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 = AddPNode::make_with_base(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(AddPNode::make_with_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 void ConnectionGraph::add_proj(Node* n, Unique_Node_List* delayed_worklist) {
1590 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() && n->in(0)->as_Call()->returns_pointer()) {
1591 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1592 } else if (n->in(0)->is_LoadFlat()) {
1593 // Treat LoadFlat outputs similar to a call return value
1594 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1595 } else if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() && n->bottom_type()->isa_ptr()) {
1596 CallNode* call = n->in(0)->as_Call();
1597 assert(call->tf()->returns_inline_type_as_fields(), "");
1598 if (n->as_Proj()->_con == TypeFunc::Parms || !returns_an_argument(call)) {
1599 // either:
1600 // - not an argument returned
1601 // - the returned buffer for a returned scalarized argument
1602 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1603 } else {
1604 add_local_var(n, PointsToNode::NoEscape);
1605 }
1606 }
1607 }
1608
1609 // Populate Connection Graph with PointsTo nodes and create simple
1610 // connection graph edges.
1611 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1612 assert(!_verify, "this method should not be called for verification");
1613 PhaseGVN* igvn = _igvn;
1614 uint n_idx = n->_idx;
1615 PointsToNode* n_ptn = ptnode_adr(n_idx);
1616 if (n_ptn != nullptr) {
1617 return; // No need to redefine PointsTo node during first iteration.
1618 }
1619 int opcode = n->Opcode();
1620 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1621 if (gc_handled) {
1622 return; // Ignore node if already handled by GC.
1623 }
1624
1625 if (n->is_Call()) {
1626 // Arguments to allocation and locking don't escape.
1627 if (n->is_AbstractLock()) {
1628 // Put Lock and Unlock nodes on IGVN worklist to process them during
1629 // first IGVN optimization when escape information is still available.
1630 record_for_optimizer(n);
1631 } else if (n->is_Allocate()) {
1632 add_call_node(n->as_Call());
1633 record_for_optimizer(n);
1634 } else {
1635 if (n->is_CallStaticJava()) {
1636 const char* name = n->as_CallStaticJava()->_name;
1637 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1638 return; // Skip uncommon traps
1639 }
1640 }
1641 // Don't mark as processed since call's arguments have to be processed.
1642 delayed_worklist->push(n);
1643 // Check if a call returns an object.
1644 if ((n->as_Call()->returns_pointer() &&
1645 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1646 (n->is_CallStaticJava() &&
1647 n->as_CallStaticJava()->is_boxing_method())) {
1648 add_call_node(n->as_Call());
1649 } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1650 bool returns_oop = false;
1651 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1652 ProjNode* pn = n->fast_out(i)->as_Proj();
1653 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1654 returns_oop = true;
1655 }
1656 }
1657 if (returns_oop) {
1658 add_call_node(n->as_Call());
1659 }
1660 }
1661 }
1662 return;
1663 }
1664 // Put this check here to process call arguments since some call nodes
1665 // point to phantom_obj.
1666 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1667 return; // Skip predefined nodes.
1668 }
1669 switch (opcode) {
1670 case Op_AddP: {
1671 Node* base = get_addp_base(n);
1672 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1673 // Field nodes are created for all field types. They are used in
1674 // adjust_scalar_replaceable_state() and split_unique_types().
1675 // Note, non-oop fields will have only base edges in Connection
1676 // Graph because such fields are not used for oop loads and stores.
1677 int offset = address_offset(n, igvn);
1678 add_field(n, PointsToNode::NoEscape, offset);
1679 if (ptn_base == nullptr) {
1680 delayed_worklist->push(n); // Process it later.
1681 } else {
1682 n_ptn = ptnode_adr(n_idx);
1683 add_base(n_ptn->as_Field(), ptn_base);
1684 }
1685 break;
1686 }
1687 case Op_CastX2P:
1688 case Op_CastI2N: {
1689 map_ideal_node(n, phantom_obj);
1690 break;
1691 }
1692 case Op_InlineType:
1693 case Op_CastPP:
1694 case Op_CheckCastPP:
1695 case Op_EncodeP:
1696 case Op_DecodeN:
1697 case Op_EncodePKlass:
1698 case Op_DecodeNKlass: {
1699 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1700 break;
1701 }
1702 case Op_CMoveP: {
1703 add_local_var(n, PointsToNode::NoEscape);
1704 // Do not add edges during first iteration because some could be
1705 // not defined yet.
1706 delayed_worklist->push(n);
1707 break;
1708 }
1709 case Op_ConP:
1710 case Op_ConN:
1711 case Op_ConNKlass: {
1712 // assume all oop constants globally escape except for null
1713 PointsToNode::EscapeState es;
1714 const Type* t = igvn->type(n);
1715 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1716 es = PointsToNode::NoEscape;
1717 } else {
1718 es = PointsToNode::GlobalEscape;
1719 }
1720 PointsToNode* ptn_con = add_java_object(n, es);
1721 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1722 break;
1723 }
1724 case Op_CreateEx: {
1725 // assume that all exception objects globally escape
1726 map_ideal_node(n, phantom_obj);
1727 break;
1728 }
1729 case Op_LoadKlass:
1730 case Op_LoadNKlass: {
1731 // Unknown class is loaded
1732 map_ideal_node(n, phantom_obj);
1733 break;
1734 }
1735 case Op_LoadP:
1736 case Op_LoadN: {
1737 add_objload_to_connection_graph(n, delayed_worklist);
1738 break;
1739 }
1740 case Op_Parm: {
1741 map_ideal_node(n, phantom_obj);
1742 break;
1743 }
1744 case Op_PartialSubtypeCheck: {
1745 // Produces Null or notNull and is used in only in CmpP so
1746 // phantom_obj could be used.
1747 map_ideal_node(n, phantom_obj); // Result is unknown
1748 break;
1749 }
1750 case Op_Phi: {
1751 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1752 // ThreadLocal has RawPtr type.
1753 const Type* t = n->as_Phi()->type();
1754 if (t->make_ptr() != nullptr) {
1755 add_local_var(n, PointsToNode::NoEscape);
1756 // Do not add edges during first iteration because some could be
1757 // not defined yet.
1758 delayed_worklist->push(n);
1759 }
1760 break;
1761 }
1762 case Op_LoadFlat:
1763 // Treat LoadFlat similar to an unknown call that receives nothing and produces its results
1764 map_ideal_node(n, phantom_obj);
1765 break;
1766 case Op_StoreFlat:
1767 // Treat StoreFlat similar to a call that escapes the stored flattened fields
1768 delayed_worklist->push(n);
1769 break;
1770 case Op_Proj: {
1771 // we are only interested in the oop result projection from a call
1772 add_proj(n, delayed_worklist);
1773 break;
1774 }
1775 case Op_Rethrow: // Exception object escapes
1776 case Op_Return: {
1777 if (n->req() > TypeFunc::Parms &&
1778 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1779 // Treat Return value as LocalVar with GlobalEscape escape state.
1780 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1781 }
1782 break;
1783 }
1784 case Op_CompareAndExchangeP:
1785 case Op_CompareAndExchangeN:
1786 case Op_GetAndSetP:
1787 case Op_GetAndSetN: {
1788 add_objload_to_connection_graph(n, delayed_worklist);
1789 // fall-through
1790 }
1791 case Op_StoreP:
1792 case Op_StoreN:
1793 case Op_StoreNKlass:
1794 case Op_WeakCompareAndSwapP:
1795 case Op_WeakCompareAndSwapN:
1796 case Op_CompareAndSwapP:
1797 case Op_CompareAndSwapN: {
1798 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1799 break;
1800 }
1801 case Op_AryEq:
1802 case Op_CountPositives:
1803 case Op_StrComp:
1804 case Op_StrEquals:
1805 case Op_StrIndexOf:
1806 case Op_StrIndexOfChar:
1807 case Op_StrInflatedCopy:
1808 case Op_StrCompressedCopy:
1809 case Op_VectorizedHashCode:
1810 case Op_EncodeISOArray: {
1811 add_local_var(n, PointsToNode::ArgEscape);
1812 delayed_worklist->push(n); // Process it later.
1813 break;
1814 }
1815 case Op_ThreadLocal: {
1816 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1817 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1818 break;
1819 }
1820 case Op_Blackhole: {
1821 // All blackhole pointer arguments are globally escaping.
1822 // Only do this if there is at least one pointer argument.
1823 // Do not add edges during first iteration because some could be
1824 // not defined yet, defer to final step.
1825 for (uint i = 0; i < n->req(); i++) {
1826 Node* in = n->in(i);
1827 if (in != nullptr) {
1828 const Type* at = _igvn->type(in);
1829 if (!at->isa_ptr()) continue;
1830
1831 add_local_var(n, PointsToNode::GlobalEscape);
1832 delayed_worklist->push(n);
1833 break;
1834 }
1835 }
1836 break;
1837 }
1838 default:
1839 ; // Do nothing for nodes not related to EA.
1840 }
1841 return;
1842 }
1843
1844 // Add final simple edges to graph.
1845 void ConnectionGraph::add_final_edges(Node *n) {
1846 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1847 #ifdef ASSERT
1848 if (_verify && n_ptn->is_JavaObject())
1849 return; // This method does not change graph for JavaObject.
1850 #endif
1851
1852 if (n->is_Call()) {
1853 process_call_arguments(n->as_Call());
1854 return;
1855 }
1856 assert(n->is_Store() || n->is_LoadStore() || n->is_StoreFlat() ||
1857 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1858 "node should be registered already");
1859 int opcode = n->Opcode();
1860 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1861 if (gc_handled) {
1862 return; // Ignore node if already handled by GC.
1863 }
1864 switch (opcode) {
1865 case Op_AddP: {
1866 Node* base = get_addp_base(n);
1867 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1868 assert(ptn_base != nullptr, "field's base should be registered");
1869 add_base(n_ptn->as_Field(), ptn_base);
1870 break;
1871 }
1872 case Op_InlineType:
1873 case Op_CastPP:
1874 case Op_CheckCastPP:
1875 case Op_EncodeP:
1876 case Op_DecodeN:
1877 case Op_EncodePKlass:
1878 case Op_DecodeNKlass: {
1879 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1880 break;
1881 }
1882 case Op_CMoveP: {
1883 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1884 Node* in = n->in(i);
1885 if (in == nullptr) {
1886 continue; // ignore null
1887 }
1888 Node* uncast_in = in->uncast();
1889 if (uncast_in->is_top() || uncast_in == n) {
1890 continue; // ignore top or inputs which go back this node
1891 }
1892 PointsToNode* ptn = ptnode_adr(in->_idx);
1893 assert(ptn != nullptr, "node should be registered");
1894 add_edge(n_ptn, ptn);
1895 }
1896 break;
1897 }
1898 case Op_LoadP:
1899 case Op_LoadN: {
1900 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1901 // ThreadLocal has RawPtr type.
1902 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1903 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1904 break;
1905 }
1906 case Op_Phi: {
1907 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1908 // ThreadLocal has RawPtr type.
1909 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1910 for (uint i = 1; i < n->req(); i++) {
1911 Node* in = n->in(i);
1912 if (in == nullptr) {
1913 continue; // ignore null
1914 }
1915 Node* uncast_in = in->uncast();
1916 if (uncast_in->is_top() || uncast_in == n) {
1917 continue; // ignore top or inputs which go back this node
1918 }
1919 PointsToNode* ptn = ptnode_adr(in->_idx);
1920 assert(ptn != nullptr, "node should be registered");
1921 add_edge(n_ptn, ptn);
1922 }
1923 break;
1924 }
1925 case Op_StoreFlat: {
1926 // StoreFlat globally escapes its stored flattened fields
1927 InlineTypeNode* value = n->as_StoreFlat()->value();
1928 ciInlineKlass* vk = _igvn->type(value)->inline_klass();
1929 for (int i = 0; i < vk->nof_nonstatic_fields(); i++) {
1930 ciField* field = vk->nonstatic_field_at(i);
1931 if (field->type()->is_primitive_type()) {
1932 continue;
1933 }
1934
1935 Node* field_value = value->field_value_by_offset(field->offset_in_bytes(), true);
1936 PointsToNode* field_value_ptn = ptnode_adr(field_value->_idx);
1937 set_escape_state(field_value_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "store into a flat field"));
1938 }
1939 break;
1940 }
1941 case Op_Proj: {
1942 add_proj(n, nullptr);
1943 break;
1944 }
1945 case Op_Rethrow: // Exception object escapes
1946 case Op_Return: {
1947 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1948 "Unexpected node type");
1949 // Treat Return value as LocalVar with GlobalEscape escape state.
1950 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1951 break;
1952 }
1953 case Op_CompareAndExchangeP:
1954 case Op_CompareAndExchangeN:
1955 case Op_GetAndSetP:
1956 case Op_GetAndSetN:{
1957 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1958 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1959 // fall-through
1960 }
1961 case Op_CompareAndSwapP:
1962 case Op_CompareAndSwapN:
1963 case Op_WeakCompareAndSwapP:
1964 case Op_WeakCompareAndSwapN:
1965 case Op_StoreP:
1966 case Op_StoreN:
1967 case Op_StoreNKlass:{
1968 add_final_edges_unsafe_access(n, opcode);
1969 break;
1970 }
1971 case Op_VectorizedHashCode:
1972 case Op_AryEq:
1973 case Op_CountPositives:
1974 case Op_StrComp:
1975 case Op_StrEquals:
1976 case Op_StrIndexOf:
1977 case Op_StrIndexOfChar:
1978 case Op_StrInflatedCopy:
1979 case Op_StrCompressedCopy:
1980 case Op_EncodeISOArray: {
1981 // char[]/byte[] arrays passed to string intrinsic do not escape but
1982 // they are not scalar replaceable. Adjust escape state for them.
1983 // Start from in(2) edge since in(1) is memory edge.
1984 for (uint i = 2; i < n->req(); i++) {
1985 Node* adr = n->in(i);
1986 const Type* at = _igvn->type(adr);
1987 if (!adr->is_top() && at->isa_ptr()) {
1988 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1989 at->isa_ptr() != nullptr, "expecting a pointer");
1990 if (adr->is_AddP()) {
1991 adr = get_addp_base(adr);
1992 }
1993 PointsToNode* ptn = ptnode_adr(adr->_idx);
1994 assert(ptn != nullptr, "node should be registered");
1995 add_edge(n_ptn, ptn);
1996 }
1997 }
1998 break;
1999 }
2000 case Op_Blackhole: {
2001 // All blackhole pointer arguments are globally escaping.
2002 for (uint i = 0; i < n->req(); i++) {
2003 Node* in = n->in(i);
2004 if (in != nullptr) {
2005 const Type* at = _igvn->type(in);
2006 if (!at->isa_ptr()) continue;
2007
2008 if (in->is_AddP()) {
2009 in = get_addp_base(in);
2010 }
2011
2012 PointsToNode* ptn = ptnode_adr(in->_idx);
2013 assert(ptn != nullptr, "should be defined already");
2014 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
2015 add_edge(n_ptn, ptn);
2016 }
2017 }
2018 break;
2019 }
2020 default: {
2021 // This method should be called only for EA specific nodes which may
2022 // miss some edges when they were created.
2023 #ifdef ASSERT
2024 n->dump(1);
2025 #endif
2026 guarantee(false, "unknown node");
2027 }
2028 }
2029 return;
2030 }
2031
2032 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
2033 Node* adr = n->in(MemNode::Address);
2034 const Type* adr_type = _igvn->type(adr);
2035 adr_type = adr_type->make_ptr();
2036 if (adr_type == nullptr) {
2037 return; // skip dead nodes
2038 }
2039 if (adr_type->isa_oopptr()
2040 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2041 && adr_type == TypeRawPtr::NOTNULL
2042 && is_captured_store_address(adr))) {
2043 delayed_worklist->push(n); // Process it later.
2044 #ifdef ASSERT
2045 assert (adr->is_AddP(), "expecting an AddP");
2046 if (adr_type == TypeRawPtr::NOTNULL) {
2047 // Verify a raw address for a store captured by Initialize node.
2048 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2049 assert(offs != Type::OffsetBot, "offset must be a constant");
2050 }
2051 #endif
2052 } else {
2053 // Ignore copy the displaced header to the BoxNode (OSR compilation).
2054 if (adr->is_BoxLock()) {
2055 return;
2056 }
2057 // Stored value escapes in unsafe access.
2058 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2059 delayed_worklist->push(n); // Process unsafe access later.
2060 return;
2061 }
2062 #ifdef ASSERT
2063 n->dump(1);
2064 assert(false, "not unsafe");
2065 #endif
2066 }
2067 }
2068
2069 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
2070 Node* adr = n->in(MemNode::Address);
2071 const Type *adr_type = _igvn->type(adr);
2072 adr_type = adr_type->make_ptr();
2073 #ifdef ASSERT
2074 if (adr_type == nullptr) {
2075 n->dump(1);
2076 assert(adr_type != nullptr, "dead node should not be on list");
2077 return true;
2078 }
2079 #endif
2080
2081 if (adr_type->isa_oopptr()
2082 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2083 && adr_type == TypeRawPtr::NOTNULL
2084 && is_captured_store_address(adr))) {
2085 // Point Address to Value
2086 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2087 assert(adr_ptn != nullptr &&
2088 adr_ptn->as_Field()->is_oop(), "node should be registered");
2089 Node* val = n->in(MemNode::ValueIn);
2090 PointsToNode* ptn = ptnode_adr(val->_idx);
2091 assert(ptn != nullptr, "node should be registered");
2092 add_edge(adr_ptn, ptn);
2093 return true;
2094 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2095 // Stored value escapes in unsafe access.
