1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/bcEscapeAnalyzer.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/c2compiler.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "utilities/macros.hpp"
46
47 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
48 // If ReduceAllocationMerges is enabled we might call split_through_phi during
49 // split_unique_types and that will create additional nodes that need to be
50 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
51 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
52 // the array will be reallocated.
53 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
54 _in_worklist(C->comp_arena()),
55 _next_pidx(0),
56 _collecting(true),
57 _verify(false),
58 _compile(C),
59 _igvn(igvn),
60 _invocation(invocation),
61 _build_iterations(0),
62 _build_time(0.),
63 _node_map(C->comp_arena()) {
64 // Add unknown java object.
65 add_java_object(C->top(), PointsToNode::GlobalEscape);
66 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
67 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
68 // Add ConP and ConN null oop nodes
69 Node* oop_null = igvn->zerocon(T_OBJECT);
70 assert(oop_null->_idx < nodes_size(), "should be created already");
71 add_java_object(oop_null, PointsToNode::NoEscape);
72 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
73 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
74 if (UseCompressedOops) {
75 Node* noop_null = igvn->zerocon(T_NARROWOOP);
76 assert(noop_null->_idx < nodes_size(), "should be created already");
77 map_ideal_node(noop_null, null_obj);
78 }
79 }
80
81 bool ConnectionGraph::has_candidates(Compile *C) {
82 // EA brings benefits only when the code has allocations and/or locks which
83 // are represented by ideal Macro nodes.
84 int cnt = C->macro_count();
85 for (int i = 0; i < cnt; i++) {
86 Node *n = C->macro_node(i);
87 if (n->is_Allocate()) {
88 return true;
89 }
90 if (n->is_Lock()) {
91 Node* obj = n->as_Lock()->obj_node()->uncast();
92 if (!(obj->is_Parm() || obj->is_Con())) {
93 return true;
94 }
95 }
96 if (n->is_CallStaticJava() &&
97 n->as_CallStaticJava()->is_boxing_method()) {
98 return true;
99 }
100 }
101 return false;
102 }
103
104 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
105 Compile::TracePhase tp(Phase::_t_escapeAnalysis);
106 ResourceMark rm;
107
108 // Add ConP and ConN null oop nodes before ConnectionGraph construction
109 // to create space for them in ConnectionGraph::_nodes[].
110 Node* oop_null = igvn->zerocon(T_OBJECT);
111 Node* noop_null = igvn->zerocon(T_NARROWOOP);
112 int invocation = 0;
113 if (C->congraph() != nullptr) {
114 invocation = C->congraph()->_invocation + 1;
115 }
116 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
117 // Perform escape analysis
118 if (congraph->compute_escape()) {
119 // There are non escaping objects.
120 C->set_congraph(congraph);
121 }
122 // Cleanup.
123 if (oop_null->outcnt() == 0) {
124 igvn->hash_delete(oop_null);
125 }
126 if (noop_null->outcnt() == 0) {
127 igvn->hash_delete(noop_null);
128 }
129 }
130
131 bool ConnectionGraph::compute_escape() {
132 Compile* C = _compile;
133 PhaseGVN* igvn = _igvn;
134
135 // Worklists used by EA.
136 Unique_Node_List delayed_worklist;
137 Unique_Node_List reducible_merges;
138 GrowableArray<Node*> alloc_worklist;
139 GrowableArray<Node*> ptr_cmp_worklist;
140 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
141 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
142 GrowableArray<PointsToNode*> ptnodes_worklist;
143 GrowableArray<JavaObjectNode*> java_objects_worklist;
144 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
145 GrowableArray<FieldNode*> oop_fields_worklist;
146 GrowableArray<SafePointNode*> sfn_worklist;
147 GrowableArray<MergeMemNode*> mergemem_worklist;
148 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
149
150 { Compile::TracePhase tp(Phase::_t_connectionGraph);
151
152 // 1. Populate Connection Graph (CG) with PointsTo nodes.
153 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
154 // Initialize worklist
155 if (C->root() != nullptr) {
156 ideal_nodes.push(C->root());
157 }
158 // Processed ideal nodes are unique on ideal_nodes list
159 // but several ideal nodes are mapped to the phantom_obj.
160 // To avoid duplicated entries on the following worklists
161 // add the phantom_obj only once to them.
162 ptnodes_worklist.append(phantom_obj);
163 java_objects_worklist.append(phantom_obj);
164 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
165 Node* n = ideal_nodes.at(next);
166 // Create PointsTo nodes and add them to Connection Graph. Called
167 // only once per ideal node since ideal_nodes is Unique_Node list.
168 add_node_to_connection_graph(n, &delayed_worklist);
169 PointsToNode* ptn = ptnode_adr(n->_idx);
170 if (ptn != nullptr && ptn != phantom_obj) {
171 ptnodes_worklist.append(ptn);
172 if (ptn->is_JavaObject()) {
173 java_objects_worklist.append(ptn->as_JavaObject());
174 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
175 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
176 // Only allocations and java static calls results are interesting.
177 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
178 }
179 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
180 oop_fields_worklist.append(ptn->as_Field());
181 }
182 }
183 // Collect some interesting nodes for further use.
184 switch (n->Opcode()) {
185 case Op_MergeMem:
186 // Collect all MergeMem nodes to add memory slices for
187 // scalar replaceable objects in split_unique_types().
188 mergemem_worklist.append(n->as_MergeMem());
189 break;
190 case Op_CmpP:
191 case Op_CmpN:
192 // Collect compare pointers nodes.
193 if (OptimizePtrCompare) {
194 ptr_cmp_worklist.append(n);
195 }
196 break;
197 case Op_MemBarStoreStore:
198 // Collect all MemBarStoreStore nodes so that depending on the
199 // escape status of the associated Allocate node some of them
200 // may be eliminated.
201 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
202 storestore_worklist.append(n->as_MemBarStoreStore());
203 }
204 break;
205 case Op_MemBarRelease:
206 if (n->req() > MemBarNode::Precedent) {
207 record_for_optimizer(n);
208 }
209 break;
210 #ifdef ASSERT
211 case Op_AddP:
212 // Collect address nodes for graph verification.
213 addp_worklist.append(n);
214 break;
215 #endif
216 case Op_ArrayCopy:
217 // Keep a list of ArrayCopy nodes so if one of its input is non
218 // escaping, we can record a unique type
219 arraycopy_worklist.append(n->as_ArrayCopy());
220 break;
221 default:
222 // not interested now, ignore...
223 break;
224 }
225 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
226 Node* m = n->fast_out(i); // Get user
227 ideal_nodes.push(m);
228 }
229 if (n->is_SafePoint()) {
230 sfn_worklist.append(n->as_SafePoint());
231 }
232 }
233
234 #ifndef PRODUCT
235 if (_compile->directive()->TraceEscapeAnalysisOption) {
236 tty->print("+++++ Initial worklist for ");
237 _compile->method()->print_name();
238 tty->print_cr(" (ea_inv=%d)", _invocation);
239 for (int i = 0; i < ptnodes_worklist.length(); i++) {
240 PointsToNode* ptn = ptnodes_worklist.at(i);
241 ptn->dump();
242 }
243 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
244 }
245 #endif
246
247 if (non_escaped_allocs_worklist.length() == 0) {
248 _collecting = false;
249 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
250 return false; // Nothing to do.
251 }
252 // Add final simple edges to graph.
253 while(delayed_worklist.size() > 0) {
254 Node* n = delayed_worklist.pop();
255 add_final_edges(n);
256 }
257
258 #ifdef ASSERT
259 if (VerifyConnectionGraph) {
260 // Verify that no new simple edges could be created and all
261 // local vars has edges.
262 _verify = true;
263 int ptnodes_length = ptnodes_worklist.length();
264 for (int next = 0; next < ptnodes_length; ++next) {
265 PointsToNode* ptn = ptnodes_worklist.at(next);
266 add_final_edges(ptn->ideal_node());
267 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
268 ptn->dump();
269 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
270 }
271 }
272 _verify = false;
273 }
274 #endif
275 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
276 // processing, calls to CI to resolve symbols (types, fields, methods)
277 // referenced in bytecode. During symbol resolution VM may throw
278 // an exception which CI cleans and converts to compilation failure.
279 if (C->failing()) {
280 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
281 return false;
282 }
283
284 // 2. Finish Graph construction by propagating references to all
285 // java objects through graph.
286 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
287 java_objects_worklist, oop_fields_worklist)) {
288 // All objects escaped or hit time or iterations limits.
289 _collecting = false;
290 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
291 return false;
292 }
293
294 // 3. Adjust scalar_replaceable state of nonescaping objects and push
295 // scalar replaceable allocations on alloc_worklist for processing
296 // in split_unique_types().
297 GrowableArray<JavaObjectNode*> jobj_worklist;
298 int non_escaped_length = non_escaped_allocs_worklist.length();
299 bool found_nsr_alloc = false;
300 for (int next = 0; next < non_escaped_length; next++) {
301 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
302 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
303 Node* n = ptn->ideal_node();
304 if (n->is_Allocate()) {
305 n->as_Allocate()->_is_non_escaping = noescape;
306 }
307 if (noescape && ptn->scalar_replaceable()) {
308 adjust_scalar_replaceable_state(ptn, reducible_merges);
309 if (ptn->scalar_replaceable()) {
310 jobj_worklist.push(ptn);
311 } else {
312 found_nsr_alloc = true;
313 }
314 }
315 }
316
317 // Propagate NSR (Not Scalar Replaceable) state.
318 if (found_nsr_alloc) {
319 find_scalar_replaceable_allocs(jobj_worklist, reducible_merges);
320 }
321
322 // alloc_worklist will be processed in reverse push order.
323 // Therefore the reducible Phis will be processed for last and that's what we
324 // want because by then the scalarizable inputs of the merge will already have
325 // an unique instance type.
326 for (uint i = 0; i < reducible_merges.size(); i++ ) {
327 Node* n = reducible_merges.at(i);
328 alloc_worklist.append(n);
329 }
330
331 for (int next = 0; next < jobj_worklist.length(); ++next) {
332 JavaObjectNode* jobj = jobj_worklist.at(next);
333 if (jobj->scalar_replaceable()) {
334 alloc_worklist.append(jobj->ideal_node());
335 }
336 }
337
338 #ifdef ASSERT
339 if (VerifyConnectionGraph) {
340 // Verify that graph is complete - no new edges could be added or needed.
341 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
342 java_objects_worklist, addp_worklist);
343 }
344 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
345 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
346 null_obj->edge_count() == 0 &&
347 !null_obj->arraycopy_src() &&
348 !null_obj->arraycopy_dst(), "sanity");
349 #endif
350
351 _collecting = false;
352
353 } // TracePhase t3("connectionGraph")
354
355 // 4. Optimize ideal graph based on EA information.
356 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
357 if (has_non_escaping_obj) {
358 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
359 }
360
361 #ifndef PRODUCT
362 if (PrintEscapeAnalysis) {
363 dump(ptnodes_worklist); // Dump ConnectionGraph
364 }
365 #endif
366
367 #ifdef ASSERT
368 if (VerifyConnectionGraph) {
369 int alloc_length = alloc_worklist.length();
370 for (int next = 0; next < alloc_length; ++next) {
371 Node* n = alloc_worklist.at(next);
372 PointsToNode* ptn = ptnode_adr(n->_idx);
373 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
374 }
375 }
376
377 if (VerifyReduceAllocationMerges) {
378 for (uint i = 0; i < reducible_merges.size(); i++ ) {
379 Node* n = reducible_merges.at(i);
380 if (!can_reduce_phi(n->as_Phi())) {
381 TraceReduceAllocationMerges = true;
382 n->dump(2);
383 n->dump(-2);
384 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
385 }
386 }
387 }
388 #endif
389
390 // 5. Separate memory graph for scalar replaceable allcations.
391 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
392 if (has_scalar_replaceable_candidates && EliminateAllocations) {
393 assert(C->do_aliasing(), "Aliasing should be enabled");
394 // Now use the escape information to create unique types for
395 // scalar replaceable objects.
396 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
397 if (C->failing()) {
398 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
399 return false;
400 }
401 C->print_method(PHASE_AFTER_EA, 2);
402
403 #ifdef ASSERT
404 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
405 tty->print("=== No allocations eliminated for ");
406 C->method()->print_short_name();
407 if (!EliminateAllocations) {
408 tty->print(" since EliminateAllocations is off ===");
409 } else if(!has_scalar_replaceable_candidates) {
410 tty->print(" since there are no scalar replaceable candidates ===");
411 }
412 tty->cr();
413 #endif
414 }
415
416 // 6. Reduce allocation merges used as debug information. This is done after
417 // split_unique_types because the methods used to create SafePointScalarObject
418 // need to traverse the memory graph to find values for object fields. We also
419 // set to null the scalarized inputs of reducible Phis so that the Allocate
420 // that they point can be later scalar replaced.
421 bool delay = _igvn->delay_transform();
422 _igvn->set_delay_transform(true);
423 for (uint i = 0; i < reducible_merges.size(); i++) {
424 Node* n = reducible_merges.at(i);
425 if (n->outcnt() > 0) {
426 if (!reduce_phi_on_safepoints(n->as_Phi())) {
427 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
428 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
429 return false;
430 }
431
432 // Now we set the scalar replaceable inputs of ophi to null, which is
433 // the last piece that would prevent it from being scalar replaceable.
434 reset_scalar_replaceable_entries(n->as_Phi());
435 }
436 }
437 _igvn->set_delay_transform(delay);
438
439 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
440 // java calls if they pass ArgEscape objects as parameters.
441 if (has_non_escaping_obj &&
442 (C->env()->should_retain_local_variables() ||
443 C->env()->jvmti_can_get_owned_monitor_info() ||
444 C->env()->jvmti_can_walk_any_space() ||
445 DeoptimizeObjectsALot)) {
446 int sfn_length = sfn_worklist.length();
447 for (int next = 0; next < sfn_length; next++) {
448 SafePointNode* sfn = sfn_worklist.at(next);
449 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
450 if (sfn->is_CallJava()) {
451 CallJavaNode* call = sfn->as_CallJava();
452 call->set_arg_escape(has_arg_escape(call));
453 }
454 }
455 }
456
457 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
458 return has_non_escaping_obj;
459 }
460
461 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
462 // if at least one scalar replaceable allocation participates in the merge.
463 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
464 bool found_sr_allocate = false;
465
466 for (uint i = 1; i < ophi->req(); i++) {
467 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
468 if (ptn != nullptr && ptn->scalar_replaceable()) {
469 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
470
471 // Don't handle arrays.
472 if (alloc->Opcode() != Op_Allocate) {
473 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
474 continue;
475 }
476
477 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
478 found_sr_allocate = true;
479 } else {
480 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
481 ptn->set_scalar_replaceable(false);
482 }
483 }
484 }
485
486 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);)
487 return found_sr_allocate;
488 }
489
490 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
491 // I require the 'other' input to be a constant so that I can move the Cmp
492 // around safely.
493 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
494 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
495 Node* left = cmp->in(1);
496 Node* right = cmp->in(2);
497
498 return (left == n || right == n) &&
499 (left->is_Con() || right->is_Con()) &&
500 cmp->outcnt() == 1;
501 }
502
503 // We are going to check if any of the SafePointScalarMerge entries
504 // in the SafePoint reference the Phi that we are checking.
