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 OpaqueConstantBool 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 // Iterate over the domains for the scalarized and non scalarized calling conventions: Only move to the next element
2113 // in the non scalarized calling convention once all elements of the scalarized calling convention for that parameter
2114 // have been iterated over. So (ignoring hidden arguments such as the null marker) iterating over:
2115 // value class MyValue {
2116 // int f1;
2117 // float f2;
2118 // }
2119 // void m(Object o, MyValue v, int i)
2120 // produces the pairs:
2121 // (Object, Object), (Myvalue, int), (MyValue, float), (int, int)
2122 class DomainIterator : public StackObj {
2123 private:
2124 const TypeTuple* _domain;
2125 const TypeTuple* _domain_cc;
2126 const GrowableArray<SigEntry>* _sig_cc;
2127
2128 uint _i_domain;
2129 uint _i_domain_cc;
2130 int _i_sig_cc;
2131 uint _depth;
2132
2133 void next_helper() {
2134 if (_sig_cc == nullptr) {
2135 return;
2136 }
2137 BasicType prev_bt = _i_sig_cc > 0 ? _sig_cc->at(_i_sig_cc-1)._bt : T_ILLEGAL;
2138 while (_i_sig_cc < _sig_cc->length()) {
2139 BasicType bt = _sig_cc->at(_i_sig_cc)._bt;
2140 assert(bt != T_VOID || _sig_cc->at(_i_sig_cc-1)._bt == prev_bt, "");
2141 if (bt == T_METADATA) {
2142 _depth++;
2143 } else if (bt == T_VOID && (prev_bt != T_LONG && prev_bt != T_DOUBLE)) {
2144 _depth--;
2145 if (_depth == 0) {
2146 _i_domain++;
2147 }
2148 } else {
2149 return;
2150 }
2151 prev_bt = bt;
2152 _i_sig_cc++;
2153 }
2154 }
2155
2156 public:
2157
2158 DomainIterator(CallJavaNode* call) :
2159 _domain(call->tf()->domain_sig()),
2160 _domain_cc(call->tf()->domain_cc()),
2161 _sig_cc(call->method()->get_sig_cc()),
2162 _i_domain(TypeFunc::Parms),
2163 _i_domain_cc(TypeFunc::Parms),
2164 _i_sig_cc(0),
2165 _depth(0) {
2166 next_helper();
2167 }
2168
2169 bool has_next() const {
2170 assert(_sig_cc == nullptr || (_i_sig_cc < _sig_cc->length()) == (_i_domain < _domain->cnt()), "should reach end in sync");
2171 assert((_i_domain < _domain->cnt()) == (_i_domain_cc < _domain_cc->cnt()), "should reach end in sync");
2172 return _i_domain < _domain->cnt();
2173 }
2174
2175 void next() {
2176 assert(_depth != 0 || _domain->field_at(_i_domain) == _domain_cc->field_at(_i_domain_cc), "should produce same non scalarized elements");
2177 _i_sig_cc++;
2178 if (_depth == 0) {
2179 _i_domain++;
2180 }
2181 _i_domain_cc++;
2182 next_helper();
2183 }
2184
2185 uint i_domain() const {
2186 return _i_domain;
2187 }
2188
2189 uint i_domain_cc() const {
2190 return _i_domain_cc;
2191 }
2192
2193 const Type* current_domain() const {
2194 return _domain->field_at(_i_domain);
2195 }
2196
2197 const Type* current_domain_cc() const {
2198 return _domain_cc->field_at(_i_domain_cc);
2199 }
2200 };
2201
2202 void ConnectionGraph::add_call_node(CallNode* call) {
2203 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2204 uint call_idx = call->_idx;
2205 if (call->is_Allocate()) {
2206 Node* k = call->in(AllocateNode::KlassNode);
2207 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2208 assert(kt != nullptr, "TypeKlassPtr required.");
2209 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2210 bool scalar_replaceable = true;
2211 NOT_PRODUCT(const char* nsr_reason = "");
2212 if (call->is_AllocateArray()) {
2213 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2214 es = PointsToNode::GlobalEscape;
2215 } else {
2216 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2217 if (length < 0) {
2218 // Not scalar replaceable if the length is not constant.
2219 scalar_replaceable = false;
2220 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2221 } else if (length > EliminateAllocationArraySizeLimit) {
2222 // Not scalar replaceable if the length is too big.
2223 scalar_replaceable = false;
2224 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2225 }
2226 }
2227 } else { // Allocate instance
2228 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2229 es = PointsToNode::GlobalEscape;
2230 } else {
2231 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2232 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2233 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2234 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2235 !ik->can_be_instantiated() ||
2236 ik->has_finalizer()) {
2237 es = PointsToNode::GlobalEscape;
2238 } else {
2239 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2240 if (nfields > EliminateAllocationFieldsLimit) {
2241 // Not scalar replaceable if there are too many fields.
2242 scalar_replaceable = false;
2243 NOT_PRODUCT(nsr_reason = "has too many fields");
2244 }
2245 }
2246 }
2247 }
2248 add_java_object(call, es);
2249 PointsToNode* ptn = ptnode_adr(call_idx);
2250 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2251 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2252 }
2253 } else if (call->is_CallStaticJava()) {
2254 // Call nodes could be different types:
2255 //
2256 // 1. CallDynamicJavaNode (what happened during call is unknown):
2257 //
2258 // - mapped to GlobalEscape JavaObject node if oop is returned;
2259 //
2260 // - all oop arguments are escaping globally;
2261 //
2262 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2263 //
2264 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2265 //
2266 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2267 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2268 // during call is returned;
2269 // - mapped to ArgEscape LocalVar node pointed to object arguments
2270 // which are returned and does not escape during call;
2271 //
2272 // - oop arguments escaping status is defined by bytecode analysis;
2273 //
2274 // For a static call, we know exactly what method is being called.
2275 // Use bytecode estimator to record whether the call's return value escapes.
2276 ciMethod* meth = call->as_CallJava()->method();
2277 if (meth == nullptr) {
2278 const char* name = call->as_CallStaticJava()->_name;
2279 assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
2280 strncmp(name, "load_unknown_inline", 19) == 0 ||
2281 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "TODO: add failed case check");
2282 // Returns a newly allocated non-escaped object.
2283 add_java_object(call, PointsToNode::NoEscape);
2284 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2285 } else if (meth->is_boxing_method()) {
2286 // Returns boxing object
2287 PointsToNode::EscapeState es;
2288 vmIntrinsics::ID intr = meth->intrinsic_id();
2289 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2290 // It does not escape if object is always allocated.
2291 es = PointsToNode::NoEscape;
2292 } else {
2293 // It escapes globally if object could be loaded from cache.
2294 es = PointsToNode::GlobalEscape;
2295 }
2296 add_java_object(call, es);
2297 if (es == PointsToNode::GlobalEscape) {
2298 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2299 }
2300 } else {
2301 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2302 call_analyzer->copy_dependencies(_compile->dependencies());
2303 if (call_analyzer->is_return_allocated()) {
2304 // Returns a newly allocated non-escaped object, simply
2305 // update dependency information.
2306 // Mark it as NoEscape so that objects referenced by
2307 // it's fields will be marked as NoEscape at least.
2308 add_java_object(call, PointsToNode::NoEscape);
2309 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2310 } else {
2311 bool ret_arg = false;
2312 // Determine whether any arguments are returned.
2313 for (DomainIterator di(call->as_CallJava()); di.has_next(); di.next()) {
2314 uint arg = di.i_domain() - TypeFunc::Parms;
2315 if (di.current_domain_cc()->isa_ptr() != nullptr &&
2316 call_analyzer->is_arg_returned(arg) &&
2317 !meth->is_scalarized_arg(arg)) {
2318 ret_arg = true;
2319 break;
2320 }
2321 }
2322 if (ret_arg) {
2323 add_local_var(call, PointsToNode::ArgEscape);
2324 } else {
2325 // Returns unknown object.
2326 map_ideal_node(call, phantom_obj);
2327 }
2328 }
2329 }
2330 } else {
2331 // An other type of call, assume the worst case:
2332 // returned value is unknown and globally escapes.
2333 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2334 map_ideal_node(call, phantom_obj);
2335 }
2336 }
2337
2338 void ConnectionGraph::process_call_arguments(CallNode *call) {
2339 bool is_arraycopy = false;
2340 switch (call->Opcode()) {
2341 #ifdef ASSERT
2342 case Op_Allocate:
2343 case Op_AllocateArray:
2344 case Op_Lock:
2345 case Op_Unlock:
2346 assert(false, "should be done already");
2347 break;
2348 #endif
2349 case Op_ArrayCopy:
2350 case Op_CallLeafNoFP:
2351 // Most array copies are ArrayCopy nodes at this point but there
2352 // are still a few direct calls to the copy subroutines (See
2353 // PhaseStringOpts::copy_string())
2354 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2355 call->as_CallLeaf()->is_call_to_arraycopystub();
2356 // fall through
2357 case Op_CallLeafVector:
2358 case Op_CallLeaf: {
2359 // Stub calls, objects do not escape but they are not scale replaceable.
2360 // Adjust escape state for outgoing arguments.
2361 const TypeTuple * d = call->tf()->domain_sig();
2362 bool src_has_oops = false;
2363 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2364 const Type* at = d->field_at(i);
2365 Node *arg = call->in(i);
2366 if (arg == nullptr) {
2367 continue;
2368 }
2369 const Type *aat = _igvn->type(arg);
2370 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2371 continue;
2372 }
2373 if (arg->is_AddP()) {
2374 //
2375 // The inline_native_clone() case when the arraycopy stub is called
2376 // after the allocation before Initialize and CheckCastPP nodes.
2377 // Or normal arraycopy for object arrays case.
2378 //
2379 // Set AddP's base (Allocate) as not scalar replaceable since
2380 // pointer to the base (with offset) is passed as argument.
2381 //
2382 arg = get_addp_base(arg);
2383 }
2384 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2385 assert(arg_ptn != nullptr, "should be registered");
2386 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2387 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2388 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2389 aat->isa_ptr() != nullptr, "expecting an Ptr");
2390 bool arg_has_oops = aat->isa_oopptr() &&
2391 (aat->isa_instptr() ||
2392 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2393 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2394 aat->isa_aryptr()->is_flat() &&
2395 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2396 if (i == TypeFunc::Parms) {
2397 src_has_oops = arg_has_oops;
2398 }
2399 //
2400 // src or dst could be j.l.Object when other is basic type array:
2401 //
2402 // arraycopy(char[],0,Object*,0,size);
2403 // arraycopy(Object*,0,char[],0,size);
2404 //
2405 // Don't add edges in such cases.
2406 //
2407 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2408 arg_has_oops && (i > TypeFunc::Parms);
2409 #ifdef ASSERT
2410 if (!(is_arraycopy ||
2411 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2412 (call->as_CallLeaf()->_name != nullptr &&
2413 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2414 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2415 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2416 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2417 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2418 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2419 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2420 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2421 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2422 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2423 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2424 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2425 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2426 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2427 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2428 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2429 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2430 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2431 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2432 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2433 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2434 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2435 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2436 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2437 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2438 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2439 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2440 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2441 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2442 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2443 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2444 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2445 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2446 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2447 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2448 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2449 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2450 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2451 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2452 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2453 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2454 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2455 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2456 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2457 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2458 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2459 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2460 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2461 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2462 strcmp(call->as_CallLeaf()->_name, "store_inline_type_fields_to_buf") == 0 ||
2463 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2464 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2465 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2466 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2467 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2468 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2469 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2470 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2471 ))) {
2472 call->dump();
2473 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2474 }
2475 #endif
2476 // Always process arraycopy's destination object since
2477 // we need to add all possible edges to references in
2478 // source object.
2479 if (arg_esc >= PointsToNode::ArgEscape &&
2480 !arg_is_arraycopy_dest) {
2481 continue;
2482 }
2483 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2484 if (call->is_ArrayCopy()) {
2485 ArrayCopyNode* ac = call->as_ArrayCopy();
2486 if (ac->is_clonebasic() ||
2487 ac->is_arraycopy_validated() ||
2488 ac->is_copyof_validated() ||
2489 ac->is_copyofrange_validated()) {
2490 es = PointsToNode::NoEscape;
2491 }
2492 }
2493 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2494 if (arg_is_arraycopy_dest) {
2495 Node* src = call->in(TypeFunc::Parms);
2496 if (src->is_AddP()) {
2497 src = get_addp_base(src);
2498 }
2499 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2500 assert(src_ptn != nullptr, "should be registered");
2501 // Special arraycopy edge:
2502 // Only escape state of destination object's fields affects
2503 // escape state of fields in source object.
2504 add_arraycopy(call, es, src_ptn, arg_ptn);
2505 }
2506 }
2507 }
2508 break;
2509 }
2510 case Op_CallStaticJava: {
2511 // For a static call, we know exactly what method is being called.
2512 // Use bytecode estimator to record the call's escape affects
2513 #ifdef ASSERT
2514 const char* name = call->as_CallStaticJava()->_name;
2515 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2516 #endif
2517 ciMethod* meth = call->as_CallJava()->method();
2518 if ((meth != nullptr) && meth->is_boxing_method()) {
2519 break; // Boxing methods do not modify any oops.
2520 }
2521 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2522 // fall-through if not a Java method or no analyzer information
2523 if (call_analyzer != nullptr) {
2524 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2525 for (DomainIterator di(call->as_CallJava()); di.has_next(); di.next()) {
2526 int k = di.i_domain() - TypeFunc::Parms;
2527 const Type* at = di.current_domain_cc();
2528 Node* arg = call->in(di.i_domain_cc());
2529 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2530 if (at->isa_ptr() != nullptr &&
2531 call_analyzer->is_arg_returned(k) &&
2532 !meth->is_scalarized_arg(k)) {
2533 // The call returns arguments.
2534 if (call_ptn != nullptr) { // Is call's result used?
2535 assert(call_ptn->is_LocalVar(), "node should be registered");
2536 assert(arg_ptn != nullptr, "node should be registered");
2537 add_edge(call_ptn, arg_ptn);
2538 }
2539 }
2540 if (at->isa_oopptr() != nullptr &&
2541 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2542 if (!call_analyzer->is_arg_stack(k)) {
2543 // The argument global escapes
2544 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2545 } else {
2546 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2547 if (!call_analyzer->is_arg_local(k)) {
2548 // The argument itself doesn't escape, but any fields might
2549 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2550 }
2551 }
2552 }
2553 }
2554 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2555 // The call returns arguments.