2096 Node* val = n->in(MemNode::ValueIn);
2097 PointsToNode* ptn = ptnode_adr(val->_idx);
2098 assert(ptn != nullptr, "node should be registered");
2099 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2100 // Add edge to object for unsafe access with offset.
2101 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2102 assert(adr_ptn != nullptr, "node should be registered");
2103 if (adr_ptn->is_Field()) {
2104 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2105 add_edge(adr_ptn, ptn);
2106 }
2107 return true;
2108 }
2109 #ifdef ASSERT
2110 n->dump(1);
2111 assert(false, "not unsafe");
2112 #endif
2113 return false;
2114 }
2115
2116 // Iterate over the domains for the scalarized and non scalarized calling conventions: Only move to the next element
2117 // in the non scalarized calling convention once all elements of the scalarized calling convention for that parameter
2118 // have been iterated over. So (ignoring hidden arguments such as the null marker) iterating over:
2119 // value class MyValue {
2120 // int f1;
2121 // float f2;
2122 // }
2123 // void m(Object o, MyValue v, int i)
2124 // produces the pairs:
2125 // (Object, Object), (Myvalue, int), (MyValue, float), (int, int)
2126 class DomainIterator : public StackObj {
2127 private:
2128 const TypeTuple* _domain;
2129 const TypeTuple* _domain_cc;
2130 const GrowableArray<SigEntry>* _sig_cc;
2131
2132 uint _i_domain;
2133 uint _i_domain_cc;
2134 int _i_sig_cc;
2135 uint _depth;
2136 uint _first_field_pos;
2137 const bool _is_static;
2138
2139 void next_helper() {
2140 if (_sig_cc == nullptr) {
2141 return;
2142 }
2143 BasicType prev_bt = _i_sig_cc > 0 ? _sig_cc->at(_i_sig_cc-1)._bt : T_ILLEGAL;
2144 BasicType prev_prev_bt = _i_sig_cc > 1 ? _sig_cc->at(_i_sig_cc-2)._bt : T_ILLEGAL;
2145 while (_i_sig_cc < _sig_cc->length()) {
2146 BasicType bt = _sig_cc->at(_i_sig_cc)._bt;
2147 assert(bt != T_VOID || _sig_cc->at(_i_sig_cc-1)._bt == prev_bt, "incorrect prev bt");
2148 if (bt == T_METADATA) {
2149 if (_depth == 0) {
2150 _first_field_pos = _i_domain_cc;
2151 }
2152 _depth++;
2153 } else if (bt == T_VOID && (prev_bt != T_LONG && prev_bt != T_DOUBLE)) {
2154 _depth--;
2155 if (_depth == 0) {
2156 _i_domain++;
2157 }
2158 } else if (bt == T_OBJECT && prev_bt == T_METADATA && (_is_static || _i_domain > 0) && _sig_cc->at(_i_sig_cc)._offset == 0) {
2159 assert(_sig_cc->at(_i_sig_cc)._vt_oop, "buffer expected right after T_METADATA");
2160 assert(_depth == 1, "only root value has buffer");
2161 _i_domain_cc++;
2162 _first_field_pos = _i_domain_cc;
2163 } else if (bt == T_BOOLEAN && prev_prev_bt == T_METADATA && (_is_static || _i_domain > 0) && _sig_cc->at(_i_sig_cc)._offset == -1) {
2164 assert(_sig_cc->at(_i_sig_cc)._null_marker, "null marker expected right after T_METADATA");
2165 assert(_depth == 1, "only root value null marker");
2166 _i_domain_cc++;
2167 _first_field_pos = _i_domain_cc;
2168 } else {
2169 return;
2170 }
2171 prev_prev_bt = prev_bt;
2172 prev_bt = bt;
2173 _i_sig_cc++;
2174 }
2175 }
2176
2177 public:
2178
2179 DomainIterator(CallJavaNode* call) :
2180 _domain(call->tf()->domain_sig()),
2181 _domain_cc(call->tf()->domain_cc()),
2182 _sig_cc(call->method()->get_sig_cc()),
2183 _i_domain(TypeFunc::Parms),
2184 _i_domain_cc(TypeFunc::Parms),
2185 _i_sig_cc(0),
2186 _depth(0),
2187 _first_field_pos(0),
2188 _is_static(call->method()->is_static()) {
2189 next_helper();
2190 }
2191
2192 bool has_next() const {
2193 assert(_sig_cc == nullptr || (_i_sig_cc < _sig_cc->length()) == (_i_domain < _domain->cnt()), "should reach end in sync");
2194 assert((_i_domain < _domain->cnt()) == (_i_domain_cc < _domain_cc->cnt()), "should reach end in sync");
2195 return _i_domain < _domain->cnt();
2196 }
2197
2198 void next() {
2199 assert(_depth != 0 || _domain->field_at(_i_domain) == _domain_cc->field_at(_i_domain_cc), "should produce same non scalarized elements");
2200 _i_sig_cc++;
2201 if (_depth == 0) {
2202 _i_domain++;
2203 }
2204 _i_domain_cc++;
2205 next_helper();
2206 }
2207
2208 uint i_domain() const {
2209 return _i_domain;
2210 }
2211
2212 uint i_domain_cc() const {
2213 return _i_domain_cc;
2214 }
2215
2216 const Type* current_domain() const {
2217 return _domain->field_at(_i_domain);
2218 }
2219
2220 const Type* current_domain_cc() const {
2221 return _domain_cc->field_at(_i_domain_cc);
2222 }
2223
2224 uint first_field_pos() const {
2225 assert(_first_field_pos >= TypeFunc::Parms, "not yet updated?");
2226 return _first_field_pos;
2227 }
2228 };
2229
2230 // Determine whether any arguments are returned.
2231 bool ConnectionGraph::returns_an_argument(CallNode* call) {
2232 ciMethod* meth = call->as_CallJava()->method();
2233 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2234 if (call_analyzer == nullptr) {
2235 return false;
2236 }
2237
2238 const TypeTuple* d = call->tf()->domain_sig();
2239 bool ret_arg = false;
2240 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2241 if (d->field_at(i)->isa_ptr() != nullptr &&
2242 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2243 if (meth->is_scalarized_arg(i - TypeFunc::Parms) && !compatible_return(call->as_CallJava(), i)) {
2244 return false;
2245 }
2246 if (call->tf()->returns_inline_type_as_fields() != meth->is_scalarized_arg(i - TypeFunc::Parms)) {
2247 return false;
2248 }
2249 ret_arg = true;
2250 }
2251 }
2252 return ret_arg;
2253 }
2254
2255 void ConnectionGraph::add_call_node(CallNode* call) {
2256 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2257 uint call_idx = call->_idx;
2258 if (call->is_Allocate()) {
2259 Node* k = call->in(AllocateNode::KlassNode);
2260 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2261 assert(kt != nullptr, "TypeKlassPtr required.");
2262 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2263 bool scalar_replaceable = true;
2264 NOT_PRODUCT(const char* nsr_reason = "");
2265 if (call->is_AllocateArray()) {
2266 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2267 es = PointsToNode::GlobalEscape;
2268 } else {
2269 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2270 if (length < 0) {
2271 // Not scalar replaceable if the length is not constant.
2272 scalar_replaceable = false;
2273 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2274 } else if (length > EliminateAllocationArraySizeLimit) {
2275 // Not scalar replaceable if the length is too big.
2276 scalar_replaceable = false;
2277 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2278 }
2279 }
2280 } else { // Allocate instance
2281 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2282 es = PointsToNode::GlobalEscape;
2283 } else {
2284 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2285 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2286 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2287 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2288 !ik->can_be_instantiated() ||
2289 ik->has_finalizer()) {
2290 es = PointsToNode::GlobalEscape;
2291 } else {
2292 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2293 if (nfields > EliminateAllocationFieldsLimit) {
2294 // Not scalar replaceable if there are too many fields.
2295 scalar_replaceable = false;
2296 NOT_PRODUCT(nsr_reason = "has too many fields");
2297 }
2298 }
2299 }
2300 }
2301 add_java_object(call, es);
2302 PointsToNode* ptn = ptnode_adr(call_idx);
2303 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2304 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2305 }
2306 } else if (call->is_CallStaticJava()) {
2307 // Call nodes could be different types:
2308 //
2309 // 1. CallDynamicJavaNode (what happened during call is unknown):
2310 //
2311 // - mapped to GlobalEscape JavaObject node if oop is returned;
2312 //
2313 // - all oop arguments are escaping globally;
2314 //
2315 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2316 //
2317 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2318 //
2319 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2320 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2321 // during call is returned;
2322 // - mapped to ArgEscape LocalVar node pointed to object arguments
2323 // which are returned and does not escape during call;
2324 //
2325 // - oop arguments escaping status is defined by bytecode analysis;
2326 //
2327 // For a static call, we know exactly what method is being called.
2328 // Use bytecode estimator to record whether the call's return value escapes.
2329 ciMethod* meth = call->as_CallJava()->method();
2330 if (meth == nullptr) {
2331 const char* name = call->as_CallStaticJava()->_name;
2332 assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
2333 strncmp(name, "load_unknown_inline", 19) == 0 ||
2334 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "TODO: add failed case check");
2335 // Returns a newly allocated non-escaped object.
2336 add_java_object(call, PointsToNode::NoEscape);
2337 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2338 } else if (meth->is_boxing_method()) {
2339 // Returns boxing object
2340 PointsToNode::EscapeState es;
2341 vmIntrinsics::ID intr = meth->intrinsic_id();
2342 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2343 // It does not escape if object is always allocated.
2344 es = PointsToNode::NoEscape;
2345 } else {
2346 // It escapes globally if object could be loaded from cache.
2347 es = PointsToNode::GlobalEscape;
2348 }
2349 add_java_object(call, es);
2350 if (es == PointsToNode::GlobalEscape) {
2351 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2352 }
2353 } else {
2354 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2355 call_analyzer->copy_dependencies(_compile->dependencies());
2356 if (call_analyzer->is_return_allocated()) {
2357 // Returns a newly allocated non-escaped object, simply
2358 // update dependency information.
2359 // Mark it as NoEscape so that objects referenced by
2360 // it's fields will be marked as NoEscape at least.
2361 add_java_object(call, PointsToNode::NoEscape);
2362 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2363 } else {
2364 // For non scalarized argument/return: add_proj() adds an edge between the return projection and the call,
2365 // process_call_arguments() adds an edge between the call and the argument
2366 // For scalarized argument/return: process_call_arguments() adds an edge between a call projection for a field
2367 // and the argument input to the call for that field. An edge is added between the projection for the returned
2368 // buffer and the call.
2369 if (returns_an_argument(call) && !call->tf()->returns_inline_type_as_fields()) {
2370 // returns non scalarized argument
2371 add_local_var(call, PointsToNode::ArgEscape);
2372 } else {
2373 // Returns unknown object or scalarized argument being returned
2374 map_ideal_node(call, phantom_obj);
2375 }
2376 }
2377 }
2378 } else {
2379 // An other type of call, assume the worst case:
2380 // returned value is unknown and globally escapes.
2381 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2382 map_ideal_node(call, phantom_obj);
2383 }
2384 }
2385
2386 // Check that the return type is compatible with the type of the argument being returned i.e. that there's no cast that
2387 // fails in the method
2388 bool ConnectionGraph::compatible_return(CallJavaNode* call, uint k) {
2389 return call->tf()->domain_sig()->field_at(k)->is_instptr()->instance_klass() == call->tf()->range_sig()->field_at(TypeFunc::Parms)->is_instptr()->instance_klass();
2390 }
2391
2392 void ConnectionGraph::process_call_arguments(CallNode *call) {
2393 bool is_arraycopy = false;
2394 switch (call->Opcode()) {
2395 #ifdef ASSERT
2396 case Op_Allocate:
2397 case Op_AllocateArray:
2398 case Op_Lock:
2399 case Op_Unlock:
2400 assert(false, "should be done already");
2401 break;
2402 #endif
2403 case Op_ArrayCopy:
2404 case Op_CallLeafNoFP:
2405 // Most array copies are ArrayCopy nodes at this point but there
2406 // are still a few direct calls to the copy subroutines (See
2407 // PhaseStringOpts::copy_string())
2408 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2409 call->as_CallLeaf()->is_call_to_arraycopystub();
2410 // fall through
2411 case Op_CallLeafVector:
2412 case Op_CallLeaf: {
2413 // Stub calls, objects do not escape but they are not scale replaceable.
2414 // Adjust escape state for outgoing arguments.
2415 const TypeTuple * d = call->tf()->domain_sig();
2416 bool src_has_oops = false;
2417 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2418 const Type* at = d->field_at(i);
2419 Node *arg = call->in(i);
2420 if (arg == nullptr) {
2421 continue;
2422 }
2423 const Type *aat = _igvn->type(arg);
2424 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2425 continue;
2426 }
2427 if (arg->is_AddP()) {
2428 //
2429 // The inline_native_clone() case when the arraycopy stub is called
2430 // after the allocation before Initialize and CheckCastPP nodes.
2431 // Or normal arraycopy for object arrays case.
2432 //
2433 // Set AddP's base (Allocate) as not scalar replaceable since
2434 // pointer to the base (with offset) is passed as argument.
2435 //
2436 arg = get_addp_base(arg);
2437 }
2438 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2439 assert(arg_ptn != nullptr, "should be registered");
2440 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2441 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2442 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2443 aat->isa_ptr() != nullptr, "expecting an Ptr");
2444 bool arg_has_oops = aat->isa_oopptr() &&
2445 (aat->isa_instptr() ||
2446 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2447 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2448 aat->isa_aryptr()->is_flat() &&
2449 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2450 if (i == TypeFunc::Parms) {
2451 src_has_oops = arg_has_oops;
2452 }
2453 //
2454 // src or dst could be j.l.Object when other is basic type array:
2455 //
2456 // arraycopy(char[],0,Object*,0,size);
2457 // arraycopy(Object*,0,char[],0,size);
2458 //
2459 // Don't add edges in such cases.
2460 //
2461 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2462 arg_has_oops && (i > TypeFunc::Parms);
2463 #ifdef ASSERT
2464 if (!(is_arraycopy ||
2465 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2466 (call->as_CallLeaf()->_name != nullptr &&
2467 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2468 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2469 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2470 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2471 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2472 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2473 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2474 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2475 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2476 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2477 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2478 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2479 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2480 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2481 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2482 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2483 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2484 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2485 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2486 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2487 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2488 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2489 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2490 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2491 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2492 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2493 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2494 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2495 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2496 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2497 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2498 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2499 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2500 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2501 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2502 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2503 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2504 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2505 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2506 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2507 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2508 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2509 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2510 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2511 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2512 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2513 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2514 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2515 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2516 strcmp(call->as_CallLeaf()->_name, "store_inline_type_fields_to_buf") == 0 ||
2517 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2518 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2519 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2520 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2521 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2522 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2523 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2524 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2525 ))) {
2526 call->dump();
2527 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2528 }
2529 #endif
2530 // Always process arraycopy's destination object since
2531 // we need to add all possible edges to references in
2532 // source object.
2533 if (arg_esc >= PointsToNode::ArgEscape &&
2534 !arg_is_arraycopy_dest) {
2535 continue;
2536 }
2537 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2538 if (call->is_ArrayCopy()) {
2539 ArrayCopyNode* ac = call->as_ArrayCopy();
2540 if (ac->is_clonebasic() ||
2541 ac->is_arraycopy_validated() ||
2542 ac->is_copyof_validated() ||
2543 ac->is_copyofrange_validated()) {
2544 es = PointsToNode::NoEscape;
2545 }
2546 }
2547 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2548 if (arg_is_arraycopy_dest) {
2549 Node* src = call->in(TypeFunc::Parms);
2550 if (src->is_AddP()) {
2551 src = get_addp_base(src);
2552 }
2553 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2554 assert(src_ptn != nullptr, "should be registered");
2555 // Special arraycopy edge:
2556 // Only escape state of destination object's fields affects
2557 // escape state of fields in source object.
2558 add_arraycopy(call, es, src_ptn, arg_ptn);
2559 }
2560 }
2561 }
2562 break;
2563 }
2564 case Op_CallStaticJava: {
2565 // For a static call, we know exactly what method is being called.
2566 // Use bytecode estimator to record the call's escape affects
2567 #ifdef ASSERT
2568 const char* name = call->as_CallStaticJava()->_name;
2569 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2570 #endif
2571 ciMethod* meth = call->as_CallJava()->method();
2572 if ((meth != nullptr) && meth->is_boxing_method()) {
2573 break; // Boxing methods do not modify any oops.