505 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
506 JVMState *jvms = sfpt->jvms();
507
508 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
509 Node* sfpt_in = sfpt->in(i);
510 if (sfpt_in->is_SafePointScalarMerge()) {
511 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
512 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
513 if (nsr_ptr == n) {
514 return true;
515 }
516 }
517 }
518
519 return false;
520 }
521
522 // Check if we are able to untangle the merge. The following patterns are
523 // supported:
524 // - Phi -> SafePoints
525 // - Phi -> CmpP/N
526 // - Phi -> AddP -> Load
527 // - Phi -> CastPP -> SafePoints
528 // - Phi -> CastPP -> AddP -> Load
529 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
530 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
531 Node* use = n->fast_out(i);
532
533 if (use->is_SafePoint()) {
534 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
535 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
536 return false;
537 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
538 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
539 return false;
540 }
541 } else if (use->is_AddP()) {
542 Node* addp = use;
543 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
544 Node* use_use = addp->fast_out(j);
545 const Type* load_type = _igvn->type(use_use);
546
547 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
548 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());)
549 return false;
550 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
551 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());)
552 return false;
553 }
554 }
555 } else if (nesting > 0) {
556 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);)
557 return false;
558 } else if (use->is_CastPP()) {
559 const Type* cast_t = _igvn->type(use);
560 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
561 #ifndef PRODUCT
562 if (TraceReduceAllocationMerges) {
563 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
564 use->dump();
565 }
566 #endif
567 return false;
568 }
569
570 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
571 if (!is_trivial_control) {
572 // If it's not a trivial control then we check if we can reduce the
573 // CmpP/N used by the If controlling the cast.
574 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
575 Node* iff = use->in(0)->in(0);
576 // We may have an OpaqueNotNull node between If and Bool nodes. But we could also have a sub class of IfNode,
577 // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
578 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
579 if (can_reduce) {
580 Node* iff_cmp = iff->in(1)->in(1);
581 int opc = iff_cmp->Opcode();
582 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
583 }
584 if (!can_reduce) {
585 #ifndef PRODUCT
586 if (TraceReduceAllocationMerges) {
587 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
588 n->dump(5);
589 }
590 #endif
591 return false;
592 }
593 }
594 }
595
596 if (!can_reduce_check_users(use, nesting+1)) {
597 return false;
598 }
599 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
600 if (!can_reduce_cmp(n, use)) {
601 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);)
602 return false;
603 }
604 } else {
605 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());)
606 return false;
607 }
608 }
609
610 return true;
611 }
612
613 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
614 // only used in some certain code shapes. Check comments in
615 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
616 // details.
617 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
618 // If there was an error attempting to reduce allocation merges for this
619 // method we might have disabled the compilation and be retrying with RAM
620 // disabled.
621 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
622 return false;
623 }
624
625 const Type* phi_t = _igvn->type(ophi);
626 if (phi_t == nullptr ||
627 phi_t->make_ptr() == nullptr ||
628 phi_t->make_ptr()->isa_aryptr() != nullptr) {
629 return false;
630 }
631
632 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
633 return false;
634 }
635
636 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
637 return true;
638 }
639
640 // This method will return a CmpP/N that we need to use on the If controlling a
641 // CastPP after it was split. This method is only called on bases that are
642 // nullable therefore we always need a controlling if for the splitted CastPP.
643 //
644 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
645 // If the CastPP currently doesn't have a control then the CmpP/N will be
646 // against the null constant, otherwise it will be against the constant input of
647 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
648 // case because we have constraints on it and because the CastPP has a control
649 // input.
650 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
651 const Type* t = base->bottom_type();
652 Node* con = nullptr;
653
654 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
655 con = _igvn->zerocon(t->basic_type());
656 } else {
657 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
658 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
659 Node* bol = curr_ctrl->in(0)->in(1);
660 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
661 Node* curr_cmp = bol->in(1);
662 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
663 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
664 }
665
666 return CmpNode::make(base, con, t->basic_type());
667 }
668
669 // This method 'specializes' the CastPP passed as parameter to the base passed
670 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
671 // means that the CastPP now will be specific for a given base instead of a Phi.
672 // An If-Then-Else-Region block is inserted to control the CastPP. The control
673 // of the CastPP is a copy of the current one (if there is one) or a check
674 // against null.
675 //
676 // Before:
677 //
678 // C1 C2 ... Cn
679 // \ | /
680 // \ | /
681 // \ | /
682 // \ | /
683 // \ | /
684 // \ | /
685 // \|/
686 // Region B1 B2 ... Bn
687 // | \ | /
688 // | \ | /
689 // | \ | /
690 // | \ | /
691 // | \ | /
692 // | \ | /
693 // ---------------> Phi
694 // |
695 // X |
696 // | |
697 // | |
698 // ------> CastPP
699 //
700 // After (only partial illustration; base = B2, current_control = C2):
701 //
702 // C2
703 // |
704 // If
705 // / \
706 // / \
707 // T F
708 // /\ /
709 // / \ /
710 // / \ /
711 // C1 CastPP Reg Cn
712 // | | |
713 // | | |
714 // | | |
715 // -------------- | ----------
716 // | | |
717 // Region
718 //
719 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
720 Node* control_successor = current_control->unique_ctrl_out();
721 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
722 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
723 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
724 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
725 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
726 Node* end_region = _igvn->transform(new RegionNode(3));
727
728 // Insert the new if-else-region block into the graph
729 end_region->set_req(1, not_eq_control);
730 end_region->set_req(2, yes_eq_control);
731 control_successor->replace_edge(current_control, end_region, _igvn);
732
733 _igvn->_worklist.push(current_control);
734 _igvn->_worklist.push(control_successor);
735
736 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::UnconditionalDependency, nullptr));
737 }
738
739 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
740 const Type* load_type = _igvn->type(curr_load);
741 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
742 Node* memory = curr_load->in(MemNode::Memory);
743
744 // The data_phi merging the loads needs to be nullable if
745 // we are loading pointers.
746 if (load_type->make_ptr() != nullptr) {
747 if (load_type->isa_narrowoop()) {
748 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
749 } else if (load_type->isa_ptr()) {
750 load_type = load_type->meet(TypePtr::NULL_PTR);
751 } else {
752 assert(false, "Unexpected load ptr type.");
753 }
754 }
755
756 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
757
758 for (int i = 1; i < bases_for_loads->length(); i++) {
759 Node* base = bases_for_loads->at(i);
760 Node* cmp_region = nullptr;
761 if (base != nullptr) {
762 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
763 cmp_region = base->unique_ctrl_out_or_null();
764 assert(cmp_region != nullptr, "There should be.");
765 base = base->find_out_with(Op_CastPP);
766 }
767
768 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
769 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
770 Node* load = curr_load->clone();
771 load->set_req(0, nullptr);
772 load->set_req(1, mem);
773 load->set_req(2, addr);
774
775 if (cmp_region != nullptr) { // see comment on previous if
776 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
777 intermediate_phi->set_req(1, _igvn->transform(load));
778 load = intermediate_phi;
779 }
780
781 data_phi->set_req(i, _igvn->transform(load));
782 } else {
783 // Just use the default, which is already in phi
784 }
785 }
786
787 // Takes care of updating CG and split_unique_types worklists due
788 // to cloned AddP->Load.
789 updates_after_load_split(data_phi, curr_load, alloc_worklist);
790
791 return _igvn->transform(data_phi);
792 }
793
794 // This method only reduces CastPP fields loads; SafePoints are handled
795 // separately. The idea here is basically to clone the CastPP and place copies
796 // on each input of the Phi, including non-scalar replaceable inputs.
797 // Experimentation shows that the resulting IR graph is simpler that way than if
798 // we just split the cast through scalar-replaceable inputs.
799 //
800 // The reduction process requires that CastPP's control be one of:
801 // 1) no control,
802 // 2) the same region as Ophi, or
803 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
804 //
805 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
806 // flow. If the CastPP originally had an IfTrue/False control input then we'll
807 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
808 // juse use a CmpP/N against the null constant.
809 //
810 // The If-Then-Else-Region isn't always needed. For instance, if input to
811 // splitted cast was not nullable (or if it was the null constant) then we don't
812 // need (shouldn't) use a CastPP at all.
813 //
814 // After the casts are splitted we'll split the AddP->Loads through the Phi and
815 // connect them to the just split CastPPs.
816 //
817 // Before (CastPP control is same as Phi):
818 //
819 // Region Allocate Null Call
820 // | \ | /
821 // | \ | /
822 // | \ | /
823 // | \ | /
824 // | \ | /
825 // | \ | /
826 // ------------------> Phi # Oop Phi
827 // | |
828 // | |
829 // | |
830 // | |
831 // ----------------> CastPP
832 // |
833 // AddP
834 // |
835 // Load
836 //
837 // After (Very much simplified):
838 //
839 // Call Null
840 // \ /
841 // CmpP
842 // |
843 // Bool#NE
844 // |
845 // If
846 // / \
847 // T F
848 // / \ /
849 // / R
850 // CastPP |
851 // | |
852 // AddP |
853 // | |
854 // Load |
855 // \ | 0
856 // Allocate \ | /
857 // \ \ | /
858 // AddP Phi
859 // \ /
860 // Load /
861 // \ 0 /
862 // \ | /
863 // \|/
864 // Phi # "Field" Phi
865 //
866 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
867 Node* ophi = curr_castpp->in(1);
868 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
869
870 // Identify which base should be used for AddP->Load later when spliting the
871 // CastPP->Loads through ophi. Three kind of values may be stored in this
872 // array, depending on the nullability status of the corresponding input in
873 // ophi.
874 //
875 // - nullptr: Meaning that the base is actually the null constant and therefore
876 // we won't try to load from it.
877 //
878 // - CFG Node: Meaning that the base is a CastPP that was specialized for
879 // this input of Ophi. I.e., we added an If->Then->Else-Region
880 // that will 'activate' the CastPp only when the input is not Null.
881 //
882 // - Other Node: Meaning that the base is not nullable and therefore we'll try
883 // to load directly from it.
884 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
885
886 for (uint i = 1; i < ophi->req(); i++) {
887 Node* base = ophi->in(i);
888 const Type* base_t = _igvn->type(base);
889
890 if (base_t->maybe_null()) {
891 if (base->is_Con()) {
892 // Nothing todo as bases_for_loads[i] is already null
893 } else {
894 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
895 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
896 }
897 } else {
898 bases_for_loads.at_put(i, base);
899 }
900 }
901
902 // Now let's split the CastPP->Loads through the Phi
903 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
904 Node* use = curr_castpp->raw_out(i);
905 if (use->is_AddP()) {
906 for (int j = use->outcnt()-1; j >= 0;) {
907 Node* use_use = use->raw_out(j);
908 assert(use_use->is_Load(), "Expected this to be a Load node.");
909
910 // We can't make an unconditional load from a nullable input. The
911 // 'split_castpp_load_through_phi` method will add an
912 // 'If-Then-Else-Region` around nullable bases and only load from them
913 // when the input is not null.
914 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
915 _igvn->replace_node(use_use, phi);
916
917 --j;
918 j = MIN2(j, (int)use->outcnt()-1);
919 }
920
921 _igvn->remove_dead_node(use);
922 }
923 --i;
924 i = MIN2(i, (int)curr_castpp->outcnt()-1);
925 }
926 }
927
928 // This method split a given CmpP/N through the Phi used in one of its inputs.
929 // As a result we convert a comparison with a pointer to a comparison with an
930 // integer.
931 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
932 // while the other must be a constant.
933 // The splitting process is basically just cloning the CmpP/N above the input
934 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
935 // can prove at compile time the result of the comparison.
936 //
937 // Before:
938 //
939 // in1 in2 ... inN
940 // \ | /
941 // \ | /
942 // \ | /
943 // \ | /
944 // \ | /
945 // \ | /
946 // Phi
947 // | Other
948 // | /
949 // | /
950 // | /
951 // CmpP/N
952 //
953 // After:
954 //
955 // in1 Other in2 Other inN Other
956 // | | | | | |
957 // \ | | | | |
958 // \ / | / | /
959 // CmpP/N CmpP/N CmpP/N
960 // Bool Bool Bool
961 // \ | /
962 // \ | /
963 // \ | /
964 // \ | /
965 // \ | /
966 // \ | /
967 // \ | /
968 // \ | /
969 // Phi
970 // |
971 // | Zero
972 // | /
973 // | /
974 // | /
975 // CmpI
976 //
977 //
978 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
979 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
980 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
981
982 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
983 Node* zero = _igvn->intcon(0);
984 Node* one = _igvn->intcon(1);
985 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
986
987 // This Phi will merge the result of the Cmps split through the Phi
988 Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
989
990 for (uint i=1; i<ophi->req(); i++) {
991 Node* ophi_input = ophi->in(i);
992 Node* res_phi_input = nullptr;
993
994 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
995 if (tcmp->singleton()) {
996 if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
997 (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
998 res_phi_input = one;
999 } else {
1000 res_phi_input = zero;
1001 }
1002 } else {
1003 Node* ncmp = _igvn->transform(cmp->clone());
1004 ncmp->set_req(1, ophi_input);
1005 ncmp->set_req(2, other);
1006 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1007 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1008 }
1009
1010 res_phi->set_req(i, res_phi_input);
1011 }
1012
1013 // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1014 Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1015 _igvn->replace_node(cmp, new_cmp);
1016 }
1017
1018 // Push the newly created AddP on alloc_worklist and patch
1019 // the connection graph. Note that the changes in the CG below
1020 // won't affect the ES of objects since the new nodes have the
1021 // same status as the old ones.
1022 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1023 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1024 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1025 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1026
1027 if (data_phi == nullptr || !data_phi->is_Phi()) {
1028 // Make this a retry?
1029 return ;
1030 }
1031
1032 Node* previous_addp = previous_load->in(MemNode::Address);
1033 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1034 for (uint i = 1; i < data_phi->req(); i++) {
1035 Node* new_load = data_phi->in(i);
1036
1037 if (new_load->is_Phi()) {
1038 // new_load is currently the "intermediate_phi" from an specialized
1039 // CastPP.
1040 new_load = new_load->in(1);
1041 }
1042
1043 // "new_load" might actually be a constant, parameter, etc.
1044 if (new_load->is_Load()) {
1045 Node* new_addp = new_load->in(MemNode::Address);
1046 Node* base = get_addp_base(new_addp);
1047
1048 // The base might not be something that we can create an unique
1049 // type for. If that's the case we are done with that input.
1050 PointsToNode* jobj_ptn = unique_java_object(base);
1051 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1052 continue;
1053 }
1054
1055 // Push to alloc_worklist since the base has an unique_type
1056 alloc_worklist.append_if_missing(new_addp);
1057
1058 // Now let's add the node to the connection graph
1059 _nodes.at_grow(new_addp->_idx, nullptr);
1060 add_field(new_addp, fn->escape_state(), fn->offset());
1061 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1062
1063 // If the load doesn't load an object then it won't be
1064 // part of the connection graph
1065 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1066 if (curr_load_ptn != nullptr) {
1067 _nodes.at_grow(new_load->_idx, nullptr);
1068 add_local_var(new_load, curr_load_ptn->escape_state());
1069 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1070 }
1071 }
1072 }
1073 }
1074
1075 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1076 // We'll pass this to 'split_through_phi' so that it'll do the split even
1077 // though the load doesn't have an unique instance type.