2556 assert(call_ptn->edge_count() > 0, "sanity");
2557 if (!call_analyzer->is_return_local()) {
2558 // Returns also unknown object.
2559 add_edge(call_ptn, phantom_obj);
2560 }
2561 }
2562 break;
2563 }
2564 }
2565 default: {
2566 // Fall-through here if not a Java method or no analyzer information
2567 // or some other type of call, assume the worst case: all arguments
2568 // globally escape.
2569 const TypeTuple* d = call->tf()->domain_cc();
2570 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2571 const Type* at = d->field_at(i);
2572 if (at->isa_oopptr() != nullptr) {
2573 Node* arg = call->in(i);
2574 if (arg->is_AddP()) {
2575 arg = get_addp_base(arg);
2576 }
2577 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2578 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2579 }
2580 }
2581 }
2582 }
2583 }
2584
2585
2586 // Finish Graph construction.
2587 bool ConnectionGraph::complete_connection_graph(
2588 GrowableArray<PointsToNode*>& ptnodes_worklist,
2589 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2590 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2591 GrowableArray<FieldNode*>& oop_fields_worklist) {
2592 // Normally only 1-3 passes needed to build Connection Graph depending
2593 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2594 // Set limit to 20 to catch situation when something did go wrong and
2595 // bailout Escape Analysis.
2596 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2597 #define GRAPH_BUILD_ITER_LIMIT 20
2598
2599 // Propagate GlobalEscape and ArgEscape escape states and check that
2600 // we still have non-escaping objects. The method pushs on _worklist
2601 // Field nodes which reference phantom_object.
2602 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2603 return false; // Nothing to do.
2604 }
2605 // Now propagate references to all JavaObject nodes.
2606 int java_objects_length = java_objects_worklist.length();
2607 elapsedTimer build_time;
2608 build_time.start();
2609 elapsedTimer time;
2610 bool timeout = false;
2611 int new_edges = 1;
2612 int iterations = 0;
2613 do {
2614 while ((new_edges > 0) &&
2615 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2616 double start_time = time.seconds();
2617 time.start();
2618 new_edges = 0;
2619 // Propagate references to phantom_object for nodes pushed on _worklist
2620 // by find_non_escaped_objects() and find_field_value().
2621 new_edges += add_java_object_edges(phantom_obj, false);
2622 for (int next = 0; next < java_objects_length; ++next) {
2623 JavaObjectNode* ptn = java_objects_worklist.at(next);
2624 new_edges += add_java_object_edges(ptn, true);
2625
2626 #define SAMPLE_SIZE 4
2627 if ((next % SAMPLE_SIZE) == 0) {
2628 // Each 4 iterations calculate how much time it will take
2629 // to complete graph construction.
2630 time.stop();
2631 // Poll for requests from shutdown mechanism to quiesce compiler
2632 // because Connection graph construction may take long time.
2633 CompileBroker::maybe_block();
2634 double stop_time = time.seconds();
2635 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2636 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2637 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2638 timeout = true;
2639 break; // Timeout
2640 }
2641 start_time = stop_time;
2642 time.start();
2643 }
2644 #undef SAMPLE_SIZE
2645
2646 }
2647 if (timeout) break;
2648 if (new_edges > 0) {
2649 // Update escape states on each iteration if graph was updated.
2650 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2651 return false; // Nothing to do.
2652 }
2653 }
2654 time.stop();
2655 if (time.seconds() >= EscapeAnalysisTimeout) {
2656 timeout = true;
2657 break;
2658 }
2659 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2660 }
2661 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2662 time.start();
2663 // Find fields which have unknown value.
2664 int fields_length = oop_fields_worklist.length();
2665 for (int next = 0; next < fields_length; next++) {
2666 FieldNode* field = oop_fields_worklist.at(next);
2667 if (field->edge_count() == 0) {
2668 new_edges += find_field_value(field);
2669 // This code may added new edges to phantom_object.
2670 // Need an other cycle to propagate references to phantom_object.
2671 }
2672 }
2673 time.stop();
2674 if (time.seconds() >= EscapeAnalysisTimeout) {
2675 timeout = true;
2676 break;
2677 }
2678 } else {
2679 new_edges = 0; // Bailout
2680 }
2681 } while (new_edges > 0);
2682
2683 build_time.stop();
2684 _build_time = build_time.seconds();
2685 _build_iterations = iterations;
2686
2687 // Bailout if passed limits.
2688 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2689 Compile* C = _compile;
2690 if (C->log() != nullptr) {
2691 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2692 C->log()->text("%s", timeout ? "time" : "iterations");
2693 C->log()->end_elem(" limit'");
2694 }
2695 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2696 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2697 // Possible infinite build_connection_graph loop,
2698 // bailout (no changes to ideal graph were made).
2699 return false;
2700 }
2701
2702 #undef GRAPH_BUILD_ITER_LIMIT
2703
2704 // Find fields initialized by null for non-escaping Allocations.
2705 int non_escaped_length = non_escaped_allocs_worklist.length();
2706 for (int next = 0; next < non_escaped_length; next++) {
2707 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2708 PointsToNode::EscapeState es = ptn->escape_state();
2709 assert(es <= PointsToNode::ArgEscape, "sanity");
2710 if (es == PointsToNode::NoEscape) {
2711 if (find_init_values_null(ptn, _igvn) > 0) {
2712 // Adding references to null object does not change escape states
2713 // since it does not escape. Also no fields are added to null object.
2714 add_java_object_edges(null_obj, false);
2715 }
2716 }
2717 Node* n = ptn->ideal_node();
2718 if (n->is_Allocate()) {
2719 // The object allocated by this Allocate node will never be
2720 // seen by an other thread. Mark it so that when it is
2721 // expanded no MemBarStoreStore is added.
2722 InitializeNode* ini = n->as_Allocate()->initialization();
2723 if (ini != nullptr)
2724 ini->set_does_not_escape();
2725 }
2726 }
2727 return true; // Finished graph construction.
2728 }
2729
2730 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2731 // and check that we still have non-escaping java objects.
2732 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2733 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2734 bool print_method) {
2735 GrowableArray<PointsToNode*> escape_worklist;
2736 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2737 int ptnodes_length = ptnodes_worklist.length();
2738 for (int next = 0; next < ptnodes_length; ++next) {
2739 PointsToNode* ptn = ptnodes_worklist.at(next);
2740 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2741 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2742 escape_worklist.push(ptn);
2743 }
2744 }
2745 // Set escape states to referenced nodes (edges list).
2746 while (escape_worklist.length() > 0) {
2747 PointsToNode* ptn = escape_worklist.pop();
2748 PointsToNode::EscapeState es = ptn->escape_state();
2749 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2750 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2751 es >= PointsToNode::ArgEscape) {
2752 // GlobalEscape or ArgEscape state of field means it has unknown value.
2753 if (add_edge(ptn, phantom_obj)) {
2754 // New edge was added
2755 add_field_uses_to_worklist(ptn->as_Field());
2756 }
2757 }
2758 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2759 PointsToNode* e = i.get();
2760 if (e->is_Arraycopy()) {
2761 assert(ptn->arraycopy_dst(), "sanity");
2762 // Propagate only fields escape state through arraycopy edge.
2763 if (e->fields_escape_state() < field_es) {
2764 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2765 escape_worklist.push(e);
2766 }
2767 } else if (es >= field_es) {
2768 // fields_escape_state is also set to 'es' if it is less than 'es'.
2769 if (e->escape_state() < es) {
2770 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2771 escape_worklist.push(e);
2772 }
2773 } else {
2774 // Propagate field escape state.
2775 bool es_changed = false;
2776 if (e->fields_escape_state() < field_es) {
2777 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2778 es_changed = true;
2779 }
2780 if ((e->escape_state() < field_es) &&
2781 e->is_Field() && ptn->is_JavaObject() &&
2782 e->as_Field()->is_oop()) {
2783 // Change escape state of referenced fields.
2784 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2785 es_changed = true;
2786 } else if (e->escape_state() < es) {
2787 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2788 es_changed = true;
2789 }
2790 if (es_changed) {
2791 escape_worklist.push(e);
2792 }
2793 }
2794 if (print_method) {
2795 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2796 }
2797 }
2798 }
2799 // Remove escaped objects from non_escaped list.
2800 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2801 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2802 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2803 non_escaped_allocs_worklist.delete_at(next);
2804 }
2805 if (ptn->escape_state() == PointsToNode::NoEscape) {
2806 // Find fields in non-escaped allocations which have unknown value.
2807 find_init_values_phantom(ptn);
2808 }
2809 }
2810 return (non_escaped_allocs_worklist.length() > 0);
2811 }
2812
2813 // Add all references to JavaObject node by walking over all uses.
2814 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2815 int new_edges = 0;
2816 if (populate_worklist) {
2817 // Populate _worklist by uses of jobj's uses.
2818 for (UseIterator i(jobj); i.has_next(); i.next()) {
2819 PointsToNode* use = i.get();
2820 if (use->is_Arraycopy()) {
2821 continue;
2822 }
2823 add_uses_to_worklist(use);
2824 if (use->is_Field() && use->as_Field()->is_oop()) {
2825 // Put on worklist all field's uses (loads) and
2826 // related field nodes (same base and offset).
2827 add_field_uses_to_worklist(use->as_Field());
2828 }
2829 }
2830 }
2831 for (int l = 0; l < _worklist.length(); l++) {
2832 PointsToNode* use = _worklist.at(l);
2833 if (PointsToNode::is_base_use(use)) {
2834 // Add reference from jobj to field and from field to jobj (field's base).
2835 use = PointsToNode::get_use_node(use)->as_Field();
2836 if (add_base(use->as_Field(), jobj)) {
2837 new_edges++;
2838 }
2839 continue;
2840 }
2841 assert(!use->is_JavaObject(), "sanity");
2842 if (use->is_Arraycopy()) {
2843 if (jobj == null_obj) { // null object does not have field edges
2844 continue;
2845 }
2846 // Added edge from Arraycopy node to arraycopy's source java object
2847 if (add_edge(use, jobj)) {
2848 jobj->set_arraycopy_src();
2849 new_edges++;
2850 }
2851 // and stop here.
2852 continue;
2853 }
2854 if (!add_edge(use, jobj)) {
2855 continue; // No new edge added, there was such edge already.
2856 }
2857 new_edges++;
2858 if (use->is_LocalVar()) {
2859 add_uses_to_worklist(use);
2860 if (use->arraycopy_dst()) {
2861 for (EdgeIterator i(use); i.has_next(); i.next()) {
2862 PointsToNode* e = i.get();
2863 if (e->is_Arraycopy()) {
2864 if (jobj == null_obj) { // null object does not have field edges
2865 continue;
2866 }
2867 // Add edge from arraycopy's destination java object to Arraycopy node.
2868 if (add_edge(jobj, e)) {
2869 new_edges++;
2870 jobj->set_arraycopy_dst();
2871 }
2872 }
2873 }
2874 }
2875 } else {
2876 // Added new edge to stored in field values.
2877 // Put on worklist all field's uses (loads) and
2878 // related field nodes (same base and offset).
2879 add_field_uses_to_worklist(use->as_Field());
2880 }
2881 }
2882 _worklist.clear();
2883 _in_worklist.reset();
2884 return new_edges;
2885 }
2886
2887 // Put on worklist all related field nodes.
2888 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2889 assert(field->is_oop(), "sanity");
2890 int offset = field->offset();
2891 add_uses_to_worklist(field);
2892 // Loop over all bases of this field and push on worklist Field nodes
2893 // with the same offset and base (since they may reference the same field).
2894 for (BaseIterator i(field); i.has_next(); i.next()) {
2895 PointsToNode* base = i.get();
2896 add_fields_to_worklist(field, base);
2897 // Check if the base was source object of arraycopy and go over arraycopy's
2898 // destination objects since values stored to a field of source object are
2899 // accessible by uses (loads) of fields of destination objects.
2900 if (base->arraycopy_src()) {
2901 for (UseIterator j(base); j.has_next(); j.next()) {
2902 PointsToNode* arycp = j.get();
2903 if (arycp->is_Arraycopy()) {
2904 for (UseIterator k(arycp); k.has_next(); k.next()) {
2905 PointsToNode* abase = k.get();
2906 if (abase->arraycopy_dst() && abase != base) {
2907 // Look for the same arraycopy reference.
2908 add_fields_to_worklist(field, abase);
2909 }
2910 }
2911 }
2912 }
2913 }
2914 }
2915 }
2916
2917 // Put on worklist all related field nodes.
2918 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2919 int offset = field->offset();
2920 if (base->is_LocalVar()) {
2921 for (UseIterator j(base); j.has_next(); j.next()) {
2922 PointsToNode* f = j.get();
2923 if (PointsToNode::is_base_use(f)) { // Field
2924 f = PointsToNode::get_use_node(f);
2925 if (f == field || !f->as_Field()->is_oop()) {
2926 continue;
2927 }
2928 int offs = f->as_Field()->offset();
2929 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2930 add_to_worklist(f);
2931 }
2932 }
2933 }
2934 } else {
2935 assert(base->is_JavaObject(), "sanity");
2936 if (// Skip phantom_object since it is only used to indicate that
2937 // this field's content globally escapes.
2938 (base != phantom_obj) &&
2939 // null object node does not have fields.
2940 (base != null_obj)) {
2941 for (EdgeIterator i(base); i.has_next(); i.next()) {
2942 PointsToNode* f = i.get();
2943 // Skip arraycopy edge since store to destination object field
2944 // does not update value in source object field.
2945 if (f->is_Arraycopy()) {
2946 assert(base->arraycopy_dst(), "sanity");
2947 continue;
2948 }
2949 if (f == field || !f->as_Field()->is_oop()) {
2950 continue;
2951 }
2952 int offs = f->as_Field()->offset();
2953 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2954 add_to_worklist(f);
2955 }
2956 }
2957 }
2958 }
2959 }
2960
2961 // Find fields which have unknown value.
2962 int ConnectionGraph::find_field_value(FieldNode* field) {
2963 // Escaped fields should have init value already.
2964 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2965 int new_edges = 0;
2966 for (BaseIterator i(field); i.has_next(); i.next()) {
2967 PointsToNode* base = i.get();
2968 if (base->is_JavaObject()) {
2969 // Skip Allocate's fields which will be processed later.