2574 }
2575 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2576 // fall-through if not a Java method or no analyzer information
2577 if (call_analyzer != nullptr) {
2578 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2579 bool ret_arg = returns_an_argument(call);
2580 for (DomainIterator di(call->as_CallJava()); di.has_next(); di.next()) {
2581 int k = di.i_domain() - TypeFunc::Parms;
2582 const Type* at = di.current_domain_cc();
2583 Node* arg = call->in(di.i_domain_cc());
2584 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2585 assert(!call_analyzer->is_arg_returned(k) || !meth->is_scalarized_arg(k) ||
2586 !compatible_return(call->as_CallJava(), di.i_domain()) ||
2587 call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1) == nullptr ||
2588 _igvn->type(call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1)) == at,
2589 "scalarized return and scalarized argument should match");
2590 if (at->isa_ptr() != nullptr && call_analyzer->is_arg_returned(k) && ret_arg) {
2591 // The call returns arguments.
2592 if (meth->is_scalarized_arg(k)) {
2593 ProjNode* res_proj = call->proj_out_or_null(di.i_domain_cc() - di.first_field_pos() + TypeFunc::Parms + 1);
2594 if (res_proj != nullptr) {
2595 assert(_igvn->type(res_proj)->isa_ptr(), "scalarized return and scalarized argument should match");
2596 if (res_proj->_con != TypeFunc::Parms) {
2597 // add an edge between the result projection for a field and the argument projection for the same argument field
2598 PointsToNode* proj_ptn = ptnode_adr(res_proj->_idx);
2599 add_edge(proj_ptn, arg_ptn);
2600 if (!call_analyzer->is_return_local()) {
2601 add_edge(proj_ptn, phantom_obj);
2602 }
2603 }
2604 }
2605 } else if (call_ptn != nullptr) { // Is call's result used?
2606 assert(call_ptn->is_LocalVar(), "node should be registered");
2607 assert(arg_ptn != nullptr, "node should be registered");
2608 add_edge(call_ptn, arg_ptn);
2609 }
2610 }
2611 if (at->isa_oopptr() != nullptr &&
2612 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2613 if (!call_analyzer->is_arg_stack(k)) {
2614 // The argument global escapes
2615 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2616 } else {
2617 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2618 if (!call_analyzer->is_arg_local(k)) {
2619 // The argument itself doesn't escape, but any fields might
2620 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2621 }
2622 }
2623 }
2624 }
2625 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2626 // The call returns arguments.
2627 assert(call_ptn->edge_count() > 0, "sanity");
2628 if (!call_analyzer->is_return_local()) {
2629 // Returns also unknown object.
2630 add_edge(call_ptn, phantom_obj);
2631 }
2632 }
2633 break;
2634 }
2635 }
2636 default: {
2637 // Fall-through here if not a Java method or no analyzer information
2638 // or some other type of call, assume the worst case: all arguments
2639 // globally escape.
2640 const TypeTuple* d = call->tf()->domain_cc();
2641 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2642 const Type* at = d->field_at(i);
2643 if (at->isa_oopptr() != nullptr) {
2644 Node* arg = call->in(i);
2645 if (arg->is_AddP()) {
2646 arg = get_addp_base(arg);
2647 }
2648 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2649 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2650 }
2651 }
2652 }
2653 }
2654 }
2655
2656
2657 // Finish Graph construction.
2658 bool ConnectionGraph::complete_connection_graph(
2659 GrowableArray<PointsToNode*>& ptnodes_worklist,
2660 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2661 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2662 GrowableArray<FieldNode*>& oop_fields_worklist) {
2663 // Normally only 1-3 passes needed to build Connection Graph depending
2664 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2665 // Set limit to 20 to catch situation when something did go wrong and
2666 // bailout Escape Analysis.
2667 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2668 #define GRAPH_BUILD_ITER_LIMIT 20
2669
2670 // Propagate GlobalEscape and ArgEscape escape states and check that
2671 // we still have non-escaping objects. The method pushs on _worklist
2672 // Field nodes which reference phantom_object.
2673 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2674 return false; // Nothing to do.
2675 }
2676 // Now propagate references to all JavaObject nodes.
2677 int java_objects_length = java_objects_worklist.length();
2678 elapsedTimer build_time;
2679 build_time.start();
2680 elapsedTimer time;
2681 bool timeout = false;
2682 int new_edges = 1;
2683 int iterations = 0;
2684 do {
2685 while ((new_edges > 0) &&
2686 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2687 double start_time = time.seconds();
2688 time.start();
2689 new_edges = 0;
2690 // Propagate references to phantom_object for nodes pushed on _worklist
2691 // by find_non_escaped_objects() and find_field_value().
2692 new_edges += add_java_object_edges(phantom_obj, false);
2693 for (int next = 0; next < java_objects_length; ++next) {
2694 JavaObjectNode* ptn = java_objects_worklist.at(next);
2695 new_edges += add_java_object_edges(ptn, true);
2696
2697 #define SAMPLE_SIZE 4
2698 if ((next % SAMPLE_SIZE) == 0) {
2699 // Each 4 iterations calculate how much time it will take
2700 // to complete graph construction.
2701 time.stop();
2702 // Poll for requests from shutdown mechanism to quiesce compiler
2703 // because Connection graph construction may take long time.
2704 CompileBroker::maybe_block();
2705 double stop_time = time.seconds();
2706 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2707 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2708 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2709 timeout = true;
2710 break; // Timeout
2711 }
2712 start_time = stop_time;
2713 time.start();
2714 }
2715 #undef SAMPLE_SIZE
2716
2717 }
2718 if (timeout) break;
2719 if (new_edges > 0) {
2720 // Update escape states on each iteration if graph was updated.
2721 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2722 return false; // Nothing to do.
2723 }
2724 }
2725 time.stop();
2726 if (time.seconds() >= EscapeAnalysisTimeout) {
2727 timeout = true;
2728 break;
2729 }
2730 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2731 }
2732 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2733 time.start();
2734 // Find fields which have unknown value.
2735 int fields_length = oop_fields_worklist.length();
2736 for (int next = 0; next < fields_length; next++) {
2737 FieldNode* field = oop_fields_worklist.at(next);
2738 if (field->edge_count() == 0) {
2739 new_edges += find_field_value(field);
2740 // This code may added new edges to phantom_object.
2741 // Need an other cycle to propagate references to phantom_object.
2742 }
2743 }
2744 time.stop();
2745 if (time.seconds() >= EscapeAnalysisTimeout) {
2746 timeout = true;
2747 break;
2748 }
2749 } else {
2750 new_edges = 0; // Bailout
2751 }
2752 } while (new_edges > 0);
2753
2754 build_time.stop();
2755 _build_time = build_time.seconds();
2756 _build_iterations = iterations;
2757
2758 // Bailout if passed limits.
2759 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2760 Compile* C = _compile;
2761 if (C->log() != nullptr) {
2762 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2763 C->log()->text("%s", timeout ? "time" : "iterations");
2764 C->log()->end_elem(" limit'");
2765 }
2766 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2767 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2768 // Possible infinite build_connection_graph loop,
2769 // bailout (no changes to ideal graph were made).
2770 return false;
2771 }
2772
2773 #undef GRAPH_BUILD_ITER_LIMIT
2774
2775 // Find fields initialized by null for non-escaping Allocations.
2776 int non_escaped_length = non_escaped_allocs_worklist.length();
2777 for (int next = 0; next < non_escaped_length; next++) {
2778 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2779 PointsToNode::EscapeState es = ptn->escape_state();
2780 assert(es <= PointsToNode::ArgEscape, "sanity");
2781 if (es == PointsToNode::NoEscape) {
2782 if (find_init_values_null(ptn, _igvn) > 0) {
2783 // Adding references to null object does not change escape states
2784 // since it does not escape. Also no fields are added to null object.
2785 add_java_object_edges(null_obj, false);
2786 }
2787 }
2788 Node* n = ptn->ideal_node();
2789 if (n->is_Allocate()) {
2790 // The object allocated by this Allocate node will never be
2791 // seen by an other thread. Mark it so that when it is
2792 // expanded no MemBarStoreStore is added.
2793 InitializeNode* ini = n->as_Allocate()->initialization();
2794 if (ini != nullptr)
2795 ini->set_does_not_escape();
2796 }
2797 }
2798 return true; // Finished graph construction.
2799 }
2800
2801 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2802 // and check that we still have non-escaping java objects.
2803 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2804 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2805 bool print_method) {
2806 GrowableArray<PointsToNode*> escape_worklist;
2807 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2808 int ptnodes_length = ptnodes_worklist.length();
2809 for (int next = 0; next < ptnodes_length; ++next) {
2810 PointsToNode* ptn = ptnodes_worklist.at(next);
2811 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2812 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2813 escape_worklist.push(ptn);
2814 }
2815 }
2816 // Set escape states to referenced nodes (edges list).
2817 while (escape_worklist.length() > 0) {
2818 PointsToNode* ptn = escape_worklist.pop();
2819 PointsToNode::EscapeState es = ptn->escape_state();
2820 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2821 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2822 es >= PointsToNode::ArgEscape) {
2823 // GlobalEscape or ArgEscape state of field means it has unknown value.
2824 if (add_edge(ptn, phantom_obj)) {
2825 // New edge was added
2826 add_field_uses_to_worklist(ptn->as_Field());
2827 }
2828 }
2829 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2830 PointsToNode* e = i.get();
2831 if (e->is_Arraycopy()) {
2832 assert(ptn->arraycopy_dst(), "sanity");
2833 // Propagate only fields escape state through arraycopy edge.
2834 if (e->fields_escape_state() < field_es) {
2835 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2836 escape_worklist.push(e);
2837 }
2838 } else if (es >= field_es) {
2839 // fields_escape_state is also set to 'es' if it is less than 'es'.
2840 if (e->escape_state() < es) {
2841 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2842 escape_worklist.push(e);
2843 }
2844 } else {
2845 // Propagate field escape state.
2846 bool es_changed = false;
2847 if (e->fields_escape_state() < field_es) {
2848 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2849 es_changed = true;
2850 }
2851 if ((e->escape_state() < field_es) &&
2852 e->is_Field() && ptn->is_JavaObject() &&
2853 e->as_Field()->is_oop()) {
2854 // Change escape state of referenced fields.
2855 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2856 es_changed = true;
2857 } else if (e->escape_state() < es) {
2858 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2859 es_changed = true;
2860 }
2861 if (es_changed) {
2862 escape_worklist.push(e);
2863 }
2864 }
2865 if (print_method) {
2866 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2867 }
2868 }
2869 }
2870 // Remove escaped objects from non_escaped list.
2871 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2872 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2873 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2874 non_escaped_allocs_worklist.delete_at(next);
2875 }
2876 if (ptn->escape_state() == PointsToNode::NoEscape) {
2877 // Find fields in non-escaped allocations which have unknown value.
2878 find_init_values_phantom(ptn);
2879 }
2880 }
2881 return (non_escaped_allocs_worklist.length() > 0);
2882 }
2883
2884 // Add all references to JavaObject node by walking over all uses.
2885 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2886 int new_edges = 0;
2887 if (populate_worklist) {
2888 // Populate _worklist by uses of jobj's uses.
2889 for (UseIterator i(jobj); i.has_next(); i.next()) {
2890 PointsToNode* use = i.get();
2891 if (use->is_Arraycopy()) {
2892 continue;
2893 }
2894 add_uses_to_worklist(use);
2895 if (use->is_Field() && use->as_Field()->is_oop()) {
2896 // Put on worklist all field's uses (loads) and
2897 // related field nodes (same base and offset).
2898 add_field_uses_to_worklist(use->as_Field());
2899 }
2900 }
2901 }
2902 for (int l = 0; l < _worklist.length(); l++) {
2903 PointsToNode* use = _worklist.at(l);
2904 if (PointsToNode::is_base_use(use)) {
2905 // Add reference from jobj to field and from field to jobj (field's base).
2906 use = PointsToNode::get_use_node(use)->as_Field();
2907 if (add_base(use->as_Field(), jobj)) {
2908 new_edges++;
2909 }
2910 continue;
2911 }
2912 assert(!use->is_JavaObject(), "sanity");
2913 if (use->is_Arraycopy()) {
2914 if (jobj == null_obj) { // null object does not have field edges
2915 continue;
2916 }
2917 // Added edge from Arraycopy node to arraycopy's source java object
2918 if (add_edge(use, jobj)) {
2919 jobj->set_arraycopy_src();
2920 new_edges++;
2921 }
2922 // and stop here.
2923 continue;
2924 }
2925 if (!add_edge(use, jobj)) {
2926 continue; // No new edge added, there was such edge already.
2927 }
2928 new_edges++;
2929 if (use->is_LocalVar()) {
2930 add_uses_to_worklist(use);
2931 if (use->arraycopy_dst()) {
2932 for (EdgeIterator i(use); i.has_next(); i.next()) {
2933 PointsToNode* e = i.get();
2934 if (e->is_Arraycopy()) {
2935 if (jobj == null_obj) { // null object does not have field edges
2936 continue;
2937 }
2938 // Add edge from arraycopy's destination java object to Arraycopy node.
2939 if (add_edge(jobj, e)) {
2940 new_edges++;
2941 jobj->set_arraycopy_dst();
2942 }
2943 }
2944 }
2945 }
2946 } else {
2947 // Added new edge to stored in field values.
2948 // Put on worklist all field's uses (loads) and
2949 // related field nodes (same base and offset).
2950 add_field_uses_to_worklist(use->as_Field());
2951 }
2952 }
2953 _worklist.clear();
2954 _in_worklist.reset();
2955 return new_edges;
2956 }
2957
2958 // Put on worklist all related field nodes.
2959 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2960 assert(field->is_oop(), "sanity");
2961 int offset = field->offset();
2962 add_uses_to_worklist(field);
2963 // Loop over all bases of this field and push on worklist Field nodes
2964 // with the same offset and base (since they may reference the same field).
2965 for (BaseIterator i(field); i.has_next(); i.next()) {
2966 PointsToNode* base = i.get();
2967 add_fields_to_worklist(field, base);
2968 // Check if the base was source object of arraycopy and go over arraycopy's
2969 // destination objects since values stored to a field of source object are
2970 // accessible by uses (loads) of fields of destination objects.
2971 if (base->arraycopy_src()) {
2972 for (UseIterator j(base); j.has_next(); j.next()) {
2973 PointsToNode* arycp = j.get();
2974 if (arycp->is_Arraycopy()) {
2975 for (UseIterator k(arycp); k.has_next(); k.next()) {
2976 PointsToNode* abase = k.get();
2977 if (abase->arraycopy_dst() && abase != base) {
2978 // Look for the same arraycopy reference.
2979 add_fields_to_worklist(field, abase);
2980 }
2981 }
2982 }
2983 }
2984 }
2985 }
2986 }
2987
2988 // Put on worklist all related field nodes.
2989 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2990 int offset = field->offset();
2991 if (base->is_LocalVar()) {
2992 for (UseIterator j(base); j.has_next(); j.next()) {
2993 PointsToNode* f = j.get();
2994 if (PointsToNode::is_base_use(f)) { // Field
2995 f = PointsToNode::get_use_node(f);
2996 if (f == field || !f->as_Field()->is_oop()) {
2997 continue;
2998 }
2999 int offs = f->as_Field()->offset();
3000 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
3001 add_to_worklist(f);
3002 }
3003 }
3004 }
3005 } else {
3006 assert(base->is_JavaObject(), "sanity");
3007 if (// Skip phantom_object since it is only used to indicate that
3008 // this field's content globally escapes.
3009 (base != phantom_obj) &&
3010 // null object node does not have fields.
3011 (base != null_obj)) {
3012 for (EdgeIterator i(base); i.has_next(); i.next()) {
3013 PointsToNode* f = i.get();
3014 // Skip arraycopy edge since store to destination object field
3015 // does not update value in source object field.
3016 if (f->is_Arraycopy()) {
3017 assert(base->arraycopy_dst(), "sanity");
3018 continue;
3019 }
3020 if (f == field || !f->as_Field()->is_oop()) {
3021 continue;
3022 }
3023 int offs = f->as_Field()->offset();
3024 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
3025 add_to_worklist(f);
3026 }
3027 }
3028 }
3029 }
3030 }
3031
3032 // Find fields which have unknown value.
3033 int ConnectionGraph::find_field_value(FieldNode* field) {
3034 // Escaped fields should have init value already.
3035 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
3036 int new_edges = 0;
3037 for (BaseIterator i(field); i.has_next(); i.next()) {
3038 PointsToNode* base = i.get();
3039 if (base->is_JavaObject()) {
3040 // Skip Allocate's fields which will be processed later.
3041 if (base->ideal_node()->is_Allocate()) {
3042 return 0;
3043 }
3044 assert(base == null_obj, "only null ptr base expected here");
3045 }
3046 }
3047 if (add_edge(field, phantom_obj)) {
3048 // New edge was added
3049 new_edges++;
3050 add_field_uses_to_worklist(field);
3051 }
3052 return new_edges;
3053 }
3054
3055 // Find fields initializing values for allocations.
3056 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
3057 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
3058 PointsToNode* init_val = phantom_obj;
3059 Node* alloc = pta->ideal_node();
3060
3061 // Do nothing for Allocate nodes since its fields values are
3062 // "known" unless they are initialized by arraycopy/clone.
