; // ignore top or inputs which go back this node
1879 }
1880 PointsToNode* ptn = ptnode_adr(in->_idx);
1881 assert(ptn != nullptr, "node should be registered");
1882 add_edge(n_ptn, ptn);
1883 }
1884 break;
1885 }
1886 case Op_LoadP:
1887 case Op_LoadN: {
1888 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1889 // ThreadLocal has RawPtr type.
1890 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1891 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1892 break;
1893 }
1894 case Op_Phi: {
1895 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1896 // ThreadLocal has RawPtr type.
1897 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1898 for (uint i = 1; i < n->req(); i++) {
1899 Node* in = n->in(i);
1900 if (in == nullptr) {
1901 continue; // ignore null
1902 }
1903 Node* uncast_in = in->uncast();
1904 if (uncast_in->is_top() || uncast_in == n) {
1905 continue; // ignore top or inputs which go back this node
1906 }
1907 PointsToNode* ptn = ptnode_adr(in->_idx);
1908 assert(ptn != nullptr, "node should be registered");
1909 add_edge(n_ptn, ptn);
1910 }
1911 break;
1912 }
1913 case Op_StoreFlat: {
1914 // StoreFlat globally escapes its stored flattened fields
1915 InlineTypeNode* value = n->as_StoreFlat()->value();
1916 ciInlineKlass* vk = _igvn->type(value)->inline_klass();
1917 for (int i = 0; i < vk->nof_nonstatic_fields(); i++) {
1918 ciField* field = vk->nonstatic_field_at(i);
1919 if (field->type()->is_primitive_type()) {
1920 continue;
1921 }
1922
1923 Node* field_value = value->field_value_by_offset(field->offset_in_bytes(), true);
1924 PointsToNode* field_value_ptn = ptnode_adr(field_value->_idx);
1925 set_escape_state(field_value_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "store into a flat field"));
1926 }
1927 break;
1928 }
1929 case Op_Proj: {
1930 if (n->in(0)->is_Call()) {
1931 // we are only interested in the oop result projection from a call
1932 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1933 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1934 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1935 } else if (n->in(0)->is_LoadFlat()) {
1936 // Treat LoadFlat outputs similar to a call return value
1937 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1938 }
1939 break;
1940 }
1941 case Op_Rethrow: // Exception object escapes
1942 case Op_Return: {
1943 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1944 "Unexpected node type");
1945 // Treat Return value as LocalVar with GlobalEscape escape state.
1946 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1947 break;
1948 }
1949 case Op_CompareAndExchangeP:
1950 case Op_CompareAndExchangeN:
1951 case Op_GetAndSetP:
1952 case Op_GetAndSetN:{
1953 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1954 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1955 // fall-through
1956 }
1957 case Op_CompareAndSwapP:
1958 case Op_CompareAndSwapN:
1959 case Op_WeakCompareAndSwapP:
1960 case Op_WeakCompareAndSwapN:
1961 case Op_StoreP:
1962 case Op_StoreN:
1963 case Op_StoreNKlass:{
1964 add_final_edges_unsafe_access(n, opcode);
1965 break;
1966 }
1967 case Op_VectorizedHashCode:
1968 case Op_AryEq:
1969 case Op_CountPositives:
1970 case Op_StrComp:
1971 case Op_StrEquals:
1972 case Op_StrIndexOf:
1973 case Op_StrIndexOfChar:
1974 case Op_StrInflatedCopy:
1975 case Op_StrCompressedCopy:
1976 case Op_EncodeISOArray: {
1977 // char[]/byte[] arrays passed to string intrinsic do not escape but
1978 // they are not scalar replaceable. Adjust escape state for them.
1979 // Start from in(2) edge since in(1) is memory edge.
1980 for (uint i = 2; i < n->req(); i++) {
1981 Node* adr = n->in(i);
1982 const Type* at = _igvn->type(adr);
1983 if (!adr->is_top() && at->isa_ptr()) {
1984 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1985 at->isa_ptr() != nullptr, "expecting a pointer");
1986 if (adr->is_AddP()) {
1987 adr = get_addp_base(adr);
1988 }
1989 PointsToNode* ptn = ptnode_adr(adr->_idx);
1990 assert(ptn != nullptr, "node should be registered");
1991 add_edge(n_ptn, ptn);
1992 }
1993 }
1994 break;
1995 }
1996 case Op_Blackhole: {
1997 // All blackhole pointer arguments are globally escaping.
1998 for (uint i = 0; i < n->req(); i++) {
1999 Node* in = n->in(i);
2000 if (in != nullptr) {
2001 const Type* at = _igvn->type(in);
2002 if (!at->isa_ptr()) continue;
2003
2004 if (in->is_AddP()) {
2005 in = get_addp_base(in);
2006 }
2007
2008 PointsToNode* ptn = ptnode_adr(in->_idx);
2009 assert(ptn != nullptr, "should be defined already");
2010 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
2011 add_edge(n_ptn, ptn);
2012 }
2013 }
2014 break;
2015 }
2016 default: {
2017 // This method should be called only for EA specific nodes which may
2018 // miss some edges when they were created.
2019 #ifdef ASSERT
2020 n->dump(1);
2021 #endif
2022 guarantee(false, "unknown node");
2023 }
2024 }
2025 return;
2026 }
2027
2028 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
2029 Node* adr = n->in(MemNode::Address);
2030 const Type* adr_type = _igvn->type(adr);
2031 adr_type = adr_type->make_ptr();
2032 if (adr_type == nullptr) {
2033 return; // skip dead nodes
2034 }
2035 if (adr_type->isa_oopptr()
2036 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2037 && adr_type == TypeRawPtr::NOTNULL
2038 && is_captured_store_address(adr))) {
2039 delayed_worklist->push(n); // Process it later.
2040 #ifdef ASSERT
2041 assert (adr->is_AddP(), "expecting an AddP");
2042 if (adr_type == TypeRawPtr::NOTNULL) {
2043 // Verify a raw address for a store captured by Initialize node.
2044 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2045 assert(offs != Type::OffsetBot, "offset must be a constant");
2046 }
2047 #endif
2048 } else {
2049 // Ignore copy the displaced header to the BoxNode (OSR compilation).
2050 if (adr->is_BoxLock()) {
2051 return;
2052 }
2053 // Stored value escapes in unsafe access.
2054 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2055 delayed_worklist->push(n); // Process unsafe access later.
2056 return;
2057 }
2058 #ifdef ASSERT
2059 n->dump(1);
2060 assert(false, "not unsafe");
2061 #endif
2062 }
2063 }
2064
2065 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
2066 Node* adr = n->in(MemNode::Address);
2067 const Type *adr_type = _igvn->type(adr);
2068 adr_type = adr_type->make_ptr();
2069 #ifdef ASSERT
2070 if (adr_type == nullptr) {
2071 n->dump(1);
2072 assert(adr_type != nullptr, "dead node should not be on list");
2073 return true;
2074 }
2075 #endif
2076
2077 if (adr_type->isa_oopptr()
2078 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2079 && adr_type == TypeRawPtr::NOTNULL
2080 && is_captured_store_address(adr))) {
2081 // Point Address to Value
2082 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2083 assert(adr_ptn != nullptr &&
2084 adr_ptn->as_Field()->is_oop(), "node should be registered");
2085 Node* val = n->in(MemNode::ValueIn);
2086 PointsToNode* ptn = ptnode_adr(val->_idx);
2087 assert(ptn != nullptr, "node should be registered");
2088 add_edge(adr_ptn, ptn);
2089 return true;
2090 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2091 // Stored value escapes in unsafe access.
2092 Node* val = n->in(MemNode::ValueIn);
2093 PointsToNode* ptn = ptnode_adr(val->_idx);
2094 assert(ptn != nullptr, "node should be registered");
2095 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2096 // Add edge to object for unsafe access with offset.
2097 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2098 assert(adr_ptn != nullptr, "node should be registered");
2099 if (adr_ptn->is_Field()) {
2100 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2101 add_edge(adr_ptn, ptn);
2102 }
2103 return true;
2104 }
2105 #ifdef ASSERT
2106 n->dump(1);
2107 assert(false, "not unsafe");
2108 #endif
2109 return false;
2110 }
2111
2112 void ConnectionGraph::add_call_node(CallNode* call) {
2113 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2114 uint call_idx = call->_idx;
2115 if (call->is_Allocate()) {
2116 Node* k = call->in(AllocateNode::KlassNode);
2117 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2118 assert(kt != nullptr, "TypeKlassPtr required.");
2119 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2120 bool scalar_replaceable = true;
2121 NOT_PRODUCT(const char* nsr_reason = "");
2122 if (call->is_AllocateArray()) {
2123 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2124 es = PointsToNode::GlobalEscape;
2125 } else {
2126 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2127 if (length < 0) {
2128 // Not scalar replaceable if the length is not constant.
2129 scalar_replaceable = false;
2130 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2131 } else if (length > EliminateAllocationArraySizeLimit) {
2132 // Not scalar replaceable if the length is too big.
2133 scalar_replaceable = false;
2134 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2135 }
2136 }
2137 } else { // Allocate instance
2138 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2139 es = PointsToNode::GlobalEscape;
2140 } else {
2141 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2142 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2143 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2144 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2145 !ik->can_be_instantiated() ||
2146 ik->has_finalizer()) {
2147 es = PointsToNode::GlobalEscape;
2148 } else {
2149 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2150 if (nfields > EliminateAllocationFieldsLimit) {
2151 // Not scalar replaceable if there are too many fields.
2152 scalar_replaceable = false;
2153 NOT_PRODUCT(nsr_reason = "has too many fields");
2154 }
2155 }
2156 }
2157 }
2158 add_java_object(call, es);
2159 PointsToNode* ptn = ptnode_adr(call_idx);
2160 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2161 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2162 }
2163 } else if (call->is_CallStaticJava()) {
2164 // Call nodes could be different types:
2165 //
2166 // 1. CallDynamicJavaNode (what happened during call is unknown):
2167 //
2168 // - mapped to GlobalEscape JavaObject node if oop is returned;
2169 //
2170 // - all oop arguments are escaping globally;
2171 //
2172 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2173 //
2174 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2175 //
2176 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2177 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2178 // during call is returned;
2179 // - mapped to ArgEscape LocalVar node pointed to object arguments
2180 // which are returned and does not escape during call;
2181 //
2182 // - oop arguments escaping status is defined by bytecode analysis;
2183 //
2184 // For a static call, we know exactly what method is being called.
2185 // Use bytecode estimator to record whether the call's return value escapes.
2186 ciMethod* meth = call->as_CallJava()->method();
2187 if (meth == nullptr) {
2188 const char* name = call->as_CallStaticJava()->_name;
2189 assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
2190 strncmp(name, "load_unknown_inline", 19) == 0 ||
2191 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "TODO: add failed case check");
2192 // Returns a newly allocated non-escaped object.
2193 add_java_object(call, PointsToNode::NoEscape);
2194 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2195 } else if (meth->is_boxing_method()) {
2196 // Returns boxing object
2197 PointsToNode::EscapeState es;
2198 vmIntrinsics::ID intr = meth->intrinsic_id();
2199 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2200 // It does not escape if object is always allocated.
2201 es = PointsToNode::NoEscape;
2202 } else {
2203 // It escapes globally if object could be loaded from cache.
2204 es = PointsToNode::GlobalEscape;
2205 }
2206 add_java_object(call, es);
2207 if (es == PointsToNode::GlobalEscape) {
2208 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2209 }
2210 } else {
2211 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2212 call_analyzer->copy_dependencies(_compile->dependencies());
2213 if (call_analyzer->is_return_allocated()) {
2214 // Returns a newly allocated non-escaped object, simply
2215 // update dependency information.
2216 // Mark it as NoEscape so that objects referenced by
2217 // it's fields will be marked as NoEscape at least.
2218 add_java_object(call, PointsToNode::NoEscape);
2219 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2220 } else {
2221 // Determine whether any arguments are returned.
2222 const TypeTuple* d = call->tf()->domain_cc();
2223 bool ret_arg = false;
2224 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2225 if (d->field_at(i)->isa_ptr() != nullptr &&
2226 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2227 ret_arg = true;
2228 break;
2229 }
2230 }
2231 if (ret_arg) {
2232 add_local_var(call, PointsToNode::ArgEscape);
2233 } else {
2234 // Returns unknown object.
2235 map_ideal_node(call, phantom_obj);
2236 }
2237 }
2238 }
2239 } else {
2240 // An other type of call, assume the worst case:
2241 // returned value is unknown and globally escapes.
2242 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2243 map_ideal_node(call, phantom_obj);
2244 }
2245 }
2246
2247 void ConnectionGraph::process_call_arguments(CallNode *call) {
2248 bool is_arraycopy = false;
2249 switch (call->Opcode()) {
2250 #ifdef ASSERT
2251 case Op_Allocate:
2252 case Op_AllocateArray:
2253 case Op_Lock:
2254 case Op_Unlock:
2255 assert(false, "should be done already");
2256 break;
2257 #endif
2258 case Op_ArrayCopy:
2259 case Op_CallLeafNoFP:
2260 // Most array copies are ArrayCopy nodes at this point but there
2261 // are still a few direct calls to the copy subroutines (See
2262 // PhaseStringOpts::copy_string())
2263 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2264 call->as_CallLeaf()->is_call_to_arraycopystub();
2265 // fall through
2266 case Op_CallLeafVector:
2267 case Op_CallLeaf: {
2268 // Stub calls, objects do not escape but they are not scale replaceable.
2269 // Adjust escape state for outgoing arguments.
2270 const TypeTuple * d = call->tf()->domain_sig();
2271 bool src_has_oops = false;
2272 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2273 const Type* at = d->field_at(i);
2274 Node *arg = call->in(i);
2275 if (arg == nullptr) {
2276 continue;
2277 }
2278 const Type *aat = _igvn->type(arg);
2279 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2280 continue;
2281 }
2282 if (arg->is_AddP()) {
2283 //
2284 // The inline_native_clone() case when the arraycopy stub is called
2285 // after the allocation before Initialize and CheckCastPP nodes.
2286 // Or normal arraycopy for object arrays case.
2287 //
2288 // Set AddP's base (Allocate) as not scalar replaceable since
2289 // pointer to the base (with offset) is passed as argument.
2290 //
2291 arg = get_addp_base(arg);
2292 }
2293 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2294 assert(arg_ptn != nullptr, "should be registered");
2295 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2296 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2297 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2298 aat->isa_ptr() != nullptr, "expecting an Ptr");
2299 bool arg_has_oops = aat->isa_oopptr() &&
2300 (aat->isa_instptr() ||
2301 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2302 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2303 aat->isa_aryptr()->is_flat() &&
2304 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2305 if (i == TypeFunc::Parms) {
2306 src_has_oops = arg_has_oops;
2307 }
2308 //
2309 // src or dst could be j.l.Object when other is basic type array:
2310 //
2311 // arraycopy(char[],0,Object*,0,size);
2312 // arraycopy(Object*,0,char[],0,size);
2313 //
2314 // Don't add edges in such cases.