2970 if (base->ideal_node()->is_Allocate()) {
2971 return 0;
2972 }
2973 assert(base == null_obj, "only null ptr base expected here");
2974 }
2975 }
2976 if (add_edge(field, phantom_obj)) {
2977 // New edge was added
2978 new_edges++;
2979 add_field_uses_to_worklist(field);
2980 }
2981 return new_edges;
2982 }
2983
2984 // Find fields initializing values for allocations.
2985 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2986 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2987 PointsToNode* init_val = phantom_obj;
2988 Node* alloc = pta->ideal_node();
2989
2990 // Do nothing for Allocate nodes since its fields values are
2991 // "known" unless they are initialized by arraycopy/clone.
2992 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2993 if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2994 // Null-free inline type arrays are initialized with an init value instead of null
2995 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
2996 assert(init_val != nullptr, "init value should be registered");
2997 } else {
2998 return 0;
2999 }
3000 }
3001 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
3002 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
3003 #ifdef ASSERT
3004 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
3005 const char* name = alloc->as_CallStaticJava()->_name;
3006 assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
3007 strncmp(name, "load_unknown_inline", 19) == 0 ||
3008 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "sanity");
3009 }
3010 #endif
3011 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
3012 int new_edges = 0;
3013 for (EdgeIterator i(pta); i.has_next(); i.next()) {
3014 PointsToNode* field = i.get();
3015 if (field->is_Field() && field->as_Field()->is_oop()) {
3016 if (add_edge(field, init_val)) {
3017 // New edge was added
3018 new_edges++;
3019 add_field_uses_to_worklist(field->as_Field());
3020 }
3021 }
3022 }
3023 return new_edges;
3024 }
3025
3026 // Find fields initializing values for allocations.
3027 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
3028 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
3029 Node* alloc = pta->ideal_node();
3030 // Do nothing for Call nodes since its fields values are unknown.
3031 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
3032 return 0;
3033 }
3034 InitializeNode* ini = alloc->as_Allocate()->initialization();
3035 bool visited_bottom_offset = false;
3036 GrowableArray<int> offsets_worklist;
3037 int new_edges = 0;
3038
3039 // Check if an oop field's initializing value is recorded and add
3040 // a corresponding null if field's value if it is not recorded.
3041 // Connection Graph does not record a default initialization by null
3042 // captured by Initialize node.
3043 //
3044 for (EdgeIterator i(pta); i.has_next(); i.next()) {
3045 PointsToNode* field = i.get(); // Field (AddP)
3046 if (!field->is_Field() || !field->as_Field()->is_oop()) {
3047 continue; // Not oop field
3048 }
3049 int offset = field->as_Field()->offset();
3050 if (offset == Type::OffsetBot) {
3051 if (!visited_bottom_offset) {
3052 // OffsetBot is used to reference array's element,
3053 // always add reference to null to all Field nodes since we don't
3054 // known which element is referenced.
3055 if (add_edge(field, null_obj)) {
3056 // New edge was added
3057 new_edges++;
3058 add_field_uses_to_worklist(field->as_Field());
3059 visited_bottom_offset = true;
3060 }
3061 }
3062 } else {
3063 // Check only oop fields.
3064 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
3065 if (adr_type->isa_rawptr()) {
3066 #ifdef ASSERT
3067 // Raw pointers are used for initializing stores so skip it
3068 // since it should be recorded already
3069 Node* base = get_addp_base(field->ideal_node());
3070 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
3071 #endif
3072 continue;
3073 }
3074 if (!offsets_worklist.contains(offset)) {
3075 offsets_worklist.append(offset);
3076 Node* value = nullptr;
3077 if (ini != nullptr) {
3078 // StoreP::value_basic_type() == T_ADDRESS
3079 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
3080 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
3081 // Make sure initializing store has the same type as this AddP.
3082 // This AddP may reference non existing field because it is on a
3083 // dead branch of bimorphic call which is not eliminated yet.
3084 if (store != nullptr && store->is_Store() &&
3085 store->as_Store()->value_basic_type() == ft) {
3086 value = store->in(MemNode::ValueIn);
3087 #ifdef ASSERT
3088 if (VerifyConnectionGraph) {
3089 // Verify that AddP already points to all objects the value points to.
3090 PointsToNode* val = ptnode_adr(value->_idx);
3091 assert((val != nullptr), "should be processed already");
3092 PointsToNode* missed_obj = nullptr;
3093 if (val->is_JavaObject()) {
3094 if (!field->points_to(val->as_JavaObject())) {
3095 missed_obj = val;
3096 }
3097 } else {
3098 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
3099 tty->print_cr("----------init store has invalid value -----");
3100 store->dump();
3101 val->dump();
3102 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
3103 }
3104 for (EdgeIterator j(val); j.has_next(); j.next()) {
3105 PointsToNode* obj = j.get();
3106 if (obj->is_JavaObject()) {
3107 if (!field->points_to(obj->as_JavaObject())) {
3108 missed_obj = obj;
3109 break;
3110 }
3111 }
3112 }
3113 }
3114 if (missed_obj != nullptr) {
3115 tty->print_cr("----------field---------------------------------");
3116 field->dump();
3117 tty->print_cr("----------missed reference to object------------");
3118 missed_obj->dump();
3119 tty->print_cr("----------object referenced by init store-------");
3120 store->dump();
3121 val->dump();
3122 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
3123 }
3124 }
3125 #endif
3126 } else {
3127 // There could be initializing stores which follow allocation.
3128 // For example, a volatile field store is not collected
3129 // by Initialize node.
3130 //
3131 // Need to check for dependent loads to separate such stores from
3132 // stores which follow loads. For now, add initial value null so
3133 // that compare pointers optimization works correctly.
3134 }
3135 }
3136 if (value == nullptr) {
3137 // A field's initializing value was not recorded. Add null.
3138 if (add_edge(field, null_obj)) {
3139 // New edge was added
3140 new_edges++;
3141 add_field_uses_to_worklist(field->as_Field());
3142 }
3143 }
3144 }
3145 }
3146 }
3147 return new_edges;
3148 }
3149
3150 // Adjust scalar_replaceable state after Connection Graph is built.
3151 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
3152 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
3153 // returns true. If one of the constraints in this method set 'jobj' to NSR
3154 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
3155 // input, 'adjust_scalar_replaceable_state' will eventually be called with
3156 // that other object and the Phi will become a reducible Phi.
3157 // There could be multiple merges involving the same jobj.
3158 Unique_Node_List candidates;
3159
3160 // Search for non-escaping objects which are not scalar replaceable
3161 // and mark them to propagate the state to referenced objects.
3162
3163 for (UseIterator i(jobj); i.has_next(); i.next()) {
3164 PointsToNode* use = i.get();
3165 if (use->is_Arraycopy()) {
3166 continue;
3167 }
3168 if (use->is_Field()) {
3169 FieldNode* field = use->as_Field();
3170 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3171 // 1. An object is not scalar replaceable if the field into which it is
3172 // stored has unknown offset (stored into unknown element of an array).
3173 if (field->offset() == Type::OffsetBot) {
3174 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
3175 return;
3176 }
3177 for (BaseIterator i(field); i.has_next(); i.next()) {
3178 PointsToNode* base = i.get();
3179 // 2. An object is not scalar replaceable if the field into which it is
3180 // stored has multiple bases one of which is null.
3181 if ((base == null_obj) && (field->base_count() > 1)) {
3182 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3183 return;
3184 }
3185 // 2.5. An object is not scalar replaceable if the field into which it is
3186 // stored has NSR base.
3187 if (!base->scalar_replaceable()) {
3188 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3189 return;
3190 }
3191 }
3192 }
3193 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3194 // 3. An object is not scalar replaceable if it is merged with other objects
3195 // and we can't remove the merge
3196 for (EdgeIterator j(use); j.has_next(); j.next()) {
3197 PointsToNode* ptn = j.get();
3198 if (ptn->is_JavaObject() && ptn != jobj) {
3199 Node* use_n = use->ideal_node();
3200
3201 // These other local vars may point to multiple objects through a Phi
3202 // In this case we skip them and see if we can reduce the Phi.
3203 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3204 use_n = use_n->in(1);
3205 }
3206
3207 // If it's already a candidate or confirmed reducible merge we can skip verification
3208 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3209 continue;
3210 }
3211
3212 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3213 candidates.push(use_n);
3214 } else {
3215 // Mark all objects as NSR if we can't remove the merge
3216 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3217 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3218 }
3219 }
3220 }
3221 if (!jobj->scalar_replaceable()) {
3222 return;
3223 }
3224 }
3225
3226 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3227 if (j.get()->is_Arraycopy()) {
3228 continue;
3229 }
3230
3231 // Non-escaping object node should point only to field nodes.
3232 FieldNode* field = j.get()->as_Field();
3233 int offset = field->as_Field()->offset();
3234
3235 // 4. An object is not scalar replaceable if it has a field with unknown
3236 // offset (array's element is accessed in loop).
3237 if (offset == Type::OffsetBot) {
3238 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3239 return;
3240 }
3241 // 5. Currently an object is not scalar replaceable if a LoadStore node
3242 // access its field since the field value is unknown after it.
3243 //
3244 Node* n = field->ideal_node();
3245
3246 // Test for an unsafe access that was parsed as maybe off heap
3247 // (with a CheckCastPP to raw memory).
3248 assert(n->is_AddP(), "expect an address computation");
3249 if (n->in(AddPNode::Base)->is_top() &&
3250 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3251 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3252 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3253 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3254 return;
3255 }
3256
3257 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3258 Node* u = n->fast_out(i);
3259 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3260 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3261 return;
3262 }
3263 }
3264
3265 // 6. Or the address may point to more then one object. This may produce
3266 // the false positive result (set not scalar replaceable)
3267 // since the flow-insensitive escape analysis can't separate
3268 // the case when stores overwrite the field's value from the case
3269 // when stores happened on different control branches.
3270 //
3271 // Note: it will disable scalar replacement in some cases:
3272 //
3273 // Point p[] = new Point[1];
3274 // p[0] = new Point(); // Will be not scalar replaced
3275 //
3276 // but it will save us from incorrect optimizations in next cases:
3277 //
3278 // Point p[] = new Point[1];
3279 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3280 //
3281 if (field->base_count() > 1 && candidates.size() == 0) {
3282 if (has_non_reducible_merge(field, reducible_merges)) {
3283 for (BaseIterator i(field); i.has_next(); i.next()) {
3284 PointsToNode* base = i.get();
3285 // Don't take into account LocalVar nodes which
3286 // may point to only one object which should be also
3287 // this field's base by now.
3288 if (base->is_JavaObject() && base != jobj) {
3289 // Mark all bases.
3290 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3291 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3292 }
3293 }
3294
3295 if (!jobj->scalar_replaceable()) {
3296 return;
3297 }
3298 }
3299 }
3300 }
3301
3302 // The candidate is truly a reducible merge only if none of the other
3303 // constraints ruled it as NSR. There could be multiple merges involving the
3304 // same jobj.
3305 assert(jobj->scalar_replaceable(), "sanity");
3306 for (uint i = 0; i < candidates.size(); i++ ) {
3307 Node* candidate = candidates.at(i);
3308 reducible_merges.push(candidate);
3309 }
3310 }
3311
3312 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3313 for (BaseIterator i(field); i.has_next(); i.next()) {
3314 Node* base = i.get()->ideal_node();
3315 if (base->is_Phi() && !reducible_merges.member(base)) {
3316 return true;
3317 }
3318 }
3319 return false;
3320 }
3321
3322 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3323 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3324
3325 // Look for 'phis' that refer to 'jobj' as the last
3326 // remaining scalar replaceable input.
3327 uint reducible_merges_cnt = reducible_merges.size();
3328 for (uint i = 0; i < reducible_merges_cnt; i++) {
3329 Node* phi = reducible_merges.at(i);
3330
3331 // This 'Phi' will be a 'good' if it still points to
3332 // at least one scalar replaceable object. Note that 'obj'
3333 // was/should be marked as NSR before calling this function.
3334 bool good_phi = false;
3335
3336 for (uint j = 1; j < phi->req(); j++) {
3337 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3338 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3339 good_phi = true;
3340 break;
3341 }
3342 }
3343
3344 if (!good_phi) {
3345 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3346 reducible_merges.remove(i);
3347
3348 // Decrement the index because the 'remove' call above actually
3349 // moves the last entry of the list to position 'i'.
3350 i--;
3351
3352 reducible_merges_cnt--;
3353 }
3354 }
3355 }
3356
3357 // Propagate NSR (Not scalar replaceable) state.
3358 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3359 int jobj_length = jobj_worklist.length();
3360 bool found_nsr_alloc = true;
3361 while (found_nsr_alloc) {
3362 found_nsr_alloc = false;
3363 for (int next = 0; next < jobj_length; ++next) {
3364 JavaObjectNode* jobj = jobj_worklist.at(next);
3365 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3366 PointsToNode* use = i.get();
3367 if (use->is_Field()) {
3368 FieldNode* field = use->as_Field();
3369 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3370 assert(field->offset() != Type::OffsetBot, "sanity");
3371 for (BaseIterator i(field); i.has_next(); i.next()) {
3372 PointsToNode* base = i.get();
3373 // An object is not scalar replaceable if the field into which
3374 // it is stored has NSR base.
3375 if ((base != null_obj) && !base->scalar_replaceable()) {
3376 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3377 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3378 // because there is no point in reducing a Phi that won't improve the number of SR
3379 // objects.
3380 revisit_reducible_phi_status(jobj, reducible_merges);
3381 found_nsr_alloc = true;
3382 break;
3383 }
3384 }
3385 } else if (use->is_LocalVar()) {
3386 Node* phi = use->ideal_node();
3387 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3388 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3389 reducible_merges.yank(phi);
3390 found_nsr_alloc = true;
3391 break;
3392 }
3393 }
3394 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3395 }
3396 }
3397 }
3398 }
3399
3400 #ifdef ASSERT
3401 void ConnectionGraph::verify_connection_graph(
3402 GrowableArray<PointsToNode*>& ptnodes_worklist,
3403 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3404 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3405 GrowableArray<Node*>& addp_worklist) {
3406 // Verify that graph is complete - no new edges could be added.
3407 int java_objects_length = java_objects_worklist.length();
3408 int non_escaped_length = non_escaped_allocs_worklist.length();
3409 int new_edges = 0;
3410 for (int next = 0; next < java_objects_length; ++next) {
3411 JavaObjectNode* ptn = java_objects_worklist.at(next);
3412 new_edges += add_java_object_edges(ptn, true);
3413 }
3414 assert(new_edges == 0, "graph was not complete");
3415 // Verify that escape state is final.