1078 bool ignore_missing_instance_id = true;
1079
1080 // All AddPs are present in the connection graph
1081 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1082
1083 // Iterate over AddP looking for a Load
1084 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1085 Node* previous_load = previous_addp->raw_out(k);
1086 if (previous_load->is_Load()) {
1087 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1088
1089 // Takes care of updating CG and split_unique_types worklists due to cloned
1090 // AddP->Load.
1091 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1092
1093 _igvn->replace_node(previous_load, data_phi);
1094 }
1095 --k;
1096 k = MIN2(k, (int)previous_addp->outcnt()-1);
1097 }
1098
1099 // Remove the old AddP from the processing list because it's dead now
1100 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1101 alloc_worklist.remove_if_existing(previous_addp);
1102 }
1103
1104 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1105 // selector is:
1106 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1107 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1108 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1109 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1110 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1111 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1112 uint number_of_sr_objects = 0;
1113 for (uint i = 1; i < ophi->req(); i++) {
1114 Node* base = ophi->in(i);
1115 JavaObjectNode* ptn = unique_java_object(base);
1116
1117 if (ptn != nullptr && ptn->scalar_replaceable()) {
1118 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1119 selector->set_req(i, sr_obj_idx);
1120 number_of_sr_objects++;
1121 }
1122 }
1123
1124 return selector->as_Phi();
1125 }
1126
1127 // Returns true if the AddP node 'n' has at least one base that is a reducible
1128 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1129 // checked instead.
1130 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1131 PointsToNode* ptn = ptnode_adr(n->_idx);
1132 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1133 return false;
1134 }
1135
1136 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1137 Node* base = i.get()->ideal_node();
1138
1139 if (reducible_merges.member(base)) {
1140 return true;
1141 }
1142
1143 if (base->is_CastPP() || base->is_CheckCastPP()) {
1144 base = base->in(1);
1145 if (reducible_merges.member(base)) {
1146 return true;
1147 }
1148 }
1149 }
1150
1151 return false;
1152 }
1153
1154 // This method will call its helper method to reduce SafePoint nodes that use
1155 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1156 // "version" of Phi use the same debug information (regarding the Phi).
1157 // Therefore, I collect all safepoints and patch them all at once.
1158 //
1159 // The safepoints using the Phi node have to be processed before safepoints of
1160 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1161 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1162 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1163 // algorithm that process Phi's safepoints will think that the added Phi
1164 // reference is a regular reference.
1165 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1166 PhiNode* selector = create_selector(ophi);
1167 Unique_Node_List safepoints;
1168 Unique_Node_List casts;
1169
1170 // Just collect the users of the Phis for later processing
1171 // in the needed order.
1172 for (uint i = 0; i < ophi->outcnt(); i++) {
1173 Node* use = ophi->raw_out(i);
1174 if (use->is_SafePoint()) {
1175 safepoints.push(use);
1176 } else if (use->is_CastPP()) {
1177 casts.push(use);
1178 } else {
1179 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1180 }
1181 }
1182
1183 // Need to process safepoints using the Phi first
1184 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1185 return false;
1186 }
1187
1188 // Now process CastPP->safepoints
1189 for (uint i = 0; i < casts.size(); i++) {
1190 Node* cast = casts.at(i);
1191 Unique_Node_List cast_sfpts;
1192
1193 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1194 Node* use_use = cast->fast_out(j);
1195 if (use_use->is_SafePoint()) {
1196 cast_sfpts.push(use_use);
1197 } else {
1198 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1199 }
1200 }
1201
1202 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1203 return false;
1204 }
1205 }
1206
1207 return true;
1208 }
1209
1210 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1211 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1212 // SafePointScalarObjectNode for each scalar replaceable input. Each
1213 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1214 // check detailed description in SafePointScalarMergeNode class header.
1215 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1216 PhaseMacroExpand mexp(*_igvn);
1217 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1218 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1219
1220 Node* nsr_merge_pointer = ophi;
1221 if (cast != nullptr) {
1222 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1223 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::RegularDependency, nullptr));
1224 }
1225
1226 for (uint spi = 0; spi < safepoints.size(); spi++) {
1227 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1228 JVMState *jvms = sfpt->jvms();
1229 uint merge_idx = (sfpt->req() - jvms->scloff());
1230 int debug_start = jvms->debug_start();
1231
1232 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1233 smerge->init_req(0, _compile->root());
1234 _igvn->register_new_node_with_optimizer(smerge);
1235
1236 // The next two inputs are:
1237 // (1) A copy of the original pointer to NSR objects.
1238 // (2) A selector, used to decide if we need to rematerialize an object
1239 // or use the pointer to a NSR object.
1240 // See more details of these fields in the declaration of SafePointScalarMergeNode
1241 sfpt->add_req(nsr_merge_pointer);
1242 sfpt->add_req(selector);
1243
1244 for (uint i = 1; i < ophi->req(); i++) {
1245 Node* base = ophi->in(i);
1246 JavaObjectNode* ptn = unique_java_object(base);
1247
1248 // If the base is not scalar replaceable we don't need to register information about
1249 // it at this time.
1250 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1251 continue;
1252 }
1253
1254 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1255 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt);
1256 if (sobj == nullptr) {
1257 return false;
1258 }
1259
1260 // Now make a pass over the debug information replacing any references
1261 // to the allocated object with "sobj"
1262 Node* ccpp = alloc->result_cast();
1263 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1264
1265 // Register the scalarized object as a candidate for reallocation
1266 smerge->add_req(sobj);
1267 }
1268
1269 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1270 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1271
1272 // The call to 'replace_edges_in_range' above might have removed the
1273 // reference to ophi that we need at _merge_pointer_idx. The line below make
1274 // sure the reference is maintained.
1275 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1276 _igvn->_worklist.push(sfpt);
1277 }
1278
1279 return true;
1280 }
1281
1282 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
1283 bool delay = _igvn->delay_transform();
1284 _igvn->set_delay_transform(true);
1285 _igvn->hash_delete(ophi);
1286
1287 // Copying all users first because some will be removed and others won't.
1288 // Ophi also may acquire some new users as part of Cast reduction.
1289 // CastPPs also need to be processed before CmpPs.
1290 Unique_Node_List castpps;
1291 Unique_Node_List others;
1292 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1293 Node* use = ophi->fast_out(i);
1294
1295 if (use->is_CastPP()) {
1296 castpps.push(use);
1297 } else if (use->is_AddP() || use->is_Cmp()) {
1298 others.push(use);
1299 } else if (use->is_SafePoint()) {
1300 // processed later
1301 } else {
1302 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1303 }
1304 }
1305
1306 // CastPPs need to be processed before Cmps because during the process of
1307 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1308 // by the If controlling the CastPP.
1309 for (uint i = 0; i < castpps.size(); i++) {
1310 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist, memnode_worklist);
1311 }
1312
1313 for (uint i = 0; i < others.size(); i++) {
1314 Node* use = others.at(i);
1315
1316 if (use->is_AddP()) {
1317 reduce_phi_on_field_access(use, alloc_worklist);
1318 } else if(use->is_Cmp()) {
1319 reduce_phi_on_cmp(use);
1320 }
1321 }
1322
1323 _igvn->set_delay_transform(delay);
1324 }
1325
1326 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1327 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1328 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1329 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1330 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1331
1332 for (uint i = 1; i < ophi->req(); i++) {
1333 Node* base = ophi->in(i);
1334 JavaObjectNode* ptn = unique_java_object(base);
1335
1336 if (ptn != nullptr && ptn->scalar_replaceable()) {
1337 new_phi->set_req(i, null_ptr);
1338 } else {
1339 new_phi->set_req(i, ophi->in(i));
1340 }
1341 }
1342
1343 for (int i = ophi->outcnt()-1; i >= 0;) {
1344 Node* out = ophi->raw_out(i);
1345
1346 if (out->is_ConstraintCast()) {
1347 const Type* out_t = _igvn->type(out)->make_ptr();
1348 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1349 bool change = out_new_t != out_t;
1350
1351 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1352 Node* out2 = out->raw_out(j);
1353 if (!out2->is_SafePoint()) {
1354 change = false;
1355 break;
1356 }
1357 }
1358
1359 if (change) {
1360 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::StrongDependency, nullptr);
1361 _igvn->replace_node(out, new_cast);
1362 _igvn->register_new_node_with_optimizer(new_cast);
1363 }
1364 }
1365
1366 --i;
1367 i = MIN2(i, (int)ophi->outcnt()-1);
1368 }
1369
1370 _igvn->replace_node(ophi, new_phi);
1371 }
1372
1373 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1374 if (!C->do_reduce_allocation_merges()) return;
1375
1376 Unique_Node_List ideal_nodes;
1377 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1378 ideal_nodes.push(root);
1379
1380 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1381 Node* n = ideal_nodes.at(next);
1382
1383 if (n->is_SafePointScalarMerge()) {
1384 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1385
1386 // Validate inputs of merge
1387 for (uint i = 1; i < merge->req(); i++) {
1388 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1389 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1390 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1391 }
1392 }
1393
1394 // Validate users of merge
1395 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1396 Node* sfpt = merge->fast_out(i);
1397 if (sfpt->is_SafePoint()) {
1398 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1399
1400 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1401 assert(false, "SafePointScalarMerge nodes can't be nested.");
1402 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1403 }
1404 } else {
1405 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1406 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1407 }
1408 }
1409 }
1410
1411 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1412 Node* m = n->fast_out(i);
1413 ideal_nodes.push(m);
1414 }
1415 }
1416 }
1417
1418 // Returns true if there is an object in the scope of sfn that does not escape globally.
1419 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1420 Compile* C = _compile;
1421 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1422 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1423 DeoptimizeObjectsALot) {
1424 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1425 int num_locs = jvms->loc_size();
1426 for (int idx = 0; idx < num_locs; idx++) {
1427 Node* l = sfn->local(jvms, idx);
1428 if (not_global_escape(l)) {
1429 return true;
1430 }
1431 }
1432 }
1433 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1434 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1435 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1436 int num_mon = jvms->nof_monitors();
1437 for (int idx = 0; idx < num_mon; idx++) {
1438 Node* m = sfn->monitor_obj(jvms, idx);
1439 if (m != nullptr && not_global_escape(m)) {
1440 return true;
1441 }
1442 }
1443 }
1444 }
1445 return false;
1446 }
1447
1448 // Returns true if at least one of the arguments to the call is an object
1449 // that does not escape globally.
1450 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1451 if (call->method() != nullptr) {
1452 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1453 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1454 Node* p = call->in(idx);
1455 if (not_global_escape(p)) {
1456 return true;
1457 }
1458 }
1459 } else {
1460 const char* name = call->as_CallStaticJava()->_name;
1461 assert(name != nullptr, "no name");
1462 // no arg escapes through uncommon traps
1463 if (strcmp(name, "uncommon_trap") != 0) {
1464 // process_call_arguments() assumes that all arguments escape globally
1465 const TypeTuple* d = call->tf()->domain();
1466 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1467 const Type* at = d->field_at(i);
1468 if (at->isa_oopptr() != nullptr) {
1469 return true;
1470 }
1471 }
1472 }
1473 }
1474 return false;
1475 }
1476
1477
1478
1479 // Utility function for nodes that load an object
1480 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1481 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1482 // ThreadLocal has RawPtr type.
1483 const Type* t = _igvn->type(n);
1484 if (t->make_ptr() != nullptr) {
1485 Node* adr = n->in(MemNode::Address);
1486 #ifdef ASSERT
1487 if (!adr->is_AddP()) {
1488 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1489 } else {
1490 assert((ptnode_adr(adr->_idx) == nullptr ||
1491 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1492 }
1493 #endif
1494 add_local_var_and_edge(n, PointsToNode::NoEscape,
1495 adr, delayed_worklist);
1496 }
1497 }
1498
1499 // Populate Connection Graph with PointsTo nodes and create simple
1500 // connection graph edges.
1501 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1502 assert(!_verify, "this method should not be called for verification");
1503 PhaseGVN* igvn = _igvn;
1504 uint n_idx = n->_idx;
1505 PointsToNode* n_ptn = ptnode_adr(n_idx);
1506 if (n_ptn != nullptr) {
1507 return; // No need to redefine PointsTo node during first iteration.
1508 }
1509 int opcode = n->Opcode();
1510 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1511 if (gc_handled) {
1512 return; // Ignore node if already handled by GC.
1513 }
1514
1515 if (n->is_Call()) {
1516 // Arguments to allocation and locking don't escape.
1517 if (n->is_AbstractLock()) {
1518 // Put Lock and Unlock nodes on IGVN worklist to process them during
1519 // first IGVN optimization when escape information is still available.
1520 record_for_optimizer(n);
1521 } else if (n->is_Allocate()) {
1522 add_call_node(n->as_Call());
1523 record_for_optimizer(n);
1524 } else {
1525 if (n->is_CallStaticJava()) {
1526 const char* name = n->as_CallStaticJava()->_name;
1527 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1528 return; // Skip uncommon traps
1529 }
1530 }
1531 // Don't mark as processed since call's arguments have to be processed.
1532 delayed_worklist->push(n);
1533 // Check if a call returns an object.
1534 if ((n->as_Call()->returns_pointer() &&
1535 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1536 (n->is_CallStaticJava() &&
1537 n->as_CallStaticJava()->is_boxing_method())) {
1538 add_call_node(n->as_Call());
1539 }
1540 }
1541 return;
1542 }
1543 // Put this check here to process call arguments since some call nodes
1544 // point to phantom_obj.
1545 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1546 return; // Skip predefined nodes.
1547 }
1548 switch (opcode) {
1549 case Op_AddP: {
1550 Node* base = get_addp_base(n);
1551 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1552 // Field nodes are created for all field types. They are used in
1553 // adjust_scalar_replaceable_state() and split_unique_types().
1554 // Note, non-oop fields will have only base edges in Connection
1555 // Graph because such fields are not used for oop loads and stores.
1556 int offset = address_offset(n, igvn);
1557 add_field(n, PointsToNode::NoEscape, offset);
1558 if (ptn_base == nullptr) {
1559 delayed_worklist->push(n); // Process it later.
1560 } else {
1561 n_ptn = ptnode_adr(n_idx);
1562 add_base(n_ptn->as_Field(), ptn_base);
1563 }
1564 break;
1565 }
1566 case Op_CastX2P: {
1567 map_ideal_node(n, phantom_obj);
1568 break;
1569 }
1570 case Op_CastPP:
1571 case Op_CheckCastPP:
1572 case Op_EncodeP:
1573 case Op_DecodeN:
1574 case Op_EncodePKlass:
1575 case Op_DecodeNKlass: {
1576 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1577 break;
1578 }
1579 case Op_CMoveP: {
1580 add_local_var(n, PointsToNode::NoEscape);
1581 // Do not add edges during first iteration because some could be
1582 // not defined yet.
1583 delayed_worklist->push(n);
1584 break;
1585 }
1586 case Op_ConP:
1587 case Op_ConN:
1588 case Op_ConNKlass: {
1589 // assume all oop constants globally escape except for null
1590 PointsToNode::EscapeState es;
1591 const Type* t = igvn->type(n);
1592 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1593 es = PointsToNode::NoEscape;
1594 } else {
1595 es = PointsToNode::GlobalEscape;
1596 }
1597 PointsToNode* ptn_con = add_java_object(n, es);
1598 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1599 break;
1600 }
1601 case Op_CreateEx: {
1602 // assume that all exception objects globally escape
1603 map_ideal_node(n, phantom_obj);
1604 break;
1605 }
1606 case Op_LoadKlass:
1607 case Op_LoadNKlass: {
1608 // Unknown class is loaded
1609 map_ideal_node(n, phantom_obj);
1610 break;
1611 }
1612 case Op_LoadP:
1613 case Op_LoadN: {
1614 add_objload_to_connection_graph(n, delayed_worklist);
1615 break;
1616 }
1617 case Op_Parm: {
1618 map_ideal_node(n, phantom_obj);
1619 break;
1620 }
1621 case Op_PartialSubtypeCheck: {
1622 // Produces Null or notNull and is used in only in CmpP so
1623 // phantom_obj could be used.