3063 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
3064 if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
3065 // Null-free inline type arrays are initialized with an init value instead of null
3066 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
3067 assert(init_val != nullptr, "init value should be registered");
3068 } else {
3069 return 0;
3070 }
3071 }
3072 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
3073 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
3074 #ifdef ASSERT
3075 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
3076 const char* name = alloc->as_CallStaticJava()->_name;
3077 assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub() ||
3078 strncmp(name, "load_unknown_inline", 19) == 0 ||
3079 strncmp(name, "store_inline_type_fields_to_buf", 31) == 0, "sanity");
3080 }
3081 #endif
3082 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
3083 int new_edges = 0;
3084 for (EdgeIterator i(pta); i.has_next(); i.next()) {
3085 PointsToNode* field = i.get();
3086 if (field->is_Field() && field->as_Field()->is_oop()) {
3087 if (add_edge(field, init_val)) {
3088 // New edge was added
3089 new_edges++;
3090 add_field_uses_to_worklist(field->as_Field());
3091 }
3092 }
3093 }
3094 return new_edges;
3095 }
3096
3097 // Find fields initializing values for allocations.
3098 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
3099 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
3100 Node* alloc = pta->ideal_node();
3101 // Do nothing for Call nodes since its fields values are unknown.
3102 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
3103 return 0;
3104 }
3105 InitializeNode* ini = alloc->as_Allocate()->initialization();
3106 bool visited_bottom_offset = false;
3107 GrowableArray<int> offsets_worklist;
3108 int new_edges = 0;
3109
3110 // Check if an oop field's initializing value is recorded and add
3111 // a corresponding null if field's value if it is not recorded.
3112 // Connection Graph does not record a default initialization by null
3113 // captured by Initialize node.
3114 //
3115 for (EdgeIterator i(pta); i.has_next(); i.next()) {
3116 PointsToNode* field = i.get(); // Field (AddP)
3117 if (!field->is_Field() || !field->as_Field()->is_oop()) {
3118 continue; // Not oop field
3119 }
3120 int offset = field->as_Field()->offset();
3121 if (offset == Type::OffsetBot) {
3122 if (!visited_bottom_offset) {
3123 // OffsetBot is used to reference array's element,
3124 // always add reference to null to all Field nodes since we don't
3125 // known which element is referenced.
3126 if (add_edge(field, null_obj)) {
3127 // New edge was added
3128 new_edges++;
3129 add_field_uses_to_worklist(field->as_Field());
3130 visited_bottom_offset = true;
3131 }
3132 }
3133 } else {
3134 // Check only oop fields.
3135 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
3136 if (adr_type->isa_rawptr()) {
3137 #ifdef ASSERT
3138 // Raw pointers are used for initializing stores so skip it
3139 // since it should be recorded already
3140 Node* base = get_addp_base(field->ideal_node());
3141 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
3142 #endif
3143 continue;
3144 }
3145 if (!offsets_worklist.contains(offset)) {
3146 offsets_worklist.append(offset);
3147 Node* value = nullptr;
3148 if (ini != nullptr) {
3149 // StoreP::value_basic_type() == T_ADDRESS
3150 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
3151 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
3152 // Make sure initializing store has the same type as this AddP.
3153 // This AddP may reference non existing field because it is on a
3154 // dead branch of bimorphic call which is not eliminated yet.
3155 if (store != nullptr && store->is_Store() &&
3156 store->as_Store()->value_basic_type() == ft) {
3157 value = store->in(MemNode::ValueIn);
3158 #ifdef ASSERT
3159 if (VerifyConnectionGraph) {
3160 // Verify that AddP already points to all objects the value points to.
3161 PointsToNode* val = ptnode_adr(value->_idx);
3162 assert((val != nullptr), "should be processed already");
3163 PointsToNode* missed_obj = nullptr;
3164 if (val->is_JavaObject()) {
3165 if (!field->points_to(val->as_JavaObject())) {
3166 missed_obj = val;
3167 }
3168 } else {
3169 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
3170 tty->print_cr("----------init store has invalid value -----");
3171 store->dump();
3172 val->dump();
3173 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
3174 }
3175 for (EdgeIterator j(val); j.has_next(); j.next()) {
3176 PointsToNode* obj = j.get();
3177 if (obj->is_JavaObject()) {
3178 if (!field->points_to(obj->as_JavaObject())) {
3179 missed_obj = obj;
3180 break;
3181 }
3182 }
3183 }
3184 }
3185 if (missed_obj != nullptr) {
3186 tty->print_cr("----------field---------------------------------");
3187 field->dump();
3188 tty->print_cr("----------missed reference to object------------");
3189 missed_obj->dump();
3190 tty->print_cr("----------object referenced by init store-------");
3191 store->dump();
3192 val->dump();
3193 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
3194 }
3195 }
3196 #endif
3197 } else {
3198 // There could be initializing stores which follow allocation.
3199 // For example, a volatile field store is not collected
3200 // by Initialize node.
3201 //
3202 // Need to check for dependent loads to separate such stores from
3203 // stores which follow loads. For now, add initial value null so
3204 // that compare pointers optimization works correctly.
3205 }
3206 }
3207 if (value == nullptr) {
3208 // A field's initializing value was not recorded. Add null.
3209 if (add_edge(field, null_obj)) {
3210 // New edge was added
3211 new_edges++;
3212 add_field_uses_to_worklist(field->as_Field());
3213 }
3214 }
3215 }
3216 }
3217 }
3218 return new_edges;
3219 }
3220
3221 // Adjust scalar_replaceable state after Connection Graph is built.
3222 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
3223 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
3224 // returns true. If one of the constraints in this method set 'jobj' to NSR
3225 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
3226 // input, 'adjust_scalar_replaceable_state' will eventually be called with
3227 // that other object and the Phi will become a reducible Phi.
3228 // There could be multiple merges involving the same jobj.
3229 Unique_Node_List candidates;
3230
3231 // Search for non-escaping objects which are not scalar replaceable
3232 // and mark them to propagate the state to referenced objects.
3233
3234 for (UseIterator i(jobj); i.has_next(); i.next()) {
3235 PointsToNode* use = i.get();
3236 if (use->is_Arraycopy()) {
3237 continue;
3238 }
3239 if (use->is_Field()) {
3240 FieldNode* field = use->as_Field();
3241 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3242 // 1. An object is not scalar replaceable if the field into which it is
3243 // stored has unknown offset (stored into unknown element of an array).
3244 if (field->offset() == Type::OffsetBot) {
3245 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
3246 return;
3247 }
3248 for (BaseIterator i(field); i.has_next(); i.next()) {
3249 PointsToNode* base = i.get();
3250 // 2. An object is not scalar replaceable if the field into which it is
3251 // stored has multiple bases one of which is null.
3252 if ((base == null_obj) && (field->base_count() > 1)) {
3253 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3254 return;
3255 }
3256 // 2.5. An object is not scalar replaceable if the field into which it is
3257 // stored has NSR base.
3258 if (!base->scalar_replaceable()) {
3259 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3260 return;
3261 }
3262 }
3263 }
3264 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3265 // 3. An object is not scalar replaceable if it is merged with other objects
3266 // and we can't remove the merge
3267 for (EdgeIterator j(use); j.has_next(); j.next()) {
3268 PointsToNode* ptn = j.get();
3269 if (ptn->is_JavaObject() && ptn != jobj) {
3270 Node* use_n = use->ideal_node();
3271
3272 // These other local vars may point to multiple objects through a Phi
3273 // In this case we skip them and see if we can reduce the Phi.
3274 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3275 use_n = use_n->in(1);
3276 }
3277
3278 // If it's already a candidate or confirmed reducible merge we can skip verification
3279 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3280 continue;
3281 }
3282
3283 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3284 candidates.push(use_n);
3285 } else {
3286 // Mark all objects as NSR if we can't remove the merge
3287 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3288 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3289 }
3290 }
3291 }
3292 if (!jobj->scalar_replaceable()) {
3293 return;
3294 }
3295 }
3296
3297 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3298 if (j.get()->is_Arraycopy()) {
3299 continue;
3300 }
3301
3302 // Non-escaping object node should point only to field nodes.
3303 FieldNode* field = j.get()->as_Field();
3304 int offset = field->as_Field()->offset();
3305
3306 // 4. An object is not scalar replaceable if it has a field with unknown
3307 // offset (array's element is accessed in loop).
3308 if (offset == Type::OffsetBot) {
3309 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3310 return;
3311 }
3312 // 5. Currently an object is not scalar replaceable if a LoadStore node
3313 // access its field since the field value is unknown after it.
3314 //
3315 Node* n = field->ideal_node();
3316
3317 // Test for an unsafe access that was parsed as maybe off heap
3318 // (with a CheckCastPP to raw memory).
3319 assert(n->is_AddP(), "expect an address computation");
3320 if (n->in(AddPNode::Base)->is_top() &&
3321 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3322 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3323 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3324 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3325 return;
3326 }
3327
3328 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3329 Node* u = n->fast_out(i);
3330 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3331 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3332 return;
3333 }
3334 }
3335
3336 // 6. Or the address may point to more then one object. This may produce
3337 // the false positive result (set not scalar replaceable)
3338 // since the flow-insensitive escape analysis can't separate
3339 // the case when stores overwrite the field's value from the case
3340 // when stores happened on different control branches.
3341 //
3342 // Note: it will disable scalar replacement in some cases:
3343 //
3344 // Point p[] = new Point[1];
3345 // p[0] = new Point(); // Will be not scalar replaced
3346 //
3347 // but it will save us from incorrect optimizations in next cases:
3348 //
3349 // Point p[] = new Point[1];
3350 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3351 //
3352 if (field->base_count() > 1 && candidates.size() == 0) {
3353 if (has_non_reducible_merge(field, reducible_merges)) {
3354 for (BaseIterator i(field); i.has_next(); i.next()) {
3355 PointsToNode* base = i.get();
3356 // Don't take into account LocalVar nodes which
3357 // may point to only one object which should be also
3358 // this field's base by now.
3359 if (base->is_JavaObject() && base != jobj) {
3360 // Mark all bases.
3361 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3362 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3363 }
3364 }
3365
3366 if (!jobj->scalar_replaceable()) {
3367 return;
3368 }
3369 }
3370 }
3371 }
3372
3373 // The candidate is truly a reducible merge only if none of the other
3374 // constraints ruled it as NSR. There could be multiple merges involving the
3375 // same jobj.
3376 assert(jobj->scalar_replaceable(), "sanity");
3377 for (uint i = 0; i < candidates.size(); i++ ) {
3378 Node* candidate = candidates.at(i);
3379 reducible_merges.push(candidate);
3380 }
3381 }
3382
3383 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3384 for (BaseIterator i(field); i.has_next(); i.next()) {
3385 Node* base = i.get()->ideal_node();
3386 if (base->is_Phi() && !reducible_merges.member(base)) {
3387 return true;
3388 }
3389 }
3390 return false;
3391 }
3392
3393 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3394 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3395
3396 // Look for 'phis' that refer to 'jobj' as the last
3397 // remaining scalar replaceable input.
3398 uint reducible_merges_cnt = reducible_merges.size();
3399 for (uint i = 0; i < reducible_merges_cnt; i++) {
3400 Node* phi = reducible_merges.at(i);
3401
3402 // This 'Phi' will be a 'good' if it still points to
3403 // at least one scalar replaceable object. Note that 'obj'
3404 // was/should be marked as NSR before calling this function.
3405 bool good_phi = false;
3406
3407 for (uint j = 1; j < phi->req(); j++) {
3408 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3409 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3410 good_phi = true;
3411 break;
3412 }
3413 }
3414
3415 if (!good_phi) {
3416 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3417 reducible_merges.remove(i);
3418
3419 // Decrement the index because the 'remove' call above actually
3420 // moves the last entry of the list to position 'i'.
3421 i--;
3422
3423 reducible_merges_cnt--;
3424 }
3425 }
3426 }
3427
3428 // Propagate NSR (Not scalar replaceable) state.
3429 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3430 int jobj_length = jobj_worklist.length();
3431 bool found_nsr_alloc = true;
3432 while (found_nsr_alloc) {
3433 found_nsr_alloc = false;
3434 for (int next = 0; next < jobj_length; ++next) {
3435 JavaObjectNode* jobj = jobj_worklist.at(next);
3436 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3437 PointsToNode* use = i.get();
3438 if (use->is_Field()) {
3439 FieldNode* field = use->as_Field();
3440 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3441 assert(field->offset() != Type::OffsetBot, "sanity");
3442 for (BaseIterator i(field); i.has_next(); i.next()) {
3443 PointsToNode* base = i.get();
3444 // An object is not scalar replaceable if the field into which
3445 // it is stored has NSR base.
3446 if ((base != null_obj) && !base->scalar_replaceable()) {
3447 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3448 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3449 // because there is no point in reducing a Phi that won't improve the number of SR
3450 // objects.
3451 revisit_reducible_phi_status(jobj, reducible_merges);
3452 found_nsr_alloc = true;
3453 break;
3454 }
3455 }
3456 } else if (use->is_LocalVar()) {
3457 Node* phi = use->ideal_node();
3458 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3459 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3460 reducible_merges.yank(phi);
3461 found_nsr_alloc = true;
3462 break;
3463 }
3464 }
3465 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3466 }
3467 }
3468 }
3469 }
3470
3471 #ifdef ASSERT
3472 void ConnectionGraph::verify_connection_graph(
3473 GrowableArray<PointsToNode*>& ptnodes_worklist,
3474 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3475 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3476 GrowableArray<Node*>& addp_worklist) {
3477 // Verify that graph is complete - no new edges could be added.
3478 int java_objects_length = java_objects_worklist.length();
3479 int non_escaped_length = non_escaped_allocs_worklist.length();
3480 int new_edges = 0;
3481 for (int next = 0; next < java_objects_length; ++next) {
3482 JavaObjectNode* ptn = java_objects_worklist.at(next);
3483 new_edges += add_java_object_edges(ptn, true);
3484 }
3485 assert(new_edges == 0, "graph was not complete");
3486 // Verify that escape state is final.
3487 int length = non_escaped_allocs_worklist.length();
3488 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3489 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3490 (non_escaped_length == length) &&
3491 (_worklist.length() == 0), "escape state was not final");
3492
3493 // Verify fields information.
3494 int addp_length = addp_worklist.length();
3495 for (int next = 0; next < addp_length; ++next ) {
3496 Node* n = addp_worklist.at(next);
3497 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3498 if (field->is_oop()) {
3499 // Verify that field has all bases
3500 Node* base = get_addp_base(n);
3501 PointsToNode* ptn = ptnode_adr(base->_idx);
3502 if (ptn->is_JavaObject()) {
3503 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3504 } else {
3505 assert(ptn->is_LocalVar(), "sanity");
3506 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3507 PointsToNode* e = i.get();
3508 if (e->is_JavaObject()) {
3509 assert(field->has_base(e->as_JavaObject()), "sanity");
3510 }
3511 }
3512 }
3513 // Verify that all fields have initializing values.
3514 if (field->edge_count() == 0) {
3515 tty->print_cr("----------field does not have references----------");
3516 field->dump();
3517 for (BaseIterator i(field); i.has_next(); i.next()) {
3518 PointsToNode* base = i.get();
3519 tty->print_cr("----------field has next base---------------------");
3520 base->dump();
3521 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3522 tty->print_cr("----------base has fields-------------------------");
3523 for (EdgeIterator j(base); j.has_next(); j.next()) {
3524 j.get()->dump();
3525 }
3526 tty->print_cr("----------base has references---------------------");
3527 for (UseIterator j(base); j.has_next(); j.next()) {
3528 j.get()->dump();
3529 }
3530 }
3531 }
3532 for (UseIterator i(field); i.has_next(); i.next()) {
3533 i.get()->dump();
3534 }
3535 assert(field->edge_count() > 0, "sanity");
3536 }
3537 }
3538 }
3539 }
3540 #endif
3541
3542 // Optimize ideal graph.
3543 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3544 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3545 Compile* C = _compile;
3546 PhaseIterGVN* igvn = _igvn;
3547 if (EliminateLocks) {
3548 // Mark locks before changing ideal graph.
3549 int cnt = C->macro_count();
3550 for (int i = 0; i < cnt; i++) {
3551 Node *n = C->macro_node(i);
3552 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3553 AbstractLockNode* alock = n->as_AbstractLock();
3554 if (!alock->is_non_esc_obj()) {
3555 const Type* obj_type = igvn->type(alock->obj_node());
3556 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3557 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3558 // The lock could be marked eliminated by lock coarsening
3559 // code during first IGVN before EA. Replace coarsened flag
3560 // to eliminate all associated locks/unlocks.
3561 #ifdef ASSERT
3562 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3563 #endif
3564 alock->set_non_esc_obj();
3565 }
3566 }
3567 }
3568 }
3569 }
3570
3571 if (OptimizePtrCompare) {
3572 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3573 Node *n = ptr_cmp_worklist.at(i);
3574 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3575 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3576 if (tcmp->singleton()) {
3577 Node* cmp = igvn->makecon(tcmp);
3578 #ifndef PRODUCT
3579 if (PrintOptimizePtrCompare) {
3580 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"));
3581 if (Verbose) {
3582 n->dump(1);
3583 }
3584 }
3585 #endif
3586 igvn->replace_node(n, cmp);
3587 }
3588 }
3589 }
3590
3591 // For MemBarStoreStore nodes added in library_call.cpp, check
3592 // escape status of associated AllocateNode and optimize out
3593 // MemBarStoreStore node if the allocated object never escapes.