3416 int length = non_escaped_allocs_worklist.length();
3417 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3418 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3419 (non_escaped_length == length) &&
3420 (_worklist.length() == 0), "escape state was not final");
3421
3422 // Verify fields information.
3423 int addp_length = addp_worklist.length();
3424 for (int next = 0; next < addp_length; ++next ) {
3425 Node* n = addp_worklist.at(next);
3426 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3427 if (field->is_oop()) {
3428 // Verify that field has all bases
3429 Node* base = get_addp_base(n);
3430 PointsToNode* ptn = ptnode_adr(base->_idx);
3431 if (ptn->is_JavaObject()) {
3432 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3433 } else {
3434 assert(ptn->is_LocalVar(), "sanity");
3435 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3436 PointsToNode* e = i.get();
3437 if (e->is_JavaObject()) {
3438 assert(field->has_base(e->as_JavaObject()), "sanity");
3439 }
3440 }
3441 }
3442 // Verify that all fields have initializing values.
3443 if (field->edge_count() == 0) {
3444 tty->print_cr("----------field does not have references----------");
3445 field->dump();
3446 for (BaseIterator i(field); i.has_next(); i.next()) {
3447 PointsToNode* base = i.get();
3448 tty->print_cr("----------field has next base---------------------");
3449 base->dump();
3450 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3451 tty->print_cr("----------base has fields-------------------------");
3452 for (EdgeIterator j(base); j.has_next(); j.next()) {
3453 j.get()->dump();
3454 }
3455 tty->print_cr("----------base has references---------------------");
3456 for (UseIterator j(base); j.has_next(); j.next()) {
3457 j.get()->dump();
3458 }
3459 }
3460 }
3461 for (UseIterator i(field); i.has_next(); i.next()) {
3462 i.get()->dump();
3463 }
3464 assert(field->edge_count() > 0, "sanity");
3465 }
3466 }
3467 }
3468 }
3469 #endif
3470
3471 // Optimize ideal graph.
3472 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3473 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3474 Compile* C = _compile;
3475 PhaseIterGVN* igvn = _igvn;
3476 if (EliminateLocks) {
3477 // Mark locks before changing ideal graph.
3478 int cnt = C->macro_count();
3479 for (int i = 0; i < cnt; i++) {
3480 Node *n = C->macro_node(i);
3481 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3482 AbstractLockNode* alock = n->as_AbstractLock();
3483 if (!alock->is_non_esc_obj()) {
3484 const Type* obj_type = igvn->type(alock->obj_node());
3485 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3486 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3487 // The lock could be marked eliminated by lock coarsening
3488 // code during first IGVN before EA. Replace coarsened flag
3489 // to eliminate all associated locks/unlocks.
3490 #ifdef ASSERT
3491 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3492 #endif
3493 alock->set_non_esc_obj();
3494 }
3495 }
3496 }
3497 }
3498 }
3499
3500 if (OptimizePtrCompare) {
3501 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3502 Node *n = ptr_cmp_worklist.at(i);
3503 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3504 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3505 if (tcmp->singleton()) {
3506 Node* cmp = igvn->makecon(tcmp);
3507 #ifndef PRODUCT
3508 if (PrintOptimizePtrCompare) {
3509 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"));
3510 if (Verbose) {
3511 n->dump(1);
3512 }
3513 }
3514 #endif
3515 igvn->replace_node(n, cmp);
3516 }
3517 }
3518 }
3519
3520 // For MemBarStoreStore nodes added in library_call.cpp, check
3521 // escape status of associated AllocateNode and optimize out
3522 // MemBarStoreStore node if the allocated object never escapes.
3523 for (int i = 0; i < storestore_worklist.length(); i++) {
3524 Node* storestore = storestore_worklist.at(i);
3525 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3526 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3527 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3528 // Non-escaping inline type buffer allocations don't require a membar
3529 storestore->as_MemBar()->remove(_igvn);
3530 } else {
3531 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3532 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3533 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3534 igvn->register_new_node_with_optimizer(mb);
3535 igvn->replace_node(storestore, mb);
3536 }
3537 }
3538 }
3539 }
3540
3541 // Atomic flat accesses on non-escaping objects can be optimized to non-atomic accesses
3542 void ConnectionGraph::optimize_flat_accesses(GrowableArray<SafePointNode*>& sfn_worklist) {
3543 PhaseIterGVN& igvn = *_igvn;
3544 bool delay = igvn.delay_transform();
3545 igvn.set_delay_transform(true);
3546 igvn.C->for_each_flat_access([&](Node* n) {
3547 Node* base = n->is_LoadFlat() ? n->as_LoadFlat()->base() : n->as_StoreFlat()->base();
3548 if (!not_global_escape(base)) {
3549 return;
3550 }
3551
3552 bool expanded;
3553 if (n->is_LoadFlat()) {
3554 expanded = n->as_LoadFlat()->expand_non_atomic(igvn);
3555 } else {
3556 expanded = n->as_StoreFlat()->expand_non_atomic(igvn);
3557 }
3558 if (expanded) {
3559 sfn_worklist.remove(n->as_SafePoint());
3560 igvn.C->remove_flat_access(n);
3561 }
3562 });
3563 igvn.set_delay_transform(delay);
3564 }
3565
3566 // Optimize objects compare.
3567 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3568 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3569 if (!OptimizePtrCompare) {
3570 return UNKNOWN;
3571 }
3572 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3573 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3574
3575 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3576 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3577 JavaObjectNode* jobj1 = unique_java_object(left);
3578 JavaObjectNode* jobj2 = unique_java_object(right);
3579
3580 // The use of this method during allocation merge reduction may cause 'left'
3581 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3582 // that doesn't reference an unique java object.
3583 if (ptn1 == nullptr || ptn2 == nullptr ||
3584 jobj1 == nullptr || jobj2 == nullptr) {
3585 return UNKNOWN;
3586 }
3587
3588 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3589 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3590
3591 // Check simple cases first.
3592 if (jobj1 != nullptr) {
3593 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3594 if (jobj1 == jobj2) {
3595 // Comparing the same not escaping object.
3596 return EQ;
3597 }
3598 Node* obj = jobj1->ideal_node();
3599 // Comparing not escaping allocation.
3600 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3601 !ptn2->points_to(jobj1)) {
3602 return NE; // This includes nullness check.
3603 }
3604 }
3605 }
3606 if (jobj2 != nullptr) {
3607 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3608 Node* obj = jobj2->ideal_node();
3609 // Comparing not escaping allocation.
3610 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3611 !ptn1->points_to(jobj2)) {
3612 return NE; // This includes nullness check.
3613 }
3614 }
3615 }
3616 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3617 jobj2 != nullptr && jobj2 != phantom_obj &&
3618 jobj1->ideal_node()->is_Con() &&
3619 jobj2->ideal_node()->is_Con()) {
3620 // Klass or String constants compare. Need to be careful with
3621 // compressed pointers - compare types of ConN and ConP instead of nodes.
3622 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3623 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3624 if (t1->make_ptr() == t2->make_ptr()) {
3625 return EQ;
3626 } else {
3627 return NE;
3628 }
3629 }
3630 if (ptn1->meet(ptn2)) {
3631 return UNKNOWN; // Sets are not disjoint
3632 }
3633
3634 // Sets are disjoint.
3635 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3636 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3637 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3638 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3639 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3640 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3641 // Check nullness of unknown object.
3642 return UNKNOWN;
3643 }
3644
3645 // Disjointness by itself is not sufficient since
3646 // alias analysis is not complete for escaped objects.
3647 // Disjoint sets are definitely unrelated only when
3648 // at least one set has only not escaping allocations.
3649 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3650 if (ptn1->non_escaping_allocation()) {
3651 return NE;
3652 }
3653 }
3654 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3655 if (ptn2->non_escaping_allocation()) {
3656 return NE;
3657 }
3658 }
3659 return UNKNOWN;
3660 }
3661
3662 // Connection Graph construction functions.
3663
3664 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3665 PointsToNode* ptadr = _nodes.at(n->_idx);
3666 if (ptadr != nullptr) {
3667 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3668 return;
3669 }
3670 Compile* C = _compile;
3671 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3672 map_ideal_node(n, ptadr);
3673 }
3674
3675 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3676 PointsToNode* ptadr = _nodes.at(n->_idx);
3677 if (ptadr != nullptr) {
3678 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3679 return ptadr;
3680 }
3681 Compile* C = _compile;
3682 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3683 map_ideal_node(n, ptadr);
3684 return ptadr;
3685 }
3686
3687 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3688 PointsToNode* ptadr = _nodes.at(n->_idx);
3689 if (ptadr != nullptr) {
3690 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3691 return;
3692 }
3693 bool unsafe = false;
3694 bool is_oop = is_oop_field(n, offset, &unsafe);
3695 if (unsafe) {
3696 es = PointsToNode::GlobalEscape;
3697 }
3698 Compile* C = _compile;
3699 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3700 map_ideal_node(n, field);
3701 }
3702
3703 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3704 PointsToNode* src, PointsToNode* dst) {
3705 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3706 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3707 PointsToNode* ptadr = _nodes.at(n->_idx);
3708 if (ptadr != nullptr) {
3709 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3710 return;
3711 }
3712 Compile* C = _compile;
3713 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3714 map_ideal_node(n, ptadr);
3715 // Add edge from arraycopy node to source object.
3716 (void)add_edge(ptadr, src);
3717 src->set_arraycopy_src();
3718 // Add edge from destination object to arraycopy node.
3719 (void)add_edge(dst, ptadr);
3720 dst->set_arraycopy_dst();
3721 }
3722
3723 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3724 const Type* adr_type = n->as_AddP()->bottom_type();
3725 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3726 BasicType bt = T_INT;
3727 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3728 // Check only oop fields.
3729 if (!adr_type->isa_aryptr() ||
3730 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3731 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3732 // OffsetBot is used to reference array's element. Ignore first AddP.
3733 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3734 bt = T_OBJECT;
3735 }
3736 }
3737 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3738 if (adr_type->isa_instptr()) {
3739 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3740 if (field != nullptr) {
3741 bt = field->layout_type();
3742 } else {
3743 // Check for unsafe oop field access
3744 if (has_oop_node_outs(n)) {
3745 bt = T_OBJECT;
3746 (*unsafe) = true;
3747 }
3748 }
3749 } else if (adr_type->isa_aryptr()) {
3750 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3751 // Ignore array length load.
3752 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3753 // Ignore first AddP.
3754 } else {
3755 const Type* elemtype = adr_type->is_aryptr()->elem();
3756 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3757 ciInlineKlass* vk = elemtype->inline_klass();
3758 field_offset += vk->payload_offset();
3759 ciField* field = vk->get_field_by_offset(field_offset, false);
3760 if (field != nullptr) {
3761 bt = field->layout_type();
3762 } else {
3763 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);
3764 bt = T_BOOLEAN;
3765 }
3766 } else {
3767 bt = elemtype->array_element_basic_type();
3768 }
3769 }
3770 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3771 // Allocation initialization, ThreadLocal field access, unsafe access
3772 if (has_oop_node_outs(n)) {
3773 bt = T_OBJECT;
3774 }
3775 }
3776 }
3777 // Note: T_NARROWOOP is not classed as a real reference type
3778 bool res = (is_reference_type(bt) || bt == T_NARROWOOP);
3779 assert(!has_oop_node_outs(n) || res, "sanity: AddP has oop outs, needs to be treated as oop field");
3780 return res;
3781 }
3782
3783 bool ConnectionGraph::has_oop_node_outs(Node* n) {
3784 return n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3785 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3786 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3787 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n);
3788 }
3789
3790 // Returns unique pointed java object or null.
3791 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3792 // If the node was created after the escape computation we can't answer.
3793 uint idx = n->_idx;
3794 if (idx >= nodes_size()) {
3795 return nullptr;
3796 }
3797 PointsToNode* ptn = ptnode_adr(idx);
3798 if (ptn == nullptr) {
3799 return nullptr;
3800 }
3801 if (ptn->is_JavaObject()) {
3802 return ptn->as_JavaObject();
3803 }
3804 assert(ptn->is_LocalVar(), "sanity");
3805 // Check all java objects it points to.
3806 JavaObjectNode* jobj = nullptr;
3807 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3808 PointsToNode* e = i.get();
3809 if (e->is_JavaObject()) {
3810 if (jobj == nullptr) {
3811 jobj = e->as_JavaObject();
3812 } else if (jobj != e) {
3813 return nullptr;
3814 }
3815 }
3816 }
3817 return jobj;
3818 }
3819
3820 // Return true if this node points only to non-escaping allocations.
3821 bool PointsToNode::non_escaping_allocation() {
3822 if (is_JavaObject()) {
3823 Node* n = ideal_node();
3824 if (n->is_Allocate() || n->is_CallStaticJava()) {
3825 return (escape_state() == PointsToNode::NoEscape);
3826 } else {
3827 return false;
3828 }
3829 }
3830 assert(is_LocalVar(), "sanity");
3831 // Check all java objects it points to.
3832 for (EdgeIterator i(this); i.has_next(); i.next()) {
3833 PointsToNode* e = i.get();
3834 if (e->is_JavaObject()) {
3835 Node* n = e->ideal_node();
3836 if ((e->escape_state() != PointsToNode::NoEscape) ||
3837 !(n->is_Allocate() || n->is_CallStaticJava())) {
3838 return false;
3839 }
3840 }
3841 }
3842 return true;
3843 }
3844
3845 // Return true if we know the node does not escape globally.
3846 bool ConnectionGraph::not_global_escape(Node *n) {
3847 assert(!_collecting, "should not call during graph construction");
3848 // If the node was created after the escape computation we can't answer.
3849 uint idx = n->_idx;
3850 if (idx >= nodes_size()) {
3851 return false;
3852 }
3853 PointsToNode* ptn = ptnode_adr(idx);
3854 if (ptn == nullptr) {
3855 return false; // not in congraph (e.g. ConI)
3856 }
3857 PointsToNode::EscapeState es = ptn->escape_state();
3858 // If we have already computed a value, return it.
3859 if (es >= PointsToNode::GlobalEscape) {
3860 return false;
3861 }
3862 if (ptn->is_JavaObject()) {
3863 return true; // (es < PointsToNode::GlobalEscape);
3864 }
3865 assert(ptn->is_LocalVar(), "sanity");
3866 // Check all java objects it points to.