2315 //
2316 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2317 arg_has_oops && (i > TypeFunc::Parms);
2318 #ifdef ASSERT
2319 if (!(is_arraycopy ||
2320 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2321 (call->as_CallLeaf()->_name != nullptr &&
2322 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2323 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2324 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2325 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2326 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2327 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2328 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2329 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2330 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2331 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2332 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2333 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2334 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2335 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2336 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2337 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2338 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2339 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2340 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2341 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2342 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2343 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2344 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2345 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2346 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2347 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2348 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2349 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2350 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2351 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2352 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2353 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2354 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2355 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2356 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2357 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2358 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2359 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2360 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2361 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2362 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2363 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2364 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2365 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2366 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2367 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2368 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2369 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2370 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2371 strcmp(call->as_CallLeaf()->_name, "store_inline_type_fields_to_buf") == 0 ||
2372 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2373 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2374 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2375 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2376 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2377 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2378 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2379 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2380 ))) {
2381 call->dump();
2382 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2383 }
2384 #endif
2385 // Always process arraycopy's destination object since
2386 // we need to add all possible edges to references in
2387 // source object.
2388 if (arg_esc >= PointsToNode::ArgEscape &&
2389 !arg_is_arraycopy_dest) {
2390 continue;
2391 }
2392 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2393 if (call->is_ArrayCopy()) {
2394 ArrayCopyNode* ac = call->as_ArrayCopy();
2395 if (ac->is_clonebasic() ||
2396 ac->is_arraycopy_validated() ||
2397 ac->is_copyof_validated() ||
2398 ac->is_copyofrange_validated()) {
2399 es = PointsToNode::NoEscape;
2400 }
2401 }
2402 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2403 if (arg_is_arraycopy_dest) {
2404 Node* src = call->in(TypeFunc::Parms);
2405 if (src->is_AddP()) {
2406 src = get_addp_base(src);
2407 }
2408 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2409 assert(src_ptn != nullptr, "should be registered");
2410 // Special arraycopy edge:
2411 // Only escape state of destination object's fields affects
2412 // escape state of fields in source object.
2413 add_arraycopy(call, es, src_ptn, arg_ptn);
2414 }
2415 }
2416 }
2417 break;
2418 }
2419 case Op_CallStaticJava: {
2420 // For a static call, we know exactly what method is being called.
2421 // Use bytecode estimator to record the call's escape affects
2422 #ifdef ASSERT
2423 const char* name = call->as_CallStaticJava()->_name;
2424 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2425 #endif
2426 ciMethod* meth = call->as_CallJava()->method();
2427 if ((meth != nullptr) && meth->is_boxing_method()) {
2428 break; // Boxing methods do not modify any oops.
2429 }
2430 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2431 // fall-through if not a Java method or no analyzer information
2432 if (call_analyzer != nullptr) {
2433 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2434 const TypeTuple* d = call->tf()->domain_cc();
2435 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2436 const Type* at = d->field_at(i);
2437 int k = i - TypeFunc::Parms;
2438 Node* arg = call->in(i);
2439 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2440 if (at->isa_ptr() != nullptr &&
2441 call_analyzer->is_arg_returned(k)) {
2442 // The call returns arguments.
2443 if (call_ptn != nullptr) { // Is call's result used?
2444 assert(call_ptn->is_LocalVar(), "node should be registered");
2445 assert(arg_ptn != nullptr, "node should be registered");
2446 add_edge(call_ptn, arg_ptn);
2447 }
2448 }
2449 if (at->isa_oopptr() != nullptr &&
2450 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2451 if (!call_analyzer->is_arg_stack(k)) {
2452 // The argument global escapes
2453 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2454 } else {
2455 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2456 if (!call_analyzer->is_arg_local(k)) {
2457 // The argument itself doesn't escape, but any fields might
2458 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2459 }
2460 }
2461 }
2462 }
2463 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2464 // The call returns arguments.
2465 assert(call_ptn->edge_count() > 0, "sanity");
2466 if (!call_analyzer->is_return_local()) {
2467 // Returns also unknown object.
2468 add_edge(call_ptn, phantom_obj);
2469 }
2470 }
2471 break;
2472 }
2473 }
2474 default: {
2475 // Fall-through here if not a Java method or no analyzer information
2476 // or some other type of call, assume the worst case: all arguments
2477 // globally escape.
2478 const TypeTuple* d = call->tf()->domain_cc();
2479 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2480 const Type* at = d->field_at(i);
2481 if (at->isa_oopptr() != nullptr) {
2482 Node* arg = call->in(i);
2483 if (arg->is_AddP()) {
2484 arg = get_addp_base(arg);
2485 }
2486 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2487 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2488 }
2489 }
2490 }
2491 }
2492 }
2493
2494
2495 // Finish Graph construction.
2496 bool ConnectionGraph::complete_connection_graph(
2497 GrowableArray<PointsToNode*>& ptnodes_worklist,
2498 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2499 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2500 GrowableArray<FieldNode*>& oop_fields_worklist) {
2501 // Normally only 1-3 passes needed to build Connection Graph depending
2502 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2503 // Set limit to 20 to catch situation when something did go wrong and
2504 // bailout Escape Analysis.
2505 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2506 #define GRAPH_BUILD_ITER_LIMIT 20
2507
2508 // Propagate GlobalEscape and ArgEscape escape states and check that
2509 // we still have non-escaping objects. The method pushs on _worklist
2510 // Field nodes which reference phantom_object.
2511 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2512 return false; // Nothing to do.
2513 }
2514 // Now propagate references to all JavaObject nodes.
2515 int java_objects_length = java_objects_worklist.length();
2516 elapsedTimer build_time;
2517 build_time.start();
2518 elapsedTimer time;
2519 bool timeout = false;
2520 int new_edges = 1;
2521 int iterations = 0;
2522 do {
2523 while ((new_edges > 0) &&
2524 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2525 double start_time = time.seconds();
2526 time.start();
2527 new_edges = 0;
2528 // Propagate references to phantom_object for nodes pushed on _worklist
2529 // by find_non_escaped_objects() and find_field_value().
2530 new_edges += add_java_object_edges(phantom_obj, false);
2531 for (int next = 0; next < java_objects_length; ++next) {
2532 JavaObjectNode* ptn = java_objects_worklist.at(next);
2533 new_edges += add_java_object_edges(ptn, true);
2534
2535 #define SAMPLE_SIZE 4
2536 if ((next % SAMPLE_SIZE) == 0) {
2537 // Each 4 iterations calculate how much time it will take
2538 // to complete graph construction.
2539 time.stop();
2540 // Poll for requests from shutdown mechanism to quiesce compiler
2541 // because Connection graph construction may take long time.
2542 CompileBroker::maybe_block();
2543 double stop_time = time.seconds();
2544 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2545 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2546 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2547 timeout = true;
2548 break; // Timeout
2549 }
2550 start_time = stop_time;
2551 time.start();
2552 }
2553 #undef SAMPLE_SIZE
2554
2555 }
2556 if (timeout) break;
2557 if (new_edges > 0) {
2558 // Update escape states on each iteration if graph was updated.
2559 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2560 return false; // Nothing to do.
2561 }
2562 }
2563 time.stop();
2564 if (time.seconds() >= EscapeAnalysisTimeout) {
2565 timeout = true;
2566 break;
2567 }
2568 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2569 }
2570 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2571 time.start();
2572 // Find fields which have unknown value.
2573 int fields_length = oop_fields_worklist.length();
2574 for (int next = 0; next < fields_length; next++) {
2575 FieldNode* field = oop_fields_worklist.at(next);
2576 if (field->edge_count() == 0) {
2577 new_edges += find_field_value(field);
2578 // This code may added new edges to phantom_object.
2579 // Need an other cycle to propagate references to phantom_object.
2580 }
2581 }
2582 time.stop();
2583 if (time.seconds() >= EscapeAnalysisTimeout) {
2584 timeout = true;
2585 break;
2586 }
2587 } else {
2588 new_edges = 0; // Bailout
2589 }
2590 } while (new_edges > 0);
2591
2592 build_time.stop();
2593 _build_time = build_time.seconds();
2594 _build_iterations = iterations;
2595
2596 // Bailout if passed limits.
2597 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2598 Compile* C = _compile;
2599 if (C->log() != nullptr) {
2600 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2601 C->log()->text("%s", timeout ? "time" : "iterations");
2602 C->log()->end_elem(" limit'");
2603 }
2604 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2605 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2606 // Possible infinite build_connection_graph loop,
2607 // bailout (no changes to ideal graph were made).
2608 return false;
2609 }
2610
2611 #undef GRAPH_BUILD_ITER_LIMIT
2612
2613 // Find fields initialized by null for non-escaping Allocations.
2614 int non_escaped_length = non_escaped_allocs_worklist.length();
2615 for (int next = 0; next < non_escaped_length; next++) {
2616 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2617 PointsToNode::EscapeState es = ptn->escape_state();
2618 assert(es <= PointsToNode::ArgEscape, "sanity");
2619 if (es == PointsToNode::NoEscape) {
2620 if (find_init_values_null(ptn, _igvn) > 0) {
2621 // Adding references to null object does not change escape states
2622 // since it does not escape. Also no fields are added to null object.
2623 add_java_object_edges(null_obj, false);
2624 }
2625 }
2626 Node* n = ptn->ideal_node();
2627 if (n->is_Allocate()) {
2628 // The object allocated by this Allocate node will never be
2629 // seen by an other thread. Mark it so that when it is
2630 // expanded no MemBarStoreStore is added.
2631 InitializeNode* ini = n->as_Allocate()->initialization();
2632 if (ini != nullptr)
2633 ini->set_does_not_escape();
2634 }
2635 }
2636 return true; // Finished graph construction.
2637 }
2638
2639 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2640 // and check that we still have non-escaping java objects.
2641 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2642 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2643 bool print_method) {
2644 GrowableArray<PointsToNode*> escape_worklist;
2645 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2646 int ptnodes_length = ptnodes_worklist.length();
2647 for (int next = 0; next < ptnodes_length; ++next) {
2648 PointsToNode* ptn = ptnodes_worklist.at(next);
2649 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2650 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2651 escape_worklist.push(ptn);
2652 }
2653 }
2654 // Set escape states to referenced nodes (edges list).
2655 while (escape_worklist.length() > 0) {
2656 PointsToNode* ptn = escape_worklist.pop();
2657 PointsToNode::EscapeState es = ptn->escape_state();
2658 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2659 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2660 es >= PointsToNode::ArgEscape) {
2661 // GlobalEscape or ArgEscape state of field means it has unknown value.
2662 if (add_edge(ptn, phantom_obj)) {
2663 // New edge was added
2664 add_field_uses_to_worklist(ptn->as_Field());
2665 }
2666 }
2667 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2668 PointsToNode* e = i.get();
2669 if (e->is_Arraycopy()) {
2670 assert(ptn->arraycopy_dst(), "sanity");
2671 // Propagate only fields escape state through arraycopy edge.
2672 if (e->fields_escape_state() < field_es) {
2673 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2674 escape_worklist.push(e);
2675 }
2676 } else if (es >= field_es) {
2677 // fields_escape_state is also set to 'es' if it is less than 'es'.
2678 if (e->escape_state() < es) {
2679 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2680 escape_worklist.push(e);
2681 }
2682 } else {
2683 // Propagate field escape state.
2684 bool es_changed = false;
2685 if (e->fields_escape_state() < field_es) {
2686 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2687 es_changed = true;
2688 }
2689 if ((e->escape_state() < field_es) &&
2690 e->is_Field() && ptn->is_JavaObject() &&
2691 e->as_Field()->is_oop()) {
2692 // Change escape state of referenced fields.
2693 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2694 es_changed = true;
2695 } else if (e->escape_state() < es) {
2696 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2697 es_changed = true;
2698 }
2699 if (es_changed) {
2700 escape_worklist.push(e);
2701 }
2702 }
2703 if (print_method) {
2704 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2705 }
2706 }
2707 }
2708 // Remove escaped objects from non_escaped list.
2709 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2710 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2711 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2712 non_escaped_allocs_worklist.delete_at(next);
2713 }
2714 if (ptn->escape_state() == PointsToNode::NoEscape) {
2715 // Find fields in non-escaped allocations which have unknown value.
2716 find_init_values_phantom(ptn);
2717 }
2718 }
2719 return (non_escaped_allocs_worklist.length() > 0);
2720 }
2721
2722 // Add all references to JavaObject node by walking over all uses.
2723 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2724 int new_edges = 0;
2725 if (populate_worklist) {
2726 // Populate _worklist by uses of jobj's uses.
2727 for (UseIterator i(jobj); i.has_next(); i.next()) {
2728 PointsToNode* use = i.get();
2729 if (use->is_Arraycopy()) {
2730 continue;
2731 }
2732 add_uses_to_worklist(use);
2733 if (use->is_Field() && use->as_Field()->is_oop()) {
2734 // Put on worklist all field's uses (loads) and
2735 // related field nodes (same base and offset).
2736 add_field_uses_to_worklist(use->as_Field());
2737 }
2738 }
2739 }
2740 for (int l = 0; l < _worklist.length(); l++) {
2741 PointsToNode* use = _worklist.at(l);
2742 if (PointsToNode::is_base_use(use)) {
2743 // Add reference from jobj to field and from field to jobj (field's base).
2744 use = PointsToNode::get_use_node(use)->as_Field();
2745 if (add_base(use->as_Field(), jobj)) {
2746 new_edges++;
2747 }
2748 continue;
2749 }
2750 assert(!use->is_JavaObject(), "sanity");
2751 if (use->is_Arraycopy()) {
2752 if (jobj == null_obj) { // null object does not have field edges
2753 continue;
2754 }
2755 // Added edge from Arraycopy node to arraycopy's source java object
2756 if (add_edge(use, jobj)) {
2757 jobj->set_arraycopy_src();
2758 new_edges++;
2759 }
2760 // and stop here.
2761 continue;
2762 }
2763 if (!add_edge(use, jobj)) {
2764 continue; // No new edge added, there was such edge already.
2765 }
2766 new_edges++;
2767 if (use->is_LocalVar()) {
2768 add_uses_to_worklist(use);
2769 if (use->arraycopy_dst()) {
2770 for (EdgeIterator i(use); i.has_next(); i.next()) {
2771 PointsToNode* e = i.get();
2772 if (e->is_Arraycopy()) {
2773 if (jobj == null_obj) { // null object does not have field edges
2774 continue;
2775 }
2776 // Add edge from arraycopy's destination java object to Arraycopy node.