3594 for (int i = 0; i < storestore_worklist.length(); i++) {
3595 Node* storestore = storestore_worklist.at(i);
3596 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3597 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3598 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3599 // Non-escaping inline type buffer allocations don't require a membar
3600 storestore->as_MemBar()->remove(_igvn);
3601 } else {
3602 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3603 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3604 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3605 igvn->register_new_node_with_optimizer(mb);
3606 igvn->replace_node(storestore, mb);
3607 }
3608 }
3609 }
3610 }
3611
3612 // Atomic flat accesses on non-escaping objects can be optimized to non-atomic accesses
3613 void ConnectionGraph::optimize_flat_accesses(GrowableArray<SafePointNode*>& sfn_worklist) {
3614 PhaseIterGVN& igvn = *_igvn;
3615 bool delay = igvn.delay_transform();
3616 igvn.set_delay_transform(true);
3617 igvn.C->for_each_flat_access([&](Node* n) {
3618 Node* base = n->is_LoadFlat() ? n->as_LoadFlat()->base() : n->as_StoreFlat()->base();
3619 if (!not_global_escape(base)) {
3620 return;
3621 }
3622
3623 bool expanded;
3624 if (n->is_LoadFlat()) {
3625 expanded = n->as_LoadFlat()->expand_non_atomic(igvn);
3626 } else {
3627 expanded = n->as_StoreFlat()->expand_non_atomic(igvn);
3628 }
3629 if (expanded) {
3630 sfn_worklist.remove(n->as_SafePoint());
3631 igvn.C->remove_flat_access(n);
3632 }
3633 });
3634 igvn.set_delay_transform(delay);
3635 }
3636
3637 // Optimize objects compare.
3638 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3639 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3640 if (!OptimizePtrCompare) {
3641 return UNKNOWN;
3642 }
3643 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3644 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3645
3646 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3647 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3648 JavaObjectNode* jobj1 = unique_java_object(left);
3649 JavaObjectNode* jobj2 = unique_java_object(right);
3650
3651 // The use of this method during allocation merge reduction may cause 'left'
3652 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3653 // that doesn't reference an unique java object.
3654 if (ptn1 == nullptr || ptn2 == nullptr ||
3655 jobj1 == nullptr || jobj2 == nullptr) {
3656 return UNKNOWN;
3657 }
3658
3659 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3660 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3661
3662 // Check simple cases first.
3663 if (jobj1 != nullptr) {
3664 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3665 if (jobj1 == jobj2) {
3666 // Comparing the same not escaping object.
3667 return EQ;
3668 }
3669 Node* obj = jobj1->ideal_node();
3670 // Comparing not escaping allocation.
3671 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3672 !ptn2->points_to(jobj1)) {
3673 return NE; // This includes nullness check.
3674 }
3675 }
3676 }
3677 if (jobj2 != nullptr) {
3678 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3679 Node* obj = jobj2->ideal_node();
3680 // Comparing not escaping allocation.
3681 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3682 !ptn1->points_to(jobj2)) {
3683 return NE; // This includes nullness check.
3684 }
3685 }
3686 }
3687 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3688 jobj2 != nullptr && jobj2 != phantom_obj &&
3689 jobj1->ideal_node()->is_Con() &&
3690 jobj2->ideal_node()->is_Con()) {
3691 // Klass or String constants compare. Need to be careful with
3692 // compressed pointers - compare types of ConN and ConP instead of nodes.
3693 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3694 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3695 if (t1->make_ptr() == t2->make_ptr()) {
3696 return EQ;
3697 } else {
3698 return NE;
3699 }
3700 }
3701 if (ptn1->meet(ptn2)) {
3702 return UNKNOWN; // Sets are not disjoint
3703 }
3704
3705 // Sets are disjoint.
3706 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3707 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3708 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3709 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3710 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3711 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3712 // Check nullness of unknown object.
3713 return UNKNOWN;
3714 }
3715
3716 // Disjointness by itself is not sufficient since
3717 // alias analysis is not complete for escaped objects.
3718 // Disjoint sets are definitely unrelated only when
3719 // at least one set has only not escaping allocations.
3720 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3721 if (ptn1->non_escaping_allocation()) {
3722 return NE;
3723 }
3724 }
3725 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3726 if (ptn2->non_escaping_allocation()) {
3727 return NE;
3728 }
3729 }
3730 return UNKNOWN;
3731 }
3732
3733 // Connection Graph construction functions.
3734
3735 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3736 PointsToNode* ptadr = _nodes.at(n->_idx);
3737 if (ptadr != nullptr) {
3738 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3739 return;
3740 }
3741 Compile* C = _compile;
3742 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3743 map_ideal_node(n, ptadr);
3744 }
3745
3746 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3747 PointsToNode* ptadr = _nodes.at(n->_idx);
3748 if (ptadr != nullptr) {
3749 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3750 return ptadr;
3751 }
3752 Compile* C = _compile;
3753 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3754 map_ideal_node(n, ptadr);
3755 return ptadr;
3756 }
3757
3758 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3759 PointsToNode* ptadr = _nodes.at(n->_idx);
3760 if (ptadr != nullptr) {
3761 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3762 return;
3763 }
3764 bool unsafe = false;
3765 bool is_oop = is_oop_field(n, offset, &unsafe);
3766 if (unsafe) {
3767 es = PointsToNode::GlobalEscape;
3768 }
3769 Compile* C = _compile;
3770 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3771 map_ideal_node(n, field);
3772 }
3773
3774 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3775 PointsToNode* src, PointsToNode* dst) {
3776 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3777 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3778 PointsToNode* ptadr = _nodes.at(n->_idx);
3779 if (ptadr != nullptr) {
3780 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3781 return;
3782 }
3783 Compile* C = _compile;
3784 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3785 map_ideal_node(n, ptadr);
3786 // Add edge from arraycopy node to source object.
3787 (void)add_edge(ptadr, src);
3788 src->set_arraycopy_src();
3789 // Add edge from destination object to arraycopy node.
3790 (void)add_edge(dst, ptadr);
3791 dst->set_arraycopy_dst();
3792 }
3793
3794 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3795 const Type* adr_type = n->as_AddP()->bottom_type();
3796 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3797 BasicType bt = T_INT;
3798 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3799 // Check only oop fields.
3800 if (!adr_type->isa_aryptr() ||
3801 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3802 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3803 // OffsetBot is used to reference array's element. Ignore first AddP.
3804 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3805 bt = T_OBJECT;
3806 }
3807 }
3808 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3809 if (adr_type->isa_instptr()) {
3810 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3811 if (field != nullptr) {
3812 bt = field->layout_type();
3813 } else {
3814 // Check for unsafe oop field access
3815 if (has_oop_node_outs(n)) {
3816 bt = T_OBJECT;
3817 (*unsafe) = true;
3818 }
3819 }
3820 } else if (adr_type->isa_aryptr()) {
3821 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3822 // Ignore array length load.
3823 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3824 // Ignore first AddP.
3825 } else {
3826 const Type* elemtype = adr_type->is_aryptr()->elem();
3827 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3828 ciInlineKlass* vk = elemtype->inline_klass();
3829 field_offset += vk->payload_offset();
3830 ciField* field = vk->get_field_by_offset(field_offset, false);
3831 if (field != nullptr) {
3832 bt = field->layout_type();
3833 } else {
3834 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);
3835 bt = T_BOOLEAN;
3836 }
3837 } else {
3838 bt = elemtype->array_element_basic_type();
3839 }
3840 }
3841 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3842 // Allocation initialization, ThreadLocal field access, unsafe access
3843 if (has_oop_node_outs(n)) {
3844 bt = T_OBJECT;
3845 }
3846 }
3847 }
3848 // Note: T_NARROWOOP is not classed as a real reference type
3849 bool res = (is_reference_type(bt) || bt == T_NARROWOOP);
3850 assert(!has_oop_node_outs(n) || res, "sanity: AddP has oop outs, needs to be treated as oop field");
3851 return res;
3852 }
3853
3854 bool ConnectionGraph::has_oop_node_outs(Node* n) {
3855 return n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3856 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3857 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3858 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n);
3859 }
3860
3861 // Returns unique pointed java object or null.
3862 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3863 // If the node was created after the escape computation we can't answer.
3864 uint idx = n->_idx;
3865 if (idx >= nodes_size()) {
3866 return nullptr;
3867 }
3868 PointsToNode* ptn = ptnode_adr(idx);
3869 if (ptn == nullptr) {
3870 return nullptr;
3871 }
3872 if (ptn->is_JavaObject()) {
3873 return ptn->as_JavaObject();
3874 }
3875 assert(ptn->is_LocalVar(), "sanity");
3876 // Check all java objects it points to.
3877 JavaObjectNode* jobj = nullptr;
3878 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3879 PointsToNode* e = i.get();
3880 if (e->is_JavaObject()) {
3881 if (jobj == nullptr) {
3882 jobj = e->as_JavaObject();
3883 } else if (jobj != e) {
3884 return nullptr;
3885 }
3886 }
3887 }
3888 return jobj;
3889 }
3890
3891 // Return true if this node points only to non-escaping allocations.
3892 bool PointsToNode::non_escaping_allocation() {
3893 if (is_JavaObject()) {
3894 Node* n = ideal_node();
3895 if (n->is_Allocate() || n->is_CallStaticJava()) {
3896 return (escape_state() == PointsToNode::NoEscape);
3897 } else {
3898 return false;
3899 }
3900 }
3901 assert(is_LocalVar(), "sanity");
3902 // Check all java objects it points to.
3903 for (EdgeIterator i(this); i.has_next(); i.next()) {
3904 PointsToNode* e = i.get();
3905 if (e->is_JavaObject()) {
3906 Node* n = e->ideal_node();
3907 if ((e->escape_state() != PointsToNode::NoEscape) ||
3908 !(n->is_Allocate() || n->is_CallStaticJava())) {
3909 return false;
3910 }
3911 }
3912 }
3913 return true;
3914 }
3915
3916 // Return true if we know the node does not escape globally.
3917 bool ConnectionGraph::not_global_escape(Node *n) {
3918 assert(!_collecting, "should not call during graph construction");
3919 // If the node was created after the escape computation we can't answer.
3920 uint idx = n->_idx;
3921 if (idx >= nodes_size()) {
3922 return false;
3923 }
3924 PointsToNode* ptn = ptnode_adr(idx);
3925 if (ptn == nullptr) {
3926 return false; // not in congraph (e.g. ConI)
3927 }
3928 PointsToNode::EscapeState es = ptn->escape_state();
3929 // If we have already computed a value, return it.
3930 if (es >= PointsToNode::GlobalEscape) {
3931 return false;
3932 }
3933 if (ptn->is_JavaObject()) {
3934 return true; // (es < PointsToNode::GlobalEscape);
3935 }
3936 assert(ptn->is_LocalVar(), "sanity");
3937 // Check all java objects it points to.
3938 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3939 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3940 return false;
3941 }
3942 }
3943 return true;
3944 }
3945
3946 // Return true if locked object does not escape globally
3947 // and locked code region (identified by BoxLockNode) is balanced:
3948 // all compiled code paths have corresponding Lock/Unlock pairs.
3949 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3950 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3951 if (EliminateNestedLocks) {
3952 // We can mark whole locking region as Local only when only
3953 // one object is used for locking.
3954 alock->box_node()->as_BoxLock()->set_local();
3955 }
3956 return true;
3957 }
3958 return false;
3959 }
3960
3961 // Helper functions
3962
3963 // Return true if this node points to specified node or nodes it points to.
3964 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3965 if (is_JavaObject()) {
3966 return (this == ptn);
3967 }
3968 assert(is_LocalVar() || is_Field(), "sanity");
3969 for (EdgeIterator i(this); i.has_next(); i.next()) {
3970 if (i.get() == ptn) {
3971 return true;
3972 }
3973 }
3974 return false;
3975 }
3976
3977 // Return true if one node points to an other.
3978 bool PointsToNode::meet(PointsToNode* ptn) {
3979 if (this == ptn) {
3980 return true;
3981 } else if (ptn->is_JavaObject()) {
3982 return this->points_to(ptn->as_JavaObject());
3983 } else if (this->is_JavaObject()) {
3984 return ptn->points_to(this->as_JavaObject());
3985 }
3986 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3987 int ptn_count = ptn->edge_count();
3988 for (EdgeIterator i(this); i.has_next(); i.next()) {
3989 PointsToNode* this_e = i.get();
3990 for (int j = 0; j < ptn_count; j++) {
3991 if (this_e == ptn->edge(j)) {
3992 return true;
3993 }
3994 }
3995 }
3996 return false;
3997 }
3998
3999 #ifdef ASSERT
4000 // Return true if bases point to this java object.
4001 bool FieldNode::has_base(JavaObjectNode* jobj) const {
4002 for (BaseIterator i(this); i.has_next(); i.next()) {
4003 if (i.get() == jobj) {
4004 return true;
4005 }
4006 }
4007 return false;
4008 }
4009 #endif
4010
4011 bool ConnectionGraph::is_captured_store_address(Node* addp) {
4012 // Handle simple case first.
4013 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
4014 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
4015 return true;
4016 } else if (addp->in(AddPNode::Address)->is_Phi()) {
4017 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
4018 Node* addp_use = addp->fast_out(i);
4019 if (addp_use->is_Store()) {
4020 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
4021 if (addp_use->fast_out(j)->is_Initialize()) {
4022 return true;
4023 }
4024 }
4025 }
4026 }
4027 }
4028 return false;
4029 }
4030
4031 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
4032 const Type *adr_type = phase->type(adr);
4033 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
4034 // We are computing a raw address for a store captured by an Initialize
4035 // compute an appropriate address type. AddP cases #3 and #5 (see below).
4036 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
4037 assert(offs != Type::OffsetBot ||
4038 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
4039 "offset must be a constant or it is initialization of array");
4040 return offs;
4041 }
4042 return adr_type->is_ptr()->flat_offset();
4043 }
4044
4045 Node* ConnectionGraph::get_addp_base(Node *addp) {
4046 assert(addp->is_AddP(), "must be AddP");
4047 //
4048 // AddP cases for Base and Address inputs:
4049 // case #1. Direct object's field reference:
4050 // Allocate
4051 // |
4052 // Proj #5 ( oop result )
4053 // |
4054 // CheckCastPP (cast to instance type)
4055 // | |
4056 // AddP ( base == address )
4057 //
4058 // case #2. Indirect object's field reference:
4059 // Phi
4060 // |
4061 // CastPP (cast to instance type)
4062 // | |
4063 // AddP ( base == address )
4064 //
4065 // case #3. Raw object's field reference for Initialize node.
4066 // Could have an additional Phi merging multiple allocations.
4067 // Allocate
4068 // |
4069 // Proj #5 ( oop result )
4070 // top |
4071 // \ |
4072 // AddP ( base == top )
4073 //
4074 // case #4. Array's element reference:
4075 // {CheckCastPP | CastPP}
4076 // | | |
4077 // | AddP ( array's element offset )
4078 // | |
4079 // AddP ( array's offset )
4080 //
4081 // case #5. Raw object's field reference for arraycopy stub call:
4082 // The inline_native_clone() case when the arraycopy stub is called
4083 // after the allocation before Initialize and CheckCastPP nodes.
4084 // Allocate
4085 // |
4086 // Proj #5 ( oop result )
4087 // | |
4088 // AddP ( base == address )
4089 //
4090 // case #6. Constant Pool, ThreadLocal, CastX2P, Klass, OSR buffer buf or
4091 // Raw object's field reference:
4092 // {ConP, ThreadLocal, CastX2P, raw Load, Parm0}
4093 // top |
4094 // \ |
4095 // AddP ( base == top )
4096 //
4097 // case #7. Klass's field reference.
4098 // LoadKlass
4099 // | |
4100 // AddP ( base == address )
4101 //
4102 // case #8. narrow Klass's field reference.
4103 // LoadNKlass
4104 // |
4105 // DecodeN
4106 // | |
4107 // AddP ( base == address )
4108 //
4109 // case #9. Mixed unsafe access
4110 // {instance}
4111 // |
4112 // CheckCastPP (raw)
4113 // top |
4114 // \ |
4115 // AddP ( base == top )
4116 //
4117 // case #10. Klass fetched with
4118 // LibraryCallKit::load_*_refined_array_klass()
4119 // which has en extra Phi.
4120 // LoadKlass LoadKlass
4121 // | |
4122 // CastPP CastPP
4123 // \ /
4124 // Phi
4125 // top |
4126 // \ |
4127 // AddP ( base == top )
4128 //
4129 Node *base = addp->in(AddPNode::Base);
4130 if (base->uncast()->is_top()) { // The AddP case #3, #6, #9, and #10.
4131 base = addp->in(AddPNode::Address);
4132 while (base->is_AddP()) {
4133 // Case #6 (unsafe access) may have several chained AddP nodes.