3867 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3868 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3869 return false;
3870 }
3871 }
3872 return true;
3873 }
3874
3875 // Return true if locked object does not escape globally
3876 // and locked code region (identified by BoxLockNode) is balanced:
3877 // all compiled code paths have corresponding Lock/Unlock pairs.
3878 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3879 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3880 if (EliminateNestedLocks) {
3881 // We can mark whole locking region as Local only when only
3882 // one object is used for locking.
3883 alock->box_node()->as_BoxLock()->set_local();
3884 }
3885 return true;
3886 }
3887 return false;
3888 }
3889
3890 // Helper functions
3891
3892 // Return true if this node points to specified node or nodes it points to.
3893 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3894 if (is_JavaObject()) {
3895 return (this == ptn);
3896 }
3897 assert(is_LocalVar() || is_Field(), "sanity");
3898 for (EdgeIterator i(this); i.has_next(); i.next()) {
3899 if (i.get() == ptn) {
3900 return true;
3901 }
3902 }
3903 return false;
3904 }
3905
3906 // Return true if one node points to an other.
3907 bool PointsToNode::meet(PointsToNode* ptn) {
3908 if (this == ptn) {
3909 return true;
3910 } else if (ptn->is_JavaObject()) {
3911 return this->points_to(ptn->as_JavaObject());
3912 } else if (this->is_JavaObject()) {
3913 return ptn->points_to(this->as_JavaObject());
3914 }
3915 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3916 int ptn_count = ptn->edge_count();
3917 for (EdgeIterator i(this); i.has_next(); i.next()) {
3918 PointsToNode* this_e = i.get();
3919 for (int j = 0; j < ptn_count; j++) {
3920 if (this_e == ptn->edge(j)) {
3921 return true;
3922 }
3923 }
3924 }
3925 return false;
3926 }
3927
3928 #ifdef ASSERT
3929 // Return true if bases point to this java object.
3930 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3931 for (BaseIterator i(this); i.has_next(); i.next()) {
3932 if (i.get() == jobj) {
3933 return true;
3934 }
3935 }
3936 return false;
3937 }
3938 #endif
3939
3940 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3941 // Handle simple case first.
3942 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3943 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3944 return true;
3945 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3946 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3947 Node* addp_use = addp->fast_out(i);
3948 if (addp_use->is_Store()) {
3949 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3950 if (addp_use->fast_out(j)->is_Initialize()) {
3951 return true;
3952 }
3953 }
3954 }
3955 }
3956 }
3957 return false;
3958 }
3959
3960 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3961 const Type *adr_type = phase->type(adr);
3962 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3963 // We are computing a raw address for a store captured by an Initialize
3964 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3965 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3966 assert(offs != Type::OffsetBot ||
3967 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3968 "offset must be a constant or it is initialization of array");
3969 return offs;
3970 }
3971 return adr_type->is_ptr()->flat_offset();
3972 }
3973
3974 Node* ConnectionGraph::get_addp_base(Node *addp) {
3975 assert(addp->is_AddP(), "must be AddP");
3976 //
3977 // AddP cases for Base and Address inputs:
3978 // case #1. Direct object's field reference:
3979 // Allocate
3980 // |
3981 // Proj #5 ( oop result )
3982 // |
3983 // CheckCastPP (cast to instance type)
3984 // | |
3985 // AddP ( base == address )
3986 //
3987 // case #2. Indirect object's field reference:
3988 // Phi
3989 // |
3990 // CastPP (cast to instance type)
3991 // | |
3992 // AddP ( base == address )
3993 //
3994 // case #3. Raw object's field reference for Initialize node.
3995 // Could have an additional Phi merging multiple allocations.
3996 // Allocate
3997 // |
3998 // Proj #5 ( oop result )
3999 // top |
4000 // \ |
4001 // AddP ( base == top )
4002 //
4003 // case #4. Array's element reference:
4004 // {CheckCastPP | CastPP}
4005 // | | |
4006 // | AddP ( array's element offset )
4007 // | |
4008 // AddP ( array's offset )
4009 //
4010 // case #5. Raw object's field reference for arraycopy stub call:
4011 // The inline_native_clone() case when the arraycopy stub is called
4012 // after the allocation before Initialize and CheckCastPP nodes.
4013 // Allocate
4014 // |
4015 // Proj #5 ( oop result )
4016 // | |
4017 // AddP ( base == address )
4018 //
4019 // case #6. Constant Pool, ThreadLocal, CastX2P, Klass, OSR buffer buf or
4020 // Raw object's field reference:
4021 // {ConP, ThreadLocal, CastX2P, raw Load, Parm0}
4022 // top |
4023 // \ |
4024 // AddP ( base == top )
4025 //
4026 // case #7. Klass's field reference.
4027 // LoadKlass
4028 // | |
4029 // AddP ( base == address )
4030 //
4031 // case #8. narrow Klass's field reference.
4032 // LoadNKlass
4033 // |
4034 // DecodeN
4035 // | |
4036 // AddP ( base == address )
4037 //
4038 // case #9. Mixed unsafe access
4039 // {instance}
4040 // |
4041 // CheckCastPP (raw)
4042 // top |
4043 // \ |
4044 // AddP ( base == top )
4045 //
4046 // case #10. Klass fetched with
4047 // LibraryCallKit::load_*_refined_array_klass()
4048 // which has en extra Phi.
4049 // LoadKlass LoadKlass
4050 // | |
4051 // CastPP CastPP
4052 // \ /
4053 // Phi
4054 // top |
4055 // \ |
4056 // AddP ( base == top )
4057 //
4058 Node *base = addp->in(AddPNode::Base);
4059 if (base->uncast()->is_top()) { // The AddP case #3, #6, #9, and #10.
4060 base = addp->in(AddPNode::Address);
4061 while (base->is_AddP()) {
4062 // Case #6 (unsafe access) may have several chained AddP nodes.
4063 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
4064 base = base->in(AddPNode::Address);
4065 }
4066 if (base->Opcode() == Op_CheckCastPP &&
4067 base->bottom_type()->isa_rawptr() &&
4068 _igvn->type(base->in(1))->isa_oopptr()) {
4069 base = base->in(1); // Case #9
4070 } else {
4071 // Case #3, #6, and #10
4072 Node* uncast_base = base->uncast();
4073 int opcode = uncast_base->Opcode();
4074 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
4075 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
4076 (_igvn->C->is_osr_compilation() && uncast_base->is_Parm() && uncast_base->as_Parm()->_con == TypeFunc::Parms)||
4077 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
4078 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_klassptr() != nullptr)) ||
4079 is_captured_store_address(addp) ||
4080 is_load_array_klass_related(uncast_base), "sanity");
4081 }
4082 }
4083 return base;
4084 }
4085
4086 #ifdef ASSERT
4087 // Case #10
4088 bool ConnectionGraph::is_load_array_klass_related(const Node* uncast_base) {
4089 if (!uncast_base->is_Phi() || uncast_base->req() != 3) {
4090 return false;
4091 }
4092 Node* in1 = uncast_base->in(1);
4093 Node* in2 = uncast_base->in(2);
4094 return in1->uncast()->Opcode() == Op_LoadKlass &&
4095 in2->uncast()->Opcode() == Op_LoadKlass;
4096 }
4097 #endif
4098
4099 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
4100 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
4101 Node* addp2 = addp->raw_out(0);
4102 if (addp->outcnt() == 1 && addp2->is_AddP() &&
4103 addp2->in(AddPNode::Base) == n &&
4104 addp2->in(AddPNode::Address) == addp) {
4105 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
4106 //
4107 // Find array's offset to push it on worklist first and
4108 // as result process an array's element offset first (pushed second)
4109 // to avoid CastPP for the array's offset.
4110 // Otherwise the inserted CastPP (LocalVar) will point to what
4111 // the AddP (Field) points to. Which would be wrong since
4112 // the algorithm expects the CastPP has the same point as
4113 // as AddP's base CheckCastPP (LocalVar).
4114 //
4115 // ArrayAllocation
4116 // |
4117 // CheckCastPP
4118 // |
4119 // memProj (from ArrayAllocation CheckCastPP)
4120 // | ||
4121 // | || Int (element index)
4122 // | || | ConI (log(element size))
4123 // | || | /
4124 // | || LShift
4125 // | || /
4126 // | AddP (array's element offset)
4127 // | |
4128 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
4129 // | / /
4130 // AddP (array's offset)
4131 // |
4132 // Load/Store (memory operation on array's element)
4133 //
4134 return addp2;
4135 }
4136 return nullptr;
4137 }
4138
4139 //
4140 // Adjust the type and inputs of an AddP which computes the
4141 // address of a field of an instance
4142 //
4143 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
4144 PhaseGVN* igvn = _igvn;
4145 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
4146 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
4147 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
4148 if (t == nullptr) {
4149 // We are computing a raw address for a store captured by an Initialize
4150 // compute an appropriate address type (cases #3 and #5).
4151 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
4152 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
4153 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
4154 assert(offs != Type::OffsetBot, "offset must be a constant");
4155 if (base_t->isa_aryptr() != nullptr) {
4156 // In the case of a flat inline type array, each field has its
4157 // own slice so we need to extract the field being accessed from
4158 // the address computation
4159 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
4160 } else {
4161 t = base_t->add_offset(offs)->is_oopptr();
4162 }
4163 }
4164 int inst_id = base_t->instance_id();
4165 assert(!t->is_known_instance() || t->instance_id() == inst_id,
4166 "old type must be non-instance or match new type");
4167
4168 // The type 't' could be subclass of 'base_t'.
4169 // As result t->offset() could be large then base_t's size and it will
4170 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
4171 // constructor verifies correctness of the offset.
4172 //
4173 // It could happened on subclass's branch (from the type profiling
4174 // inlining) which was not eliminated during parsing since the exactness
4175 // of the allocation type was not propagated to the subclass type check.
4176 //
4177 // Or the type 't' could be not related to 'base_t' at all.
4178 // It could happen when CHA type is different from MDO type on a dead path
4179 // (for example, from instanceof check) which is not collapsed during parsing.
4180 //
4181 // Do nothing for such AddP node and don't process its users since
4182 // this code branch will go away.
4183 //
4184 if (!t->is_known_instance() &&
4185 !base_t->maybe_java_subtype_of(t)) {
4186 return false; // bail out
4187 }
4188 const TypePtr* tinst = base_t->add_offset(t->offset());
4189 if (tinst->isa_aryptr() && t->isa_aryptr()) {
4190 // In the case of a flat inline type array, each field has its
4191 // own slice so we need to keep track of the field being accessed.
4192 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
4193 // Keep array properties (not flat/null-free)
4194 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
4195 if (tinst == nullptr) {
4196 return false; // Skip dead path with inconsistent properties
4197 }
4198 }
4199
4200 // Do NOT remove the next line: ensure a new alias index is allocated
4201 // for the instance type. Note: C++ will not remove it since the call
4202 // has side effect.
4203 int alias_idx = _compile->get_alias_index(tinst);
4204 igvn->set_type(addp, tinst);
4205 // record the allocation in the node map
4206 set_map(addp, get_map(base->_idx));
4207 // Set addp's Base and Address to 'base'.
4208 Node *abase = addp->in(AddPNode::Base);
4209 Node *adr = addp->in(AddPNode::Address);
4210 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
4211 adr->in(0)->_idx == (uint)inst_id) {
4212 // Skip AddP cases #3 and #5.
4213 } else {
4214 assert(!abase->is_top(), "sanity"); // AddP case #3
4215 if (abase != base) {
4216 igvn->hash_delete(addp);
4217 addp->set_req(AddPNode::Base, base);
4218 if (abase == adr) {
4219 addp->set_req(AddPNode::Address, base);
4220 } else {
4221 // AddP case #4 (adr is array's element offset AddP node)
4222 #ifdef ASSERT
4223 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
4224 assert(adr->is_AddP() && atype != nullptr &&
4225 atype->instance_id() == inst_id, "array's element offset should be processed first");
4226 #endif
4227 }
4228 igvn->hash_insert(addp);
4229 }
4230 }
4231 // Put on IGVN worklist since at least addp's type was changed above.
4232 record_for_optimizer(addp);
4233 return true;
4234 }
4235
4236 //
4237 // Create a new version of orig_phi if necessary. Returns either the newly
4238 // created phi or an existing phi. Sets create_new to indicate whether a new
4239 // phi was created. Cache the last newly created phi in the node map.
4240 //
4241 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
4242 Compile *C = _compile;
4243 PhaseGVN* igvn = _igvn;
4244 new_created = false;
4245 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
4246 // nothing to do if orig_phi is bottom memory or matches alias_idx
4247 if (phi_alias_idx == alias_idx) {
4248 return orig_phi;
4249 }
4250 // Have we recently created a Phi for this alias index?
4251 PhiNode *result = get_map_phi(orig_phi->_idx);
4252 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
4253 return result;
4254 }
4255 // Previous check may fail when the same wide memory Phi was split into Phis
4256 // for different memory slices. Search all Phis for this region.
4257 if (result != nullptr) {
4258 Node* region = orig_phi->in(0);
4259 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4260 Node* phi = region->fast_out(i);
4261 if (phi->is_Phi() &&
4262 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4263 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4264 return phi->as_Phi();
4265 }
4266 }
4267 }
4268 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4269 if (C->do_escape_analysis() == true && !C->failing()) {
4270 // Retry compilation without escape analysis.
4271 // If this is the first failure, the sentinel string will "stick"
4272 // to the Compile object, and the C2Compiler will see it and retry.
4273 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4274 }
4275 return nullptr;
4276 }
4277 orig_phi_worklist.append_if_missing(orig_phi);
4278 const TypePtr *atype = C->get_adr_type(alias_idx);
4279 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4280 C->copy_node_notes_to(result, orig_phi);
4281 igvn->set_type(result, result->bottom_type());
4282 record_for_optimizer(result);
4283 set_map(orig_phi, result);
4284 new_created = true;
4285 return result;
4286 }
4287
4288 //
4289 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4290 // specified alias index.