2777 if (add_edge(jobj, e)) {
2778 new_edges++;
2779 jobj->set_arraycopy_dst();
2780 }
2781 }
2782 }
2783 }
2784 } else {
2785 // Added new edge to stored in field values.
2786 // Put on worklist all field's uses (loads) and
2787 // related field nodes (same base and offset).
2788 add_field_uses_to_worklist(use->as_Field());
2789 }
2790 }
2791 _worklist.clear();
2792 _in_worklist.reset();
2793 return new_edges;
2794 }
2795
2796 // Put on worklist all related field nodes.
2797 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2798 assert(field->is_oop(), "sanity");
2799 int offset = field->offset();
2800 add_uses_to_worklist(field);
2801 // Loop over all bases of this field and push on worklist Field nodes
2802 // with the same offset and base (since they may reference the same field).
2803 for (BaseIterator i(field); i.has_next(); i.next()) {
2804 PointsToNode* base = i.get();
2805 add_fields_to_worklist(field, base);
2806 // Check if the base was source object of arraycopy and go over arraycopy's
2807 // destination objects since values stored to a field of source object are
2808 // accessible by uses (loads) of fields of destination objects.
2809 if (base->arraycopy_src()) {
2810 for (UseIterator j(base); j.has_next(); j.next()) {
2811 PointsToNode* arycp = j.get();
2812 if (arycp->is_Arraycopy()) {
2813 for (UseIterator k(arycp); k.has_next(); k.next()) {
2814 PointsToNode* abase = k.get();
2815 if (abase->arraycopy_dst() && abase != base) {
2816 // Look for the same arraycopy reference.
2817 add_fields_to_worklist(field, abase);
2818 }
2819 }
2820 }
2821 }
2822 }
2823 }
2824 }
2825
2826 // Put on worklist all related field nodes.
2827 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2828 int offset = field->offset();
2829 if (base->is_LocalVar()) {
2830 for (UseIterator j(base); j.has_next(); j.next()) {
2831 PointsToNode* f = j.get();
2832 if (PointsToNode::is_base_use(f)) { // Field
2833 f = PointsToNode::get_use_node(f);
2834 if (f == field || !f->as_Field()->is_oop()) {
2835 continue;
2836 }
2837 int offs = f->as_Field()->offset();
2838 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2839 add_to_worklist(f);
2840 }
2841 }
2842 }
2843 } else {
2844 assert(base->is_JavaObject(), "sanity");
2845 if (// Skip phantom_object since it is only used to indicate that
2846 // this field's content globally escapes.
2847 (base != phantom_obj) &&
2848 // null object node does not have fields.
2849 (base != null_obj)) {
2850 for (EdgeIterator i(base); i.has_next(); i.next()) {
2851 PointsToNode* f = i.get();
2852 // Skip arraycopy edge since store to destination object field
2853 // does not update value in source object field.
2854 if (f->is_Arraycopy()) {
2855 assert(base->arraycopy_dst(), "sanity");
2856 continue;
2857 }
2858 if (f == field || !f->as_Field()->is_oop()) {
2859 continue;
2860 }
2861 int offs = f->as_Field()->offset();
2862 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2863 add_to_worklist(f);
2864 }
2865 }
2866 }
2867 }
2868 }
2869
2870 // Find fields which have unknown value.
2871 int ConnectionGraph::find_field_value(FieldNode* field) {
2872 // Escaped fields should have init value already.
2873 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2874 int new_edges = 0;
2875 for (BaseIterator i(field); i.has_next(); i.next()) {
2876 PointsToNode* base = i.get();
2877 if (base->is_JavaObject()) {
2878 // Skip Allocate's fields which will be processed later.
2879 if (base->ideal_node()->is_Allocate()) {
2880 return 0;
2881 }
2882 assert(base == null_obj, "only null ptr base expected here");
2883 }
2884 }
2885 if (add_edge(field, phantom_obj)) {
2886 // New edge was added
2887 new_edges++;
2888 add_field_uses_to_worklist(field);
2889 }
2890 return new_edges;
2891 }
2892
2893 // Find fields initializing values for allocations.
2894 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2895 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2896 PointsToNode* init_val = phantom_obj;
2897 Node* alloc = pta->ideal_node();
2898
2899 // Do nothing for Allocate nodes since its fields values are
2900 // "known" unless they are initialized by arraycopy/clone.
2901 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2902 if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2903 // Null-free inline type arrays are initialized with an init value instead of null
2904 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
2905 assert(init_val != nullptr, "init value should be registered");
2906 } else {
2907 return 0;
2908 }
2909 }
2910 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
2911 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
2912 #ifdef ASSERT
2913 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
2914 const char* name = alloc->as_CallStaticJava()->_name;
2915 assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
2916 strncmp(name, "load_unknown_inline", 19) == 0 ||
2917 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "sanity");
2918 }
2919 #endif
2920 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2921 int new_edges = 0;
2922 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2923 PointsToNode* field = i.get();
2924 if (field->is_Field() && field->as_Field()->is_oop()) {
2925 if (add_edge(field, init_val)) {
2926 // New edge was added
2927 new_edges++;
2928 add_field_uses_to_worklist(field->as_Field());
2929 }
2930 }
2931 }
2932 return new_edges;
2933 }
2934
2935 // Find fields initializing values for allocations.
2936 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2937 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2938 Node* alloc = pta->ideal_node();
2939 // Do nothing for Call nodes since its fields values are unknown.
2940 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2941 return 0;
2942 }
2943 InitializeNode* ini = alloc->as_Allocate()->initialization();
2944 bool visited_bottom_offset = false;
2945 GrowableArray<int> offsets_worklist;
2946 int new_edges = 0;
2947
2948 // Check if an oop field's initializing value is recorded and add
2949 // a corresponding null if field's value if it is not recorded.
2950 // Connection Graph does not record a default initialization by null
2951 // captured by Initialize node.
2952 //
2953 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2954 PointsToNode* field = i.get(); // Field (AddP)
2955 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2956 continue; // Not oop field
2957 }
2958 int offset = field->as_Field()->offset();
2959 if (offset == Type::OffsetBot) {
2960 if (!visited_bottom_offset) {
2961 // OffsetBot is used to reference array's element,
2962 // always add reference to null to all Field nodes since we don't
2963 // known which element is referenced.
2964 if (add_edge(field, null_obj)) {
2965 // New edge was added
2966 new_edges++;
2967 add_field_uses_to_worklist(field->as_Field());
2968 visited_bottom_offset = true;
2969 }
2970 }
2971 } else {
2972 // Check only oop fields.
2973 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2974 if (adr_type->isa_rawptr()) {
2975 #ifdef ASSERT
2976 // Raw pointers are used for initializing stores so skip it
2977 // since it should be recorded already
2978 Node* base = get_addp_base(field->ideal_node());
2979 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2980 #endif
2981 continue;
2982 }
2983 if (!offsets_worklist.contains(offset)) {
2984 offsets_worklist.append(offset);
2985 Node* value = nullptr;
2986 if (ini != nullptr) {
2987 // StoreP::value_basic_type() == T_ADDRESS
2988 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2989 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2990 // Make sure initializing store has the same type as this AddP.
2991 // This AddP may reference non existing field because it is on a
2992 // dead branch of bimorphic call which is not eliminated yet.
2993 if (store != nullptr && store->is_Store() &&
2994 store->as_Store()->value_basic_type() == ft) {
2995 value = store->in(MemNode::ValueIn);
2996 #ifdef ASSERT
2997 if (VerifyConnectionGraph) {
2998 // Verify that AddP already points to all objects the value points to.
2999 PointsToNode* val = ptnode_adr(value->_idx);
3000 assert((val != nullptr), "should be processed already");
3001 PointsToNode* missed_obj = nullptr;
3002 if (val->is_JavaObject()) {
3003 if (!field->points_to(val->as_JavaObject())) {
3004 missed_obj = val;
3005 }
3006 } else {
3007 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
3008 tty->print_cr("----------init store has invalid value -----");
3009 store->dump();
3010 val->dump();
3011 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
3012 }
3013 for (EdgeIterator j(val); j.has_next(); j.next()) {
3014 PointsToNode* obj = j.get();
3015 if (obj->is_JavaObject()) {
3016 if (!field->points_to(obj->as_JavaObject())) {
3017 missed_obj = obj;
3018 break;
3019 }
3020 }
3021 }
3022 }
3023 if (missed_obj != nullptr) {
3024 tty->print_cr("----------field---------------------------------");
3025 field->dump();
3026 tty->print_cr("----------missed reference to object------------");
3027 missed_obj->dump();
3028 tty->print_cr("----------object referenced by init store-------");
3029 store->dump();
3030 val->dump();
3031 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
3032 }
3033 }
3034 #endif
3035 } else {
3036 // There could be initializing stores which follow allocation.
3037 // For example, a volatile field store is not collected
3038 // by Initialize node.
3039 //
3040 // Need to check for dependent loads to separate such stores from
3041 // stores which follow loads. For now, add initial value null so
3042 // that compare pointers optimization works correctly.
3043 }
3044 }
3045 if (value == nullptr) {
3046 // A field's initializing value was not recorded. Add null.
3047 if (add_edge(field, null_obj)) {
3048 // New edge was added
3049 new_edges++;
3050 add_field_uses_to_worklist(field->as_Field());
3051 }
3052 }
3053 }
3054 }
3055 }
3056 return new_edges;
3057 }
3058
3059 // Adjust scalar_replaceable state after Connection Graph is built.
3060 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
3061 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
3062 // returns true. If one of the constraints in this method set 'jobj' to NSR
3063 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
3064 // input, 'adjust_scalar_replaceable_state' will eventually be called with
3065 // that other object and the Phi will become a reducible Phi.
3066 // There could be multiple merges involving the same jobj.
3067 Unique_Node_List candidates;
3068
3069 // Search for non-escaping objects which are not scalar replaceable
3070 // and mark them to propagate the state to referenced objects.
3071
3072 for (UseIterator i(jobj); i.has_next(); i.next()) {
3073 PointsToNode* use = i.get();
3074 if (use->is_Arraycopy()) {
3075 continue;
3076 }
3077 if (use->is_Field()) {
3078 FieldNode* field = use->as_Field();
3079 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3080 // 1. An object is not scalar replaceable if the field into which it is
3081 // stored has unknown offset (stored into unknown element of an array).
3082 if (field->offset() == Type::OffsetBot) {
3083 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
3084 return;
3085 }
3086 for (BaseIterator i(field); i.has_next(); i.next()) {
3087 PointsToNode* base = i.get();
3088 // 2. An object is not scalar replaceable if the field into which it is
3089 // stored has multiple bases one of which is null.
3090 if ((base == null_obj) && (field->base_count() > 1)) {
3091 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3092 return;
3093 }
3094 // 2.5. An object is not scalar replaceable if the field into which it is
3095 // stored has NSR base.
3096 if (!base->scalar_replaceable()) {
3097 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3098 return;
3099 }
3100 }
3101 }
3102 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3103 // 3. An object is not scalar replaceable if it is merged with other objects
3104 // and we can't remove the merge
3105 for (EdgeIterator j(use); j.has_next(); j.next()) {
3106 PointsToNode* ptn = j.get();
3107 if (ptn->is_JavaObject() && ptn != jobj) {
3108 Node* use_n = use->ideal_node();
3109
3110 // These other local vars may point to multiple objects through a Phi
3111 // In this case we skip them and see if we can reduce the Phi.
3112 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3113 use_n = use_n->in(1);
3114 }
3115
3116 // If it's already a candidate or confirmed reducible merge we can skip verification
3117 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3118 continue;
3119 }
3120
3121 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3122 candidates.push(use_n);
3123 } else {
3124 // Mark all objects as NSR if we can't remove the merge
3125 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3126 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3127 }
3128 }
3129 }
3130 if (!jobj->scalar_replaceable()) {
3131 return;
3132 }
3133 }
3134
3135 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3136 if (j.get()->is_Arraycopy()) {
3137 continue;
3138 }
3139
3140 // Non-escaping object node should point only to field nodes.
3141 FieldNode* field = j.get()->as_Field();
3142 int offset = field->as_Field()->offset();
3143
3144 // 4. An object is not scalar replaceable if it has a field with unknown
3145 // offset (array's element is accessed in loop).
3146 if (offset == Type::OffsetBot) {
3147 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3148 return;
3149 }
3150 // 5. Currently an object is not scalar replaceable if a LoadStore node
3151 // access its field since the field value is unknown after it.
3152 //
3153 Node* n = field->ideal_node();
3154
3155 // Test for an unsafe access that was parsed as maybe off heap
3156 // (with a CheckCastPP to raw memory).
3157 assert(n->is_AddP(), "expect an address computation");
3158 if (n->in(AddPNode::Base)->is_top() &&
3159 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3160 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3161 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3162 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3163 return;
3164 }
3165
3166 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3167 Node* u = n->fast_out(i);
3168 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3169 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3170 return;
3171 }
3172 }
3173
3174 // 6. Or the address may point to more then one object. This may produce
3175 // the false positive result (set not scalar replaceable)
3176 // since the flow-insensitive escape analysis can't separate
3177 // the case when stores overwrite the field's value from the case
3178 // when stores happened on different control branches.
3179 //
3180 // Note: it will disable scalar replacement in some cases:
3181 //
3182 // Point p[] = new Point[1];
3183 // p[0] = new Point(); // Will be not scalar replaced
3184 //
3185 // but it will save us from incorrect optimizations in next cases:
3186 //
3187 // Point p[] = new Point[1];
3188 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3189 //
3190 if (field->base_count() > 1 && candidates.size() == 0) {
3191 if (has_non_reducible_merge(field, reducible_merges)) {
3192 for (BaseIterator i(field); i.has_next(); i.next()) {
3193 PointsToNode* base = i.get();
3194 // Don't take into account LocalVar nodes which
3195 // may point to only one object which should be also
3196 // this field's base by now.
3197 if (base->is_JavaObject() && base != jobj) {
3198 // Mark all bases.
3199 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3200 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3201 }
3202 }
3203
3204 if (!jobj->scalar_replaceable()) {
3205 return;
3206 }
3207 }
3208 }
3209 }
3210
3211 // The candidate is truly a reducible merge only if none of the other
3212 // constraints ruled it as NSR. There could be multiple merges involving the
3213 // same jobj.