1624 map_ideal_node(n, phantom_obj); // Result is unknown
1625 break;
1626 }
1627 case Op_Phi: {
1628 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1629 // ThreadLocal has RawPtr type.
1630 const Type* t = n->as_Phi()->type();
1631 if (t->make_ptr() != nullptr) {
1632 add_local_var(n, PointsToNode::NoEscape);
1633 // Do not add edges during first iteration because some could be
1634 // not defined yet.
1635 delayed_worklist->push(n);
1636 }
1637 break;
1638 }
1639 case Op_Proj: {
1640 // we are only interested in the oop result projection from a call
1641 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1642 n->in(0)->as_Call()->returns_pointer()) {
1643 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1644 }
1645 break;
1646 }
1647 case Op_Rethrow: // Exception object escapes
1648 case Op_Return: {
1649 if (n->req() > TypeFunc::Parms &&
1650 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1651 // Treat Return value as LocalVar with GlobalEscape escape state.
1652 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1653 }
1654 break;
1655 }
1656 case Op_CompareAndExchangeP:
1657 case Op_CompareAndExchangeN:
1658 case Op_GetAndSetP:
1659 case Op_GetAndSetN: {
1660 add_objload_to_connection_graph(n, delayed_worklist);
1661 // fall-through
1662 }
1663 case Op_StoreP:
1664 case Op_StoreN:
1665 case Op_StoreNKlass:
1666 case Op_WeakCompareAndSwapP:
1667 case Op_WeakCompareAndSwapN:
1668 case Op_CompareAndSwapP:
1669 case Op_CompareAndSwapN: {
1670 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1671 break;
1672 }
1673 case Op_AryEq:
1674 case Op_CountPositives:
1675 case Op_StrComp:
1676 case Op_StrEquals:
1677 case Op_StrIndexOf:
1678 case Op_StrIndexOfChar:
1679 case Op_StrInflatedCopy:
1680 case Op_StrCompressedCopy:
1681 case Op_VectorizedHashCode:
1682 case Op_EncodeISOArray: {
1683 add_local_var(n, PointsToNode::ArgEscape);
1684 delayed_worklist->push(n); // Process it later.
1685 break;
1686 }
1687 case Op_ThreadLocal: {
1688 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1689 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1690 break;
1691 }
1692 case Op_Blackhole: {
1693 // All blackhole pointer arguments are globally escaping.
1694 // Only do this if there is at least one pointer argument.
1695 // Do not add edges during first iteration because some could be
1696 // not defined yet, defer to final step.
1697 for (uint i = 0; i < n->req(); i++) {
1698 Node* in = n->in(i);
1699 if (in != nullptr) {
1700 const Type* at = _igvn->type(in);
1701 if (!at->isa_ptr()) continue;
1702
1703 add_local_var(n, PointsToNode::GlobalEscape);
1704 delayed_worklist->push(n);
1705 break;
1706 }
1707 }
1708 break;
1709 }
1710 default:
1711 ; // Do nothing for nodes not related to EA.
1712 }
1713 return;
1714 }
1715
1716 // Add final simple edges to graph.
1717 void ConnectionGraph::add_final_edges(Node *n) {
1718 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1719 #ifdef ASSERT
1720 if (_verify && n_ptn->is_JavaObject())
1721 return; // This method does not change graph for JavaObject.
1722 #endif
1723
1724 if (n->is_Call()) {
1725 process_call_arguments(n->as_Call());
1726 return;
1727 }
1728 assert(n->is_Store() || n->is_LoadStore() ||
1729 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1730 "node should be registered already");
1731 int opcode = n->Opcode();
1732 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1733 if (gc_handled) {
1734 return; // Ignore node if already handled by GC.
1735 }
1736 switch (opcode) {
1737 case Op_AddP: {
1738 Node* base = get_addp_base(n);
1739 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1740 assert(ptn_base != nullptr, "field's base should be registered");
1741 add_base(n_ptn->as_Field(), ptn_base);
1742 break;
1743 }
1744 case Op_CastPP:
1745 case Op_CheckCastPP:
1746 case Op_EncodeP:
1747 case Op_DecodeN:
1748 case Op_EncodePKlass:
1749 case Op_DecodeNKlass: {
1750 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1751 break;
1752 }
1753 case Op_CMoveP: {
1754 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1755 Node* in = n->in(i);
1756 if (in == nullptr) {
1757 continue; // ignore null
1758 }
1759 Node* uncast_in = in->uncast();
1760 if (uncast_in->is_top() || uncast_in == n) {
1761 continue; // ignore top or inputs which go back this node
1762 }
1763 PointsToNode* ptn = ptnode_adr(in->_idx);
1764 assert(ptn != nullptr, "node should be registered");
1765 add_edge(n_ptn, ptn);
1766 }
1767 break;
1768 }
1769 case Op_LoadP:
1770 case Op_LoadN: {
1771 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1772 // ThreadLocal has RawPtr type.
1773 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1774 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1775 break;
1776 }
1777 case Op_Phi: {
1778 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1779 // ThreadLocal has RawPtr type.
1780 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1781 for (uint i = 1; i < n->req(); i++) {
1782 Node* in = n->in(i);
1783 if (in == nullptr) {
1784 continue; // ignore null
1785 }
1786 Node* uncast_in = in->uncast();
1787 if (uncast_in->is_top() || uncast_in == n) {
1788 continue; // ignore top or inputs which go back this node
1789 }
1790 PointsToNode* ptn = ptnode_adr(in->_idx);
1791 assert(ptn != nullptr, "node should be registered");
1792 add_edge(n_ptn, ptn);
1793 }
1794 break;
1795 }
1796 case Op_Proj: {
1797 // we are only interested in the oop result projection from a call
1798 assert(n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1799 n->in(0)->as_Call()->returns_pointer(), "Unexpected node type");
1800 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1801 break;
1802 }
1803 case Op_Rethrow: // Exception object escapes
1804 case Op_Return: {
1805 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1806 "Unexpected node type");
1807 // Treat Return value as LocalVar with GlobalEscape escape state.
1808 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1809 break;
1810 }
1811 case Op_CompareAndExchangeP:
1812 case Op_CompareAndExchangeN:
1813 case Op_GetAndSetP:
1814 case Op_GetAndSetN:{
1815 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1816 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1817 // fall-through
1818 }
1819 case Op_CompareAndSwapP:
1820 case Op_CompareAndSwapN:
1821 case Op_WeakCompareAndSwapP:
1822 case Op_WeakCompareAndSwapN:
1823 case Op_StoreP:
1824 case Op_StoreN:
1825 case Op_StoreNKlass:{
1826 add_final_edges_unsafe_access(n, opcode);
1827 break;
1828 }
1829 case Op_VectorizedHashCode:
1830 case Op_AryEq:
1831 case Op_CountPositives:
1832 case Op_StrComp:
1833 case Op_StrEquals:
1834 case Op_StrIndexOf:
1835 case Op_StrIndexOfChar:
1836 case Op_StrInflatedCopy:
1837 case Op_StrCompressedCopy:
1838 case Op_EncodeISOArray: {
1839 // char[]/byte[] arrays passed to string intrinsic do not escape but
1840 // they are not scalar replaceable. Adjust escape state for them.
1841 // Start from in(2) edge since in(1) is memory edge.
1842 for (uint i = 2; i < n->req(); i++) {
1843 Node* adr = n->in(i);
1844 const Type* at = _igvn->type(adr);
1845 if (!adr->is_top() && at->isa_ptr()) {
1846 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1847 at->isa_ptr() != nullptr, "expecting a pointer");
1848 if (adr->is_AddP()) {
1849 adr = get_addp_base(adr);
1850 }
1851 PointsToNode* ptn = ptnode_adr(adr->_idx);
1852 assert(ptn != nullptr, "node should be registered");
1853 add_edge(n_ptn, ptn);
1854 }
1855 }
1856 break;
1857 }
1858 case Op_Blackhole: {
1859 // All blackhole pointer arguments are globally escaping.
1860 for (uint i = 0; i < n->req(); i++) {
1861 Node* in = n->in(i);
1862 if (in != nullptr) {
1863 const Type* at = _igvn->type(in);
1864 if (!at->isa_ptr()) continue;
1865
1866 if (in->is_AddP()) {
1867 in = get_addp_base(in);
1868 }
1869
1870 PointsToNode* ptn = ptnode_adr(in->_idx);
1871 assert(ptn != nullptr, "should be defined already");
1872 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1873 add_edge(n_ptn, ptn);
1874 }
1875 }
1876 break;
1877 }
1878 default: {
1879 // This method should be called only for EA specific nodes which may
1880 // miss some edges when they were created.
1881 #ifdef ASSERT
1882 n->dump(1);
1883 #endif
1884 guarantee(false, "unknown node");
1885 }
1886 }
1887 return;
1888 }
1889
1890 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1891 Node* adr = n->in(MemNode::Address);
1892 const Type* adr_type = _igvn->type(adr);
1893 adr_type = adr_type->make_ptr();
1894 if (adr_type == nullptr) {
1895 return; // skip dead nodes
1896 }
1897 if (adr_type->isa_oopptr()
1898 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1899 && adr_type == TypeRawPtr::NOTNULL
1900 && is_captured_store_address(adr))) {
1901 delayed_worklist->push(n); // Process it later.
1902 #ifdef ASSERT
1903 assert (adr->is_AddP(), "expecting an AddP");
1904 if (adr_type == TypeRawPtr::NOTNULL) {
1905 // Verify a raw address for a store captured by Initialize node.
1906 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1907 assert(offs != Type::OffsetBot, "offset must be a constant");
1908 }
1909 #endif
1910 } else {
1911 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1912 if (adr->is_BoxLock()) {
1913 return;
1914 }
1915 // Stored value escapes in unsafe access.
1916 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1917 delayed_worklist->push(n); // Process unsafe access later.
1918 return;
1919 }
1920 #ifdef ASSERT
1921 n->dump(1);
1922 assert(false, "not unsafe");
1923 #endif
1924 }
1925 }
1926
1927 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1928 Node* adr = n->in(MemNode::Address);
1929 const Type *adr_type = _igvn->type(adr);
1930 adr_type = adr_type->make_ptr();
1931 #ifdef ASSERT
1932 if (adr_type == nullptr) {
1933 n->dump(1);
1934 assert(adr_type != nullptr, "dead node should not be on list");
1935 return true;
1936 }
1937 #endif
1938
1939 if (adr_type->isa_oopptr()
1940 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1941 && adr_type == TypeRawPtr::NOTNULL
1942 && is_captured_store_address(adr))) {
1943 // Point Address to Value
1944 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1945 assert(adr_ptn != nullptr &&
1946 adr_ptn->as_Field()->is_oop(), "node should be registered");
1947 Node* val = n->in(MemNode::ValueIn);
1948 PointsToNode* ptn = ptnode_adr(val->_idx);
1949 assert(ptn != nullptr, "node should be registered");
1950 add_edge(adr_ptn, ptn);
1951 return true;
1952 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1953 // Stored value escapes in unsafe access.
1954 Node* val = n->in(MemNode::ValueIn);
1955 PointsToNode* ptn = ptnode_adr(val->_idx);
1956 assert(ptn != nullptr, "node should be registered");
1957 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
1958 // Add edge to object for unsafe access with offset.
1959 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1960 assert(adr_ptn != nullptr, "node should be registered");
1961 if (adr_ptn->is_Field()) {
1962 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
1963 add_edge(adr_ptn, ptn);
1964 }
1965 return true;
1966 }
1967 #ifdef ASSERT
1968 n->dump(1);
1969 assert(false, "not unsafe");
1970 #endif
1971 return false;
1972 }
1973
1974 void ConnectionGraph::add_call_node(CallNode* call) {
1975 assert(call->returns_pointer(), "only for call which returns pointer");
1976 uint call_idx = call->_idx;
1977 if (call->is_Allocate()) {
1978 Node* k = call->in(AllocateNode::KlassNode);
1979 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
1980 assert(kt != nullptr, "TypeKlassPtr required.");
1981 PointsToNode::EscapeState es = PointsToNode::NoEscape;
1982 bool scalar_replaceable = true;
1983 NOT_PRODUCT(const char* nsr_reason = "");
1984 if (call->is_AllocateArray()) {
1985 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
1986 es = PointsToNode::GlobalEscape;
1987 } else {
1988 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
1989 if (length < 0) {
1990 // Not scalar replaceable if the length is not constant.
1991 scalar_replaceable = false;
1992 NOT_PRODUCT(nsr_reason = "has a non-constant length");
1993 } else if (length > EliminateAllocationArraySizeLimit) {
1994 // Not scalar replaceable if the length is too big.
1995 scalar_replaceable = false;
1996 NOT_PRODUCT(nsr_reason = "has a length that is too big");
1997 }
1998 }
1999 } else { // Allocate instance
2000 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2001 es = PointsToNode::GlobalEscape;
2002 } else {
2003 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2004 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2005 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2006 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2007 !ik->can_be_instantiated() ||
2008 ik->has_finalizer()) {
2009 es = PointsToNode::GlobalEscape;
2010 } else {
2011 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2012 if (nfields > EliminateAllocationFieldsLimit) {
2013 // Not scalar replaceable if there are too many fields.
2014 scalar_replaceable = false;
2015 NOT_PRODUCT(nsr_reason = "has too many fields");
2016 }
2017 }
2018 }
2019 }
2020 add_java_object(call, es);
2021 PointsToNode* ptn = ptnode_adr(call_idx);
2022 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2023 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2024 }
2025 } else if (call->is_CallStaticJava()) {
2026 // Call nodes could be different types:
2027 //
2028 // 1. CallDynamicJavaNode (what happened during call is unknown):
2029 //
2030 // - mapped to GlobalEscape JavaObject node if oop is returned;
2031 //
2032 // - all oop arguments are escaping globally;
2033 //
2034 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2035 //
2036 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2037 //
2038 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2039 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2040 // during call is returned;
2041 // - mapped to ArgEscape LocalVar node pointed to object arguments
2042 // which are returned and does not escape during call;
2043 //
2044 // - oop arguments escaping status is defined by bytecode analysis;
2045 //
2046 // For a static call, we know exactly what method is being called.
2047 // Use bytecode estimator to record whether the call's return value escapes.
2048 ciMethod* meth = call->as_CallJava()->method();
2049 if (meth == nullptr) {
2050 const char* name = call->as_CallStaticJava()->_name;
2051 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0, "TODO: add failed case check");
2052 // Returns a newly allocated non-escaped object.
2053 add_java_object(call, PointsToNode::NoEscape);
2054 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2055 } else if (meth->is_boxing_method()) {
2056 // Returns boxing object
2057 PointsToNode::EscapeState es;
2058 vmIntrinsics::ID intr = meth->intrinsic_id();
2059 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2060 // It does not escape if object is always allocated.
2061 es = PointsToNode::NoEscape;
2062 } else {
2063 // It escapes globally if object could be loaded from cache.