4291 //
4292 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4293 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4294 Compile *C = _compile;
4295 PhaseGVN* igvn = _igvn;
4296 bool new_phi_created;
4297 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4298 if (!new_phi_created) {
4299 return result;
4300 }
4301 GrowableArray<PhiNode *> phi_list;
4302 GrowableArray<uint> cur_input;
4303 PhiNode *phi = orig_phi;
4304 uint idx = 1;
4305 bool finished = false;
4306 while(!finished) {
4307 while (idx < phi->req()) {
4308 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4309 if (mem != nullptr && mem->is_Phi()) {
4310 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4311 if (new_phi_created) {
4312 // found an phi for which we created a new split, push current one on worklist and begin
4313 // processing new one
4314 phi_list.push(phi);
4315 cur_input.push(idx);
4316 phi = mem->as_Phi();
4317 result = newphi;
4318 idx = 1;
4319 continue;
4320 } else {
4321 mem = newphi;
4322 }
4323 }
4324 if (C->failing()) {
4325 return nullptr;
4326 }
4327 result->set_req(idx++, mem);
4328 }
4329 #ifdef ASSERT
4330 // verify that the new Phi has an input for each input of the original
4331 assert( phi->req() == result->req(), "must have same number of inputs.");
4332 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4333 #endif
4334 // Check if all new phi's inputs have specified alias index.
4335 // Otherwise use old phi.
4336 for (uint i = 1; i < phi->req(); i++) {
4337 Node* in = result->in(i);
4338 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4339 }
4340 // we have finished processing a Phi, see if there are any more to do
4341 finished = (phi_list.length() == 0 );
4342 if (!finished) {
4343 phi = phi_list.pop();
4344 idx = cur_input.pop();
4345 PhiNode *prev_result = get_map_phi(phi->_idx);
4346 prev_result->set_req(idx++, result);
4347 result = prev_result;
4348 }
4349 }
4350 return result;
4351 }
4352
4353 //
4354 // The next methods are derived from methods in MemNode.
4355 //
4356 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4357 Node *mem = mmem;
4358 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4359 // means an array I have not precisely typed yet. Do not do any
4360 // alias stuff with it any time soon.
4361 if (toop->base() != Type::AnyPtr &&
4362 !(toop->isa_instptr() &&
4363 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4364 toop->offset() == Type::OffsetBot)) {
4365 mem = mmem->memory_at(alias_idx);
4366 // Update input if it is progress over what we have now
4367 }
4368 return mem;
4369 }
4370
4371 //
4372 // Move memory users to their memory slices.
4373 //
4374 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4375 Compile* C = _compile;
4376 PhaseGVN* igvn = _igvn;
4377 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4378 assert(tp != nullptr, "ptr type");
4379 int alias_idx = C->get_alias_index(tp);
4380 int general_idx = C->get_general_index(alias_idx);
4381
4382 // Move users first
4383 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4384 Node* use = n->fast_out(i);
4385 if (use->is_MergeMem()) {
4386 MergeMemNode* mmem = use->as_MergeMem();
4387 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4388 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4389 continue; // Nothing to do
4390 }
4391 // Replace previous general reference to mem node.
4392 uint orig_uniq = C->unique();
4393 Node* m = find_inst_mem(n, general_idx, orig_phis);
4394 assert(orig_uniq == C->unique(), "no new nodes");
4395 mmem->set_memory_at(general_idx, m);
4396 --imax;
4397 --i;
4398 } else if (use->is_MemBar()) {
4399 assert(!use->is_Initialize(), "initializing stores should not be moved");
4400 if (use->req() > MemBarNode::Precedent &&
4401 use->in(MemBarNode::Precedent) == n) {
4402 // Don't move related membars.
4403 record_for_optimizer(use);
4404 continue;
4405 }
4406 tp = use->as_MemBar()->adr_type()->isa_ptr();
4407 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4408 alias_idx == general_idx) {
4409 continue; // Nothing to do
4410 }
4411 // Move to general memory slice.
4412 uint orig_uniq = C->unique();
4413 Node* m = find_inst_mem(n, general_idx, orig_phis);
4414 assert(orig_uniq == C->unique(), "no new nodes");
4415 igvn->hash_delete(use);
4416 imax -= use->replace_edge(n, m, igvn);
4417 igvn->hash_insert(use);
4418 record_for_optimizer(use);
4419 --i;
4420 #ifdef ASSERT
4421 } else if (use->is_Mem()) {
4422 // Memory nodes should have new memory input.
4423 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4424 assert(tp != nullptr, "ptr type");
4425 int idx = C->get_alias_index(tp);
4426 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4427 "Following memory nodes should have new memory input or be on the same memory slice");
4428 } else if (use->is_Phi()) {
4429 // Phi nodes should be split and moved already.
4430 tp = use->as_Phi()->adr_type()->isa_ptr();
4431 assert(tp != nullptr, "ptr type");
4432 int idx = C->get_alias_index(tp);
4433 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4434 } else {
4435 use->dump();
4436 assert(false, "should not be here");
4437 #endif
4438 }
4439 }
4440 }
4441
4442 //
4443 // Search memory chain of "mem" to find a MemNode whose address
4444 // is the specified alias index.
4445 //
4446 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4447 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4448 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4449 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4450 return nullptr;
4451 }
4452 if (orig_mem == nullptr) {
4453 return orig_mem;
4454 }
4455 Compile* C = _compile;
4456 PhaseGVN* igvn = _igvn;
4457 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4458 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4459 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4460 Node *prev = nullptr;
4461 Node *result = orig_mem;
4462 while (prev != result) {
4463 prev = result;
4464 if (result == start_mem) {
4465 break; // hit one of our sentinels
4466 }
4467 if (result->is_Mem()) {
4468 const Type *at = igvn->type(result->in(MemNode::Address));
4469 if (at == Type::TOP) {
4470 break; // Dead
4471 }
4472 assert (at->isa_ptr() != nullptr, "pointer type required.");
4473 int idx = C->get_alias_index(at->is_ptr());
4474 if (idx == alias_idx) {
4475 break; // Found
4476 }
4477 if (!is_instance && (at->isa_oopptr() == nullptr ||
4478 !at->is_oopptr()->is_known_instance())) {
4479 break; // Do not skip store to general memory slice.
4480 }
4481 result = result->in(MemNode::Memory);
4482 }
4483 if (!is_instance) {
4484 continue; // don't search further for non-instance types
4485 }
4486 // skip over a call which does not affect this memory slice
4487 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4488 Node *proj_in = result->in(0);
4489 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4490 break; // hit one of our sentinels
4491 } else if (proj_in->is_Call()) {
4492 // ArrayCopy node processed here as well
4493 CallNode *call = proj_in->as_Call();
4494 if (!call->may_modify(toop, igvn)) {
4495 result = call->in(TypeFunc::Memory);
4496 }
4497 } else if (proj_in->is_Initialize()) {
4498 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4499 // Stop if this is the initialization for the object instance which
4500 // which contains this memory slice, otherwise skip over it.
4501 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4502 result = proj_in->in(TypeFunc::Memory);
4503 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4504 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4505 result = get_map(result->_idx);
4506 assert(result != nullptr, "new projection should have been allocated");
4507 break;
4508 }
4509 } else if (proj_in->is_MemBar()) {
4510 // Check if there is an array copy for a clone
4511 // Step over GC barrier when ReduceInitialCardMarks is disabled
4512 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4513 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4514
4515 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4516 // Stop if it is a clone
4517 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4518 if (ac->may_modify(toop, igvn)) {
4519 break;
4520 }
4521 }
4522 result = proj_in->in(TypeFunc::Memory);
4523 }
4524 } else if (result->is_MergeMem()) {
4525 MergeMemNode *mmem = result->as_MergeMem();
4526 result = step_through_mergemem(mmem, alias_idx, toop);
4527 if (result == mmem->base_memory()) {
4528 // Didn't find instance memory, search through general slice recursively.
4529 result = mmem->memory_at(C->get_general_index(alias_idx));
4530 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4531 if (C->failing()) {
4532 return nullptr;
4533 }
4534 mmem->set_memory_at(alias_idx, result);
4535 }
4536 } else if (result->is_Phi() &&
4537 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4538 Node *un = result->as_Phi()->unique_input(igvn);
4539 if (un != nullptr) {
4540 orig_phis.append_if_missing(result->as_Phi());
4541 result = un;
4542 } else {
4543 break;
4544 }
4545 } else if (result->is_ClearArray()) {
4546 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4547 // Can not bypass initialization of the instance
4548 // we are looking for.
4549 break;
4550 }
4551 // Otherwise skip it (the call updated 'result' value).
4552 } else if (result->Opcode() == Op_SCMemProj) {
4553 Node* mem = result->in(0);
4554 Node* adr = nullptr;
4555 if (mem->is_LoadStore()) {
4556 adr = mem->in(MemNode::Address);
4557 } else {
4558 assert(mem->Opcode() == Op_EncodeISOArray ||
4559 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4560 adr = mem->in(3); // Memory edge corresponds to destination array
4561 }
4562 const Type *at = igvn->type(adr);
4563 if (at != Type::TOP) {
4564 assert(at->isa_ptr() != nullptr, "pointer type required.");
4565 int idx = C->get_alias_index(at->is_ptr());
4566 if (idx == alias_idx) {
4567 // Assert in debug mode
4568 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4569 break; // In product mode return SCMemProj node
4570 }
4571 }
4572 result = mem->in(MemNode::Memory);
4573 } else if (result->Opcode() == Op_StrInflatedCopy) {
4574 Node* adr = result->in(3); // Memory edge corresponds to destination array
4575 const Type *at = igvn->type(adr);
4576 if (at != Type::TOP) {
4577 assert(at->isa_ptr() != nullptr, "pointer type required.");
4578 int idx = C->get_alias_index(at->is_ptr());
4579 if (idx == alias_idx) {
4580 // Assert in debug mode
4581 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4582 break; // In product mode return SCMemProj node
4583 }
4584 }
4585 result = result->in(MemNode::Memory);
4586 }
4587 }
4588 if (result->is_Phi()) {
4589 PhiNode *mphi = result->as_Phi();
4590 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4591 const TypePtr *t = mphi->adr_type();
4592 if (!is_instance) {
4593 // Push all non-instance Phis on the orig_phis worklist to update inputs
4594 // during Phase 4 if needed.
4595 orig_phis.append_if_missing(mphi);
4596 } else if (C->get_alias_index(t) != alias_idx) {
4597 // Create a new Phi with the specified alias index type.
4598 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4599 }
4600 }
4601 // the result is either MemNode, PhiNode, InitializeNode.
4602 return result;
4603 }
4604
4605 //
4606 // Convert the types of non-escaped object to instance types where possible,
4607 // propagate the new type information through the graph, and update memory
4608 // edges and MergeMem inputs to reflect the new type.
4609 //
4610 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4611 // The processing is done in 4 phases:
4612 //
4613 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4614 // types for the CheckCastPP for allocations where possible.
4615 // Propagate the new types through users as follows:
4616 // casts and Phi: push users on alloc_worklist
4617 // AddP: cast Base and Address inputs to the instance type
4618 // push any AddP users on alloc_worklist and push any memnode
4619 // users onto memnode_worklist.
4620 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4621 // search the Memory chain for a store with the appropriate type
4622 // address type. If a Phi is found, create a new version with
4623 // the appropriate memory slices from each of the Phi inputs.
4624 // For stores, process the users as follows:
4625 // MemNode: push on memnode_worklist
4626 // MergeMem: push on mergemem_worklist
4627 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4628 // moving the first node encountered of each instance type to the
4629 // the input corresponding to its alias index.
4630 // appropriate memory slice.
4631 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4632 //
4633 // In the following example, the CheckCastPP nodes are the cast of allocation
4634 // results and the allocation of node 29 is non-escaped and eligible to be an
4635 // instance type.
4636 //
4637 // We start with:
4638 //
4639 // 7 Parm #memory
4640 // 10 ConI "12"
4641 // 19 CheckCastPP "Foo"
4642 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4643 // 29 CheckCastPP "Foo"
4644 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4645 //
4646 // 40 StoreP 25 7 20 ... alias_index=4
4647 // 50 StoreP 35 40 30 ... alias_index=4
4648 // 60 StoreP 45 50 20 ... alias_index=4
4649 // 70 LoadP _ 60 30 ... alias_index=4
4650 // 80 Phi 75 50 60 Memory alias_index=4
4651 // 90 LoadP _ 80 30 ... alias_index=4
4652 // 100 LoadP _ 80 20 ... alias_index=4
4653 //
4654 //
4655 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4656 // and creating a new alias index for node 30. This gives:
4657 //
4658 // 7 Parm #memory
4659 // 10 ConI "12"
4660 // 19 CheckCastPP "Foo"
4661 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4662 // 29 CheckCastPP "Foo" iid=24
4663 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4664 //
4665 // 40 StoreP 25 7 20 ... alias_index=4
4666 // 50 StoreP 35 40 30 ... alias_index=6
4667 // 60 StoreP 45 50 20 ... alias_index=4
4668 // 70 LoadP _ 60 30 ... alias_index=6
4669 // 80 Phi 75 50 60 Memory alias_index=4
4670 // 90 LoadP _ 80 30 ... alias_index=6
4671 // 100 LoadP _ 80 20 ... alias_index=4
4672 //
4673 // In phase 2, new memory inputs are computed for the loads and stores,
4674 // And a new version of the phi is created. In phase 4, the inputs to
4675 // node 80 are updated and then the memory nodes are updated with the
4676 // values computed in phase 2. This results in:
4677 //
4678 // 7 Parm #memory
4679 // 10 ConI "12"
4680 // 19 CheckCastPP "Foo"
4681 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4682 // 29 CheckCastPP "Foo" iid=24
4683 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4684 //
4685 // 40 StoreP 25 7 20 ... alias_index=4
4686 // 50 StoreP 35 7 30 ... alias_index=6
4687 // 60 StoreP 45 40 20 ... alias_index=4
4688 // 70 LoadP _ 50 30 ... alias_index=6
4689 // 80 Phi 75 40 60 Memory alias_index=4
4690 // 120 Phi 75 50 50 Memory alias_index=6
4691 // 90 LoadP _ 120 30 ... alias_index=6
4692 // 100 LoadP _ 80 20 ... alias_index=4
4693 //
4694 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4695 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4696 GrowableArray<MergeMemNode*> &mergemem_worklist,
4697 Unique_Node_List &reducible_merges) {
4698 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4699 GrowableArray<Node *> memnode_worklist;
4700 GrowableArray<PhiNode *> orig_phis;
4701 PhaseIterGVN *igvn = _igvn;
4702 uint new_index_start = (uint) _compile->num_alias_types();
4703 VectorSet visited;
4704 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4705
4706 // Phase 1: Process possible allocations from alloc_worklist.