3214 assert(jobj->scalar_replaceable(), "sanity");
3215 for (uint i = 0; i < candidates.size(); i++ ) {
3216 Node* candidate = candidates.at(i);
3217 reducible_merges.push(candidate);
3218 }
3219 }
3220
3221 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3222 for (BaseIterator i(field); i.has_next(); i.next()) {
3223 Node* base = i.get()->ideal_node();
3224 if (base->is_Phi() && !reducible_merges.member(base)) {
3225 return true;
3226 }
3227 }
3228 return false;
3229 }
3230
3231 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3232 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3233
3234 // Look for 'phis' that refer to 'jobj' as the last
3235 // remaining scalar replaceable input.
3236 uint reducible_merges_cnt = reducible_merges.size();
3237 for (uint i = 0; i < reducible_merges_cnt; i++) {
3238 Node* phi = reducible_merges.at(i);
3239
3240 // This 'Phi' will be a 'good' if it still points to
3241 // at least one scalar replaceable object. Note that 'obj'
3242 // was/should be marked as NSR before calling this function.
3243 bool good_phi = false;
3244
3245 for (uint j = 1; j < phi->req(); j++) {
3246 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3247 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3248 good_phi = true;
3249 break;
3250 }
3251 }
3252
3253 if (!good_phi) {
3254 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3255 reducible_merges.remove(i);
3256
3257 // Decrement the index because the 'remove' call above actually
3258 // moves the last entry of the list to position 'i'.
3259 i--;
3260
3261 reducible_merges_cnt--;
3262 }
3263 }
3264 }
3265
3266 // Propagate NSR (Not scalar replaceable) state.
3267 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3268 int jobj_length = jobj_worklist.length();
3269 bool found_nsr_alloc = true;
3270 while (found_nsr_alloc) {
3271 found_nsr_alloc = false;
3272 for (int next = 0; next < jobj_length; ++next) {
3273 JavaObjectNode* jobj = jobj_worklist.at(next);
3274 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3275 PointsToNode* use = i.get();
3276 if (use->is_Field()) {
3277 FieldNode* field = use->as_Field();
3278 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3279 assert(field->offset() != Type::OffsetBot, "sanity");
3280 for (BaseIterator i(field); i.has_next(); i.next()) {
3281 PointsToNode* base = i.get();
3282 // An object is not scalar replaceable if the field into which
3283 // it is stored has NSR base.
3284 if ((base != null_obj) && !base->scalar_replaceable()) {
3285 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3286 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3287 // because there is no point in reducing a Phi that won't improve the number of SR
3288 // objects.
3289 revisit_reducible_phi_status(jobj, reducible_merges);
3290 found_nsr_alloc = true;
3291 break;
3292 }
3293 }
3294 } else if (use->is_LocalVar()) {
3295 Node* phi = use->ideal_node();
3296 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3297 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3298 reducible_merges.yank(phi);
3299 found_nsr_alloc = true;
3300 break;
3301 }
3302 }
3303 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3304 }
3305 }
3306 }
3307 }
3308
3309 #ifdef ASSERT
3310 void ConnectionGraph::verify_connection_graph(
3311 GrowableArray<PointsToNode*>& ptnodes_worklist,
3312 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3313 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3314 GrowableArray<Node*>& addp_worklist) {
3315 // Verify that graph is complete - no new edges could be added.
3316 int java_objects_length = java_objects_worklist.length();
3317 int non_escaped_length = non_escaped_allocs_worklist.length();
3318 int new_edges = 0;
3319 for (int next = 0; next < java_objects_length; ++next) {
3320 JavaObjectNode* ptn = java_objects_worklist.at(next);
3321 new_edges += add_java_object_edges(ptn, true);
3322 }
3323 assert(new_edges == 0, "graph was not complete");
3324 // Verify that escape state is final.
3325 int length = non_escaped_allocs_worklist.length();
3326 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3327 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3328 (non_escaped_length == length) &&
3329 (_worklist.length() == 0), "escape state was not final");
3330
3331 // Verify fields information.
3332 int addp_length = addp_worklist.length();
3333 for (int next = 0; next < addp_length; ++next ) {
3334 Node* n = addp_worklist.at(next);
3335 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3336 if (field->is_oop()) {
3337 // Verify that field has all bases
3338 Node* base = get_addp_base(n);
3339 PointsToNode* ptn = ptnode_adr(base->_idx);
3340 if (ptn->is_JavaObject()) {
3341 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3342 } else {
3343 assert(ptn->is_LocalVar(), "sanity");
3344 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3345 PointsToNode* e = i.get();
3346 if (e->is_JavaObject()) {
3347 assert(field->has_base(e->as_JavaObject()), "sanity");
3348 }
3349 }
3350 }
3351 // Verify that all fields have initializing values.
3352 if (field->edge_count() == 0) {
3353 tty->print_cr("----------field does not have references----------");
3354 field->dump();
3355 for (BaseIterator i(field); i.has_next(); i.next()) {
3356 PointsToNode* base = i.get();
3357 tty->print_cr("----------field has next base---------------------");
3358 base->dump();
3359 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3360 tty->print_cr("----------base has fields-------------------------");
3361 for (EdgeIterator j(base); j.has_next(); j.next()) {
3362 j.get()->dump();
3363 }
3364 tty->print_cr("----------base has references---------------------");
3365 for (UseIterator j(base); j.has_next(); j.next()) {
3366 j.get()->dump();
3367 }
3368 }
3369 }
3370 for (UseIterator i(field); i.has_next(); i.next()) {
3371 i.get()->dump();
3372 }
3373 assert(field->edge_count() > 0, "sanity");
3374 }
3375 }
3376 }
3377 }
3378 #endif
3379
3380 // Optimize ideal graph.
3381 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3382 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3383 Compile* C = _compile;
3384 PhaseIterGVN* igvn = _igvn;
3385 if (EliminateLocks) {
3386 // Mark locks before changing ideal graph.
3387 int cnt = C->macro_count();
3388 for (int i = 0; i < cnt; i++) {
3389 Node *n = C->macro_node(i);
3390 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3391 AbstractLockNode* alock = n->as_AbstractLock();
3392 if (!alock->is_non_esc_obj()) {
3393 const Type* obj_type = igvn->type(alock->obj_node());
3394 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3395 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3396 // The lock could be marked eliminated by lock coarsening
3397 // code during first IGVN before EA. Replace coarsened flag
3398 // to eliminate all associated locks/unlocks.
3399 #ifdef ASSERT
3400 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3401 #endif
3402 alock->set_non_esc_obj();
3403 }
3404 }
3405 }
3406 }
3407 }
3408
3409 if (OptimizePtrCompare) {
3410 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3411 Node *n = ptr_cmp_worklist.at(i);
3412 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3413 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3414 if (tcmp->singleton()) {
3415 Node* cmp = igvn->makecon(tcmp);
3416 #ifndef PRODUCT
3417 if (PrintOptimizePtrCompare) {
3418 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"));
3419 if (Verbose) {
3420 n->dump(1);
3421 }
3422 }
3423 #endif
3424 igvn->replace_node(n, cmp);
3425 }
3426 }
3427 }
3428
3429 // For MemBarStoreStore nodes added in library_call.cpp, check
3430 // escape status of associated AllocateNode and optimize out
3431 // MemBarStoreStore node if the allocated object never escapes.
3432 for (int i = 0; i < storestore_worklist.length(); i++) {
3433 Node* storestore = storestore_worklist.at(i);
3434 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3435 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3436 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3437 // Non-escaping inline type buffer allocations don't require a membar
3438 storestore->as_MemBar()->remove(_igvn);
3439 } else {
3440 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3441 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3442 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3443 igvn->register_new_node_with_optimizer(mb);
3444 igvn->replace_node(storestore, mb);
3445 }
3446 }
3447 }
3448 }
3449
3450 // Atomic flat accesses on non-escaping objects can be optimized to non-atomic accesses
3451 void ConnectionGraph::optimize_flat_accesses(GrowableArray<SafePointNode*>& sfn_worklist) {
3452 PhaseIterGVN& igvn = *_igvn;
3453 bool delay = igvn.delay_transform();
3454 igvn.set_delay_transform(true);
3455 igvn.C->for_each_flat_access([&](Node* n) {
3456 Node* base = n->is_LoadFlat() ? n->as_LoadFlat()->base() : n->as_StoreFlat()->base();
3457 if (!not_global_escape(base)) {
3458 return;
3459 }
3460
3461 bool expanded;
3462 if (n->is_LoadFlat()) {
3463 expanded = n->as_LoadFlat()->expand_non_atomic(igvn);
3464 } else {
3465 expanded = n->as_StoreFlat()->expand_non_atomic(igvn);
3466 }
3467 if (expanded) {
3468 sfn_worklist.remove(n->as_SafePoint());
3469 igvn.C->remove_flat_access(n);
3470 }
3471 });
3472 igvn.set_delay_transform(delay);
3473 }
3474
3475 // Optimize objects compare.
3476 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3477 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3478 if (!OptimizePtrCompare) {
3479 return UNKNOWN;
3480 }
3481 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3482 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3483
3484 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3485 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3486 JavaObjectNode* jobj1 = unique_java_object(left);
3487 JavaObjectNode* jobj2 = unique_java_object(right);
3488
3489 // The use of this method during allocation merge reduction may cause 'left'
3490 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3491 // that doesn't reference an unique java object.
3492 if (ptn1 == nullptr || ptn2 == nullptr ||
3493 jobj1 == nullptr || jobj2 == nullptr) {
3494 return UNKNOWN;
3495 }
3496
3497 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3498 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3499
3500 // Check simple cases first.
3501 if (jobj1 != nullptr) {
3502 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3503 if (jobj1 == jobj2) {
3504 // Comparing the same not escaping object.
3505 return EQ;
3506 }
3507 Node* obj = jobj1->ideal_node();
3508 // Comparing not escaping allocation.
3509 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3510 !ptn2->points_to(jobj1)) {
3511 return NE; // This includes nullness check.
3512 }
3513 }
3514 }
3515 if (jobj2 != nullptr) {
3516 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3517 Node* obj = jobj2->ideal_node();
3518 // Comparing not escaping allocation.
3519 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3520 !ptn1->points_to(jobj2)) {
3521 return NE; // This includes nullness check.
3522 }
3523 }
3524 }
3525 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3526 jobj2 != nullptr && jobj2 != phantom_obj &&
3527 jobj1->ideal_node()->is_Con() &&
3528 jobj2->ideal_node()->is_Con()) {
3529 // Klass or String constants compare. Need to be careful with
3530 // compressed pointers - compare types of ConN and ConP instead of nodes.
3531 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3532 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3533 if (t1->make_ptr() == t2->make_ptr()) {
3534 return EQ;
3535 } else {
3536 return NE;
3537 }
3538 }
3539 if (ptn1->meet(ptn2)) {
3540 return UNKNOWN; // Sets are not disjoint
3541 }
3542
3543 // Sets are disjoint.
3544 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3545 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3546 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3547 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3548 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3549 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3550 // Check nullness of unknown object.
3551 return UNKNOWN;
3552 }
3553
3554 // Disjointness by itself is not sufficient since
3555 // alias analysis is not complete for escaped objects.
3556 // Disjoint sets are definitely unrelated only when
3557 // at least one set has only not escaping allocations.
3558 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3559 if (ptn1->non_escaping_allocation()) {
3560 return NE;
3561 }
3562 }
3563 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3564 if (ptn2->non_escaping_allocation()) {
3565 return NE;
3566 }
3567 }
3568 return UNKNOWN;
3569 }
3570
3571 // Connection Graph construction functions.
3572
3573 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3574 PointsToNode* ptadr = _nodes.at(n->_idx);
3575 if (ptadr != nullptr) {
3576 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3577 return;
3578 }
3579 Compile* C = _compile;
3580 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3581 map_ideal_node(n, ptadr);
3582 }
3583
3584 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3585 PointsToNode* ptadr = _nodes.at(n->_idx);
3586 if (ptadr != nullptr) {
3587 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3588 return ptadr;
3589 }
3590 Compile* C = _compile;
3591 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3592 map_ideal_node(n, ptadr);
3593 return ptadr;
3594 }
3595
3596 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3597 PointsToNode* ptadr = _nodes.at(n->_idx);
3598 if (ptadr != nullptr) {
3599 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3600 return;
3601 }
3602 bool unsafe = false;
3603 bool is_oop = is_oop_field(n, offset, &unsafe);
3604 if (unsafe) {
3605 es = PointsToNode::GlobalEscape;
3606 }
3607 Compile* C = _compile;
3608 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3609 map_ideal_node(n, field);
3610 }
3611
3612 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3613 PointsToNode* src, PointsToNode* dst) {
3614 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3615 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3616 PointsToNode* ptadr = _nodes.at(n->_idx);
3617 if (ptadr != nullptr) {
3618 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3619 return;
3620 }
3621 Compile* C = _compile;
3622 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3623 map_ideal_node(n, ptadr);
3624 // Add edge from arraycopy node to source object.
3625 (void)add_edge(ptadr, src);
3626 src->set_arraycopy_src();
3627 // Add edge from destination object to arraycopy node.
3628 (void)add_edge(dst, ptadr);
3629 dst->set_arraycopy_dst();
3630 }
3631
3632 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3633 const Type* adr_type = n->as_AddP()->bottom_type();
3634 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3635 BasicType bt = T_INT;
3636 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3637 // Check only oop fields.
3638 if (!adr_type->isa_aryptr() ||
3639 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3640 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3641 // OffsetBot is used to reference array's element. Ignore first AddP.
3642 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3643 bt = T_OBJECT;
3644 }
3645 }
3646 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3647 if (adr_type->isa_instptr()) {
3648 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3649 if (field != nullptr) {
3650 bt = field->layout_type();
3651 } else {
3652 // Check for unsafe oop field access
3653 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3654 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3655 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3656 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3657 bt = T_OBJECT;
3658 (*unsafe) = true;
3659 }
3660 }
3661 } else if (adr_type->isa_aryptr()) {
3662 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3663 // Ignore array length load.
3664 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3665 // Ignore first AddP.
3666 } else {
3667 const Type* elemtype = adr_type->is_aryptr()->elem();
3668 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3669 ciInlineKlass* vk = elemtype->inline_klass();
3670 field_offset += vk->payload_offset();
3671 ciField* field = vk->get_field_by_offset(field_offset, false);
3672 if (field != nullptr) {
3673 bt = field->layout_type();
3674 } else {
3675 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);
3676 bt = T_BOOLEAN;
3677 }
3678 } else {
3679 bt = elemtype->array_element_basic_type();
3680 }
3681 }
3682 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3683 // Allocation initialization, ThreadLocal field access, unsafe access
3684 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3685 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3686 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3687 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3688 bt = T_OBJECT;
3689 }
3690 }
3691 }
3692 // Note: T_NARROWOOP is not classed as a real reference type
3693 return (is_reference_type(bt) || bt == T_NARROWOOP);
3694 }
3695
3696 // Returns unique pointed java object or null.