2064 es = PointsToNode::GlobalEscape;
2065 }
2066 add_java_object(call, es);
2067 if (es == PointsToNode::GlobalEscape) {
2068 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2069 }
2070 } else {
2071 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2072 call_analyzer->copy_dependencies(_compile->dependencies());
2073 if (call_analyzer->is_return_allocated()) {
2074 // Returns a newly allocated non-escaped object, simply
2075 // update dependency information.
2076 // Mark it as NoEscape so that objects referenced by
2077 // it's fields will be marked as NoEscape at least.
2078 add_java_object(call, PointsToNode::NoEscape);
2079 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2080 } else {
2081 // Determine whether any arguments are returned.
2082 const TypeTuple* d = call->tf()->domain();
2083 bool ret_arg = false;
2084 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2085 if (d->field_at(i)->isa_ptr() != nullptr &&
2086 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2087 ret_arg = true;
2088 break;
2089 }
2090 }
2091 if (ret_arg) {
2092 add_local_var(call, PointsToNode::ArgEscape);
2093 } else {
2094 // Returns unknown object.
2095 map_ideal_node(call, phantom_obj);
2096 }
2097 }
2098 }
2099 } else {
2100 // An other type of call, assume the worst case:
2101 // returned value is unknown and globally escapes.
2102 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2103 map_ideal_node(call, phantom_obj);
2104 }
2105 }
2106
2107 void ConnectionGraph::process_call_arguments(CallNode *call) {
2108 bool is_arraycopy = false;
2109 switch (call->Opcode()) {
2110 #ifdef ASSERT
2111 case Op_Allocate:
2112 case Op_AllocateArray:
2113 case Op_Lock:
2114 case Op_Unlock:
2115 assert(false, "should be done already");
2116 break;
2117 #endif
2118 case Op_ArrayCopy:
2119 case Op_CallLeafNoFP:
2120 // Most array copies are ArrayCopy nodes at this point but there
2121 // are still a few direct calls to the copy subroutines (See
2122 // PhaseStringOpts::copy_string())
2123 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2124 call->as_CallLeaf()->is_call_to_arraycopystub();
2125 // fall through
2126 case Op_CallLeafVector:
2127 case Op_CallLeaf: {
2128 // Stub calls, objects do not escape but they are not scale replaceable.
2129 // Adjust escape state for outgoing arguments.
2130 const TypeTuple * d = call->tf()->domain();
2131 bool src_has_oops = false;
2132 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2133 const Type* at = d->field_at(i);
2134 Node *arg = call->in(i);
2135 if (arg == nullptr) {
2136 continue;
2137 }
2138 const Type *aat = _igvn->type(arg);
2139 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2140 continue;
2141 }
2142 if (arg->is_AddP()) {
2143 //
2144 // The inline_native_clone() case when the arraycopy stub is called
2145 // after the allocation before Initialize and CheckCastPP nodes.
2146 // Or normal arraycopy for object arrays case.
2147 //
2148 // Set AddP's base (Allocate) as not scalar replaceable since
2149 // pointer to the base (with offset) is passed as argument.
2150 //
2151 arg = get_addp_base(arg);
2152 }
2153 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2154 assert(arg_ptn != nullptr, "should be registered");
2155 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2156 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2157 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2158 aat->isa_ptr() != nullptr, "expecting an Ptr");
2159 bool arg_has_oops = aat->isa_oopptr() &&
2160 (aat->isa_instptr() ||
2161 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)));
2162 if (i == TypeFunc::Parms) {
2163 src_has_oops = arg_has_oops;
2164 }
2165 //
2166 // src or dst could be j.l.Object when other is basic type array:
2167 //
2168 // arraycopy(char[],0,Object*,0,size);
2169 // arraycopy(Object*,0,char[],0,size);
2170 //
2171 // Don't add edges in such cases.
2172 //
2173 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2174 arg_has_oops && (i > TypeFunc::Parms);
2175 #ifdef ASSERT
2176 if (!(is_arraycopy ||
2177 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2178 (call->as_CallLeaf()->_name != nullptr &&
2179 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2180 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2181 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2182 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2183 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2184 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2185 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2186 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2187 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2188 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2189 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2190 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2191 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2192 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2193 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2194 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2195 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2196 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2197 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2198 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2199 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2200 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2201 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2202 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2203 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2204 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2205 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2206 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2207 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2208 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2209 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2210 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2211 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2212 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2213 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2214 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2215 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2216 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2217 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2218 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2219 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2220 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2221 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2222 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2223 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2224 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2225 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2226 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2227 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2228 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2229 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2230 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2231 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2232 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2233 ))) {
2234 call->dump();
2235 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2236 }
2237 #endif
2238 // Always process arraycopy's destination object since
2239 // we need to add all possible edges to references in
2240 // source object.
2241 if (arg_esc >= PointsToNode::ArgEscape &&
2242 !arg_is_arraycopy_dest) {
2243 continue;
2244 }
2245 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2246 if (call->is_ArrayCopy()) {
2247 ArrayCopyNode* ac = call->as_ArrayCopy();
2248 if (ac->is_clonebasic() ||
2249 ac->is_arraycopy_validated() ||
2250 ac->is_copyof_validated() ||
2251 ac->is_copyofrange_validated()) {
2252 es = PointsToNode::NoEscape;
2253 }
2254 }
2255 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2256 if (arg_is_arraycopy_dest) {
2257 Node* src = call->in(TypeFunc::Parms);
2258 if (src->is_AddP()) {
2259 src = get_addp_base(src);
2260 }
2261 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2262 assert(src_ptn != nullptr, "should be registered");
2263 // Special arraycopy edge:
2264 // Only escape state of destination object's fields affects
2265 // escape state of fields in source object.
2266 add_arraycopy(call, es, src_ptn, arg_ptn);
2267 }
2268 }
2269 }
2270 break;
2271 }
2272 case Op_CallStaticJava: {
2273 // For a static call, we know exactly what method is being called.
2274 // Use bytecode estimator to record the call's escape affects
2275 #ifdef ASSERT
2276 const char* name = call->as_CallStaticJava()->_name;
2277 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2278 #endif
2279 ciMethod* meth = call->as_CallJava()->method();
2280 if ((meth != nullptr) && meth->is_boxing_method()) {
2281 break; // Boxing methods do not modify any oops.
2282 }
2283 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2284 // fall-through if not a Java method or no analyzer information
2285 if (call_analyzer != nullptr) {
2286 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2287 const TypeTuple* d = call->tf()->domain();
2288 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2289 const Type* at = d->field_at(i);
2290 int k = i - TypeFunc::Parms;
2291 Node* arg = call->in(i);
2292 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2293 if (at->isa_ptr() != nullptr &&
2294 call_analyzer->is_arg_returned(k)) {
2295 // The call returns arguments.
2296 if (call_ptn != nullptr) { // Is call's result used?
2297 assert(call_ptn->is_LocalVar(), "node should be registered");
2298 assert(arg_ptn != nullptr, "node should be registered");
2299 add_edge(call_ptn, arg_ptn);
2300 }
2301 }
2302 if (at->isa_oopptr() != nullptr &&
2303 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2304 if (!call_analyzer->is_arg_stack(k)) {
2305 // The argument global escapes
2306 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2307 } else {
2308 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2309 if (!call_analyzer->is_arg_local(k)) {
2310 // The argument itself doesn't escape, but any fields might
2311 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2312 }
2313 }
2314 }
2315 }
2316 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2317 // The call returns arguments.
2318 assert(call_ptn->edge_count() > 0, "sanity");
2319 if (!call_analyzer->is_return_local()) {
2320 // Returns also unknown object.
2321 add_edge(call_ptn, phantom_obj);
2322 }
2323 }
2324 break;
2325 }
2326 }
2327 default: {
2328 // Fall-through here if not a Java method or no analyzer information
2329 // or some other type of call, assume the worst case: all arguments
2330 // globally escape.
2331 const TypeTuple* d = call->tf()->domain();
2332 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2333 const Type* at = d->field_at(i);
2334 if (at->isa_oopptr() != nullptr) {
2335 Node* arg = call->in(i);
2336 if (arg->is_AddP()) {
2337 arg = get_addp_base(arg);
2338 }
2339 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2340 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2341 }
2342 }
2343 }
2344 }
2345 }
2346
2347
2348 // Finish Graph construction.
2349 bool ConnectionGraph::complete_connection_graph(
2350 GrowableArray<PointsToNode*>& ptnodes_worklist,
2351 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2352 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2353 GrowableArray<FieldNode*>& oop_fields_worklist) {
2354 // Normally only 1-3 passes needed to build Connection Graph depending
2355 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2356 // Set limit to 20 to catch situation when something did go wrong and
2357 // bailout Escape Analysis.
2358 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2359 #define GRAPH_BUILD_ITER_LIMIT 20
2360
2361 // Propagate GlobalEscape and ArgEscape escape states and check that
2362 // we still have non-escaping objects. The method pushs on _worklist
2363 // Field nodes which reference phantom_object.
2364 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2365 return false; // Nothing to do.
2366 }
2367 // Now propagate references to all JavaObject nodes.
2368 int java_objects_length = java_objects_worklist.length();
2369 elapsedTimer build_time;
2370 build_time.start();
2371 elapsedTimer time;
2372 bool timeout = false;
2373 int new_edges = 1;
2374 int iterations = 0;
2375 do {
2376 while ((new_edges > 0) &&
2377 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2378 double start_time = time.seconds();
2379 time.start();
2380 new_edges = 0;
2381 // Propagate references to phantom_object for nodes pushed on _worklist
2382 // by find_non_escaped_objects() and find_field_value().
2383 new_edges += add_java_object_edges(phantom_obj, false);
2384 for (int next = 0; next < java_objects_length; ++next) {
2385 JavaObjectNode* ptn = java_objects_worklist.at(next);
2386 new_edges += add_java_object_edges(ptn, true);
2387
2388 #define SAMPLE_SIZE 4
2389 if ((next % SAMPLE_SIZE) == 0) {
2390 // Each 4 iterations calculate how much time it will take
2391 // to complete graph construction.
2392 time.stop();
2393 // Poll for requests from shutdown mechanism to quiesce compiler
2394 // because Connection graph construction may take long time.
2395 CompileBroker::maybe_block();
2396 double stop_time = time.seconds();
2397 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2398 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2399 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2400 timeout = true;
2401 break; // Timeout
2402 }
2403 start_time = stop_time;
2404 time.start();
2405 }
2406 #undef SAMPLE_SIZE
2407
2408 }
2409 if (timeout) break;
2410 if (new_edges > 0) {
2411 // Update escape states on each iteration if graph was updated.
2412 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2413 return false; // Nothing to do.
2414 }
2415 }
2416 time.stop();
2417 if (time.seconds() >= EscapeAnalysisTimeout) {
2418 timeout = true;
2419 break;
2420 }
2421 }
2422 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2423 time.start();
2424 // Find fields which have unknown value.
2425 int fields_length = oop_fields_worklist.length();
2426 for (int next = 0; next < fields_length; next++) {
2427 FieldNode* field = oop_fields_worklist.at(next);
2428 if (field->edge_count() == 0) {
2429 new_edges += find_field_value(field);
2430 // This code may added new edges to phantom_object.
2431 // Need an other cycle to propagate references to phantom_object.
2432 }
2433 }
2434 time.stop();
2435 if (time.seconds() >= EscapeAnalysisTimeout) {
2436 timeout = true;
2437 break;
2438 }
2439 } else {
2440 new_edges = 0; // Bailout
2441 }
2442 } while (new_edges > 0);
2443
2444 build_time.stop();
2445 _build_time = build_time.seconds();
2446 _build_iterations = iterations;
2447
2448 // Bailout if passed limits.
2449 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2450 Compile* C = _compile;
2451 if (C->log() != nullptr) {
2452 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2453 C->log()->text("%s", timeout ? "time" : "iterations");
2454 C->log()->end_elem(" limit'");
2455 }
2456 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2457 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2458 // Possible infinite build_connection_graph loop,
2459 // bailout (no changes to ideal graph were made).
2460 return false;
2461 }
2462
2463 #undef GRAPH_BUILD_ITER_LIMIT
2464
2465 // Find fields initialized by null for non-escaping Allocations.
2466 int non_escaped_length = non_escaped_allocs_worklist.length();
2467 for (int next = 0; next < non_escaped_length; next++) {
2468 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2469 PointsToNode::EscapeState es = ptn->escape_state();
2470 assert(es <= PointsToNode::ArgEscape, "sanity");
2471 if (es == PointsToNode::NoEscape) {
2472 if (find_init_values_null(ptn, _igvn) > 0) {
2473 // Adding references to null object does not change escape states
2474 // since it does not escape. Also no fields are added to null object.
2475 add_java_object_edges(null_obj, false);
2476 }
2477 }
2478 Node* n = ptn->ideal_node();
2479 if (n->is_Allocate()) {
2480 // The object allocated by this Allocate node will never be
2481 // seen by an other thread. Mark it so that when it is
2482 // expanded no MemBarStoreStore is added.
2483 InitializeNode* ini = n->as_Allocate()->initialization();
2484 if (ini != nullptr)
2485 ini->set_does_not_escape();
2486 }
2487 }
2488 return true; // Finished graph construction.
2489 }
2490
2491 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2492 // and check that we still have non-escaping java objects.
2493 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2494 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist) {
2495 GrowableArray<PointsToNode*> escape_worklist;
2496 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2497 int ptnodes_length = ptnodes_worklist.length();
2498 for (int next = 0; next < ptnodes_length; ++next) {
2499 PointsToNode* ptn = ptnodes_worklist.at(next);
2500 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2501 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2502 escape_worklist.push(ptn);
2503 }
2504 }
2505 // Set escape states to referenced nodes (edges list).
2506 while (escape_worklist.length() > 0) {
2507 PointsToNode* ptn = escape_worklist.pop();
2508 PointsToNode::EscapeState es = ptn->escape_state();
2509 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2510 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2511 es >= PointsToNode::ArgEscape) {
2512 // GlobalEscape or ArgEscape state of field means it has unknown value.
2513 if (add_edge(ptn, phantom_obj)) {
2514 // New edge was added
2515 add_field_uses_to_worklist(ptn->as_Field());
2516 }
2517 }
2518 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2519 PointsToNode* e = i.get();
2520 if (e->is_Arraycopy()) {
2521 assert(ptn->arraycopy_dst(), "sanity");
2522 // Propagate only fields escape state through arraycopy edge.
2523 if (e->fields_escape_state() < field_es) {
2524 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2525 escape_worklist.push(e);
2526 }
2527 } else if (es >= field_es) {
2528 // fields_escape_state is also set to 'es' if it is less than 'es'.
2529 if (e->escape_state() < es) {
2530 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2531 escape_worklist.push(e);
2532 }
2533 } else {
2534 // Propagate field escape state.
2535 bool es_changed = false;
2536 if (e->fields_escape_state() < field_es) {
2537 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2538 es_changed = true;
2539 }
2540 if ((e->escape_state() < field_es) &&
2541 e->is_Field() && ptn->is_JavaObject() &&
2542 e->as_Field()->is_oop()) {
2543 // Change escape state of referenced fields.
2544 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2545 es_changed = true;
2546 } else if (e->escape_state() < es) {
2547 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2548 es_changed = true;
2549 }
2550 if (es_changed) {
2551 escape_worklist.push(e);
2552 }
2553 }
2554 }
2555 }
2556 // Remove escaped objects from non_escaped list.