4134 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
4135 base = base->in(AddPNode::Address);
4136 }
4137 if (base->Opcode() == Op_CheckCastPP &&
4138 base->bottom_type()->isa_rawptr() &&
4139 _igvn->type(base->in(1))->isa_oopptr()) {
4140 base = base->in(1); // Case #9
4141 } else {
4142 // Case #3, #6, and #10
4143 Node* uncast_base = base->uncast();
4144 int opcode = uncast_base->Opcode();
4145 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
4146 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
4147 (_igvn->C->is_osr_compilation() && uncast_base->is_Parm() && uncast_base->as_Parm()->_con == TypeFunc::Parms)||
4148 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
4149 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_klassptr() != nullptr)) ||
4150 is_captured_store_address(addp) ||
4151 is_load_array_klass_related(uncast_base), "sanity");
4152 }
4153 }
4154 return base;
4155 }
4156
4157 #ifdef ASSERT
4158 // Case #10
4159 bool ConnectionGraph::is_load_array_klass_related(const Node* uncast_base) {
4160 if (!uncast_base->is_Phi() || uncast_base->req() != 3) {
4161 return false;
4162 }
4163 Node* in1 = uncast_base->in(1);
4164 Node* in2 = uncast_base->in(2);
4165 return in1->uncast()->Opcode() == Op_LoadKlass &&
4166 in2->uncast()->Opcode() == Op_LoadKlass;
4167 }
4168 #endif
4169
4170 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
4171 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
4172 Node* addp2 = addp->raw_out(0);
4173 if (addp->outcnt() == 1 && addp2->is_AddP() &&
4174 addp2->in(AddPNode::Base) == n &&
4175 addp2->in(AddPNode::Address) == addp) {
4176 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
4177 //
4178 // Find array's offset to push it on worklist first and
4179 // as result process an array's element offset first (pushed second)
4180 // to avoid CastPP for the array's offset.
4181 // Otherwise the inserted CastPP (LocalVar) will point to what
4182 // the AddP (Field) points to. Which would be wrong since
4183 // the algorithm expects the CastPP has the same point as
4184 // as AddP's base CheckCastPP (LocalVar).
4185 //
4186 // ArrayAllocation
4187 // |
4188 // CheckCastPP
4189 // |
4190 // memProj (from ArrayAllocation CheckCastPP)
4191 // | ||
4192 // | || Int (element index)
4193 // | || | ConI (log(element size))
4194 // | || | /
4195 // | || LShift
4196 // | || /
4197 // | AddP (array's element offset)
4198 // | |
4199 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
4200 // | / /
4201 // AddP (array's offset)
4202 // |
4203 // Load/Store (memory operation on array's element)
4204 //
4205 return addp2;
4206 }
4207 return nullptr;
4208 }
4209
4210 //
4211 // Adjust the type and inputs of an AddP which computes the
4212 // address of a field of an instance
4213 //
4214 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
4215 PhaseGVN* igvn = _igvn;
4216 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
4217 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
4218 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
4219 if (t == nullptr) {
4220 // We are computing a raw address for a store captured by an Initialize
4221 // compute an appropriate address type (cases #3 and #5).
4222 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
4223 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
4224 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
4225 assert(offs != Type::OffsetBot, "offset must be a constant");
4226 if (base_t->isa_aryptr() != nullptr) {
4227 // In the case of a flat inline type array, each field has its
4228 // own slice so we need to extract the field being accessed from
4229 // the address computation
4230 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
4231 } else {
4232 t = base_t->add_offset(offs)->is_oopptr();
4233 }
4234 }
4235 int inst_id = base_t->instance_id();
4236 assert(!t->is_known_instance() || t->instance_id() == inst_id,
4237 "old type must be non-instance or match new type");
4238
4239 // The type 't' could be subclass of 'base_t'.
4240 // As result t->offset() could be large then base_t's size and it will
4241 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
4242 // constructor verifies correctness of the offset.
4243 //
4244 // It could happened on subclass's branch (from the type profiling
4245 // inlining) which was not eliminated during parsing since the exactness
4246 // of the allocation type was not propagated to the subclass type check.
4247 //
4248 // Or the type 't' could be not related to 'base_t' at all.
4249 // It could happen when CHA type is different from MDO type on a dead path
4250 // (for example, from instanceof check) which is not collapsed during parsing.
4251 //
4252 // Do nothing for such AddP node and don't process its users since
4253 // this code branch will go away.
4254 //
4255 if (!t->is_known_instance() &&
4256 !base_t->maybe_java_subtype_of(t)) {
4257 return false; // bail out
4258 }
4259 const TypePtr* tinst = base_t->add_offset(t->offset());
4260 if (tinst->isa_aryptr() && t->isa_aryptr()) {
4261 // In the case of a flat inline type array, each field has its
4262 // own slice so we need to keep track of the field being accessed.
4263 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
4264 // Keep array properties (not flat/null-free)
4265 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
4266 if (tinst == nullptr) {
4267 return false; // Skip dead path with inconsistent properties
4268 }
4269 }
4270
4271 // Do NOT remove the next line: ensure a new alias index is allocated
4272 // for the instance type. Note: C++ will not remove it since the call
4273 // has side effect.
4274 int alias_idx = _compile->get_alias_index(tinst);
4275 igvn->set_type(addp, tinst);
4276 // record the allocation in the node map
4277 set_map(addp, get_map(base->_idx));
4278 // Set addp's Base and Address to 'base'.
4279 Node *abase = addp->in(AddPNode::Base);
4280 Node *adr = addp->in(AddPNode::Address);
4281 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
4282 adr->in(0)->_idx == (uint)inst_id) {
4283 // Skip AddP cases #3 and #5.
4284 } else {
4285 assert(!abase->is_top(), "sanity"); // AddP case #3
4286 if (abase != base) {
4287 igvn->hash_delete(addp);
4288 addp->set_req(AddPNode::Base, base);
4289 if (abase == adr) {
4290 addp->set_req(AddPNode::Address, base);
4291 } else {
4292 // AddP case #4 (adr is array's element offset AddP node)
4293 #ifdef ASSERT
4294 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
4295 assert(adr->is_AddP() && atype != nullptr &&
4296 atype->instance_id() == inst_id, "array's element offset should be processed first");
4297 #endif
4298 }
4299 igvn->hash_insert(addp);
4300 }
4301 }
4302 // Put on IGVN worklist since at least addp's type was changed above.
4303 record_for_optimizer(addp);
4304 return true;
4305 }
4306
4307 //
4308 // Create a new version of orig_phi if necessary. Returns either the newly
4309 // created phi or an existing phi. Sets create_new to indicate whether a new
4310 // phi was created. Cache the last newly created phi in the node map.
4311 //
4312 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
4313 Compile *C = _compile;
4314 PhaseGVN* igvn = _igvn;
4315 new_created = false;
4316 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
4317 // nothing to do if orig_phi is bottom memory or matches alias_idx
4318 if (phi_alias_idx == alias_idx) {
4319 return orig_phi;
4320 }
4321 // Have we recently created a Phi for this alias index?
4322 PhiNode *result = get_map_phi(orig_phi->_idx);
4323 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
4324 return result;
4325 }
4326 // Previous check may fail when the same wide memory Phi was split into Phis
4327 // for different memory slices. Search all Phis for this region.
4328 if (result != nullptr) {
4329 Node* region = orig_phi->in(0);
4330 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4331 Node* phi = region->fast_out(i);
4332 if (phi->is_Phi() &&
4333 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4334 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4335 return phi->as_Phi();
4336 }
4337 }
4338 }
4339 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4340 if (C->do_escape_analysis() == true && !C->failing()) {
4341 // Retry compilation without escape analysis.
4342 // If this is the first failure, the sentinel string will "stick"
4343 // to the Compile object, and the C2Compiler will see it and retry.
4344 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4345 }
4346 return nullptr;
4347 }
4348 orig_phi_worklist.append_if_missing(orig_phi);
4349 const TypePtr *atype = C->get_adr_type(alias_idx);
4350 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4351 C->copy_node_notes_to(result, orig_phi);
4352 igvn->set_type(result, result->bottom_type());
4353 record_for_optimizer(result);
4354 set_map(orig_phi, result);
4355 new_created = true;
4356 return result;
4357 }
4358
4359 //
4360 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4361 // specified alias index.
4362 //
4363 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4364 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4365 Compile *C = _compile;
4366 PhaseGVN* igvn = _igvn;
4367 bool new_phi_created;
4368 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4369 if (!new_phi_created) {
4370 return result;
4371 }
4372 GrowableArray<PhiNode *> phi_list;
4373 GrowableArray<uint> cur_input;
4374 PhiNode *phi = orig_phi;
4375 uint idx = 1;
4376 bool finished = false;
4377 while(!finished) {
4378 while (idx < phi->req()) {
4379 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4380 if (mem != nullptr && mem->is_Phi()) {
4381 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4382 if (new_phi_created) {
4383 // found an phi for which we created a new split, push current one on worklist and begin
4384 // processing new one
4385 phi_list.push(phi);
4386 cur_input.push(idx);
4387 phi = mem->as_Phi();
4388 result = newphi;
4389 idx = 1;
4390 continue;
4391 } else {
4392 mem = newphi;
4393 }
4394 }
4395 if (C->failing()) {
4396 return nullptr;
4397 }
4398 result->set_req(idx++, mem);
4399 }
4400 #ifdef ASSERT
4401 // verify that the new Phi has an input for each input of the original
4402 assert( phi->req() == result->req(), "must have same number of inputs.");
4403 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4404 #endif
4405 // Check if all new phi's inputs have specified alias index.
4406 // Otherwise use old phi.
4407 for (uint i = 1; i < phi->req(); i++) {
4408 Node* in = result->in(i);
4409 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4410 }
4411 // we have finished processing a Phi, see if there are any more to do
4412 finished = (phi_list.length() == 0 );
4413 if (!finished) {
4414 phi = phi_list.pop();
4415 idx = cur_input.pop();
4416 PhiNode *prev_result = get_map_phi(phi->_idx);
4417 prev_result->set_req(idx++, result);
4418 result = prev_result;
4419 }
4420 }
4421 return result;
4422 }
4423
4424 //
4425 // The next methods are derived from methods in MemNode.
4426 //
4427 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4428 Node *mem = mmem;
4429 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4430 // means an array I have not precisely typed yet. Do not do any
4431 // alias stuff with it any time soon.
4432 if (toop->base() != Type::AnyPtr &&
4433 !(toop->isa_instptr() &&
4434 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4435 toop->offset() == Type::OffsetBot)) {
4436 mem = mmem->memory_at(alias_idx);
4437 // Update input if it is progress over what we have now
4438 }
4439 return mem;
4440 }
4441
4442 //
4443 // Move memory users to their memory slices.
4444 //
4445 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4446 Compile* C = _compile;
4447 PhaseGVN* igvn = _igvn;
4448 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4449 assert(tp != nullptr, "ptr type");
4450 int alias_idx = C->get_alias_index(tp);
4451 int general_idx = C->get_general_index(alias_idx);
4452
4453 // Move users first
4454 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4455 Node* use = n->fast_out(i);
4456 if (use->is_MergeMem()) {
4457 MergeMemNode* mmem = use->as_MergeMem();
4458 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4459 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4460 continue; // Nothing to do
4461 }
4462 // Replace previous general reference to mem node.
4463 uint orig_uniq = C->unique();
4464 Node* m = find_inst_mem(n, general_idx, orig_phis);
4465 assert(orig_uniq == C->unique(), "no new nodes");
4466 mmem->set_memory_at(general_idx, m);
4467 --imax;
4468 --i;
4469 } else if (use->is_MemBar()) {
4470 assert(!use->is_Initialize(), "initializing stores should not be moved");
4471 if (use->req() > MemBarNode::Precedent &&
4472 use->in(MemBarNode::Precedent) == n) {
4473 // Don't move related membars.
4474 record_for_optimizer(use);
4475 continue;
4476 }
4477 tp = use->as_MemBar()->adr_type()->isa_ptr();
4478 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4479 alias_idx == general_idx) {
4480 continue; // Nothing to do
4481 }
4482 // Move to general memory slice.
4483 uint orig_uniq = C->unique();
4484 Node* m = find_inst_mem(n, general_idx, orig_phis);
4485 assert(orig_uniq == C->unique(), "no new nodes");
4486 igvn->hash_delete(use);
4487 imax -= use->replace_edge(n, m, igvn);
4488 igvn->hash_insert(use);
4489 record_for_optimizer(use);
4490 --i;
4491 #ifdef ASSERT
4492 } else if (use->is_Mem()) {
4493 // Memory nodes should have new memory input.
4494 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4495 assert(tp != nullptr, "ptr type");
4496 int idx = C->get_alias_index(tp);
4497 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4498 "Following memory nodes should have new memory input or be on the same memory slice");
4499 } else if (use->is_Phi()) {
4500 // Phi nodes should be split and moved already.
4501 tp = use->as_Phi()->adr_type()->isa_ptr();
4502 assert(tp != nullptr, "ptr type");
4503 int idx = C->get_alias_index(tp);
4504 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4505 } else {
4506 use->dump();
4507 assert(false, "should not be here");
4508 #endif
4509 }
4510 }
4511 }
4512
4513 //
4514 // Search memory chain of "mem" to find a MemNode whose address
4515 // is the specified alias index.
4516 //
4517 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4518 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4519 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4520 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4521 return nullptr;
4522 }
4523 if (orig_mem == nullptr) {
4524 return orig_mem;
4525 }
4526 Compile* C = _compile;
4527 PhaseGVN* igvn = _igvn;
4528 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4529 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4530 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4531 Node *prev = nullptr;
4532 Node *result = orig_mem;
4533 while (prev != result) {
4534 prev = result;
4535 if (result == start_mem) {
4536 break; // hit one of our sentinels
4537 }
4538 if (result->is_Mem()) {
4539 const Type *at = igvn->type(result->in(MemNode::Address));
4540 if (at == Type::TOP) {
4541 break; // Dead
4542 }
4543 assert (at->isa_ptr() != nullptr, "pointer type required.");
4544 int idx = C->get_alias_index(at->is_ptr());
4545 if (idx == alias_idx) {
4546 break; // Found
4547 }
4548 if (!is_instance && (at->isa_oopptr() == nullptr ||
4549 !at->is_oopptr()->is_known_instance())) {
4550 break; // Do not skip store to general memory slice.
4551 }
4552 result = result->in(MemNode::Memory);
4553 }
4554 if (!is_instance) {
4555 continue; // don't search further for non-instance types
4556 }
4557 // skip over a call which does not affect this memory slice
4558 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4559 Node *proj_in = result->in(0);
4560 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4561 break; // hit one of our sentinels
4562 } else if (proj_in->is_Call()) {
4563 // ArrayCopy node processed here as well
4564 CallNode *call = proj_in->as_Call();
4565 if (!call->may_modify(toop, igvn)) {
4566 result = call->in(TypeFunc::Memory);
4567 }
4568 } else if (proj_in->is_Initialize()) {
4569 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4570 // Stop if this is the initialization for the object instance which
4571 // which contains this memory slice, otherwise skip over it.
4572 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4573 result = proj_in->in(TypeFunc::Memory);
4574 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4575 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4576 result = get_map(result->_idx);
4577 assert(result != nullptr, "new projection should have been allocated");
4578 break;
4579 }
4580 } else if (proj_in->is_MemBar()) {
4581 // Check if there is an array copy for a clone
4582 // Step over GC barrier when ReduceInitialCardMarks is disabled
4583 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4584 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4585
4586 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4587 // Stop if it is a clone
4588 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4589 if (ac->may_modify(toop, igvn)) {
4590 break;
4591 }
4592 }
4593 result = proj_in->in(TypeFunc::Memory);
4594 }
4595 } else if (result->is_MergeMem()) {
4596 MergeMemNode *mmem = result->as_MergeMem();
4597 result = step_through_mergemem(mmem, alias_idx, toop);
4598 if (result == mmem->base_memory()) {
4599 // Didn't find instance memory, search through general slice recursively.
4600 result = mmem->memory_at(C->get_general_index(alias_idx));
4601 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4602 if (C->failing()) {
4603 return nullptr;
4604 }
4605 mmem->set_memory_at(alias_idx, result);
4606 }
4607 } else if (result->is_Phi() &&
4608 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4609 Node *un = result->as_Phi()->unique_input(igvn);
4610 if (un != nullptr) {
4611 orig_phis.append_if_missing(result->as_Phi());
4612 result = un;
4613 } else {
4614 break;
4615 }
4616 } else if (result->is_ClearArray()) {
4617 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4618 // Can not bypass initialization of the instance
4619 // we are looking for.
4620 break;
4621 }
4622 // Otherwise skip it (the call updated 'result' value).