3697 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3698 // If the node was created after the escape computation we can't answer.
3699 uint idx = n->_idx;
3700 if (idx >= nodes_size()) {
3701 return nullptr;
3702 }
3703 PointsToNode* ptn = ptnode_adr(idx);
3704 if (ptn == nullptr) {
3705 return nullptr;
3706 }
3707 if (ptn->is_JavaObject()) {
3708 return ptn->as_JavaObject();
3709 }
3710 assert(ptn->is_LocalVar(), "sanity");
3711 // Check all java objects it points to.
3712 JavaObjectNode* jobj = nullptr;
3713 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3714 PointsToNode* e = i.get();
3715 if (e->is_JavaObject()) {
3716 if (jobj == nullptr) {
3717 jobj = e->as_JavaObject();
3718 } else if (jobj != e) {
3719 return nullptr;
3720 }
3721 }
3722 }
3723 return jobj;
3724 }
3725
3726 // Return true if this node points only to non-escaping allocations.
3727 bool PointsToNode::non_escaping_allocation() {
3728 if (is_JavaObject()) {
3729 Node* n = ideal_node();
3730 if (n->is_Allocate() || n->is_CallStaticJava()) {
3731 return (escape_state() == PointsToNode::NoEscape);
3732 } else {
3733 return false;
3734 }
3735 }
3736 assert(is_LocalVar(), "sanity");
3737 // Check all java objects it points to.
3738 for (EdgeIterator i(this); i.has_next(); i.next()) {
3739 PointsToNode* e = i.get();
3740 if (e->is_JavaObject()) {
3741 Node* n = e->ideal_node();
3742 if ((e->escape_state() != PointsToNode::NoEscape) ||
3743 !(n->is_Allocate() || n->is_CallStaticJava())) {
3744 return false;
3745 }
3746 }
3747 }
3748 return true;
3749 }
3750
3751 // Return true if we know the node does not escape globally.
3752 bool ConnectionGraph::not_global_escape(Node *n) {
3753 assert(!_collecting, "should not call during graph construction");
3754 // If the node was created after the escape computation we can't answer.
3755 uint idx = n->_idx;
3756 if (idx >= nodes_size()) {
3757 return false;
3758 }
3759 PointsToNode* ptn = ptnode_adr(idx);
3760 if (ptn == nullptr) {
3761 return false; // not in congraph (e.g. ConI)
3762 }
3763 PointsToNode::EscapeState es = ptn->escape_state();
3764 // If we have already computed a value, return it.
3765 if (es >= PointsToNode::GlobalEscape) {
3766 return false;
3767 }
3768 if (ptn->is_JavaObject()) {
3769 return true; // (es < PointsToNode::GlobalEscape);
3770 }
3771 assert(ptn->is_LocalVar(), "sanity");
3772 // Check all java objects it points to.
3773 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3774 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3775 return false;
3776 }
3777 }
3778 return true;
3779 }
3780
3781 // Return true if locked object does not escape globally
3782 // and locked code region (identified by BoxLockNode) is balanced:
3783 // all compiled code paths have corresponding Lock/Unlock pairs.
3784 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3785 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3786 if (EliminateNestedLocks) {
3787 // We can mark whole locking region as Local only when only
3788 // one object is used for locking.
3789 alock->box_node()->as_BoxLock()->set_local();
3790 }
3791 return true;
3792 }
3793 return false;
3794 }
3795
3796 // Helper functions
3797
3798 // Return true if this node points to specified node or nodes it points to.
3799 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3800 if (is_JavaObject()) {
3801 return (this == ptn);
3802 }
3803 assert(is_LocalVar() || is_Field(), "sanity");
3804 for (EdgeIterator i(this); i.has_next(); i.next()) {
3805 if (i.get() == ptn) {
3806 return true;
3807 }
3808 }
3809 return false;
3810 }
3811
3812 // Return true if one node points to an other.
3813 bool PointsToNode::meet(PointsToNode* ptn) {
3814 if (this == ptn) {
3815 return true;
3816 } else if (ptn->is_JavaObject()) {
3817 return this->points_to(ptn->as_JavaObject());
3818 } else if (this->is_JavaObject()) {
3819 return ptn->points_to(this->as_JavaObject());
3820 }
3821 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3822 int ptn_count = ptn->edge_count();
3823 for (EdgeIterator i(this); i.has_next(); i.next()) {
3824 PointsToNode* this_e = i.get();
3825 for (int j = 0; j < ptn_count; j++) {
3826 if (this_e == ptn->edge(j)) {
3827 return true;
3828 }
3829 }
3830 }
3831 return false;
3832 }
3833
3834 #ifdef ASSERT
3835 // Return true if bases point to this java object.
3836 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3837 for (BaseIterator i(this); i.has_next(); i.next()) {
3838 if (i.get() == jobj) {
3839 return true;
3840 }
3841 }
3842 return false;
3843 }
3844 #endif
3845
3846 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3847 // Handle simple case first.
3848 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3849 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3850 return true;
3851 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3852 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3853 Node* addp_use = addp->fast_out(i);
3854 if (addp_use->is_Store()) {
3855 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3856 if (addp_use->fast_out(j)->is_Initialize()) {
3857 return true;
3858 }
3859 }
3860 }
3861 }
3862 }
3863 return false;
3864 }
3865
3866 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3867 const Type *adr_type = phase->type(adr);
3868 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3869 // We are computing a raw address for a store captured by an Initialize
3870 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3871 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3872 assert(offs != Type::OffsetBot ||
3873 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3874 "offset must be a constant or it is initialization of array");
3875 return offs;
3876 }
3877 return adr_type->is_ptr()->flat_offset();
3878 }
3879
3880 Node* ConnectionGraph::get_addp_base(Node *addp) {
3881 assert(addp->is_AddP(), "must be AddP");
3882 //
3883 // AddP cases for Base and Address inputs:
3884 // case #1. Direct object's field reference:
3885 // Allocate
3886 // |
3887 // Proj #5 ( oop result )
3888 // |
3889 // CheckCastPP (cast to instance type)
3890 // | |
3891 // AddP ( base == address )
3892 //
3893 // case #2. Indirect object's field reference:
3894 // Phi
3895 // |
3896 // CastPP (cast to instance type)
3897 // | |
3898 // AddP ( base == address )
3899 //
3900 // case #3. Raw object's field reference for Initialize node:
3901 // Allocate
3902 // |
3903 // Proj #5 ( oop result )
3904 // top |
3905 // \ |
3906 // AddP ( base == top )
3907 //
3908 // case #4. Array's element reference:
3909 // {CheckCastPP | CastPP}
3910 // | | |
3911 // | AddP ( array's element offset )
3912 // | |
3913 // AddP ( array's offset )
3914 //
3915 // case #5. Raw object's field reference for arraycopy stub call:
3916 // The inline_native_clone() case when the arraycopy stub is called
3917 // after the allocation before Initialize and CheckCastPP nodes.
3918 // Allocate
3919 // |
3920 // Proj #5 ( oop result )
3921 // | |
3922 // AddP ( base == address )
3923 //
3924 // case #6. Constant Pool, ThreadLocal, CastX2P or
3925 // Raw object's field reference:
3926 // {ConP, ThreadLocal, CastX2P, raw Load}
3927 // top |
3928 // \ |
3929 // AddP ( base == top )
3930 //
3931 // case #7. Klass's field reference.
3932 // LoadKlass
3933 // | |
3934 // AddP ( base == address )
3935 //
3936 // case #8. narrow Klass's field reference.
3937 // LoadNKlass
3938 // |
3939 // DecodeN
3940 // | |
3941 // AddP ( base == address )
3942 //
3943 // case #9. Mixed unsafe access
3944 // {instance}
3945 // |
3946 // CheckCastPP (raw)
3947 // top |
3948 // \ |
3949 // AddP ( base == top )
3950 //
3951 Node *base = addp->in(AddPNode::Base);
3952 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3953 base = addp->in(AddPNode::Address);
3954 while (base->is_AddP()) {
3955 // Case #6 (unsafe access) may have several chained AddP nodes.
3956 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3957 base = base->in(AddPNode::Address);
3958 }
3959 if (base->Opcode() == Op_CheckCastPP &&
3960 base->bottom_type()->isa_rawptr() &&
3961 _igvn->type(base->in(1))->isa_oopptr()) {
3962 base = base->in(1); // Case #9
3963 } else {
3964 Node* uncast_base = base->uncast();
3965 int opcode = uncast_base->Opcode();
3966 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3967 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3968 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3969 is_captured_store_address(addp), "sanity");
3970 }
3971 }
3972 return base;
3973 }
3974
3975 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3976 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3977 Node* addp2 = addp->raw_out(0);
3978 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3979 addp2->in(AddPNode::Base) == n &&
3980 addp2->in(AddPNode::Address) == addp) {
3981 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3982 //
3983 // Find array's offset to push it on worklist first and
3984 // as result process an array's element offset first (pushed second)
3985 // to avoid CastPP for the array's offset.
3986 // Otherwise the inserted CastPP (LocalVar) will point to what
3987 // the AddP (Field) points to. Which would be wrong since
3988 // the algorithm expects the CastPP has the same point as
3989 // as AddP's base CheckCastPP (LocalVar).
3990 //
3991 // ArrayAllocation
3992 // |
3993 // CheckCastPP
3994 // |
3995 // memProj (from ArrayAllocation CheckCastPP)
3996 // | ||
3997 // | || Int (element index)
3998 // | || | ConI (log(element size))
3999 // | || | /
4000 // | || LShift
4001 // | || /
4002 // | AddP (array's element offset)
4003 // | |
4004 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
4005 // | / /
4006 // AddP (array's offset)
4007 // |
4008 // Load/Store (memory operation on array's element)
4009 //
4010 return addp2;
4011 }
4012 return nullptr;
4013 }
4014
4015 //
4016 // Adjust the type and inputs of an AddP which computes the
4017 // address of a field of an instance
4018 //
4019 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
4020 PhaseGVN* igvn = _igvn;
4021 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
4022 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
4023 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
4024 if (t == nullptr) {
4025 // We are computing a raw address for a store captured by an Initialize
4026 // compute an appropriate address type (cases #3 and #5).
4027 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
4028 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
4029 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
4030 assert(offs != Type::OffsetBot, "offset must be a constant");
4031 if (base_t->isa_aryptr() != nullptr) {
4032 // In the case of a flat inline type array, each field has its
4033 // own slice so we need to extract the field being accessed from
4034 // the address computation
4035 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
4036 } else {
4037 t = base_t->add_offset(offs)->is_oopptr();
4038 }
4039 }
4040 int inst_id = base_t->instance_id();
4041 assert(!t->is_known_instance() || t->instance_id() == inst_id,
4042 "old type must be non-instance or match new type");
4043
4044 // The type 't' could be subclass of 'base_t'.
4045 // As result t->offset() could be large then base_t's size and it will
4046 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
4047 // constructor verifies correctness of the offset.
4048 //
4049 // It could happened on subclass's branch (from the type profiling
4050 // inlining) which was not eliminated during parsing since the exactness
4051 // of the allocation type was not propagated to the subclass type check.
4052 //
4053 // Or the type 't' could be not related to 'base_t' at all.
4054 // It could happen when CHA type is different from MDO type on a dead path
4055 // (for example, from instanceof check) which is not collapsed during parsing.
4056 //
4057 // Do nothing for such AddP node and don't process its users since
4058 // this code branch will go away.
4059 //
4060 if (!t->is_known_instance() &&
4061 !base_t->maybe_java_subtype_of(t)) {
4062 return false; // bail out
4063 }
4064 const TypePtr* tinst = base_t->add_offset(t->offset());
4065 if (tinst->isa_aryptr() && t->isa_aryptr()) {
4066 // In the case of a flat inline type array, each field has its
4067 // own slice so we need to keep track of the field being accessed.
4068 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
4069 // Keep array properties (not flat/null-free)
4070 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
4071 if (tinst == nullptr) {
4072 return false; // Skip dead path with inconsistent properties
4073 }
4074 }
4075
4076 // Do NOT remove the next line: ensure a new alias index is allocated
4077 // for the instance type. Note: C++ will not remove it since the call
4078 // has side effect.
4079 int alias_idx = _compile->get_alias_index(tinst);
4080 igvn->set_type(addp, tinst);
4081 // record the allocation in the node map
4082 set_map(addp, get_map(base->_idx));
4083 // Set addp's Base and Address to 'base'.
4084 Node *abase = addp->in(AddPNode::Base);
4085 Node *adr = addp->in(AddPNode::Address);
4086 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
4087 adr->in(0)->_idx == (uint)inst_id) {
4088 // Skip AddP cases #3 and #5.
4089 } else {
4090 assert(!abase->is_top(), "sanity"); // AddP case #3
4091 if (abase != base) {
4092 igvn->hash_delete(addp);
4093 addp->set_req(AddPNode::Base, base);
4094 if (abase == adr) {
4095 addp->set_req(AddPNode::Address, base);
4096 } else {
4097 // AddP case #4 (adr is array's element offset AddP node)
4098 #ifdef ASSERT
4099 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
4100 assert(adr->is_AddP() && atype != nullptr &&
4101 atype->instance_id() == inst_id, "array's element offset should be processed first");
4102 #endif
4103 }
4104 igvn->hash_insert(addp);
4105 }
4106 }
4107 // Put on IGVN worklist since at least addp's type was changed above.
4108 record_for_optimizer(addp);
4109 return true;
4110 }
4111
4112 //
4113 // Create a new version of orig_phi if necessary. Returns either the newly
4114 // created phi or an existing phi. Sets create_new to indicate whether a new
4115 // phi was created. Cache the last newly created phi in the node map.
4116 //
4117 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
4118 Compile *C = _compile;
4119 PhaseGVN* igvn = _igvn;
4120 new_created = false;
4121 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
4122 // nothing to do if orig_phi is bottom memory or matches alias_idx
4123 if (phi_alias_idx == alias_idx) {
4124 return orig_phi;
4125 }
4126 // Have we recently created a Phi for this alias index?
4127 PhiNode *result = get_map_phi(orig_phi->_idx);
4128 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
4129 return result;
4130 }
4131 // Previous check may fail when the same wide memory Phi was split into Phis
4132 // for different memory slices. Search all Phis for this region.
4133 if (result != nullptr) {
4134 Node* region = orig_phi->in(0);
4135 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4136 Node* phi = region->fast_out(i);
4137 if (phi->is_Phi() &&
4138 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4139 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4140 return phi->as_Phi();
4141 }
4142 }
4143 }
4144 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4145 if (C->do_escape_analysis() == true && !C->failing()) {
4146 // Retry compilation without escape analysis.