2557 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2558 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2559 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2560 non_escaped_allocs_worklist.delete_at(next);
2561 }
2562 if (ptn->escape_state() == PointsToNode::NoEscape) {
2563 // Find fields in non-escaped allocations which have unknown value.
2564 find_init_values_phantom(ptn);
2565 }
2566 }
2567 return (non_escaped_allocs_worklist.length() > 0);
2568 }
2569
2570 // Add all references to JavaObject node by walking over all uses.
2571 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2572 int new_edges = 0;
2573 if (populate_worklist) {
2574 // Populate _worklist by uses of jobj's uses.
2575 for (UseIterator i(jobj); i.has_next(); i.next()) {
2576 PointsToNode* use = i.get();
2577 if (use->is_Arraycopy()) {
2578 continue;
2579 }
2580 add_uses_to_worklist(use);
2581 if (use->is_Field() && use->as_Field()->is_oop()) {
2582 // Put on worklist all field's uses (loads) and
2583 // related field nodes (same base and offset).
2584 add_field_uses_to_worklist(use->as_Field());
2585 }
2586 }
2587 }
2588 for (int l = 0; l < _worklist.length(); l++) {
2589 PointsToNode* use = _worklist.at(l);
2590 if (PointsToNode::is_base_use(use)) {
2591 // Add reference from jobj to field and from field to jobj (field's base).
2592 use = PointsToNode::get_use_node(use)->as_Field();
2593 if (add_base(use->as_Field(), jobj)) {
2594 new_edges++;
2595 }
2596 continue;
2597 }
2598 assert(!use->is_JavaObject(), "sanity");
2599 if (use->is_Arraycopy()) {
2600 if (jobj == null_obj) { // null object does not have field edges
2601 continue;
2602 }
2603 // Added edge from Arraycopy node to arraycopy's source java object
2604 if (add_edge(use, jobj)) {
2605 jobj->set_arraycopy_src();
2606 new_edges++;
2607 }
2608 // and stop here.
2609 continue;
2610 }
2611 if (!add_edge(use, jobj)) {
2612 continue; // No new edge added, there was such edge already.
2613 }
2614 new_edges++;
2615 if (use->is_LocalVar()) {
2616 add_uses_to_worklist(use);
2617 if (use->arraycopy_dst()) {
2618 for (EdgeIterator i(use); i.has_next(); i.next()) {
2619 PointsToNode* e = i.get();
2620 if (e->is_Arraycopy()) {
2621 if (jobj == null_obj) { // null object does not have field edges
2622 continue;
2623 }
2624 // Add edge from arraycopy's destination java object to Arraycopy node.
2625 if (add_edge(jobj, e)) {
2626 new_edges++;
2627 jobj->set_arraycopy_dst();
2628 }
2629 }
2630 }
2631 }
2632 } else {
2633 // Added new edge to stored in field values.
2634 // Put on worklist all field's uses (loads) and
2635 // related field nodes (same base and offset).
2636 add_field_uses_to_worklist(use->as_Field());
2637 }
2638 }
2639 _worklist.clear();
2640 _in_worklist.reset();
2641 return new_edges;
2642 }
2643
2644 // Put on worklist all related field nodes.
2645 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2646 assert(field->is_oop(), "sanity");
2647 int offset = field->offset();
2648 add_uses_to_worklist(field);
2649 // Loop over all bases of this field and push on worklist Field nodes
2650 // with the same offset and base (since they may reference the same field).
2651 for (BaseIterator i(field); i.has_next(); i.next()) {
2652 PointsToNode* base = i.get();
2653 add_fields_to_worklist(field, base);
2654 // Check if the base was source object of arraycopy and go over arraycopy's
2655 // destination objects since values stored to a field of source object are
2656 // accessible by uses (loads) of fields of destination objects.
2657 if (base->arraycopy_src()) {
2658 for (UseIterator j(base); j.has_next(); j.next()) {
2659 PointsToNode* arycp = j.get();
2660 if (arycp->is_Arraycopy()) {
2661 for (UseIterator k(arycp); k.has_next(); k.next()) {
2662 PointsToNode* abase = k.get();
2663 if (abase->arraycopy_dst() && abase != base) {
2664 // Look for the same arraycopy reference.
2665 add_fields_to_worklist(field, abase);
2666 }
2667 }
2668 }
2669 }
2670 }
2671 }
2672 }
2673
2674 // Put on worklist all related field nodes.
2675 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2676 int offset = field->offset();
2677 if (base->is_LocalVar()) {
2678 for (UseIterator j(base); j.has_next(); j.next()) {
2679 PointsToNode* f = j.get();
2680 if (PointsToNode::is_base_use(f)) { // Field
2681 f = PointsToNode::get_use_node(f);
2682 if (f == field || !f->as_Field()->is_oop()) {
2683 continue;
2684 }
2685 int offs = f->as_Field()->offset();
2686 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2687 add_to_worklist(f);
2688 }
2689 }
2690 }
2691 } else {
2692 assert(base->is_JavaObject(), "sanity");
2693 if (// Skip phantom_object since it is only used to indicate that
2694 // this field's content globally escapes.
2695 (base != phantom_obj) &&
2696 // null object node does not have fields.
2697 (base != null_obj)) {
2698 for (EdgeIterator i(base); i.has_next(); i.next()) {
2699 PointsToNode* f = i.get();
2700 // Skip arraycopy edge since store to destination object field
2701 // does not update value in source object field.
2702 if (f->is_Arraycopy()) {
2703 assert(base->arraycopy_dst(), "sanity");
2704 continue;
2705 }
2706 if (f == field || !f->as_Field()->is_oop()) {
2707 continue;
2708 }
2709 int offs = f->as_Field()->offset();
2710 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2711 add_to_worklist(f);
2712 }
2713 }
2714 }
2715 }
2716 }
2717
2718 // Find fields which have unknown value.
2719 int ConnectionGraph::find_field_value(FieldNode* field) {
2720 // Escaped fields should have init value already.
2721 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2722 int new_edges = 0;
2723 for (BaseIterator i(field); i.has_next(); i.next()) {
2724 PointsToNode* base = i.get();
2725 if (base->is_JavaObject()) {
2726 // Skip Allocate's fields which will be processed later.
2727 if (base->ideal_node()->is_Allocate()) {
2728 return 0;
2729 }
2730 assert(base == null_obj, "only null ptr base expected here");
2731 }
2732 }
2733 if (add_edge(field, phantom_obj)) {
2734 // New edge was added
2735 new_edges++;
2736 add_field_uses_to_worklist(field);
2737 }
2738 return new_edges;
2739 }
2740
2741 // Find fields initializing values for allocations.
2742 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2743 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2744 Node* alloc = pta->ideal_node();
2745
2746 // Do nothing for Allocate nodes since its fields values are
2747 // "known" unless they are initialized by arraycopy/clone.
2748 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2749 return 0;
2750 }
2751 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
2752 #ifdef ASSERT
2753 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == nullptr) {
2754 const char* name = alloc->as_CallStaticJava()->_name;
2755 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0, "sanity");
2756 }
2757 #endif
2758 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2759 int new_edges = 0;
2760 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2761 PointsToNode* field = i.get();
2762 if (field->is_Field() && field->as_Field()->is_oop()) {
2763 if (add_edge(field, phantom_obj)) {
2764 // New edge was added
2765 new_edges++;
2766 add_field_uses_to_worklist(field->as_Field());
2767 }
2768 }
2769 }
2770 return new_edges;
2771 }
2772
2773 // Find fields initializing values for allocations.
2774 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2775 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2776 Node* alloc = pta->ideal_node();
2777 // Do nothing for Call nodes since its fields values are unknown.
2778 if (!alloc->is_Allocate()) {
2779 return 0;
2780 }
2781 InitializeNode* ini = alloc->as_Allocate()->initialization();
2782 bool visited_bottom_offset = false;
2783 GrowableArray<int> offsets_worklist;
2784 int new_edges = 0;
2785
2786 // Check if an oop field's initializing value is recorded and add
2787 // a corresponding null if field's value if it is not recorded.
2788 // Connection Graph does not record a default initialization by null
2789 // captured by Initialize node.
2790 //
2791 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2792 PointsToNode* field = i.get(); // Field (AddP)
2793 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2794 continue; // Not oop field
2795 }
2796 int offset = field->as_Field()->offset();
2797 if (offset == Type::OffsetBot) {
2798 if (!visited_bottom_offset) {
2799 // OffsetBot is used to reference array's element,
2800 // always add reference to null to all Field nodes since we don't
2801 // known which element is referenced.
2802 if (add_edge(field, null_obj)) {
2803 // New edge was added
2804 new_edges++;
2805 add_field_uses_to_worklist(field->as_Field());
2806 visited_bottom_offset = true;
2807 }
2808 }
2809 } else {
2810 // Check only oop fields.
2811 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2812 if (adr_type->isa_rawptr()) {
2813 #ifdef ASSERT
2814 // Raw pointers are used for initializing stores so skip it
2815 // since it should be recorded already
2816 Node* base = get_addp_base(field->ideal_node());
2817 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2818 #endif
2819 continue;
2820 }
2821 if (!offsets_worklist.contains(offset)) {
2822 offsets_worklist.append(offset);
2823 Node* value = nullptr;
2824 if (ini != nullptr) {
2825 // StoreP::value_basic_type() == T_ADDRESS
2826 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2827 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2828 // Make sure initializing store has the same type as this AddP.
2829 // This AddP may reference non existing field because it is on a
2830 // dead branch of bimorphic call which is not eliminated yet.
2831 if (store != nullptr && store->is_Store() &&
2832 store->as_Store()->value_basic_type() == ft) {
2833 value = store->in(MemNode::ValueIn);
2834 #ifdef ASSERT
2835 if (VerifyConnectionGraph) {
2836 // Verify that AddP already points to all objects the value points to.
2837 PointsToNode* val = ptnode_adr(value->_idx);
2838 assert((val != nullptr), "should be processed already");
2839 PointsToNode* missed_obj = nullptr;
2840 if (val->is_JavaObject()) {
2841 if (!field->points_to(val->as_JavaObject())) {
2842 missed_obj = val;
2843 }
2844 } else {
2845 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2846 tty->print_cr("----------init store has invalid value -----");
2847 store->dump();
2848 val->dump();
2849 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2850 }
2851 for (EdgeIterator j(val); j.has_next(); j.next()) {
2852 PointsToNode* obj = j.get();
2853 if (obj->is_JavaObject()) {
2854 if (!field->points_to(obj->as_JavaObject())) {
2855 missed_obj = obj;
2856 break;
2857 }
2858 }
2859 }
2860 }
2861 if (missed_obj != nullptr) {
2862 tty->print_cr("----------field---------------------------------");
2863 field->dump();
2864 tty->print_cr("----------missed referernce to object-----------");
2865 missed_obj->dump();
2866 tty->print_cr("----------object referernced by init store -----");
2867 store->dump();
2868 val->dump();
2869 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2870 }
2871 }
2872 #endif
2873 } else {
2874 // There could be initializing stores which follow allocation.
2875 // For example, a volatile field store is not collected
2876 // by Initialize node.
2877 //
2878 // Need to check for dependent loads to separate such stores from
2879 // stores which follow loads. For now, add initial value null so
2880 // that compare pointers optimization works correctly.
2881 }
2882 }
2883 if (value == nullptr) {
2884 // A field's initializing value was not recorded. Add null.
2885 if (add_edge(field, null_obj)) {
2886 // New edge was added
2887 new_edges++;
2888 add_field_uses_to_worklist(field->as_Field());
2889 }
2890 }
2891 }
2892 }
2893 }
2894 return new_edges;
2895 }
2896
2897 // Adjust scalar_replaceable state after Connection Graph is built.
2898 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2899 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2900 // returns true. If one of the constraints in this method set 'jobj' to NSR
2901 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2902 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2903 // that other object and the Phi will become a reducible Phi.
2904 // There could be multiple merges involving the same jobj.
2905 Unique_Node_List candidates;
2906
2907 // Search for non-escaping objects which are not scalar replaceable
2908 // and mark them to propagate the state to referenced objects.
2909
2910 for (UseIterator i(jobj); i.has_next(); i.next()) {
2911 PointsToNode* use = i.get();
2912 if (use->is_Arraycopy()) {
2913 continue;
2914 }
2915 if (use->is_Field()) {
2916 FieldNode* field = use->as_Field();
2917 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2918 // 1. An object is not scalar replaceable if the field into which it is
2919 // stored has unknown offset (stored into unknown element of an array).
2920 if (field->offset() == Type::OffsetBot) {
2921 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2922 return;
2923 }
2924 for (BaseIterator i(field); i.has_next(); i.next()) {
2925 PointsToNode* base = i.get();
2926 // 2. An object is not scalar replaceable if the field into which it is
2927 // stored has multiple bases one of which is null.
2928 if ((base == null_obj) && (field->base_count() > 1)) {
2929 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2930 return;
2931 }
2932 // 2.5. An object is not scalar replaceable if the field into which it is
2933 // stored has NSR base.
2934 if (!base->scalar_replaceable()) {
2935 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2936 return;
2937 }
2938 }
2939 }
2940 assert(use->is_Field() || use->is_LocalVar(), "sanity");
2941 // 3. An object is not scalar replaceable if it is merged with other objects
2942 // and we can't remove the merge
2943 for (EdgeIterator j(use); j.has_next(); j.next()) {
2944 PointsToNode* ptn = j.get();
2945 if (ptn->is_JavaObject() && ptn != jobj) {
2946 Node* use_n = use->ideal_node();
2947
2948 // These other local vars may point to multiple objects through a Phi
2949 // In this case we skip them and see if we can reduce the Phi.
2950 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
2951 use_n = use_n->in(1);
2952 }
2953
2954 // If it's already a candidate or confirmed reducible merge we can skip verification
2955 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
2956 continue;
2957 }
2958
2959 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
2960 candidates.push(use_n);
2961 } else {
2962 // Mark all objects as NSR if we can't remove the merge
2963 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
2964 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
2965 }
2966 }
2967 }
2968 if (!jobj->scalar_replaceable()) {
2969 return;
2970 }
2971 }
2972
2973 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
2974 if (j.get()->is_Arraycopy()) {
2975 continue;
2976 }
2977
2978 // Non-escaping object node should point only to field nodes.
2979 FieldNode* field = j.get()->as_Field();
2980 int offset = field->as_Field()->offset();
2981
2982 // 4. An object is not scalar replaceable if it has a field with unknown
2983 // offset (array's element is accessed in loop).
2984 if (offset == Type::OffsetBot) {
2985 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
2986 return;
2987 }
2988 // 5. Currently an object is not scalar replaceable if a LoadStore node
2989 // access its field since the field value is unknown after it.
2990 //
2991 Node* n = field->ideal_node();
2992
2993 // Test for an unsafe access that was parsed as maybe off heap
2994 // (with a CheckCastPP to raw memory).
2995 assert(n->is_AddP(), "expect an address computation");
2996 if (n->in(AddPNode::Base)->is_top() &&
2997 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
2998 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
2999 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3000 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3001 return;
3002 }
3003
3004 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3005 Node* u = n->fast_out(i);
3006 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3007 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3008 return;
3009 }
3010 }
3011
3012 // 6. Or the address may point to more then one object. This may produce
3013 // the false positive result (set not scalar replaceable)
3014 // since the flow-insensitive escape analysis can't separate
3015 // the case when stores overwrite the field's value from the case
3016 // when stores happened on different control branches.