4707 // Create instance types for the CheckCastPP for allocations where possible.
4708 //
4709 // (Note: don't forget to change the order of the second AddP node on
4710 // the alloc_worklist if the order of the worklist processing is changed,
4711 // see the comment in find_second_addp().)
4712 //
4713 while (alloc_worklist.length() != 0) {
4714 Node *n = alloc_worklist.pop();
4715 uint ni = n->_idx;
4716 if (n->is_Call()) {
4717 CallNode *alloc = n->as_Call();
4718 // copy escape information to call node
4719 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4720 PointsToNode::EscapeState es = ptn->escape_state();
4721 // We have an allocation or call which returns a Java object,
4722 // see if it is non-escaped.
4723 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4724 continue;
4725 }
4726 // Find CheckCastPP for the allocate or for the return value of a call
4727 n = alloc->result_cast();
4728 if (n == nullptr) { // No uses except Initialize node
4729 if (alloc->is_Allocate()) {
4730 // Set the scalar_replaceable flag for allocation
4731 // so it could be eliminated if it has no uses.
4732 alloc->as_Allocate()->_is_scalar_replaceable = true;
4733 }
4734 continue;
4735 }
4736 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4737 // we could reach here for allocate case if one init is associated with many allocs.
4738 if (alloc->is_Allocate()) {
4739 alloc->as_Allocate()->_is_scalar_replaceable = false;
4740 }
4741 continue;
4742 }
4743
4744 // The inline code for Object.clone() casts the allocation result to
4745 // java.lang.Object and then to the actual type of the allocated
4746 // object. Detect this case and use the second cast.
4747 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4748 // the allocation result is cast to java.lang.Object and then
4749 // to the actual Array type.
4750 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4751 && (alloc->is_AllocateArray() ||
4752 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4753 Node *cast2 = nullptr;
4754 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4755 Node *use = n->fast_out(i);
4756 if (use->is_CheckCastPP()) {
4757 cast2 = use;
4758 break;
4759 }
4760 }
4761 if (cast2 != nullptr) {
4762 n = cast2;
4763 } else {
4764 // Non-scalar replaceable if the allocation type is unknown statically
4765 // (reflection allocation), the object can't be restored during
4766 // deoptimization without precise type.
4767 continue;
4768 }
4769 }
4770
4771 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4772 if (t == nullptr) {
4773 continue; // not a TypeOopPtr
4774 }
4775 if (!t->klass_is_exact()) {
4776 continue; // not an unique type
4777 }
4778 if (alloc->is_Allocate()) {
4779 // Set the scalar_replaceable flag for allocation
4780 // so it could be eliminated.
4781 alloc->as_Allocate()->_is_scalar_replaceable = true;
4782 }
4783 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4784 // in order for an object to be scalar-replaceable, it must be:
4785 // - a direct allocation (not a call returning an object)
4786 // - non-escaping
4787 // - eligible to be a unique type
4788 // - not determined to be ineligible by escape analysis
4789 set_map(alloc, n);
4790 set_map(n, alloc);
4791 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4792 igvn->hash_delete(n);
4793 igvn->set_type(n, tinst);
4794 n->raise_bottom_type(tinst);
4795 igvn->hash_insert(n);
4796 record_for_optimizer(n);
4797 // Allocate an alias index for the header fields. Accesses to
4798 // the header emitted during macro expansion wouldn't have
4799 // correct memory state otherwise.
4800 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4801 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4802 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4803 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4804 InitializeNode* init = alloc->as_Allocate()->initialization();
4805 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4806 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4807 const TypePtr* adr_type = proj->adr_type();
4808 const TypePtr* new_adr_type = tinst->with_offset(adr_type->offset());
4809 if (adr_type->isa_aryptr()) {
4810 // In the case of a flat inline type array, each field has its own slice so we need a
4811 // NarrowMemProj for each field of the flat array elements
4812 new_adr_type = new_adr_type->is_aryptr()->with_field_offset(adr_type->is_aryptr()->field_offset().get());
4813 }
4814 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4815 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4816 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4817 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4818 igvn->set_type(new_proj, new_proj->bottom_type());
4819 record_for_optimizer(new_proj);
4820 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4821 }
4822 };
4823 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4824
4825 // First, put on the worklist all Field edges from Connection Graph
4826 // which is more accurate than putting immediate users from Ideal Graph.
4827 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4828 PointsToNode* tgt = e.get();
4829 if (tgt->is_Arraycopy()) {
4830 continue;
4831 }
4832 Node* use = tgt->ideal_node();
4833 assert(tgt->is_Field() && use->is_AddP(),
4834 "only AddP nodes are Field edges in CG");
4835 if (use->outcnt() > 0) { // Don't process dead nodes
4836 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4837 if (addp2 != nullptr) {
4838 assert(alloc->is_AllocateArray(),"array allocation was expected");
4839 alloc_worklist.append_if_missing(addp2);
4840 }
4841 alloc_worklist.append_if_missing(use);
4842 }
4843 }
4844
4845 // An allocation may have an Initialize which has raw stores. Scan
4846 // the users of the raw allocation result and push AddP users
4847 // on alloc_worklist.
4848 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4849 assert (raw_result != nullptr, "must have an allocation result");
4850 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4851 Node *use = raw_result->fast_out(i);
4852 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4853 Node* addp2 = find_second_addp(use, raw_result);
4854 if (addp2 != nullptr) {
4855 assert(alloc->is_AllocateArray(),"array allocation was expected");
4856 alloc_worklist.append_if_missing(addp2);
4857 }
4858 alloc_worklist.append_if_missing(use);
4859 } else if (use->is_MemBar()) {
4860 memnode_worklist.append_if_missing(use);
4861 }
4862 }
4863 }
4864 } else if (n->is_AddP()) {
4865 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4866 // This AddP will go away when we reduce the Phi
4867 continue;
4868 }
4869 Node* addp_base = get_addp_base(n);
4870 JavaObjectNode* jobj = unique_java_object(addp_base);
4871 if (jobj == nullptr || jobj == phantom_obj) {
4872 #ifdef ASSERT
4873 ptnode_adr(get_addp_base(n)->_idx)->dump();
4874 ptnode_adr(n->_idx)->dump();
4875 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4876 #endif
4877 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4878 return;
4879 }
4880 Node *base = get_map(jobj->idx()); // CheckCastPP node
4881 if (!split_AddP(n, base)) continue; // wrong type from dead path
4882 } else if (n->is_Phi() ||
4883 n->is_CheckCastPP() ||
4884 n->is_EncodeP() ||
4885 n->is_DecodeN() ||
4886 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4887 if (visited.test_set(n->_idx)) {
4888 assert(n->is_Phi(), "loops only through Phi's");
4889 continue; // already processed
4890 }
4891 // Reducible Phi's will be removed from the graph after split_unique_types
4892 // finishes. For now we just try to split out the SR inputs of the merge.
4893 Node* parent = n->in(1);
4894 if (reducible_merges.member(n)) {
4895 reduce_phi(n->as_Phi(), alloc_worklist);
4896 #ifdef ASSERT
4897 if (VerifyReduceAllocationMerges) {
4898 reduced_merges.push(n);
4899 }
4900 #endif
4901 continue;
4902 } else if (reducible_merges.member(parent)) {
4903 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4904 // part of reduce_merge.
4905 continue;
4906 }
4907 JavaObjectNode* jobj = unique_java_object(n);
4908 if (jobj == nullptr || jobj == phantom_obj) {
4909 #ifdef ASSERT
4910 ptnode_adr(n->_idx)->dump();
4911 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4912 #endif
4913 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4914 return;
4915 } else {
4916 Node *val = get_map(jobj->idx()); // CheckCastPP node
4917 TypeNode *tn = n->as_Type();
4918 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4919 assert(tinst != nullptr && tinst->is_known_instance() &&
4920 tinst->instance_id() == jobj->idx() , "instance type expected.");
4921
4922 const Type *tn_type = igvn->type(tn);
4923 const TypeOopPtr *tn_t;
4924 if (tn_type->isa_narrowoop()) {
4925 tn_t = tn_type->make_ptr()->isa_oopptr();
4926 } else {
4927 tn_t = tn_type->isa_oopptr();
4928 }
4929 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4930 if (tn_t->isa_aryptr()) {
4931 // Keep array properties (not flat/null-free)
4932 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
4933 if (tinst == nullptr) {
4934 continue; // Skip dead path with inconsistent properties
4935 }
4936 }
4937 if (tn_type->isa_narrowoop()) {
4938 tn_type = tinst->make_narrowoop();
4939 } else {
4940 tn_type = tinst;
4941 }
4942 igvn->hash_delete(tn);
4943 igvn->set_type(tn, tn_type);
4944 tn->set_type(tn_type);
4945 igvn->hash_insert(tn);
4946 record_for_optimizer(n);
4947 } else {
4948 assert(tn_type == TypePtr::NULL_PTR ||
4949 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4950 "unexpected type");
4951 continue; // Skip dead path with different type
4952 }
4953 }
4954 } else {
4955 DEBUG_ONLY(n->dump();)
4956 assert(false, "EA: unexpected node");
4957 continue;
4958 }
4959 // push allocation's users on appropriate worklist
4960 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4961 Node *use = n->fast_out(i);
4962 if (use->is_Mem() && use->in(MemNode::Address) == n) {
4963 // Load/store to instance's field
4964 memnode_worklist.append_if_missing(use);
4965 } else if (use->is_MemBar()) {
4966 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4967 memnode_worklist.append_if_missing(use);
4968 }
4969 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4970 Node* addp2 = find_second_addp(use, n);
4971 if (addp2 != nullptr) {
4972 alloc_worklist.append_if_missing(addp2);
4973 }
4974 alloc_worklist.append_if_missing(use);
4975 } else if (use->is_Phi() ||
4976 use->is_CheckCastPP() ||
4977 use->is_EncodeNarrowPtr() ||
4978 use->is_DecodeNarrowPtr() ||
4979 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4980 alloc_worklist.append_if_missing(use);
4981 #ifdef ASSERT
4982 } else if (use->is_Mem()) {
4983 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4984 } else if (use->is_MergeMem()) {
4985 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4986 } else if (use->is_SafePoint()) {
4987 // Look for MergeMem nodes for calls which reference unique allocation
4988 // (through CheckCastPP nodes) even for debug info.
4989 Node* m = use->in(TypeFunc::Memory);
4990 if (m->is_MergeMem()) {
4991 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4992 }
4993 } else if (use->Opcode() == Op_EncodeISOArray) {
4994 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4995 // EncodeISOArray overwrites destination array
4996 memnode_worklist.append_if_missing(use);
4997 }
4998 } else if (use->Opcode() == Op_Return) {
4999 // Allocation is referenced by field of returned inline type
5000 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
5001 } else {
5002 uint op = use->Opcode();
5003 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
5004 (use->in(MemNode::Memory) == n)) {
5005 // They overwrite memory edge corresponding to destination array,
5006 memnode_worklist.append_if_missing(use);
5007 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
5008 op == Op_CastP2X ||
5009 op == Op_FastLock || op == Op_AryEq ||
5010 op == Op_StrComp || op == Op_CountPositives ||
5011 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
5012 op == Op_StrEquals || op == Op_VectorizedHashCode ||
5013 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
5014 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
5015 op == Op_ReinterpretS2HF ||
5016 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
5017 n->dump();
5018 use->dump();
5019 assert(false, "EA: missing allocation reference path");
5020 }
5021 #endif
5022 }
5023 }
5024
5025 }
5026
5027 #ifdef ASSERT
5028 if (VerifyReduceAllocationMerges) {
5029 for (uint i = 0; i < reducible_merges.size(); i++) {
5030 Node* phi = reducible_merges.at(i);
5031
5032 if (!reduced_merges.member(phi)) {
5033 phi->dump(2);
5034 phi->dump(-2);
5035 assert(false, "This reducible merge wasn't reduced.");
5036 }
5037
5038 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
5039 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
5040 Node* use = phi->fast_out(j);
5041 if (!use->is_SafePoint() && !use->is_CastPP()) {
5042 phi->dump(2);
5043 phi->dump(-2);
5044 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
5045 }
5046 }
5047 }
5048 }
5049 #endif
5050
5051 // Go over all ArrayCopy nodes and if one of the inputs has a unique
5052 // type, record it in the ArrayCopy node so we know what memory this
5053 // node uses/modified.
5054 for (int next = 0; next < arraycopy_worklist.length(); next++) {
5055 ArrayCopyNode* ac = arraycopy_worklist.at(next);
5056 Node* dest = ac->in(ArrayCopyNode::Dest);
5057 if (dest->is_AddP()) {
5058 dest = get_addp_base(dest);
5059 }
5060 JavaObjectNode* jobj = unique_java_object(dest);
5061 if (jobj != nullptr) {
5062 Node *base = get_map(jobj->idx());
5063 if (base != nullptr) {
5064 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5065 ac->_dest_type = base_t;
5066 }
5067 }
5068 Node* src = ac->in(ArrayCopyNode::Src);
5069 if (src->is_AddP()) {
5070 src = get_addp_base(src);
5071 }
5072 jobj = unique_java_object(src);
5073 if (jobj != nullptr) {
5074 Node* base = get_map(jobj->idx());
5075 if (base != nullptr) {
5076 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5077 ac->_src_type = base_t;
5078 }
5079 }
5080 }
5081
5082 // New alias types were created in split_AddP().
5083 uint new_index_end = (uint) _compile->num_alias_types();
5084
5085 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
5086
5087 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
5088 // compute new values for Memory inputs (the Memory inputs are not
5089 // actually updated until phase 4.)
5090 if (memnode_worklist.length() == 0)
5091 return; // nothing to do
5092 while (memnode_worklist.length() != 0) {
5093 Node *n = memnode_worklist.pop();
5094 if (visited.test_set(n->_idx)) {
5095 continue;
5096 }
5097 if (n->is_Phi() || n->is_ClearArray()) {
5098 // we don't need to do anything, but the users must be pushed
5099 } else if (n->is_MemBar()) { // MemBar nodes
5100 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
5101 // we don't need to do anything, but the users must be pushed
5102 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
5103 if (n == nullptr) {
5104 continue;
5105 }
5106 }
5107 } else if (n->is_CallLeaf()) {
5108 // Runtime calls with narrow memory input (no MergeMem node)
5109 // get the memory projection
5110 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
5111 if (n == nullptr) {
5112 continue;
5113 }
5114 } else if (n->Opcode() == Op_StrInflatedCopy) {
5115 // Check direct uses of StrInflatedCopy.