4147 // If this is the first failure, the sentinel string will "stick"
4148 // to the Compile object, and the C2Compiler will see it and retry.
4149 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4150 }
4151 return nullptr;
4152 }
4153 orig_phi_worklist.append_if_missing(orig_phi);
4154 const TypePtr *atype = C->get_adr_type(alias_idx);
4155 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4156 C->copy_node_notes_to(result, orig_phi);
4157 igvn->set_type(result, result->bottom_type());
4158 record_for_optimizer(result);
4159 set_map(orig_phi, result);
4160 new_created = true;
4161 return result;
4162 }
4163
4164 //
4165 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4166 // specified alias index.
4167 //
4168 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4169 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4170 Compile *C = _compile;
4171 PhaseGVN* igvn = _igvn;
4172 bool new_phi_created;
4173 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4174 if (!new_phi_created) {
4175 return result;
4176 }
4177 GrowableArray<PhiNode *> phi_list;
4178 GrowableArray<uint> cur_input;
4179 PhiNode *phi = orig_phi;
4180 uint idx = 1;
4181 bool finished = false;
4182 while(!finished) {
4183 while (idx < phi->req()) {
4184 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4185 if (mem != nullptr && mem->is_Phi()) {
4186 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4187 if (new_phi_created) {
4188 // found an phi for which we created a new split, push current one on worklist and begin
4189 // processing new one
4190 phi_list.push(phi);
4191 cur_input.push(idx);
4192 phi = mem->as_Phi();
4193 result = newphi;
4194 idx = 1;
4195 continue;
4196 } else {
4197 mem = newphi;
4198 }
4199 }
4200 if (C->failing()) {
4201 return nullptr;
4202 }
4203 result->set_req(idx++, mem);
4204 }
4205 #ifdef ASSERT
4206 // verify that the new Phi has an input for each input of the original
4207 assert( phi->req() == result->req(), "must have same number of inputs.");
4208 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4209 #endif
4210 // Check if all new phi's inputs have specified alias index.
4211 // Otherwise use old phi.
4212 for (uint i = 1; i < phi->req(); i++) {
4213 Node* in = result->in(i);
4214 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4215 }
4216 // we have finished processing a Phi, see if there are any more to do
4217 finished = (phi_list.length() == 0 );
4218 if (!finished) {
4219 phi = phi_list.pop();
4220 idx = cur_input.pop();
4221 PhiNode *prev_result = get_map_phi(phi->_idx);
4222 prev_result->set_req(idx++, result);
4223 result = prev_result;
4224 }
4225 }
4226 return result;
4227 }
4228
4229 //
4230 // The next methods are derived from methods in MemNode.
4231 //
4232 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4233 Node *mem = mmem;
4234 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4235 // means an array I have not precisely typed yet. Do not do any
4236 // alias stuff with it any time soon.
4237 if (toop->base() != Type::AnyPtr &&
4238 !(toop->isa_instptr() &&
4239 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4240 toop->offset() == Type::OffsetBot)) {
4241 mem = mmem->memory_at(alias_idx);
4242 // Update input if it is progress over what we have now
4243 }
4244 return mem;
4245 }
4246
4247 //
4248 // Move memory users to their memory slices.
4249 //
4250 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4251 Compile* C = _compile;
4252 PhaseGVN* igvn = _igvn;
4253 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4254 assert(tp != nullptr, "ptr type");
4255 int alias_idx = C->get_alias_index(tp);
4256 int general_idx = C->get_general_index(alias_idx);
4257
4258 // Move users first
4259 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4260 Node* use = n->fast_out(i);
4261 if (use->is_MergeMem()) {
4262 MergeMemNode* mmem = use->as_MergeMem();
4263 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4264 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4265 continue; // Nothing to do
4266 }
4267 // Replace previous general reference to mem node.
4268 uint orig_uniq = C->unique();
4269 Node* m = find_inst_mem(n, general_idx, orig_phis);
4270 assert(orig_uniq == C->unique(), "no new nodes");
4271 mmem->set_memory_at(general_idx, m);
4272 --imax;
4273 --i;
4274 } else if (use->is_MemBar()) {
4275 assert(!use->is_Initialize(), "initializing stores should not be moved");
4276 if (use->req() > MemBarNode::Precedent &&
4277 use->in(MemBarNode::Precedent) == n) {
4278 // Don't move related membars.
4279 record_for_optimizer(use);
4280 continue;
4281 }
4282 tp = use->as_MemBar()->adr_type()->isa_ptr();
4283 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4284 alias_idx == general_idx) {
4285 continue; // Nothing to do
4286 }
4287 // Move to general memory slice.
4288 uint orig_uniq = C->unique();
4289 Node* m = find_inst_mem(n, general_idx, orig_phis);
4290 assert(orig_uniq == C->unique(), "no new nodes");
4291 igvn->hash_delete(use);
4292 imax -= use->replace_edge(n, m, igvn);
4293 igvn->hash_insert(use);
4294 record_for_optimizer(use);
4295 --i;
4296 #ifdef ASSERT
4297 } else if (use->is_Mem()) {
4298 // Memory nodes should have new memory input.
4299 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4300 assert(tp != nullptr, "ptr type");
4301 int idx = C->get_alias_index(tp);
4302 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4303 "Following memory nodes should have new memory input or be on the same memory slice");
4304 } else if (use->is_Phi()) {
4305 // Phi nodes should be split and moved already.
4306 tp = use->as_Phi()->adr_type()->isa_ptr();
4307 assert(tp != nullptr, "ptr type");
4308 int idx = C->get_alias_index(tp);
4309 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4310 } else {
4311 use->dump();
4312 assert(false, "should not be here");
4313 #endif
4314 }
4315 }
4316 }
4317
4318 //
4319 // Search memory chain of "mem" to find a MemNode whose address
4320 // is the specified alias index.
4321 //
4322 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4323 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4324 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4325 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4326 return nullptr;
4327 }
4328 if (orig_mem == nullptr) {
4329 return orig_mem;
4330 }
4331 Compile* C = _compile;
4332 PhaseGVN* igvn = _igvn;
4333 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4334 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4335 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4336 Node *prev = nullptr;
4337 Node *result = orig_mem;
4338 while (prev != result) {
4339 prev = result;
4340 if (result == start_mem) {
4341 break; // hit one of our sentinels
4342 }
4343 if (result->is_Mem()) {
4344 const Type *at = igvn->type(result->in(MemNode::Address));
4345 if (at == Type::TOP) {
4346 break; // Dead
4347 }
4348 assert (at->isa_ptr() != nullptr, "pointer type required.");
4349 int idx = C->get_alias_index(at->is_ptr());
4350 if (idx == alias_idx) {
4351 break; // Found
4352 }
4353 if (!is_instance && (at->isa_oopptr() == nullptr ||
4354 !at->is_oopptr()->is_known_instance())) {
4355 break; // Do not skip store to general memory slice.
4356 }
4357 result = result->in(MemNode::Memory);
4358 }
4359 if (!is_instance) {
4360 continue; // don't search further for non-instance types
4361 }
4362 // skip over a call which does not affect this memory slice
4363 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4364 Node *proj_in = result->in(0);
4365 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4366 break; // hit one of our sentinels
4367 } else if (proj_in->is_Call()) {
4368 // ArrayCopy node processed here as well
4369 CallNode *call = proj_in->as_Call();
4370 if (!call->may_modify(toop, igvn)) {
4371 result = call->in(TypeFunc::Memory);
4372 }
4373 } else if (proj_in->is_Initialize()) {
4374 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4375 // Stop if this is the initialization for the object instance which
4376 // which contains this memory slice, otherwise skip over it.
4377 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4378 result = proj_in->in(TypeFunc::Memory);
4379 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4380 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4381 result = get_map(result->_idx);
4382 assert(result != nullptr, "new projection should have been allocated");
4383 break;
4384 }
4385 } else if (proj_in->is_MemBar()) {
4386 // Check if there is an array copy for a clone
4387 // Step over GC barrier when ReduceInitialCardMarks is disabled
4388 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4389 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4390
4391 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4392 // Stop if it is a clone
4393 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4394 if (ac->may_modify(toop, igvn)) {
4395 break;
4396 }
4397 }
4398 result = proj_in->in(TypeFunc::Memory);
4399 }
4400 } else if (result->is_MergeMem()) {
4401 MergeMemNode *mmem = result->as_MergeMem();
4402 result = step_through_mergemem(mmem, alias_idx, toop);
4403 if (result == mmem->base_memory()) {
4404 // Didn't find instance memory, search through general slice recursively.
4405 result = mmem->memory_at(C->get_general_index(alias_idx));
4406 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4407 if (C->failing()) {
4408 return nullptr;
4409 }
4410 mmem->set_memory_at(alias_idx, result);
4411 }
4412 } else if (result->is_Phi() &&
4413 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4414 Node *un = result->as_Phi()->unique_input(igvn);
4415 if (un != nullptr) {
4416 orig_phis.append_if_missing(result->as_Phi());
4417 result = un;
4418 } else {
4419 break;
4420 }
4421 } else if (result->is_ClearArray()) {
4422 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4423 // Can not bypass initialization of the instance
4424 // we are looking for.
4425 break;
4426 }
4427 // Otherwise skip it (the call updated 'result' value).
4428 } else if (result->Opcode() == Op_SCMemProj) {
4429 Node* mem = result->in(0);
4430 Node* adr = nullptr;
4431 if (mem->is_LoadStore()) {
4432 adr = mem->in(MemNode::Address);
4433 } else {
4434 assert(mem->Opcode() == Op_EncodeISOArray ||
4435 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4436 adr = mem->in(3); // Memory edge corresponds to destination array
4437 }
4438 const Type *at = igvn->type(adr);
4439 if (at != Type::TOP) {
4440 assert(at->isa_ptr() != nullptr, "pointer type required.");
4441 int idx = C->get_alias_index(at->is_ptr());
4442 if (idx == alias_idx) {
4443 // Assert in debug mode
4444 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4445 break; // In product mode return SCMemProj node
4446 }
4447 }
4448 result = mem->in(MemNode::Memory);
4449 } else if (result->Opcode() == Op_StrInflatedCopy) {
4450 Node* adr = result->in(3); // Memory edge corresponds to destination array
4451 const Type *at = igvn->type(adr);
4452 if (at != Type::TOP) {
4453 assert(at->isa_ptr() != nullptr, "pointer type required.");
4454 int idx = C->get_alias_index(at->is_ptr());
4455 if (idx == alias_idx) {
4456 // Assert in debug mode
4457 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4458 break; // In product mode return SCMemProj node
4459 }
4460 }
4461 result = result->in(MemNode::Memory);
4462 }
4463 }
4464 if (result->is_Phi()) {
4465 PhiNode *mphi = result->as_Phi();
4466 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4467 const TypePtr *t = mphi->adr_type();
4468 if (!is_instance) {
4469 // Push all non-instance Phis on the orig_phis worklist to update inputs
4470 // during Phase 4 if needed.
4471 orig_phis.append_if_missing(mphi);
4472 } else if (C->get_alias_index(t) != alias_idx) {
4473 // Create a new Phi with the specified alias index type.
4474 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4475 }
4476 }
4477 // the result is either MemNode, PhiNode, InitializeNode.
4478 return result;
4479 }
4480
4481 //
4482 // Convert the types of non-escaped object to instance types where possible,
4483 // propagate the new type information through the graph, and update memory
4484 // edges and MergeMem inputs to reflect the new type.
4485 //
4486 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4487 // The processing is done in 4 phases:
4488 //
4489 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4490 // types for the CheckCastPP for allocations where possible.
4491 // Propagate the new types through users as follows:
4492 // casts and Phi: push users on alloc_worklist
4493 // AddP: cast Base and Address inputs to the instance type
4494 // push any AddP users on alloc_worklist and push any memnode
4495 // users onto memnode_worklist.
4496 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4497 // search the Memory chain for a store with the appropriate type
4498 // address type. If a Phi is found, create a new version with
4499 // the appropriate memory slices from each of the Phi inputs.
4500 // For stores, process the users as follows:
4501 // MemNode: push on memnode_worklist
4502 // MergeMem: push on mergemem_worklist
4503 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4504 // moving the first node encountered of each instance type to the
4505 // the input corresponding to its alias index.
4506 // appropriate memory slice.
4507 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4508 //
4509 // In the following example, the CheckCastPP nodes are the cast of allocation
4510 // results and the allocation of node 29 is non-escaped and eligible to be an
4511 // instance type.
4512 //
4513 // We start with:
4514 //
4515 // 7 Parm #memory
4516 // 10 ConI "12"
4517 // 19 CheckCastPP "Foo"
4518 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4519 // 29 CheckCastPP "Foo"
4520 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4521 //
4522 // 40 StoreP 25 7 20 ... alias_index=4
4523 // 50 StoreP 35 40 30 ... alias_index=4
4524 // 60 StoreP 45 50 20 ... alias_index=4
4525 // 70 LoadP _ 60 30 ... alias_index=4
4526 // 80 Phi 75 50 60 Memory alias_index=4
4527 // 90 LoadP _ 80 30 ... alias_index=4
4528 // 100 LoadP _ 80 20 ... alias_index=4
4529 //
4530 //
4531 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4532 // and creating a new alias index for node 30. This gives:
4533 //
4534 // 7 Parm #memory
4535 // 10 ConI "12"
4536 // 19 CheckCastPP "Foo"
4537 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4538 // 29 CheckCastPP "Foo" iid=24
4539 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4540 //
4541 // 40 StoreP 25 7 20 ... alias_index=4
4542 // 50 StoreP 35 40 30 ... alias_index=6
4543 // 60 StoreP 45 50 20 ... alias_index=4
4544 // 70 LoadP _ 60 30 ... alias_index=6
4545 // 80 Phi 75 50 60 Memory alias_index=4
4546 // 90 LoadP _ 80 30 ... alias_index=6
4547 // 100 LoadP _ 80 20 ... alias_index=4
4548 //
4549 // In phase 2, new memory inputs are computed for the loads and stores,
4550 // And a new version of the phi is created. In phase 4, the inputs to
4551 // node 80 are updated and then the memory nodes are updated with the
4552 // values computed in phase 2. This results in:
4553 //
4554 // 7 Parm #memory
4555 // 10 ConI "12"
4556 // 19 CheckCastPP "Foo"
4557 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4558 // 29 CheckCastPP "Foo" iid=24
4559 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4560 //
4561 // 40 StoreP 25 7 20 ... alias_index=4
4562 // 50 StoreP 35 7 30 ... alias_index=6
4563 // 60 StoreP 45 40 20 ... alias_index=4
4564 // 70 LoadP _ 50 30 ... alias_index=6
4565 // 80 Phi 75 40 60 Memory alias_index=4
4566 // 120 Phi 75 50 50 Memory alias_index=6
4567 // 90 LoadP _ 120 30 ... alias_index=6
4568 // 100 LoadP _ 80 20 ... alias_index=4
4569 //
4570 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4571 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4572 GrowableArray<MergeMemNode*> &mergemem_worklist,
4573 Unique_Node_List &reducible_merges) {
4574 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4575 GrowableArray<Node *> memnode_worklist;
4576 GrowableArray<PhiNode *> orig_phis;
4577 PhaseIterGVN *igvn = _igvn;
4578 uint new_index_start = (uint) _compile->num_alias_types();
4579 VectorSet visited;
4580 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4581 uint unique_old = _compile->unique();
4582
4583 // Phase 1: Process possible allocations from alloc_worklist.