3017 //
3018 // Note: it will disable scalar replacement in some cases:
3019 //
3020 // Point p[] = new Point[1];
3021 // p[0] = new Point(); // Will be not scalar replaced
3022 //
3023 // but it will save us from incorrect optimizations in next cases:
3024 //
3025 // Point p[] = new Point[1];
3026 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3027 //
3028 if (field->base_count() > 1 && candidates.size() == 0) {
3029 if (has_non_reducible_merge(field, reducible_merges)) {
3030 for (BaseIterator i(field); i.has_next(); i.next()) {
3031 PointsToNode* base = i.get();
3032 // Don't take into account LocalVar nodes which
3033 // may point to only one object which should be also
3034 // this field's base by now.
3035 if (base->is_JavaObject() && base != jobj) {
3036 // Mark all bases.
3037 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3038 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3039 }
3040 }
3041
3042 if (!jobj->scalar_replaceable()) {
3043 return;
3044 }
3045 }
3046 }
3047 }
3048
3049 // The candidate is truly a reducible merge only if none of the other
3050 // constraints ruled it as NSR. There could be multiple merges involving the
3051 // same jobj.
3052 assert(jobj->scalar_replaceable(), "sanity");
3053 for (uint i = 0; i < candidates.size(); i++ ) {
3054 Node* candidate = candidates.at(i);
3055 reducible_merges.push(candidate);
3056 }
3057 }
3058
3059 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3060 for (BaseIterator i(field); i.has_next(); i.next()) {
3061 Node* base = i.get()->ideal_node();
3062 if (base->is_Phi() && !reducible_merges.member(base)) {
3063 return true;
3064 }
3065 }
3066 return false;
3067 }
3068
3069 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3070 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3071
3072 // Look for 'phis' that refer to 'jobj' as the last
3073 // remaining scalar replaceable input.
3074 uint reducible_merges_cnt = reducible_merges.size();
3075 for (uint i = 0; i < reducible_merges_cnt; i++) {
3076 Node* phi = reducible_merges.at(i);
3077
3078 // This 'Phi' will be a 'good' if it still points to
3079 // at least one scalar replaceable object. Note that 'obj'
3080 // was/should be marked as NSR before calling this function.
3081 bool good_phi = false;
3082
3083 for (uint j = 1; j < phi->req(); j++) {
3084 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3085 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3086 good_phi = true;
3087 break;
3088 }
3089 }
3090
3091 if (!good_phi) {
3092 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3093 reducible_merges.remove(i);
3094
3095 // Decrement the index because the 'remove' call above actually
3096 // moves the last entry of the list to position 'i'.
3097 i--;
3098
3099 reducible_merges_cnt--;
3100 }
3101 }
3102 }
3103
3104 // Propagate NSR (Not scalar replaceable) state.
3105 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3106 int jobj_length = jobj_worklist.length();
3107 bool found_nsr_alloc = true;
3108 while (found_nsr_alloc) {
3109 found_nsr_alloc = false;
3110 for (int next = 0; next < jobj_length; ++next) {
3111 JavaObjectNode* jobj = jobj_worklist.at(next);
3112 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3113 PointsToNode* use = i.get();
3114 if (use->is_Field()) {
3115 FieldNode* field = use->as_Field();
3116 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3117 assert(field->offset() != Type::OffsetBot, "sanity");
3118 for (BaseIterator i(field); i.has_next(); i.next()) {
3119 PointsToNode* base = i.get();
3120 // An object is not scalar replaceable if the field into which
3121 // it is stored has NSR base.
3122 if ((base != null_obj) && !base->scalar_replaceable()) {
3123 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3124 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3125 // because there is no point in reducing a Phi that won't improve the number of SR
3126 // objects.
3127 revisit_reducible_phi_status(jobj, reducible_merges);
3128 found_nsr_alloc = true;
3129 break;
3130 }
3131 }
3132 }
3133 }
3134 }
3135 }
3136 }
3137
3138 #ifdef ASSERT
3139 void ConnectionGraph::verify_connection_graph(
3140 GrowableArray<PointsToNode*>& ptnodes_worklist,
3141 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3142 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3143 GrowableArray<Node*>& addp_worklist) {
3144 // Verify that graph is complete - no new edges could be added.
3145 int java_objects_length = java_objects_worklist.length();
3146 int non_escaped_length = non_escaped_allocs_worklist.length();
3147 int new_edges = 0;
3148 for (int next = 0; next < java_objects_length; ++next) {
3149 JavaObjectNode* ptn = java_objects_worklist.at(next);
3150 new_edges += add_java_object_edges(ptn, true);
3151 }
3152 assert(new_edges == 0, "graph was not complete");
3153 // Verify that escape state is final.
3154 int length = non_escaped_allocs_worklist.length();
3155 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist);
3156 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3157 (non_escaped_length == length) &&
3158 (_worklist.length() == 0), "escape state was not final");
3159
3160 // Verify fields information.
3161 int addp_length = addp_worklist.length();
3162 for (int next = 0; next < addp_length; ++next ) {
3163 Node* n = addp_worklist.at(next);
3164 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3165 if (field->is_oop()) {
3166 // Verify that field has all bases
3167 Node* base = get_addp_base(n);
3168 PointsToNode* ptn = ptnode_adr(base->_idx);
3169 if (ptn->is_JavaObject()) {
3170 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3171 } else {
3172 assert(ptn->is_LocalVar(), "sanity");
3173 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3174 PointsToNode* e = i.get();
3175 if (e->is_JavaObject()) {
3176 assert(field->has_base(e->as_JavaObject()), "sanity");
3177 }
3178 }
3179 }
3180 // Verify that all fields have initializing values.
3181 if (field->edge_count() == 0) {
3182 tty->print_cr("----------field does not have references----------");
3183 field->dump();
3184 for (BaseIterator i(field); i.has_next(); i.next()) {
3185 PointsToNode* base = i.get();
3186 tty->print_cr("----------field has next base---------------------");
3187 base->dump();
3188 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3189 tty->print_cr("----------base has fields-------------------------");
3190 for (EdgeIterator j(base); j.has_next(); j.next()) {
3191 j.get()->dump();
3192 }
3193 tty->print_cr("----------base has references---------------------");
3194 for (UseIterator j(base); j.has_next(); j.next()) {
3195 j.get()->dump();
3196 }
3197 }
3198 }
3199 for (UseIterator i(field); i.has_next(); i.next()) {
3200 i.get()->dump();
3201 }
3202 assert(field->edge_count() > 0, "sanity");
3203 }
3204 }
3205 }
3206 }
3207 #endif
3208
3209 // Optimize ideal graph.
3210 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3211 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3212 Compile* C = _compile;
3213 PhaseIterGVN* igvn = _igvn;
3214 if (EliminateLocks) {
3215 // Mark locks before changing ideal graph.
3216 int cnt = C->macro_count();
3217 for (int i = 0; i < cnt; i++) {
3218 Node *n = C->macro_node(i);
3219 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3220 AbstractLockNode* alock = n->as_AbstractLock();
3221 if (!alock->is_non_esc_obj()) {
3222 if (can_eliminate_lock(alock)) {
3223 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3224 // The lock could be marked eliminated by lock coarsening
3225 // code during first IGVN before EA. Replace coarsened flag
3226 // to eliminate all associated locks/unlocks.
3227 #ifdef ASSERT
3228 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3229 #endif
3230 alock->set_non_esc_obj();
3231 }
3232 }
3233 }
3234 }
3235 }
3236
3237 if (OptimizePtrCompare) {
3238 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3239 Node *n = ptr_cmp_worklist.at(i);
3240 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3241 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3242 if (tcmp->singleton()) {
3243 Node* cmp = igvn->makecon(tcmp);
3244 #ifndef PRODUCT
3245 if (PrintOptimizePtrCompare) {
3246 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"));
3247 if (Verbose) {
3248 n->dump(1);
3249 }
3250 }
3251 #endif
3252 igvn->replace_node(n, cmp);
3253 }
3254 }
3255 }
3256
3257 // For MemBarStoreStore nodes added in library_call.cpp, check
3258 // escape status of associated AllocateNode and optimize out
3259 // MemBarStoreStore node if the allocated object never escapes.
3260 for (int i = 0; i < storestore_worklist.length(); i++) {
3261 Node* storestore = storestore_worklist.at(i);
3262 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3263 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3264 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3265 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3266 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3267 igvn->register_new_node_with_optimizer(mb);
3268 igvn->replace_node(storestore, mb);
3269 }
3270 }
3271 }
3272
3273 // Optimize objects compare.
3274 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3275 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3276 if (!OptimizePtrCompare) {
3277 return UNKNOWN;
3278 }
3279 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3280 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3281
3282 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3283 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3284 JavaObjectNode* jobj1 = unique_java_object(left);
3285 JavaObjectNode* jobj2 = unique_java_object(right);
3286
3287 // The use of this method during allocation merge reduction may cause 'left'
3288 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3289 // that doesn't reference an unique java object.
3290 if (ptn1 == nullptr || ptn2 == nullptr ||
3291 jobj1 == nullptr || jobj2 == nullptr) {
3292 return UNKNOWN;
3293 }
3294
3295 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3296 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3297
3298 // Check simple cases first.
3299 if (jobj1 != nullptr) {
3300 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3301 if (jobj1 == jobj2) {
3302 // Comparing the same not escaping object.
3303 return EQ;
3304 }
3305 Node* obj = jobj1->ideal_node();
3306 // Comparing not escaping allocation.
3307 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3308 !ptn2->points_to(jobj1)) {
3309 return NE; // This includes nullness check.
3310 }
3311 }
3312 }
3313 if (jobj2 != nullptr) {
3314 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3315 Node* obj = jobj2->ideal_node();
3316 // Comparing not escaping allocation.
3317 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3318 !ptn1->points_to(jobj2)) {
3319 return NE; // This includes nullness check.
3320 }
3321 }
3322 }
3323 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3324 jobj2 != nullptr && jobj2 != phantom_obj &&
3325 jobj1->ideal_node()->is_Con() &&
3326 jobj2->ideal_node()->is_Con()) {
3327 // Klass or String constants compare. Need to be careful with
3328 // compressed pointers - compare types of ConN and ConP instead of nodes.
3329 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3330 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3331 if (t1->make_ptr() == t2->make_ptr()) {
3332 return EQ;
3333 } else {
3334 return NE;
3335 }
3336 }
3337 if (ptn1->meet(ptn2)) {
3338 return UNKNOWN; // Sets are not disjoint
3339 }
3340
3341 // Sets are disjoint.
3342 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3343 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3344 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3345 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3346 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3347 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3348 // Check nullness of unknown object.
3349 return UNKNOWN;
3350 }
3351
3352 // Disjointness by itself is not sufficient since
3353 // alias analysis is not complete for escaped objects.
3354 // Disjoint sets are definitely unrelated only when
3355 // at least one set has only not escaping allocations.
3356 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3357 if (ptn1->non_escaping_allocation()) {
3358 return NE;
3359 }
3360 }
3361 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3362 if (ptn2->non_escaping_allocation()) {
3363 return NE;
3364 }
3365 }
3366 return UNKNOWN;
3367 }
3368
3369 // Connection Graph construction functions.
3370
3371 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3372 PointsToNode* ptadr = _nodes.at(n->_idx);
3373 if (ptadr != nullptr) {
3374 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3375 return;
3376 }
3377 Compile* C = _compile;
3378 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3379 map_ideal_node(n, ptadr);
3380 }
3381
3382 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3383 PointsToNode* ptadr = _nodes.at(n->_idx);
3384 if (ptadr != nullptr) {
3385 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3386 return ptadr;
3387 }
3388 Compile* C = _compile;
3389 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3390 map_ideal_node(n, ptadr);
3391 return ptadr;
3392 }
3393
3394 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3395 PointsToNode* ptadr = _nodes.at(n->_idx);
3396 if (ptadr != nullptr) {
3397 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3398 return;
3399 }
3400 bool unsafe = false;
3401 bool is_oop = is_oop_field(n, offset, &unsafe);
3402 if (unsafe) {
3403 es = PointsToNode::GlobalEscape;
3404 }
3405 Compile* C = _compile;
3406 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3407 map_ideal_node(n, field);
3408 }
3409
3410 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3411 PointsToNode* src, PointsToNode* dst) {
3412 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3413 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3414 PointsToNode* ptadr = _nodes.at(n->_idx);
3415 if (ptadr != nullptr) {
3416 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3417 return;
3418 }
3419 Compile* C = _compile;
3420 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3421 map_ideal_node(n, ptadr);
3422 // Add edge from arraycopy node to source object.
3423 (void)add_edge(ptadr, src);
3424 src->set_arraycopy_src();
3425 // Add edge from destination object to arraycopy node.
3426 (void)add_edge(dst, ptadr);
3427 dst->set_arraycopy_dst();
3428 }
3429
3430 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3431 const Type* adr_type = n->as_AddP()->bottom_type();
3432 BasicType bt = T_INT;
3433 if (offset == Type::OffsetBot) {
3434 // Check only oop fields.
3435 if (!adr_type->isa_aryptr() ||
3436 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3437 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3438 // OffsetBot is used to reference array's element. Ignore first AddP.
3439 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3440 bt = T_OBJECT;
3441 }
3442 }
3443 } else if (offset != Type::klass_offset()) {
3444 if (adr_type->isa_instptr()) {
3445 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
3446 if (field != nullptr) {
3447 bt = field->layout_type();
3448 } else {
3449 // Check for unsafe oop field access
3450 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3451 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3452 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3453 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3454 bt = T_OBJECT;
3455 (*unsafe) = true;
3456 }
3457 }
3458 } else if (adr_type->isa_aryptr()) {
3459 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3460 // Ignore array length load.
3461 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3462 // Ignore first AddP.
3463 } else {
3464 const Type* elemtype = adr_type->isa_aryptr()->elem();
3465 bt = elemtype->array_element_basic_type();
3466 }
3467 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3468 // Allocation initialization, ThreadLocal field access, unsafe access
3469 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3470 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3471 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3472 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3473 bt = T_OBJECT;
3474 }
3475 }
3476 }
3477 // Note: T_NARROWOOP is not classed as a real reference type
3478 return (is_reference_type(bt) || bt == T_NARROWOOP);
3479 }
3480
3481 // Returns unique pointed java object or null.
3482 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3483 // If the node was created after the escape computation we can't answer.
3484 uint idx = n->_idx;
3485 if (idx >= nodes_size()) {
3486 return nullptr;
3487 }
3488 PointsToNode* ptn = ptnode_adr(idx);
3489 if (ptn == nullptr) {
3490 return nullptr;
3491 }
3492 if (ptn->is_JavaObject()) {
3493 return ptn->as_JavaObject();
3494 }
3495 assert(ptn->is_LocalVar(), "sanity");
3496 // Check all java objects it points to.
3497 JavaObjectNode* jobj = nullptr;
3498 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3499 PointsToNode* e = i.get();
3500 if (e->is_JavaObject()) {
3501 if (jobj == nullptr) {
3502 jobj = e->as_JavaObject();
3503 } else if (jobj != e) {
3504 return nullptr;
3505 }
3506 }
3507 }
3508 return jobj;
3509 }
3510
3511 // Return true if this node points only to non-escaping allocations.