4623 } else if (result->Opcode() == Op_SCMemProj) {
4624 Node* mem = result->in(0);
4625 Node* adr = nullptr;
4626 if (mem->is_LoadStore()) {
4627 adr = mem->in(MemNode::Address);
4628 } else {
4629 assert(mem->Opcode() == Op_EncodeISOArray ||
4630 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4631 adr = mem->in(3); // Memory edge corresponds to destination array
4632 }
4633 const Type *at = igvn->type(adr);
4634 if (at != Type::TOP) {
4635 assert(at->isa_ptr() != nullptr, "pointer type required.");
4636 int idx = C->get_alias_index(at->is_ptr());
4637 if (idx == alias_idx) {
4638 // Assert in debug mode
4639 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4640 break; // In product mode return SCMemProj node
4641 }
4642 }
4643 result = mem->in(MemNode::Memory);
4644 } else if (result->Opcode() == Op_StrInflatedCopy) {
4645 Node* adr = result->in(3); // Memory edge corresponds to destination array
4646 const Type *at = igvn->type(adr);
4647 if (at != Type::TOP) {
4648 assert(at->isa_ptr() != nullptr, "pointer type required.");
4649 int idx = C->get_alias_index(at->is_ptr());
4650 if (idx == alias_idx) {
4651 // Assert in debug mode
4652 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4653 break; // In product mode return SCMemProj node
4654 }
4655 }
4656 result = result->in(MemNode::Memory);
4657 }
4658 }
4659 if (result->is_Phi()) {
4660 PhiNode *mphi = result->as_Phi();
4661 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4662 const TypePtr *t = mphi->adr_type();
4663 if (!is_instance) {
4664 // Push all non-instance Phis on the orig_phis worklist to update inputs
4665 // during Phase 4 if needed.
4666 orig_phis.append_if_missing(mphi);
4667 } else if (C->get_alias_index(t) != alias_idx) {
4668 // Create a new Phi with the specified alias index type.
4669 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4670 }
4671 }
4672 // the result is either MemNode, PhiNode, InitializeNode.
4673 return result;
4674 }
4675
4676 //
4677 // Convert the types of non-escaped object to instance types where possible,
4678 // propagate the new type information through the graph, and update memory
4679 // edges and MergeMem inputs to reflect the new type.
4680 //
4681 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4682 // The processing is done in 4 phases:
4683 //
4684 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4685 // types for the CheckCastPP for allocations where possible.
4686 // Propagate the new types through users as follows:
4687 // casts and Phi: push users on alloc_worklist
4688 // AddP: cast Base and Address inputs to the instance type
4689 // push any AddP users on alloc_worklist and push any memnode
4690 // users onto memnode_worklist.
4691 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4692 // search the Memory chain for a store with the appropriate type
4693 // address type. If a Phi is found, create a new version with
4694 // the appropriate memory slices from each of the Phi inputs.
4695 // For stores, process the users as follows:
4696 // MemNode: push on memnode_worklist
4697 // MergeMem: push on mergemem_worklist
4698 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4699 // moving the first node encountered of each instance type to the
4700 // the input corresponding to its alias index.
4701 // appropriate memory slice.
4702 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4703 //
4704 // In the following example, the CheckCastPP nodes are the cast of allocation
4705 // results and the allocation of node 29 is non-escaped and eligible to be an
4706 // instance type.
4707 //
4708 // We start with:
4709 //
4710 // 7 Parm #memory
4711 // 10 ConI "12"
4712 // 19 CheckCastPP "Foo"
4713 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4714 // 29 CheckCastPP "Foo"
4715 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4716 //
4717 // 40 StoreP 25 7 20 ... alias_index=4
4718 // 50 StoreP 35 40 30 ... alias_index=4
4719 // 60 StoreP 45 50 20 ... alias_index=4
4720 // 70 LoadP _ 60 30 ... alias_index=4
4721 // 80 Phi 75 50 60 Memory alias_index=4
4722 // 90 LoadP _ 80 30 ... alias_index=4
4723 // 100 LoadP _ 80 20 ... alias_index=4
4724 //
4725 //
4726 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4727 // and creating a new alias index for node 30. This gives:
4728 //
4729 // 7 Parm #memory
4730 // 10 ConI "12"
4731 // 19 CheckCastPP "Foo"
4732 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4733 // 29 CheckCastPP "Foo" iid=24
4734 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4735 //
4736 // 40 StoreP 25 7 20 ... alias_index=4
4737 // 50 StoreP 35 40 30 ... alias_index=6
4738 // 60 StoreP 45 50 20 ... alias_index=4
4739 // 70 LoadP _ 60 30 ... alias_index=6
4740 // 80 Phi 75 50 60 Memory alias_index=4
4741 // 90 LoadP _ 80 30 ... alias_index=6
4742 // 100 LoadP _ 80 20 ... alias_index=4
4743 //
4744 // In phase 2, new memory inputs are computed for the loads and stores,
4745 // And a new version of the phi is created. In phase 4, the inputs to
4746 // node 80 are updated and then the memory nodes are updated with the
4747 // values computed in phase 2. This results in:
4748 //
4749 // 7 Parm #memory
4750 // 10 ConI "12"
4751 // 19 CheckCastPP "Foo"
4752 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4753 // 29 CheckCastPP "Foo" iid=24
4754 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4755 //
4756 // 40 StoreP 25 7 20 ... alias_index=4
4757 // 50 StoreP 35 7 30 ... alias_index=6
4758 // 60 StoreP 45 40 20 ... alias_index=4
4759 // 70 LoadP _ 50 30 ... alias_index=6
4760 // 80 Phi 75 40 60 Memory alias_index=4
4761 // 120 Phi 75 50 50 Memory alias_index=6
4762 // 90 LoadP _ 120 30 ... alias_index=6
4763 // 100 LoadP _ 80 20 ... alias_index=4
4764 //
4765 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4766 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4767 GrowableArray<MergeMemNode*> &mergemem_worklist,
4768 Unique_Node_List &reducible_merges) {
4769 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4770 GrowableArray<Node *> memnode_worklist;
4771 GrowableArray<PhiNode *> orig_phis;
4772 PhaseIterGVN *igvn = _igvn;
4773 uint new_index_start = (uint) _compile->num_alias_types();
4774 VectorSet visited;
4775 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4776
4777 // Phase 1: Process possible allocations from alloc_worklist.
4778 // Create instance types for the CheckCastPP for allocations where possible.
4779 //
4780 // (Note: don't forget to change the order of the second AddP node on
4781 // the alloc_worklist if the order of the worklist processing is changed,
4782 // see the comment in find_second_addp().)
4783 //
4784 while (alloc_worklist.length() != 0) {
4785 Node *n = alloc_worklist.pop();
4786 uint ni = n->_idx;
4787 if (n->is_Call()) {
4788 CallNode *alloc = n->as_Call();
4789 // copy escape information to call node
4790 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4791 PointsToNode::EscapeState es = ptn->escape_state();
4792 // We have an allocation or call which returns a Java object,
4793 // see if it is non-escaped.
4794 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4795 continue;
4796 }
4797 // Find CheckCastPP for the allocate or for the return value of a call
4798 n = alloc->result_cast();
4799 if (n == nullptr) { // No uses except Initialize node
4800 if (alloc->is_Allocate()) {
4801 // Set the scalar_replaceable flag for allocation
4802 // so it could be eliminated if it has no uses.
4803 alloc->as_Allocate()->_is_scalar_replaceable = true;
4804 }
4805 continue;
4806 }
4807 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4808 // we could reach here for allocate case if one init is associated with many allocs.
4809 if (alloc->is_Allocate()) {
4810 alloc->as_Allocate()->_is_scalar_replaceable = false;
4811 }
4812 continue;
4813 }
4814
4815 // The inline code for Object.clone() casts the allocation result to
4816 // java.lang.Object and then to the actual type of the allocated
4817 // object. Detect this case and use the second cast.
4818 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4819 // the allocation result is cast to java.lang.Object and then
4820 // to the actual Array type.
4821 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4822 && (alloc->is_AllocateArray() ||
4823 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4824 Node *cast2 = nullptr;
4825 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4826 Node *use = n->fast_out(i);
4827 if (use->is_CheckCastPP()) {
4828 cast2 = use;
4829 break;
4830 }
4831 }
4832 if (cast2 != nullptr) {
4833 n = cast2;
4834 } else {
4835 // Non-scalar replaceable if the allocation type is unknown statically
4836 // (reflection allocation), the object can't be restored during
4837 // deoptimization without precise type.
4838 continue;
4839 }
4840 }
4841
4842 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4843 if (t == nullptr) {
4844 continue; // not a TypeOopPtr
4845 }
4846 if (!t->klass_is_exact()) {
4847 continue; // not an unique type
4848 }
4849 if (alloc->is_Allocate()) {
4850 // Set the scalar_replaceable flag for allocation
4851 // so it could be eliminated.
4852 alloc->as_Allocate()->_is_scalar_replaceable = true;
4853 }
4854 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4855 // in order for an object to be scalar-replaceable, it must be:
4856 // - a direct allocation (not a call returning an object)
4857 // - non-escaping
4858 // - eligible to be a unique type
4859 // - not determined to be ineligible by escape analysis
4860 set_map(alloc, n);
4861 set_map(n, alloc);
4862 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4863 igvn->hash_delete(n);
4864 igvn->set_type(n, tinst);
4865 n->raise_bottom_type(tinst);
4866 igvn->hash_insert(n);
4867 record_for_optimizer(n);
4868 // Allocate an alias index for the header fields. Accesses to
4869 // the header emitted during macro expansion wouldn't have
4870 // correct memory state otherwise.
4871 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4872 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4873 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4874 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4875 InitializeNode* init = alloc->as_Allocate()->initialization();
4876 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4877 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4878 const TypePtr* adr_type = proj->adr_type();
4879 const TypePtr* new_adr_type = tinst->with_offset(adr_type->offset());
4880 if (adr_type->isa_aryptr()) {
4881 // In the case of a flat inline type array, each field has its own slice so we need a
4882 // NarrowMemProj for each field of the flat array elements
4883 new_adr_type = new_adr_type->is_aryptr()->with_field_offset(adr_type->is_aryptr()->field_offset().get());
4884 }
4885 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4886 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4887 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4888 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4889 igvn->set_type(new_proj, new_proj->bottom_type());
4890 record_for_optimizer(new_proj);
4891 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4892 }
4893 };
4894 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4895
4896 // First, put on the worklist all Field edges from Connection Graph
4897 // which is more accurate than putting immediate users from Ideal Graph.
4898 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4899 PointsToNode* tgt = e.get();
4900 if (tgt->is_Arraycopy()) {
4901 continue;
4902 }
4903 Node* use = tgt->ideal_node();
4904 assert(tgt->is_Field() && use->is_AddP(),
4905 "only AddP nodes are Field edges in CG");
4906 if (use->outcnt() > 0) { // Don't process dead nodes
4907 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4908 if (addp2 != nullptr) {
4909 assert(alloc->is_AllocateArray(),"array allocation was expected");
4910 alloc_worklist.append_if_missing(addp2);
4911 }
4912 alloc_worklist.append_if_missing(use);
4913 }
4914 }
4915
4916 // An allocation may have an Initialize which has raw stores. Scan
4917 // the users of the raw allocation result and push AddP users
4918 // on alloc_worklist.
4919 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4920 assert (raw_result != nullptr, "must have an allocation result");
4921 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4922 Node *use = raw_result->fast_out(i);
4923 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4924 Node* addp2 = find_second_addp(use, raw_result);
4925 if (addp2 != nullptr) {
4926 assert(alloc->is_AllocateArray(),"array allocation was expected");
4927 alloc_worklist.append_if_missing(addp2);
4928 }
4929 alloc_worklist.append_if_missing(use);
4930 } else if (use->is_MemBar()) {
4931 memnode_worklist.append_if_missing(use);
4932 }
4933 }
4934 }
4935 } else if (n->is_AddP()) {
4936 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4937 // This AddP will go away when we reduce the Phi
4938 continue;
4939 }
4940 Node* addp_base = get_addp_base(n);
4941 JavaObjectNode* jobj = unique_java_object(addp_base);
4942 if (jobj == nullptr || jobj == phantom_obj) {
4943 #ifdef ASSERT
4944 ptnode_adr(get_addp_base(n)->_idx)->dump();
4945 ptnode_adr(n->_idx)->dump();
4946 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4947 #endif
4948 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4949 return;
4950 }
4951 Node *base = get_map(jobj->idx()); // CheckCastPP node
4952 if (!split_AddP(n, base)) continue; // wrong type from dead path
4953 } else if (n->is_Phi() ||
4954 n->is_CheckCastPP() ||
4955 n->is_EncodeP() ||
4956 n->is_DecodeN() ||
4957 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4958 if (visited.test_set(n->_idx)) {
4959 assert(n->is_Phi(), "loops only through Phi's");
4960 continue; // already processed
4961 }
4962 // Reducible Phi's will be removed from the graph after split_unique_types
4963 // finishes. For now we just try to split out the SR inputs of the merge.
4964 Node* parent = n->in(1);
4965 if (reducible_merges.member(n)) {
4966 reduce_phi(n->as_Phi(), alloc_worklist);
4967 #ifdef ASSERT
4968 if (VerifyReduceAllocationMerges) {
4969 reduced_merges.push(n);
4970 }
4971 #endif
4972 continue;
4973 } else if (reducible_merges.member(parent)) {
4974 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4975 // part of reduce_merge.
4976 continue;
4977 }
4978 JavaObjectNode* jobj = unique_java_object(n);
4979 if (jobj == nullptr || jobj == phantom_obj) {
4980 #ifdef ASSERT
4981 ptnode_adr(n->_idx)->dump();
4982 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4983 #endif
4984 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4985 return;
4986 } else {
4987 Node *val = get_map(jobj->idx()); // CheckCastPP node
4988 TypeNode *tn = n->as_Type();
4989 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4990 assert(tinst != nullptr && tinst->is_known_instance() &&
4991 tinst->instance_id() == jobj->idx() , "instance type expected.");
4992
4993 const Type *tn_type = igvn->type(tn);
4994 const TypeOopPtr *tn_t;
4995 if (tn_type->isa_narrowoop()) {
4996 tn_t = tn_type->make_ptr()->isa_oopptr();
4997 } else {
4998 tn_t = tn_type->isa_oopptr();
4999 }
5000 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
5001 if (tn_t->isa_aryptr()) {
5002 // Keep array properties (not flat/null-free)
5003 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
5004 if (tinst == nullptr) {
5005 continue; // Skip dead path with inconsistent properties
5006 }
5007 }
5008 if (tn_type->isa_narrowoop()) {
5009 tn_type = tinst->make_narrowoop();
5010 } else {
5011 tn_type = tinst;
5012 }
5013 igvn->hash_delete(tn);
5014 igvn->set_type(tn, tn_type);
5015 tn->set_type(tn_type);
5016 igvn->hash_insert(tn);
5017 record_for_optimizer(n);
5018 } else {
5019 assert(tn_type == TypePtr::NULL_PTR ||
5020 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
5021 "unexpected type");
5022 continue; // Skip dead path with different type
5023 }
5024 }
5025 } else {
5026 DEBUG_ONLY(n->dump();)
5027 assert(false, "EA: unexpected node");
5028 continue;
5029 }
5030 // push allocation's users on appropriate worklist
5031 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5032 Node *use = n->fast_out(i);
5033 if (use->is_Mem() && use->in(MemNode::Address) == n) {
5034 // Load/store to instance's field
5035 memnode_worklist.append_if_missing(use);
5036 } else if (use->is_MemBar()) {
5037 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5038 memnode_worklist.append_if_missing(use);
5039 }
5040 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
5041 Node* addp2 = find_second_addp(use, n);
5042 if (addp2 != nullptr) {
5043 alloc_worklist.append_if_missing(addp2);
5044 }
5045 alloc_worklist.append_if_missing(use);
5046 } else if (use->is_Phi() ||
5047 use->is_CheckCastPP() ||
5048 use->is_EncodeNarrowPtr() ||
5049 use->is_DecodeNarrowPtr() ||
5050 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
5051 alloc_worklist.append_if_missing(use);
5052 #ifdef ASSERT
5053 } else if (use->is_Mem()) {
5054 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
5055 } else if (use->is_MergeMem()) {
5056 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5057 } else if (use->is_SafePoint()) {
5058 // Look for MergeMem nodes for calls which reference unique allocation
5059 // (through CheckCastPP nodes) even for debug info.
5060 Node* m = use->in(TypeFunc::Memory);
5061 if (m->is_MergeMem()) {
5062 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5063 }
5064 } else if (use->Opcode() == Op_EncodeISOArray) {
5065 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5066 // EncodeISOArray overwrites destination array
5067 memnode_worklist.append_if_missing(use);
5068 }
5069 } else if (use->Opcode() == Op_Return) {
5070 // Allocation is referenced by field of returned inline type
5071 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
5072 } else {
5073 uint op = use->Opcode();
5074 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
5075 (use->in(MemNode::Memory) == n)) {
5076 // They overwrite memory edge corresponding to destination array,
5077 memnode_worklist.append_if_missing(use);
5078 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
5079 op == Op_CastP2X ||
5080 op == Op_FastLock || op == Op_AryEq ||
5081 op == Op_StrComp || op == Op_CountPositives ||
5082 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
5083 op == Op_StrEquals || op == Op_VectorizedHashCode ||
5084 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
5085 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
5086 op == Op_ReinterpretS2HF ||
5087 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
5088 n->dump();
5089 use->dump();
5090 assert(false, "EA: missing allocation reference path");
5091 }
5092 #endif
5093 }
5094 }
5095
5096 }
5097
5098 #ifdef ASSERT
5099 if (VerifyReduceAllocationMerges) {
5100 for (uint i = 0; i < reducible_merges.size(); i++) {
5101 Node* phi = reducible_merges.at(i);
5102
5103 if (!reduced_merges.member(phi)) {
5104 phi->dump(2);
5105 phi->dump(-2);
5106 assert(false, "This reducible merge wasn't reduced.");
5107 }
5108
5109 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
5110 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
5111 Node* use = phi->fast_out(j);
5112 if (!use->is_SafePoint() && !use->is_CastPP()) {
5113 phi->dump(2);
5114 phi->dump(-2);
5115 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
5116 }
5117 }
5118 }
5119 }
5120 #endif
5121
5122 // Go over all ArrayCopy nodes and if one of the inputs has a unique
5123 // type, record it in the ArrayCopy node so we know what memory this
5124 // node uses/modified.