4584 // Create instance types for the CheckCastPP for allocations where possible.
4585 //
4586 // (Note: don't forget to change the order of the second AddP node on
4587 // the alloc_worklist if the order of the worklist processing is changed,
4588 // see the comment in find_second_addp().)
4589 //
4590 while (alloc_worklist.length() != 0) {
4591 Node *n = alloc_worklist.pop();
4592 uint ni = n->_idx;
4593 if (n->is_Call()) {
4594 CallNode *alloc = n->as_Call();
4595 // copy escape information to call node
4596 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4597 PointsToNode::EscapeState es = ptn->escape_state();
4598 // We have an allocation or call which returns a Java object,
4599 // see if it is non-escaped.
4600 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4601 continue;
4602 }
4603 // Find CheckCastPP for the allocate or for the return value of a call
4604 n = alloc->result_cast();
4605 if (n == nullptr) { // No uses except Initialize node
4606 if (alloc->is_Allocate()) {
4607 // Set the scalar_replaceable flag for allocation
4608 // so it could be eliminated if it has no uses.
4609 alloc->as_Allocate()->_is_scalar_replaceable = true;
4610 }
4611 continue;
4612 }
4613 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4614 // we could reach here for allocate case if one init is associated with many allocs.
4615 if (alloc->is_Allocate()) {
4616 alloc->as_Allocate()->_is_scalar_replaceable = false;
4617 }
4618 continue;
4619 }
4620
4621 // The inline code for Object.clone() casts the allocation result to
4622 // java.lang.Object and then to the actual type of the allocated
4623 // object. Detect this case and use the second cast.
4624 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4625 // the allocation result is cast to java.lang.Object and then
4626 // to the actual Array type.
4627 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4628 && (alloc->is_AllocateArray() ||
4629 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4630 Node *cast2 = nullptr;
4631 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4632 Node *use = n->fast_out(i);
4633 if (use->is_CheckCastPP()) {
4634 cast2 = use;
4635 break;
4636 }
4637 }
4638 if (cast2 != nullptr) {
4639 n = cast2;
4640 } else {
4641 // Non-scalar replaceable if the allocation type is unknown statically
4642 // (reflection allocation), the object can't be restored during
4643 // deoptimization without precise type.
4644 continue;
4645 }
4646 }
4647
4648 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4649 if (t == nullptr) {
4650 continue; // not a TypeOopPtr
4651 }
4652 if (!t->klass_is_exact()) {
4653 continue; // not an unique type
4654 }
4655 if (alloc->is_Allocate()) {
4656 // Set the scalar_replaceable flag for allocation
4657 // so it could be eliminated.
4658 alloc->as_Allocate()->_is_scalar_replaceable = true;
4659 }
4660 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4661 // in order for an object to be scalar-replaceable, it must be:
4662 // - a direct allocation (not a call returning an object)
4663 // - non-escaping
4664 // - eligible to be a unique type
4665 // - not determined to be ineligible by escape analysis
4666 set_map(alloc, n);
4667 set_map(n, alloc);
4668 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4669 igvn->hash_delete(n);
4670 igvn->set_type(n, tinst);
4671 n->raise_bottom_type(tinst);
4672 igvn->hash_insert(n);
4673 record_for_optimizer(n);
4674 // Allocate an alias index for the header fields. Accesses to
4675 // the header emitted during macro expansion wouldn't have
4676 // correct memory state otherwise.
4677 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4678 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4679 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4680 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4681 InitializeNode* init = alloc->as_Allocate()->initialization();
4682 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4683 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4684 const TypePtr* adr_type = proj->adr_type();
4685 const TypePtr* new_adr_type = tinst->with_offset(adr_type->offset());
4686 if (adr_type->isa_aryptr()) {
4687 // In the case of a flat inline type array, each field has its own slice so we need a
4688 // NarrowMemProj for each field of the flat array elements
4689 new_adr_type = new_adr_type->is_aryptr()->with_field_offset(adr_type->is_aryptr()->field_offset().get());
4690 }
4691 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4692 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4693 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4694 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4695 igvn->set_type(new_proj, new_proj->bottom_type());
4696 record_for_optimizer(new_proj);
4697 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4698 }
4699 };
4700 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4701
4702 // First, put on the worklist all Field edges from Connection Graph
4703 // which is more accurate than putting immediate users from Ideal Graph.
4704 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4705 PointsToNode* tgt = e.get();
4706 if (tgt->is_Arraycopy()) {
4707 continue;
4708 }
4709 Node* use = tgt->ideal_node();
4710 assert(tgt->is_Field() && use->is_AddP(),
4711 "only AddP nodes are Field edges in CG");
4712 if (use->outcnt() > 0) { // Don't process dead nodes
4713 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4714 if (addp2 != nullptr) {
4715 assert(alloc->is_AllocateArray(),"array allocation was expected");
4716 alloc_worklist.append_if_missing(addp2);
4717 }
4718 alloc_worklist.append_if_missing(use);
4719 }
4720 }
4721
4722 // An allocation may have an Initialize which has raw stores. Scan
4723 // the users of the raw allocation result and push AddP users
4724 // on alloc_worklist.
4725 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4726 assert (raw_result != nullptr, "must have an allocation result");
4727 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4728 Node *use = raw_result->fast_out(i);
4729 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4730 Node* addp2 = find_second_addp(use, raw_result);
4731 if (addp2 != nullptr) {
4732 assert(alloc->is_AllocateArray(),"array allocation was expected");
4733 alloc_worklist.append_if_missing(addp2);
4734 }
4735 alloc_worklist.append_if_missing(use);
4736 } else if (use->is_MemBar()) {
4737 memnode_worklist.append_if_missing(use);
4738 }
4739 }
4740 }
4741 } else if (n->is_AddP()) {
4742 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4743 // This AddP will go away when we reduce the Phi
4744 continue;
4745 }
4746 Node* addp_base = get_addp_base(n);
4747 JavaObjectNode* jobj = unique_java_object(addp_base);
4748 if (jobj == nullptr || jobj == phantom_obj) {
4749 #ifdef ASSERT
4750 ptnode_adr(get_addp_base(n)->_idx)->dump();
4751 ptnode_adr(n->_idx)->dump();
4752 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4753 #endif
4754 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4755 return;
4756 }
4757 Node *base = get_map(jobj->idx()); // CheckCastPP node
4758 if (!split_AddP(n, base)) continue; // wrong type from dead path
4759 } else if (n->is_Phi() ||
4760 n->is_CheckCastPP() ||
4761 n->is_EncodeP() ||
4762 n->is_DecodeN() ||
4763 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4764 if (visited.test_set(n->_idx)) {
4765 assert(n->is_Phi(), "loops only through Phi's");
4766 continue; // already processed
4767 }
4768 // Reducible Phi's will be removed from the graph after split_unique_types
4769 // finishes. For now we just try to split out the SR inputs of the merge.
4770 Node* parent = n->in(1);
4771 if (reducible_merges.member(n)) {
4772 reduce_phi(n->as_Phi(), alloc_worklist);
4773 #ifdef ASSERT
4774 if (VerifyReduceAllocationMerges) {
4775 reduced_merges.push(n);
4776 }
4777 #endif
4778 continue;
4779 } else if (reducible_merges.member(parent)) {
4780 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4781 // part of reduce_merge.
4782 continue;
4783 }
4784 JavaObjectNode* jobj = unique_java_object(n);
4785 if (jobj == nullptr || jobj == phantom_obj) {
4786 #ifdef ASSERT
4787 ptnode_adr(n->_idx)->dump();
4788 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4789 #endif
4790 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4791 return;
4792 } else {
4793 Node *val = get_map(jobj->idx()); // CheckCastPP node
4794 TypeNode *tn = n->as_Type();
4795 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4796 assert(tinst != nullptr && tinst->is_known_instance() &&
4797 tinst->instance_id() == jobj->idx() , "instance type expected.");
4798
4799 const Type *tn_type = igvn->type(tn);
4800 const TypeOopPtr *tn_t;
4801 if (tn_type->isa_narrowoop()) {
4802 tn_t = tn_type->make_ptr()->isa_oopptr();
4803 } else {
4804 tn_t = tn_type->isa_oopptr();
4805 }
4806 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4807 if (tn_t->isa_aryptr()) {
4808 // Keep array properties (not flat/null-free)
4809 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
4810 if (tinst == nullptr) {
4811 continue; // Skip dead path with inconsistent properties
4812 }
4813 }
4814 if (tn_type->isa_narrowoop()) {
4815 tn_type = tinst->make_narrowoop();
4816 } else {
4817 tn_type = tinst;
4818 }
4819 igvn->hash_delete(tn);
4820 igvn->set_type(tn, tn_type);
4821 tn->set_type(tn_type);
4822 igvn->hash_insert(tn);
4823 record_for_optimizer(n);
4824 } else {
4825 assert(tn_type == TypePtr::NULL_PTR ||
4826 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4827 "unexpected type");
4828 continue; // Skip dead path with different type
4829 }
4830 }
4831 } else {
4832 DEBUG_ONLY(n->dump();)
4833 assert(false, "EA: unexpected node");
4834 continue;
4835 }
4836 // push allocation's users on appropriate worklist
4837 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4838 Node *use = n->fast_out(i);
4839 if (use->is_Mem() && use->in(MemNode::Address) == n) {
4840 // Load/store to instance's field
4841 memnode_worklist.append_if_missing(use);
4842 } else if (use->is_MemBar()) {
4843 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4844 memnode_worklist.append_if_missing(use);
4845 }
4846 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4847 Node* addp2 = find_second_addp(use, n);
4848 if (addp2 != nullptr) {
4849 alloc_worklist.append_if_missing(addp2);
4850 }
4851 alloc_worklist.append_if_missing(use);
4852 } else if (use->is_Phi() ||
4853 use->is_CheckCastPP() ||
4854 use->is_EncodeNarrowPtr() ||
4855 use->is_DecodeNarrowPtr() ||
4856 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4857 alloc_worklist.append_if_missing(use);
4858 #ifdef ASSERT
4859 } else if (use->is_Mem()) {
4860 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4861 } else if (use->is_MergeMem()) {
4862 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4863 } else if (use->is_SafePoint()) {
4864 // Look for MergeMem nodes for calls which reference unique allocation
4865 // (through CheckCastPP nodes) even for debug info.
4866 Node* m = use->in(TypeFunc::Memory);
4867 if (m->is_MergeMem()) {
4868 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4869 }
4870 } else if (use->Opcode() == Op_EncodeISOArray) {
4871 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4872 // EncodeISOArray overwrites destination array
4873 memnode_worklist.append_if_missing(use);
4874 }
4875 } else if (use->Opcode() == Op_Return) {
4876 // Allocation is referenced by field of returned inline type
4877 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
4878 } else {
4879 uint op = use->Opcode();
4880 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4881 (use->in(MemNode::Memory) == n)) {
4882 // They overwrite memory edge corresponding to destination array,
4883 memnode_worklist.append_if_missing(use);
4884 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4885 op == Op_CastP2X ||
4886 op == Op_FastLock || op == Op_AryEq ||
4887 op == Op_StrComp || op == Op_CountPositives ||
4888 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4889 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4890 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4891 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
4892 op == Op_ReinterpretS2HF ||
4893 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4894 n->dump();
4895 use->dump();
4896 assert(false, "EA: missing allocation reference path");
4897 }
4898 #endif
4899 }
4900 }
4901
4902 }
4903
4904 #ifdef ASSERT
4905 if (VerifyReduceAllocationMerges) {
4906 for (uint i = 0; i < reducible_merges.size(); i++) {
4907 Node* phi = reducible_merges.at(i);
4908
4909 if (!reduced_merges.member(phi)) {
4910 phi->dump(2);
4911 phi->dump(-2);
4912 assert(false, "This reducible merge wasn't reduced.");
4913 }
4914
4915 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4916 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4917 Node* use = phi->fast_out(j);
4918 if (!use->is_SafePoint() && !use->is_CastPP()) {
4919 phi->dump(2);
4920 phi->dump(-2);
4921 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4922 }
4923 }
4924 }
4925 }
4926 #endif
4927
4928 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4929 // type, record it in the ArrayCopy node so we know what memory this
4930 // node uses/modified.
4931 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4932 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4933 Node* dest = ac->in(ArrayCopyNode::Dest);
4934 if (dest->is_AddP()) {
4935 dest = get_addp_base(dest);
4936 }
4937 JavaObjectNode* jobj = unique_java_object(dest);
4938 if (jobj != nullptr) {
4939 Node *base = get_map(jobj->idx());
4940 if (base != nullptr) {
4941 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4942 ac->_dest_type = base_t;
4943 }
4944 }
4945 Node* src = ac->in(ArrayCopyNode::Src);
4946 if (src->is_AddP()) {
4947 src = get_addp_base(src);
4948 }
4949 jobj = unique_java_object(src);
4950 if (jobj != nullptr) {
4951 Node* base = get_map(jobj->idx());
4952 if (base != nullptr) {
4953 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4954 ac->_src_type = base_t;
4955 }
4956 }
4957 }
4958
4959 // New alias types were created in split_AddP().
4960 uint new_index_end = (uint) _compile->num_alias_types();
4961
4962 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
4963
4964 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4965 // compute new values for Memory inputs (the Memory inputs are not
4966 // actually updated until phase 4.)