5116 // It is memory type Node - no special SCMemProj node.
5117 } else if (n->Opcode() == Op_StrCompressedCopy ||
5118 n->Opcode() == Op_EncodeISOArray) {
5119 // get the memory projection
5120 n = n->find_out_with(Op_SCMemProj);
5121 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5122 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
5123 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5124 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
5125 } else if (n->is_Proj()) {
5126 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
5127 } else {
5128 #ifdef ASSERT
5129 if (!n->is_Mem()) {
5130 n->dump();
5131 }
5132 assert(n->is_Mem(), "memory node required.");
5133 #endif
5134 Node *addr = n->in(MemNode::Address);
5135 const Type *addr_t = igvn->type(addr);
5136 if (addr_t == Type::TOP) {
5137 continue;
5138 }
5139 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
5140 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
5141 assert ((uint)alias_idx < new_index_end, "wrong alias index");
5142 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
5143 if (_compile->failing()) {
5144 return;
5145 }
5146 if (mem != n->in(MemNode::Memory)) {
5147 // We delay the memory edge update since we need old one in
5148 // MergeMem code below when instances memory slices are separated.
5149 set_map(n, mem);
5150 }
5151 if (n->is_Load()) {
5152 continue; // don't push users
5153 } else if (n->is_LoadStore()) {
5154 // get the memory projection
5155 n = n->find_out_with(Op_SCMemProj);
5156 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5157 }
5158 }
5159 // push user on appropriate worklist
5160 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5161 Node *use = n->fast_out(i);
5162 if (use->is_Phi() || use->is_ClearArray()) {
5163 memnode_worklist.append_if_missing(use);
5164 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
5165 memnode_worklist.append_if_missing(use);
5166 } else if (use->is_MemBar() || use->is_CallLeaf()) {
5167 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5168 memnode_worklist.append_if_missing(use);
5169 }
5170 } else if (use->is_Proj()) {
5171 assert(n->is_Initialize(), "We only push projections of Initialize");
5172 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
5173 memnode_worklist.append_if_missing(use);
5174 }
5175 #ifdef ASSERT
5176 } else if (use->is_Mem()) {
5177 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
5178 } else if (use->is_MergeMem()) {
5179 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5180 } else if (use->Opcode() == Op_EncodeISOArray) {
5181 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5182 // EncodeISOArray overwrites destination array
5183 memnode_worklist.append_if_missing(use);
5184 }
5185 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
5186 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5187 // store_unknown_inline overwrites destination array
5188 memnode_worklist.append_if_missing(use);
5189 } else {
5190 uint op = use->Opcode();
5191 if ((use->in(MemNode::Memory) == n) &&
5192 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
5193 // They overwrite memory edge corresponding to destination array,
5194 memnode_worklist.append_if_missing(use);
5195 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
5196 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
5197 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
5198 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
5199 n->dump();
5200 use->dump();
5201 assert(false, "EA: missing memory path");
5202 }
5203 #endif
5204 }
5205 }
5206 }
5207
5208 // Phase 3: Process MergeMem nodes from mergemem_worklist.
5209 // Walk each memory slice moving the first node encountered of each
5210 // instance type to the input corresponding to its alias index.
5211 uint length = mergemem_worklist.length();
5212 for( uint next = 0; next < length; ++next ) {
5213 MergeMemNode* nmm = mergemem_worklist.at(next);
5214 assert(!visited.test_set(nmm->_idx), "should not be visited before");
5215 // Note: we don't want to use MergeMemStream here because we only want to
5216 // scan inputs which exist at the start, not ones we add during processing.
5217 // Note 2: MergeMem may already contains instance memory slices added
5218 // during find_inst_mem() call when memory nodes were processed above.
5219 igvn->hash_delete(nmm);
5220 uint nslices = MIN2(nmm->req(), new_index_start);
5221 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
5222 Node* mem = nmm->in(i);
5223 Node* cur = nullptr;
5224 if (mem == nullptr || mem->is_top()) {
5225 continue;
5226 }
5227 // First, update mergemem by moving memory nodes to corresponding slices
5228 // if their type became more precise since this mergemem was created.
5229 while (mem->is_Mem()) {
5230 const Type* at = igvn->type(mem->in(MemNode::Address));
5231 if (at != Type::TOP) {
5232 assert (at->isa_ptr() != nullptr, "pointer type required.");
5233 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
5234 if (idx == i) {
5235 if (cur == nullptr) {
5236 cur = mem;
5237 }
5238 } else {
5239 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
5240 nmm->set_memory_at(idx, mem);
5241 }
5242 }
5243 }
5244 mem = mem->in(MemNode::Memory);
5245 }
5246 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
5247 // Find any instance of the current type if we haven't encountered
5248 // already a memory slice of the instance along the memory chain.
5249 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5250 if((uint)_compile->get_general_index(ni) == i) {
5251 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
5252 if (nmm->is_empty_memory(m)) {
5253 Node* result = find_inst_mem(mem, ni, orig_phis);
5254 if (_compile->failing()) {
5255 return;
5256 }
5257 nmm->set_memory_at(ni, result);
5258 }
5259 }
5260 }
5261 }
5262 // Find the rest of instances values
5263 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5264 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
5265 Node* result = step_through_mergemem(nmm, ni, tinst);
5266 if (result == nmm->base_memory()) {
5267 // Didn't find instance memory, search through general slice recursively.
5268 result = nmm->memory_at(_compile->get_general_index(ni));
5269 result = find_inst_mem(result, ni, orig_phis);
5270 if (_compile->failing()) {
5271 return;
5272 }
5273 nmm->set_memory_at(ni, result);
5274 }
5275 }
5276
5277 // If we have crossed the 3/4 point of max node limit it's too risky
5278 // to continue with EA/SR because we might hit the max node limit.
5279 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
5280 if (_compile->do_reduce_allocation_merges()) {
5281 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
5282 } else if (_invocation > 0) {
5283 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
5284 } else {
5285 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
5286 }
5287 return;
5288 }
5289
5290 igvn->hash_insert(nmm);
5291 record_for_optimizer(nmm);
5292 }
5293
5294 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5295
5296 // Phase 4: Update the inputs of non-instance memory Phis and
5297 // the Memory input of memnodes
5298 // First update the inputs of any non-instance Phi's from
5299 // which we split out an instance Phi. Note we don't have
5300 // to recursively process Phi's encountered on the input memory
5301 // chains as is done in split_memory_phi() since they will
5302 // also be processed here.
5303 for (int j = 0; j < orig_phis.length(); j++) {
5304 PhiNode *phi = orig_phis.at(j);
5305 int alias_idx = _compile->get_alias_index(phi->adr_type());
5306 igvn->hash_delete(phi);
5307 for (uint i = 1; i < phi->req(); i++) {
5308 Node *mem = phi->in(i);
5309 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5310 if (_compile->failing()) {
5311 return;
5312 }
5313 if (mem != new_mem) {
5314 phi->set_req(i, new_mem);
5315 }
5316 }
5317 igvn->hash_insert(phi);
5318 record_for_optimizer(phi);
5319 }
5320
5321 // Update the memory inputs of MemNodes with the value we computed
5322 // in Phase 2 and move stores memory users to corresponding memory slices.
5323 // Disable memory split verification code until the fix for 6984348.
5324 // Currently it produces false negative results since it does not cover all cases.
5325 #if 0 // ifdef ASSERT
5326 visited.Reset();
5327 Node_Stack old_mems(arena, _compile->unique() >> 2);
5328 #endif
5329 for (uint i = 0; i < ideal_nodes.size(); i++) {
5330 Node* n = ideal_nodes.at(i);
5331 Node* nmem = get_map(n->_idx);
5332 assert(nmem != nullptr, "sanity");
5333 if (n->is_Mem()) {
5334 #if 0 // ifdef ASSERT
5335 Node* old_mem = n->in(MemNode::Memory);
5336 if (!visited.test_set(old_mem->_idx)) {
5337 old_mems.push(old_mem, old_mem->outcnt());
5338 }
5339 #endif
5340 assert(n->in(MemNode::Memory) != nmem, "sanity");
5341 if (!n->is_Load()) {
5342 // Move memory users of a store first.
5343 move_inst_mem(n, orig_phis);
5344 }
5345 // Now update memory input
5346 igvn->hash_delete(n);
5347 n->set_req(MemNode::Memory, nmem);
5348 igvn->hash_insert(n);
5349 record_for_optimizer(n);
5350 } else {
5351 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5352 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5353 }
5354 }
5355 #if 0 // ifdef ASSERT
5356 // Verify that memory was split correctly
5357 while (old_mems.is_nonempty()) {
5358 Node* old_mem = old_mems.node();
5359 uint old_cnt = old_mems.index();
5360 old_mems.pop();
5361 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5362 }
5363 #endif
5364 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5365 }
5366
5367 #ifndef PRODUCT
5368 int ConnectionGraph::_no_escape_counter = 0;
5369 int ConnectionGraph::_arg_escape_counter = 0;
5370 int ConnectionGraph::_global_escape_counter = 0;
5371
5372 static const char *node_type_names[] = {
5373 "UnknownType",
5374 "JavaObject",
5375 "LocalVar",
5376 "Field",
5377 "Arraycopy"
5378 };
5379
5380 static const char *esc_names[] = {
5381 "UnknownEscape",
5382 "NoEscape",
5383 "ArgEscape",
5384 "GlobalEscape"
5385 };
5386
5387 const char* PointsToNode::esc_name() const {
5388 return esc_names[(int)escape_state()];
5389 }
5390
5391 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5392 NodeType nt = node_type();
5393 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5394 if (print_state) {
5395 EscapeState es = escape_state();
5396 EscapeState fields_es = fields_escape_state();
5397 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5398 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5399 out->print("NSR ");
5400 }
5401 }
5402 }
5403
5404 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5405 dump_header(print_state, out);
5406 if (is_Field()) {
5407 FieldNode* f = (FieldNode*)this;
5408 if (f->is_oop()) {
5409 out->print("oop ");
5410 }
5411 if (f->offset() > 0) {
5412 out->print("+%d ", f->offset());
5413 }
5414 out->print("(");
5415 for (BaseIterator i(f); i.has_next(); i.next()) {
5416 PointsToNode* b = i.get();
5417 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5418 }
5419 out->print(" )");
5420 }
5421 out->print("[");
5422 for (EdgeIterator i(this); i.has_next(); i.next()) {
5423 PointsToNode* e = i.get();
5424 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5425 }
5426 out->print(" [");
5427 for (UseIterator i(this); i.has_next(); i.next()) {
5428 PointsToNode* u = i.get();
5429 bool is_base = false;
5430 if (PointsToNode::is_base_use(u)) {
5431 is_base = true;
5432 u = PointsToNode::get_use_node(u)->as_Field();
5433 }
5434 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5435 }
5436 out->print(" ]] ");
5437 if (_node == nullptr) {
5438 out->print("<null>%s", newline ? "\n" : "");
5439 } else {
5440 _node->dump(newline ? "\n" : "", false, out);
5441 }
5442 }
5443
5444 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5445 bool first = true;
5446 int ptnodes_length = ptnodes_worklist.length();
5447 for (int i = 0; i < ptnodes_length; i++) {
5448 PointsToNode *ptn = ptnodes_worklist.at(i);
5449 if (ptn == nullptr || !ptn->is_JavaObject()) {
5450 continue;
5451 }
5452 PointsToNode::EscapeState es = ptn->escape_state();
5453 if ((es != PointsToNode::NoEscape) && !Verbose) {
5454 continue;
5455 }
5456 Node* n = ptn->ideal_node();
5457 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5458 n->as_CallStaticJava()->is_boxing_method())) {
5459 if (first) {
5460 tty->cr();
5461 tty->print("======== Connection graph for ");
5462 _compile->method()->print_short_name();
5463 tty->cr();
5464 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5465 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5466 tty->cr();
5467 first = false;
5468 }
5469 ptn->dump();
5470 // Print all locals and fields which reference this allocation
5471 for (UseIterator j(ptn); j.has_next(); j.next()) {
5472 PointsToNode* use = j.get();
5473 if (use->is_LocalVar()) {
5474 use->dump(Verbose);
5475 } else if (Verbose) {
5476 use->dump();
5477 }
5478 }
5479 tty->cr();
5480 }
5481 }
5482 }
5483
5484 void ConnectionGraph::print_statistics() {
5485 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));
5486 }
5487
5488 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5489 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5490 return;
5491 }
5492 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5493 JavaObjectNode* ptn = java_objects_worklist.at(next);
5494 if (ptn->ideal_node()->is_Allocate()) {
5495 if (ptn->escape_state() == PointsToNode::NoEscape) {
5496 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5497 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5498 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5499 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5500 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5501 } else {
5502 assert(false, "Unexpected Escape State");
5503 }
5504 }
5505 }
5506 }
5507
5508 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5509 if (_compile->directive()->TraceEscapeAnalysisOption) {
5510 assert(ptn != nullptr, "should not be null");
5511 assert(reason != nullptr, "should not be null");
5512 ptn->dump_header(true);
5513 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5514 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5515 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5516 }
5517 }
5518
5519 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5520 if (_compile->directive()->TraceEscapeAnalysisOption) {
5521 stringStream ss;
5522 ss.print("propagated from: ");
5523 from->dump(true, &ss, false);
5524 return ss.as_string();
5525 } else {
5526 return nullptr;
5527 }
5528 }
5529
5530 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5531 if (_compile->directive()->TraceEscapeAnalysisOption) {
5532 stringStream ss;
5533 ss.print("escapes as arg to:");
5534 call->dump("", false, &ss);
5535 return ss.as_string();
5536 } else {
5537 return nullptr;
5538 }
5539 }
5540
5541 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5542 if (_compile->directive()->TraceEscapeAnalysisOption) {
5543 stringStream ss;
5544 ss.print("is merged with other object: ");
5545 other->dump_header(true, &ss);
5546 return ss.as_string();
5547 } else {
5548 return nullptr;
5549 }
5550 }
5551
5552 #endif
5553
5554 void ConnectionGraph::record_for_optimizer(Node *n) {
5555 _igvn->_worklist.push(n);
5556 _igvn->add_users_to_worklist(n);
5557 }