5125 for (int next = 0; next < arraycopy_worklist.length(); next++) {
5126 ArrayCopyNode* ac = arraycopy_worklist.at(next);
5127 Node* dest = ac->in(ArrayCopyNode::Dest);
5128 if (dest->is_AddP()) {
5129 dest = get_addp_base(dest);
5130 }
5131 JavaObjectNode* jobj = unique_java_object(dest);
5132 if (jobj != nullptr) {
5133 Node *base = get_map(jobj->idx());
5134 if (base != nullptr) {
5135 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5136 ac->_dest_type = base_t;
5137 }
5138 }
5139 Node* src = ac->in(ArrayCopyNode::Src);
5140 if (src->is_AddP()) {
5141 src = get_addp_base(src);
5142 }
5143 jobj = unique_java_object(src);
5144 if (jobj != nullptr) {
5145 Node* base = get_map(jobj->idx());
5146 if (base != nullptr) {
5147 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
5148 ac->_src_type = base_t;
5149 }
5150 }
5151 }
5152
5153 // New alias types were created in split_AddP().
5154 uint new_index_end = (uint) _compile->num_alias_types();
5155
5156 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
5157
5158 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
5159 // compute new values for Memory inputs (the Memory inputs are not
5160 // actually updated until phase 4.)
5161 if (memnode_worklist.length() == 0)
5162 return; // nothing to do
5163 while (memnode_worklist.length() != 0) {
5164 Node *n = memnode_worklist.pop();
5165 if (visited.test_set(n->_idx)) {
5166 continue;
5167 }
5168 if (n->is_Phi() || n->is_ClearArray()) {
5169 // we don't need to do anything, but the users must be pushed
5170 } else if (n->is_MemBar()) { // MemBar nodes
5171 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
5172 // we don't need to do anything, but the users must be pushed
5173 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
5174 if (n == nullptr) {
5175 continue;
5176 }
5177 }
5178 } else if (n->is_CallLeaf()) {
5179 // Runtime calls with narrow memory input (no MergeMem node)
5180 // get the memory projection
5181 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
5182 if (n == nullptr) {
5183 continue;
5184 }
5185 } else if (n->Opcode() == Op_StrInflatedCopy) {
5186 // Check direct uses of StrInflatedCopy.
5187 // It is memory type Node - no special SCMemProj node.
5188 } else if (n->Opcode() == Op_StrCompressedCopy ||
5189 n->Opcode() == Op_EncodeISOArray) {
5190 // get the memory projection
5191 n = n->find_out_with(Op_SCMemProj);
5192 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5193 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
5194 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5195 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
5196 } else if (n->is_Proj()) {
5197 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
5198 } else {
5199 #ifdef ASSERT
5200 if (!n->is_Mem()) {
5201 n->dump();
5202 }
5203 assert(n->is_Mem(), "memory node required.");
5204 #endif
5205 Node *addr = n->in(MemNode::Address);
5206 const Type *addr_t = igvn->type(addr);
5207 if (addr_t == Type::TOP) {
5208 continue;
5209 }
5210 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
5211 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
5212 assert ((uint)alias_idx < new_index_end, "wrong alias index");
5213 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
5214 if (_compile->failing()) {
5215 return;
5216 }
5217 if (mem != n->in(MemNode::Memory)) {
5218 // We delay the memory edge update since we need old one in
5219 // MergeMem code below when instances memory slices are separated.
5220 set_map(n, mem);
5221 }
5222 if (n->is_Load()) {
5223 continue; // don't push users
5224 } else if (n->is_LoadStore()) {
5225 // get the memory projection
5226 n = n->find_out_with(Op_SCMemProj);
5227 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
5228 }
5229 }
5230 // push user on appropriate worklist
5231 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5232 Node *use = n->fast_out(i);
5233 if (use->is_Phi() || use->is_ClearArray()) {
5234 memnode_worklist.append_if_missing(use);
5235 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
5236 memnode_worklist.append_if_missing(use);
5237 } else if (use->is_MemBar() || use->is_CallLeaf()) {
5238 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
5239 memnode_worklist.append_if_missing(use);
5240 }
5241 } else if (use->is_Proj()) {
5242 assert(n->is_Initialize(), "We only push projections of Initialize");
5243 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
5244 memnode_worklist.append_if_missing(use);
5245 }
5246 #ifdef ASSERT
5247 } else if (use->is_Mem()) {
5248 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
5249 } else if (use->is_MergeMem()) {
5250 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
5251 } else if (use->Opcode() == Op_EncodeISOArray) {
5252 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
5253 // EncodeISOArray overwrites destination array
5254 memnode_worklist.append_if_missing(use);
5255 }
5256 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
5257 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
5258 // store_unknown_inline overwrites destination array
5259 memnode_worklist.append_if_missing(use);
5260 } else {
5261 uint op = use->Opcode();
5262 if ((use->in(MemNode::Memory) == n) &&
5263 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
5264 // They overwrite memory edge corresponding to destination array,
5265 memnode_worklist.append_if_missing(use);
5266 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
5267 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
5268 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
5269 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
5270 n->dump();
5271 use->dump();
5272 assert(false, "EA: missing memory path");
5273 }
5274 #endif
5275 }
5276 }
5277 }
5278
5279 // Phase 3: Process MergeMem nodes from mergemem_worklist.
5280 // Walk each memory slice moving the first node encountered of each
5281 // instance type to the input corresponding to its alias index.
5282 uint length = mergemem_worklist.length();
5283 for( uint next = 0; next < length; ++next ) {
5284 MergeMemNode* nmm = mergemem_worklist.at(next);
5285 assert(!visited.test_set(nmm->_idx), "should not be visited before");
5286 // Note: we don't want to use MergeMemStream here because we only want to
5287 // scan inputs which exist at the start, not ones we add during processing.
5288 // Note 2: MergeMem may already contains instance memory slices added
5289 // during find_inst_mem() call when memory nodes were processed above.
5290 igvn->hash_delete(nmm);
5291 uint nslices = MIN2(nmm->req(), new_index_start);
5292 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
5293 Node* mem = nmm->in(i);
5294 Node* cur = nullptr;
5295 if (mem == nullptr || mem->is_top()) {
5296 continue;
5297 }
5298 // First, update mergemem by moving memory nodes to corresponding slices
5299 // if their type became more precise since this mergemem was created.
5300 while (mem->is_Mem()) {
5301 const Type* at = igvn->type(mem->in(MemNode::Address));
5302 if (at != Type::TOP) {
5303 assert (at->isa_ptr() != nullptr, "pointer type required.");
5304 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
5305 if (idx == i) {
5306 if (cur == nullptr) {
5307 cur = mem;
5308 }
5309 } else {
5310 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
5311 nmm->set_memory_at(idx, mem);
5312 }
5313 }
5314 }
5315 mem = mem->in(MemNode::Memory);
5316 }
5317 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
5318 // Find any instance of the current type if we haven't encountered
5319 // already a memory slice of the instance along the memory chain.
5320 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5321 if((uint)_compile->get_general_index(ni) == i) {
5322 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
5323 if (nmm->is_empty_memory(m)) {
5324 Node* result = find_inst_mem(mem, ni, orig_phis);
5325 if (_compile->failing()) {
5326 return;
5327 }
5328 nmm->set_memory_at(ni, result);
5329 }
5330 }
5331 }
5332 }
5333 // Find the rest of instances values
5334 for (uint ni = new_index_start; ni < new_index_end; ni++) {
5335 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
5336 Node* result = step_through_mergemem(nmm, ni, tinst);
5337 if (result == nmm->base_memory()) {
5338 // Didn't find instance memory, search through general slice recursively.
5339 result = nmm->memory_at(_compile->get_general_index(ni));
5340 result = find_inst_mem(result, ni, orig_phis);
5341 if (_compile->failing()) {
5342 return;
5343 }
5344 nmm->set_memory_at(ni, result);
5345 }
5346 }
5347
5348 // If we have crossed the 3/4 point of max node limit it's too risky
5349 // to continue with EA/SR because we might hit the max node limit.
5350 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
5351 if (_compile->do_reduce_allocation_merges()) {
5352 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
5353 } else if (_invocation > 0) {
5354 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
5355 } else {
5356 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
5357 }
5358 return;
5359 }
5360
5361 igvn->hash_insert(nmm);
5362 record_for_optimizer(nmm);
5363 }
5364
5365 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5366
5367 // Phase 4: Update the inputs of non-instance memory Phis and
5368 // the Memory input of memnodes
5369 // First update the inputs of any non-instance Phi's from
5370 // which we split out an instance Phi. Note we don't have
5371 // to recursively process Phi's encountered on the input memory
5372 // chains as is done in split_memory_phi() since they will
5373 // also be processed here.
5374 for (int j = 0; j < orig_phis.length(); j++) {
5375 PhiNode *phi = orig_phis.at(j);
5376 int alias_idx = _compile->get_alias_index(phi->adr_type());
5377 igvn->hash_delete(phi);
5378 for (uint i = 1; i < phi->req(); i++) {
5379 Node *mem = phi->in(i);
5380 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5381 if (_compile->failing()) {
5382 return;
5383 }
5384 if (mem != new_mem) {
5385 phi->set_req(i, new_mem);
5386 }
5387 }
5388 igvn->hash_insert(phi);
5389 record_for_optimizer(phi);
5390 }
5391
5392 // Update the memory inputs of MemNodes with the value we computed
5393 // in Phase 2 and move stores memory users to corresponding memory slices.
5394 // Disable memory split verification code until the fix for 6984348.
5395 // Currently it produces false negative results since it does not cover all cases.
5396 #if 0 // ifdef ASSERT
5397 visited.Reset();
5398 Node_Stack old_mems(arena, _compile->unique() >> 2);
5399 #endif
5400 for (uint i = 0; i < ideal_nodes.size(); i++) {
5401 Node* n = ideal_nodes.at(i);
5402 Node* nmem = get_map(n->_idx);
5403 assert(nmem != nullptr, "sanity");
5404 if (n->is_Mem()) {
5405 #if 0 // ifdef ASSERT
5406 Node* old_mem = n->in(MemNode::Memory);
5407 if (!visited.test_set(old_mem->_idx)) {
5408 old_mems.push(old_mem, old_mem->outcnt());
5409 }
5410 #endif
5411 assert(n->in(MemNode::Memory) != nmem, "sanity");
5412 if (!n->is_Load()) {
5413 // Move memory users of a store first.
5414 move_inst_mem(n, orig_phis);
5415 }
5416 // Now update memory input
5417 igvn->hash_delete(n);
5418 n->set_req(MemNode::Memory, nmem);
5419 igvn->hash_insert(n);
5420 record_for_optimizer(n);
5421 } else {
5422 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5423 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5424 }
5425 }
5426 #if 0 // ifdef ASSERT
5427 // Verify that memory was split correctly
5428 while (old_mems.is_nonempty()) {
5429 Node* old_mem = old_mems.node();
5430 uint old_cnt = old_mems.index();
5431 old_mems.pop();
5432 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5433 }
5434 #endif
5435 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5436 }
5437
5438 #ifndef PRODUCT
5439 int ConnectionGraph::_no_escape_counter = 0;
5440 int ConnectionGraph::_arg_escape_counter = 0;
5441 int ConnectionGraph::_global_escape_counter = 0;
5442
5443 static const char *node_type_names[] = {
5444 "UnknownType",
5445 "JavaObject",
5446 "LocalVar",
5447 "Field",
5448 "Arraycopy"
5449 };
5450
5451 static const char *esc_names[] = {
5452 "UnknownEscape",
5453 "NoEscape",
5454 "ArgEscape",
5455 "GlobalEscape"
5456 };
5457
5458 const char* PointsToNode::esc_name() const {
5459 return esc_names[(int)escape_state()];
5460 }
5461
5462 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5463 NodeType nt = node_type();
5464 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5465 if (print_state) {
5466 EscapeState es = escape_state();
5467 EscapeState fields_es = fields_escape_state();
5468 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5469 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5470 out->print("NSR ");
5471 }
5472 }
5473 }
5474
5475 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5476 dump_header(print_state, out);
5477 if (is_Field()) {
5478 FieldNode* f = (FieldNode*)this;
5479 if (f->is_oop()) {
5480 out->print("oop ");
5481 }
5482 if (f->offset() > 0) {
5483 out->print("+%d ", f->offset());
5484 }
5485 out->print("(");
5486 for (BaseIterator i(f); i.has_next(); i.next()) {
5487 PointsToNode* b = i.get();
5488 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5489 }
5490 out->print(" )");
5491 }
5492 out->print("[");
5493 for (EdgeIterator i(this); i.has_next(); i.next()) {
5494 PointsToNode* e = i.get();
5495 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5496 }
5497 out->print(" [");
5498 for (UseIterator i(this); i.has_next(); i.next()) {
5499 PointsToNode* u = i.get();
5500 bool is_base = false;
5501 if (PointsToNode::is_base_use(u)) {
5502 is_base = true;
5503 u = PointsToNode::get_use_node(u)->as_Field();
5504 }
5505 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5506 }
5507 out->print(" ]] ");
5508 if (_node == nullptr) {
5509 out->print("<null>%s", newline ? "\n" : "");
5510 } else {
5511 _node->dump(newline ? "\n" : "", false, out);
5512 }
5513 }
5514
5515 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5516 bool first = true;
5517 int ptnodes_length = ptnodes_worklist.length();
5518 for (int i = 0; i < ptnodes_length; i++) {
5519 PointsToNode *ptn = ptnodes_worklist.at(i);
5520 if (ptn == nullptr || !ptn->is_JavaObject()) {
5521 continue;
5522 }
5523 PointsToNode::EscapeState es = ptn->escape_state();
5524 if ((es != PointsToNode::NoEscape) && !Verbose) {
5525 continue;
5526 }
5527 Node* n = ptn->ideal_node();
5528 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5529 n->as_CallStaticJava()->is_boxing_method())) {
5530 if (first) {
5531 tty->cr();
5532 tty->print("======== Connection graph for ");
5533 _compile->method()->print_short_name();
5534 tty->cr();
5535 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5536 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5537 tty->cr();
5538 first = false;
5539 }
5540 ptn->dump();
5541 // Print all locals and fields which reference this allocation
5542 for (UseIterator j(ptn); j.has_next(); j.next()) {
5543 PointsToNode* use = j.get();
5544 if (use->is_LocalVar()) {
5545 use->dump(Verbose);
5546 } else if (Verbose) {
5547 use->dump();
5548 }
5549 }
5550 tty->cr();
5551 }
5552 }
5553 }
5554
5555 void ConnectionGraph::print_statistics() {
5556 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));
5557 }
5558
5559 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5560 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5561 return;
5562 }
5563 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5564 JavaObjectNode* ptn = java_objects_worklist.at(next);
5565 if (ptn->ideal_node()->is_Allocate()) {
5566 if (ptn->escape_state() == PointsToNode::NoEscape) {
5567 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5568 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5569 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5570 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5571 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5572 } else {
5573 assert(false, "Unexpected Escape State");
5574 }
5575 }
5576 }
5577 }
5578
5579 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5580 if (_compile->directive()->TraceEscapeAnalysisOption) {
5581 assert(ptn != nullptr, "should not be null");
5582 assert(reason != nullptr, "should not be null");
5583 ptn->dump_header(true);
5584 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5585 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5586 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5587 }
5588 }
5589
5590 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5591 if (_compile->directive()->TraceEscapeAnalysisOption) {
5592 stringStream ss;
5593 ss.print("propagated from: ");
5594 from->dump(true, &ss, false);
5595 return ss.as_string();
5596 } else {
5597 return nullptr;
5598 }
5599 }
5600
5601 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5602 if (_compile->directive()->TraceEscapeAnalysisOption) {
5603 stringStream ss;
5604 ss.print("escapes as arg to:");
5605 call->dump("", false, &ss);
5606 return ss.as_string();
5607 } else {
5608 return nullptr;
5609 }
5610 }
5611
5612 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5613 if (_compile->directive()->TraceEscapeAnalysisOption) {
5614 stringStream ss;
5615 ss.print("is merged with other object: ");
5616 other->dump_header(true, &ss);
5617 return ss.as_string();
5618 } else {
5619 return nullptr;
5620 }
5621 }
5622
5623 #endif
5624
5625 void ConnectionGraph::record_for_optimizer(Node *n) {
5626 _igvn->_worklist.push(n);
5627 _igvn->add_users_to_worklist(n);
5628 }