4967 if (memnode_worklist.length() == 0)
4968 return; // nothing to do
4969 while (memnode_worklist.length() != 0) {
4970 Node *n = memnode_worklist.pop();
4971 if (visited.test_set(n->_idx)) {
4972 continue;
4973 }
4974 if (n->is_Phi() || n->is_ClearArray()) {
4975 // we don't need to do anything, but the users must be pushed
4976 } else if (n->is_MemBar()) { // MemBar nodes
4977 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
4978 // we don't need to do anything, but the users must be pushed
4979 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4980 if (n == nullptr) {
4981 continue;
4982 }
4983 }
4984 } else if (n->is_CallLeaf()) {
4985 // Runtime calls with narrow memory input (no MergeMem node)
4986 // get the memory projection
4987 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4988 if (n == nullptr) {
4989 continue;
4990 }
4991 } else if (n->Opcode() == Op_StrInflatedCopy) {
4992 // Check direct uses of StrInflatedCopy.
4993 // It is memory type Node - no special SCMemProj node.
4994 } else if (n->Opcode() == Op_StrCompressedCopy ||
4995 n->Opcode() == Op_EncodeISOArray) {
4996 // get the memory projection
4997 n = n->find_out_with(Op_SCMemProj);
4998 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4999 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
5000 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5001 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
5002 } else if (n->is_Proj()) {
5003 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
5004 } else {
5005 #ifdef ASSERT
5006 if (!n->is_Mem()) {
5007 n->dump();
5008 }
5009 assert(n->is_Mem(), "memory node required.");
5010 #endif
5011 Node *addr = n->in(MemNode::Address);
5012 const Type *addr_t = igvn->type(addr);
5013 if (addr_t == Type::TOP) {
5014 continue;
5015 }
5016 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
5017 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
5018 assert ((uint)alias_idx < new_index_end, "wrong alias index");
5019 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
5020 if (_compile->failing()) {
5021 return;
5022 }
5023 if (mem != n->in(MemNode::Memory)) {
5024 // We delay the memory edge update since we need old one in
5025 // MergeMem code below when instances memory slices are separated.
5026 set_map(n, mem);
5027 }
5028 if (n->is_Load()) {
5029 continue; // don't push users
5030 } else if (n->is_LoadStore()) {
5031 // get the memory projection
5032 n = n->find_out_with(Op_SCMemProj);
5033 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5034 }
5035 }
5036 // push user on appropriate worklist
5037 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5038 Node *use = n->fast_out(i);
5039 if (use->is_Phi() || use->is_ClearArray()) {
5040 memnode_worklist.append_if_missing(use);
5041 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
5042 memnode_worklist.append_if_missing(use);
5043 } else if (use->is_MemBar() || use->is_CallLeaf()) {
5044 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5045 memnode_worklist.append_if_missing(use);
5046 }
5047 } else if (use->is_Proj()) {
5048 assert(n->is_Initialize(), "We only push projections of Initialize");
5049 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
5050 memnode_worklist.append_if_missing(use);
5051 }
5052 #ifdef ASSERT
5053 } else if (use->is_Mem()) {
5054 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
5055 } else if (use->is_MergeMem()) {
5056 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5057 } else if (use->Opcode() == Op_EncodeISOArray) {
5058 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5059 // EncodeISOArray overwrites destination array
5060 memnode_worklist.append_if_missing(use);
5061 }
5062 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
5063 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5064 // store_unknown_inline overwrites destination array
5065 memnode_worklist.append_if_missing(use);
5066 } else {
5067 uint op = use->Opcode();
5068 if ((use->in(MemNode::Memory) == n) &&
5069 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
5070 // They overwrite memory edge corresponding to destination array,
5071 memnode_worklist.append_if_missing(use);
5072 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
5073 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
5074 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
5075 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
5076 n->dump();
5077 use->dump();
5078 assert(false, "EA: missing memory path");
5079 }
5080 #endif
5081 }
5082 }
5083 }
5084
5085 // Phase 3: Process MergeMem nodes from mergemem_worklist.
5086 // Walk each memory slice moving the first node encountered of each
5087 // instance type to the input corresponding to its alias index.
5088 uint length = mergemem_worklist.length();
5089 for( uint next = 0; next < length; ++next ) {
5090 MergeMemNode* nmm = mergemem_worklist.at(next);
5091 assert(!visited.test_set(nmm->_idx), "should not be visited before");
5092 // Note: we don't want to use MergeMemStream here because we only want to
5093 // scan inputs which exist at the start, not ones we add during processing.
5094 // Note 2: MergeMem may already contains instance memory slices added
5095 // during find_inst_mem() call when memory nodes were processed above.
5096 igvn->hash_delete(nmm);
5097 uint nslices = MIN2(nmm->req(), new_index_start);
5098 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
5099 Node* mem = nmm->in(i);
5100 Node* cur = nullptr;
5101 if (mem == nullptr || mem->is_top()) {
5102 continue;
5103 }
5104 // First, update mergemem by moving memory nodes to corresponding slices
5105 // if their type became more precise since this mergemem was created.
5106 while (mem->is_Mem()) {
5107 const Type* at = igvn->type(mem->in(MemNode::Address));
5108 if (at != Type::TOP) {
5109 assert (at->isa_ptr() != nullptr, "pointer type required.");
5110 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
5111 if (idx == i) {
5112 if (cur == nullptr) {
5113 cur = mem;
5114 }
5115 } else {
5116 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
5117 nmm->set_memory_at(idx, mem);
5118 }
5119 }
5120 }
5121 mem = mem->in(MemNode::Memory);
5122 }
5123 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
5124 // Find any instance of the current type if we haven't encountered
5125 // already a memory slice of the instance along the memory chain.
5126 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5127 if((uint)_compile->get_general_index(ni) == i) {
5128 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
5129 if (nmm->is_empty_memory(m)) {
5130 Node* result = find_inst_mem(mem, ni, orig_phis);
5131 if (_compile->failing()) {
5132 return;
5133 }
5134 nmm->set_memory_at(ni, result);
5135 }
5136 }
5137 }
5138 }
5139 // Find the rest of instances values
5140 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5141 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
5142 Node* result = step_through_mergemem(nmm, ni, tinst);
5143 if (result == nmm->base_memory()) {
5144 // Didn't find instance memory, search through general slice recursively.
5145 result = nmm->memory_at(_compile->get_general_index(ni));
5146 result = find_inst_mem(result, ni, orig_phis);
5147 if (_compile->failing()) {
5148 return;
5149 }
5150 nmm->set_memory_at(ni, result);
5151 }
5152 }
5153
5154 // If we have crossed the 3/4 point of max node limit it's too risky
5155 // to continue with EA/SR because we might hit the max node limit.
5156 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
5157 if (_compile->do_reduce_allocation_merges()) {
5158 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
5159 } else if (_invocation > 0) {
5160 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
5161 } else {
5162 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
5163 }
5164 return;
5165 }
5166
5167 igvn->hash_insert(nmm);
5168 record_for_optimizer(nmm);
5169 }
5170
5171 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5172
5173 // Phase 4: Update the inputs of non-instance memory Phis and
5174 // the Memory input of memnodes
5175 // First update the inputs of any non-instance Phi's from
5176 // which we split out an instance Phi. Note we don't have
5177 // to recursively process Phi's encountered on the input memory
5178 // chains as is done in split_memory_phi() since they will
5179 // also be processed here.
5180 for (int j = 0; j < orig_phis.length(); j++) {
5181 PhiNode *phi = orig_phis.at(j);
5182 int alias_idx = _compile->get_alias_index(phi->adr_type());
5183 igvn->hash_delete(phi);
5184 for (uint i = 1; i < phi->req(); i++) {
5185 Node *mem = phi->in(i);
5186 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5187 if (_compile->failing()) {
5188 return;
5189 }
5190 if (mem != new_mem) {
5191 phi->set_req(i, new_mem);
5192 }
5193 }
5194 igvn->hash_insert(phi);
5195 record_for_optimizer(phi);
5196 }
5197
5198 // Update the memory inputs of MemNodes with the value we computed
5199 // in Phase 2 and move stores memory users to corresponding memory slices.
5200 // Disable memory split verification code until the fix for 6984348.
5201 // Currently it produces false negative results since it does not cover all cases.
5202 #if 0 // ifdef ASSERT
5203 visited.Reset();
5204 Node_Stack old_mems(arena, _compile->unique() >> 2);
5205 #endif
5206 for (uint i = 0; i < ideal_nodes.size(); i++) {
5207 Node* n = ideal_nodes.at(i);
5208 Node* nmem = get_map(n->_idx);
5209 assert(nmem != nullptr, "sanity");
5210 if (n->is_Mem()) {
5211 #if 0 // ifdef ASSERT
5212 Node* old_mem = n->in(MemNode::Memory);
5213 if (!visited.test_set(old_mem->_idx)) {
5214 old_mems.push(old_mem, old_mem->outcnt());
5215 }
5216 #endif
5217 assert(n->in(MemNode::Memory) != nmem, "sanity");
5218 if (!n->is_Load()) {
5219 // Move memory users of a store first.
5220 move_inst_mem(n, orig_phis);
5221 }
5222 // Now update memory input
5223 igvn->hash_delete(n);
5224 n->set_req(MemNode::Memory, nmem);
5225 igvn->hash_insert(n);
5226 record_for_optimizer(n);
5227 } else {
5228 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5229 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5230 }
5231 }
5232 #if 0 // ifdef ASSERT
5233 // Verify that memory was split correctly
5234 while (old_mems.is_nonempty()) {
5235 Node* old_mem = old_mems.node();
5236 uint old_cnt = old_mems.index();
5237 old_mems.pop();
5238 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5239 }
5240 #endif
5241 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5242 }
5243
5244 #ifndef PRODUCT
5245 int ConnectionGraph::_no_escape_counter = 0;
5246 int ConnectionGraph::_arg_escape_counter = 0;
5247 int ConnectionGraph::_global_escape_counter = 0;
5248
5249 static const char *node_type_names[] = {
5250 "UnknownType",
5251 "JavaObject",
5252 "LocalVar",
5253 "Field",
5254 "Arraycopy"
5255 };
5256
5257 static const char *esc_names[] = {
5258 "UnknownEscape",
5259 "NoEscape",
5260 "ArgEscape",
5261 "GlobalEscape"
5262 };
5263
5264 const char* PointsToNode::esc_name() const {
5265 return esc_names[(int)escape_state()];
5266 }
5267
5268 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5269 NodeType nt = node_type();
5270 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5271 if (print_state) {
5272 EscapeState es = escape_state();
5273 EscapeState fields_es = fields_escape_state();
5274 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5275 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5276 out->print("NSR ");
5277 }
5278 }
5279 }
5280
5281 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5282 dump_header(print_state, out);
5283 if (is_Field()) {
5284 FieldNode* f = (FieldNode*)this;
5285 if (f->is_oop()) {
5286 out->print("oop ");
5287 }
5288 if (f->offset() > 0) {
5289 out->print("+%d ", f->offset());
5290 }
5291 out->print("(");
5292 for (BaseIterator i(f); i.has_next(); i.next()) {
5293 PointsToNode* b = i.get();
5294 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5295 }
5296 out->print(" )");
5297 }
5298 out->print("[");
5299 for (EdgeIterator i(this); i.has_next(); i.next()) {
5300 PointsToNode* e = i.get();
5301 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5302 }
5303 out->print(" [");
5304 for (UseIterator i(this); i.has_next(); i.next()) {
5305 PointsToNode* u = i.get();
5306 bool is_base = false;
5307 if (PointsToNode::is_base_use(u)) {
5308 is_base = true;
5309 u = PointsToNode::get_use_node(u)->as_Field();
5310 }
5311 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5312 }
5313 out->print(" ]] ");
5314 if (_node == nullptr) {
5315 out->print("<null>%s", newline ? "\n" : "");
5316 } else {
5317 _node->dump(newline ? "\n" : "", false, out);
5318 }
5319 }
5320
5321 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5322 bool first = true;
5323 int ptnodes_length = ptnodes_worklist.length();
5324 for (int i = 0; i < ptnodes_length; i++) {
5325 PointsToNode *ptn = ptnodes_worklist.at(i);
5326 if (ptn == nullptr || !ptn->is_JavaObject()) {
5327 continue;
5328 }
5329 PointsToNode::EscapeState es = ptn->escape_state();
5330 if ((es != PointsToNode::NoEscape) && !Verbose) {
5331 continue;
5332 }
5333 Node* n = ptn->ideal_node();
5334 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5335 n->as_CallStaticJava()->is_boxing_method())) {
5336 if (first) {
5337 tty->cr();
5338 tty->print("======== Connection graph for ");
5339 _compile->method()->print_short_name();
5340 tty->cr();
5341 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5342 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5343 tty->cr();
5344 first = false;
5345 }
5346 ptn->dump();
5347 // Print all locals and fields which reference this allocation
5348 for (UseIterator j(ptn); j.has_next(); j.next()) {
5349 PointsToNode* use = j.get();
5350 if (use->is_LocalVar()) {
5351 use->dump(Verbose);
5352 } else if (Verbose) {
5353 use->dump();
5354 }
5355 }
5356 tty->cr();
5357 }
5358 }
5359 }
5360
5361 void ConnectionGraph::print_statistics() {
5362 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));
5363 }
5364
5365 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5366 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5367 return;
5368 }
5369 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5370 JavaObjectNode* ptn = java_objects_worklist.at(next);
5371 if (ptn->ideal_node()->is_Allocate()) {
5372 if (ptn->escape_state() == PointsToNode::NoEscape) {
5373 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5374 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5375 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5376 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5377 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5378 } else {
5379 assert(false, "Unexpected Escape State");
5380 }
5381 }
5382 }
5383 }
5384
5385 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5386 if (_compile->directive()->TraceEscapeAnalysisOption) {
5387 assert(ptn != nullptr, "should not be null");
5388 assert(reason != nullptr, "should not be null");
5389 ptn->dump_header(true);
5390 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5391 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5392 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5393 }
5394 }
5395
5396 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5397 if (_compile->directive()->TraceEscapeAnalysisOption) {
5398 stringStream ss;
5399 ss.print("propagated from: ");
5400 from->dump(true, &ss, false);
5401 return ss.as_string();
5402 } else {
5403 return nullptr;
5404 }
5405 }
5406
5407 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5408 if (_compile->directive()->TraceEscapeAnalysisOption) {
5409 stringStream ss;
5410 ss.print("escapes as arg to:");
5411 call->dump("", false, &ss);
5412 return ss.as_string();
5413 } else {
5414 return nullptr;
5415 }
5416 }
5417
5418 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5419 if (_compile->directive()->TraceEscapeAnalysisOption) {
5420 stringStream ss;
5421 ss.print("is merged with other object: ");
5422 other->dump_header(true, &ss);
5423 return ss.as_string();
5424 } else {
5425 return nullptr;
5426 }
5427 }
5428
5429 #endif
5430
5431 void ConnectionGraph::record_for_optimizer(Node *n) {
5432 _igvn->_worklist.push(n);
5433 _igvn->add_users_to_worklist(n);
5434 }