3512 bool PointsToNode::non_escaping_allocation() {
3513 if (is_JavaObject()) {
3514 Node* n = ideal_node();
3515 if (n->is_Allocate() || n->is_CallStaticJava()) {
3516 return (escape_state() == PointsToNode::NoEscape);
3517 } else {
3518 return false;
3519 }
3520 }
3521 assert(is_LocalVar(), "sanity");
3522 // Check all java objects it points to.
3523 for (EdgeIterator i(this); i.has_next(); i.next()) {
3524 PointsToNode* e = i.get();
3525 if (e->is_JavaObject()) {
3526 Node* n = e->ideal_node();
3527 if ((e->escape_state() != PointsToNode::NoEscape) ||
3528 !(n->is_Allocate() || n->is_CallStaticJava())) {
3529 return false;
3530 }
3531 }
3532 }
3533 return true;
3534 }
3535
3536 // Return true if we know the node does not escape globally.
3537 bool ConnectionGraph::not_global_escape(Node *n) {
3538 assert(!_collecting, "should not call during graph construction");
3539 // If the node was created after the escape computation we can't answer.
3540 uint idx = n->_idx;
3541 if (idx >= nodes_size()) {
3542 return false;
3543 }
3544 PointsToNode* ptn = ptnode_adr(idx);
3545 if (ptn == nullptr) {
3546 return false; // not in congraph (e.g. ConI)
3547 }
3548 PointsToNode::EscapeState es = ptn->escape_state();
3549 // If we have already computed a value, return it.
3550 if (es >= PointsToNode::GlobalEscape) {
3551 return false;
3552 }
3553 if (ptn->is_JavaObject()) {
3554 return true; // (es < PointsToNode::GlobalEscape);
3555 }
3556 assert(ptn->is_LocalVar(), "sanity");
3557 // Check all java objects it points to.
3558 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3559 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3560 return false;
3561 }
3562 }
3563 return true;
3564 }
3565
3566 // Return true if locked object does not escape globally
3567 // and locked code region (identified by BoxLockNode) is balanced:
3568 // all compiled code paths have corresponding Lock/Unlock pairs.
3569 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3570 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3571 if (EliminateNestedLocks) {
3572 // We can mark whole locking region as Local only when only
3573 // one object is used for locking.
3574 alock->box_node()->as_BoxLock()->set_local();
3575 }
3576 return true;
3577 }
3578 return false;
3579 }
3580
3581 // Helper functions
3582
3583 // Return true if this node points to specified node or nodes it points to.
3584 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3585 if (is_JavaObject()) {
3586 return (this == ptn);
3587 }
3588 assert(is_LocalVar() || is_Field(), "sanity");
3589 for (EdgeIterator i(this); i.has_next(); i.next()) {
3590 if (i.get() == ptn) {
3591 return true;
3592 }
3593 }
3594 return false;
3595 }
3596
3597 // Return true if one node points to an other.
3598 bool PointsToNode::meet(PointsToNode* ptn) {
3599 if (this == ptn) {
3600 return true;
3601 } else if (ptn->is_JavaObject()) {
3602 return this->points_to(ptn->as_JavaObject());
3603 } else if (this->is_JavaObject()) {
3604 return ptn->points_to(this->as_JavaObject());
3605 }
3606 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3607 int ptn_count = ptn->edge_count();
3608 for (EdgeIterator i(this); i.has_next(); i.next()) {
3609 PointsToNode* this_e = i.get();
3610 for (int j = 0; j < ptn_count; j++) {
3611 if (this_e == ptn->edge(j)) {
3612 return true;
3613 }
3614 }
3615 }
3616 return false;
3617 }
3618
3619 #ifdef ASSERT
3620 // Return true if bases point to this java object.
3621 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3622 for (BaseIterator i(this); i.has_next(); i.next()) {
3623 if (i.get() == jobj) {
3624 return true;
3625 }
3626 }
3627 return false;
3628 }
3629 #endif
3630
3631 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3632 // Handle simple case first.
3633 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3634 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3635 return true;
3636 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3637 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3638 Node* addp_use = addp->fast_out(i);
3639 if (addp_use->is_Store()) {
3640 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3641 if (addp_use->fast_out(j)->is_Initialize()) {
3642 return true;
3643 }
3644 }
3645 }
3646 }
3647 }
3648 return false;
3649 }
3650
3651 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3652 const Type *adr_type = phase->type(adr);
3653 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3654 // We are computing a raw address for a store captured by an Initialize
3655 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3656 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3657 assert(offs != Type::OffsetBot ||
3658 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3659 "offset must be a constant or it is initialization of array");
3660 return offs;
3661 }
3662 const TypePtr *t_ptr = adr_type->isa_ptr();
3663 assert(t_ptr != nullptr, "must be a pointer type");
3664 return t_ptr->offset();
3665 }
3666
3667 Node* ConnectionGraph::get_addp_base(Node *addp) {
3668 assert(addp->is_AddP(), "must be AddP");
3669 //
3670 // AddP cases for Base and Address inputs:
3671 // case #1. Direct object's field reference:
3672 // Allocate
3673 // |
3674 // Proj #5 ( oop result )
3675 // |
3676 // CheckCastPP (cast to instance type)
3677 // | |
3678 // AddP ( base == address )
3679 //
3680 // case #2. Indirect object's field reference:
3681 // Phi
3682 // |
3683 // CastPP (cast to instance type)
3684 // | |
3685 // AddP ( base == address )
3686 //
3687 // case #3. Raw object's field reference for Initialize node:
3688 // Allocate
3689 // |
3690 // Proj #5 ( oop result )
3691 // top |
3692 // \ |
3693 // AddP ( base == top )
3694 //
3695 // case #4. Array's element reference:
3696 // {CheckCastPP | CastPP}
3697 // | | |
3698 // | AddP ( array's element offset )
3699 // | |
3700 // AddP ( array's offset )
3701 //
3702 // case #5. Raw object's field reference for arraycopy stub call:
3703 // The inline_native_clone() case when the arraycopy stub is called
3704 // after the allocation before Initialize and CheckCastPP nodes.
3705 // Allocate
3706 // |
3707 // Proj #5 ( oop result )
3708 // | |
3709 // AddP ( base == address )
3710 //
3711 // case #6. Constant Pool, ThreadLocal, CastX2P or
3712 // Raw object's field reference:
3713 // {ConP, ThreadLocal, CastX2P, raw Load}
3714 // top |
3715 // \ |
3716 // AddP ( base == top )
3717 //
3718 // case #7. Klass's field reference.
3719 // LoadKlass
3720 // | |
3721 // AddP ( base == address )
3722 //
3723 // case #8. narrow Klass's field reference.
3724 // LoadNKlass
3725 // |
3726 // DecodeN
3727 // | |
3728 // AddP ( base == address )
3729 //
3730 // case #9. Mixed unsafe access
3731 // {instance}
3732 // |
3733 // CheckCastPP (raw)
3734 // top |
3735 // \ |
3736 // AddP ( base == top )
3737 //
3738 Node *base = addp->in(AddPNode::Base);
3739 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3740 base = addp->in(AddPNode::Address);
3741 while (base->is_AddP()) {
3742 // Case #6 (unsafe access) may have several chained AddP nodes.
3743 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3744 base = base->in(AddPNode::Address);
3745 }
3746 if (base->Opcode() == Op_CheckCastPP &&
3747 base->bottom_type()->isa_rawptr() &&
3748 _igvn->type(base->in(1))->isa_oopptr()) {
3749 base = base->in(1); // Case #9
3750 } else {
3751 Node* uncast_base = base->uncast();
3752 int opcode = uncast_base->Opcode();
3753 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3754 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3755 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3756 is_captured_store_address(addp), "sanity");
3757 }
3758 }
3759 return base;
3760 }
3761
3762 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3763 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3764 Node* addp2 = addp->raw_out(0);
3765 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3766 addp2->in(AddPNode::Base) == n &&
3767 addp2->in(AddPNode::Address) == addp) {
3768 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3769 //
3770 // Find array's offset to push it on worklist first and
3771 // as result process an array's element offset first (pushed second)
3772 // to avoid CastPP for the array's offset.
3773 // Otherwise the inserted CastPP (LocalVar) will point to what
3774 // the AddP (Field) points to. Which would be wrong since
3775 // the algorithm expects the CastPP has the same point as
3776 // as AddP's base CheckCastPP (LocalVar).
3777 //
3778 // ArrayAllocation
3779 // |
3780 // CheckCastPP
3781 // |
3782 // memProj (from ArrayAllocation CheckCastPP)
3783 // | ||
3784 // | || Int (element index)
3785 // | || | ConI (log(element size))
3786 // | || | /
3787 // | || LShift
3788 // | || /
3789 // | AddP (array's element offset)
3790 // | |
3791 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3792 // | / /
3793 // AddP (array's offset)
3794 // |
3795 // Load/Store (memory operation on array's element)
3796 //
3797 return addp2;
3798 }
3799 return nullptr;
3800 }
3801
3802 //
3803 // Adjust the type and inputs of an AddP which computes the
3804 // address of a field of an instance
3805 //
3806 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3807 PhaseGVN* igvn = _igvn;
3808 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3809 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3810 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3811 if (t == nullptr) {
3812 // We are computing a raw address for a store captured by an Initialize
3813 // compute an appropriate address type (cases #3 and #5).
3814 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3815 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3816 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3817 assert(offs != Type::OffsetBot, "offset must be a constant");
3818 t = base_t->add_offset(offs)->is_oopptr();
3819 }
3820 int inst_id = base_t->instance_id();
3821 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3822 "old type must be non-instance or match new type");
3823
3824 // The type 't' could be subclass of 'base_t'.
3825 // As result t->offset() could be large then base_t's size and it will
3826 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3827 // constructor verifies correctness of the offset.
3828 //
3829 // It could happened on subclass's branch (from the type profiling
3830 // inlining) which was not eliminated during parsing since the exactness
3831 // of the allocation type was not propagated to the subclass type check.
3832 //
3833 // Or the type 't' could be not related to 'base_t' at all.
3834 // It could happened when CHA type is different from MDO type on a dead path
3835 // (for example, from instanceof check) which is not collapsed during parsing.
3836 //
3837 // Do nothing for such AddP node and don't process its users since
3838 // this code branch will go away.
3839 //
3840 if (!t->is_known_instance() &&
3841 !base_t->maybe_java_subtype_of(t)) {
3842 return false; // bail out
3843 }
3844 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
3845 // Do NOT remove the next line: ensure a new alias index is allocated
3846 // for the instance type. Note: C++ will not remove it since the call
3847 // has side effect.
3848 int alias_idx = _compile->get_alias_index(tinst);
3849 igvn->set_type(addp, tinst);
3850 // record the allocation in the node map
3851 set_map(addp, get_map(base->_idx));
3852 // Set addp's Base and Address to 'base'.
3853 Node *abase = addp->in(AddPNode::Base);
3854 Node *adr = addp->in(AddPNode::Address);
3855 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3856 adr->in(0)->_idx == (uint)inst_id) {
3857 // Skip AddP cases #3 and #5.
3858 } else {
3859 assert(!abase->is_top(), "sanity"); // AddP case #3
3860 if (abase != base) {
3861 igvn->hash_delete(addp);
3862 addp->set_req(AddPNode::Base, base);
3863 if (abase == adr) {
3864 addp->set_req(AddPNode::Address, base);
3865 } else {
3866 // AddP case #4 (adr is array's element offset AddP node)
3867 #ifdef ASSERT
3868 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3869 assert(adr->is_AddP() && atype != nullptr &&
3870 atype->instance_id() == inst_id, "array's element offset should be processed first");
3871 #endif
3872 }
3873 igvn->hash_insert(addp);
3874 }
3875 }
3876 // Put on IGVN worklist since at least addp's type was changed above.
3877 record_for_optimizer(addp);
3878 return true;
3879 }
3880
3881 //
3882 // Create a new version of orig_phi if necessary. Returns either the newly
3883 // created phi or an existing phi. Sets create_new to indicate whether a new
3884 // phi was created. Cache the last newly created phi in the node map.
3885 //
3886 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3887 Compile *C = _compile;
3888 PhaseGVN* igvn = _igvn;
3889 new_created = false;
3890 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3891 // nothing to do if orig_phi is bottom memory or matches alias_idx
3892 if (phi_alias_idx == alias_idx) {
3893 return orig_phi;
3894 }
3895 // Have we recently created a Phi for this alias index?
3896 PhiNode *result = get_map_phi(orig_phi->_idx);
3897 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3898 return result;
3899 }
3900 // Previous check may fail when the same wide memory Phi was split into Phis
3901 // for different memory slices. Search all Phis for this region.
3902 if (result != nullptr) {
3903 Node* region = orig_phi->in(0);
3904 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3905 Node* phi = region->fast_out(i);
3906 if (phi->is_Phi() &&
3907 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3908 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3909 return phi->as_Phi();
3910 }
3911 }
3912 }
3913 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
3914 if (C->do_escape_analysis() == true && !C->failing()) {
3915 // Retry compilation without escape analysis.
3916 // If this is the first failure, the sentinel string will "stick"
3917 // to the Compile object, and the C2Compiler will see it and retry.
3918 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
3919 }
3920 return nullptr;
3921 }
3922 orig_phi_worklist.append_if_missing(orig_phi);
3923 const TypePtr *atype = C->get_adr_type(alias_idx);
3924 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
3925 C->copy_node_notes_to(result, orig_phi);
3926 igvn->set_type(result, result->bottom_type());
3927 record_for_optimizer(result);
3928 set_map(orig_phi, result);
3929 new_created = true;
3930 return result;
3931 }
3932
3933 //
3934 // Return a new version of Memory Phi "orig_phi" with the inputs having the
3935 // specified alias index.
3936 //
3937 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
3938 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
3939 Compile *C = _compile;
3940 PhaseGVN* igvn = _igvn;
3941 bool new_phi_created;
3942 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
3943 if (!new_phi_created) {
3944 return result;
3945 }
3946 GrowableArray<PhiNode *> phi_list;
3947 GrowableArray<uint> cur_input;
3948 PhiNode *phi = orig_phi;
3949 uint idx = 1;
3950 bool finished = false;
3951 while(!finished) {
3952 while (idx < phi->req()) {
3953 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
3954 if (mem != nullptr && mem->is_Phi()) {
3955 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
3956 if (new_phi_created) {
3957 // found an phi for which we created a new split, push current one on worklist and begin
3958 // processing new one
3959 phi_list.push(phi);
3960 cur_input.push(idx);
3961 phi = mem->as_Phi();
3962 result = newphi;
3963 idx = 1;
3964 continue;
3965 } else {
3966 mem = newphi;
3967 }
3968 }
3969 if (C->failing()) {
3970 return nullptr;
3971 }
3972 result->set_req(idx++, mem);
3973 }
3974 #ifdef ASSERT
3975 // verify that the new Phi has an input for each input of the original
3976 assert( phi->req() == result->req(), "must have same number of inputs.");
3977 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
3978 #endif
3979 // Check if all new phi's inputs have specified alias index.
3980 // Otherwise use old phi.
3981 for (uint i = 1; i < phi->req(); i++) {
3982 Node* in = result->in(i);
3983 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
3984 }
3985 // we have finished processing a Phi, see if there are any more to do
3986 finished = (phi_list.length() == 0 );
3987 if (!finished) {
3988 phi = phi_list.pop();
3989 idx = cur_input.pop();
3990 PhiNode *prev_result = get_map_phi(phi->_idx);
3991 prev_result->set_req(idx++, result);
3992 result = prev_result;
3993 }
3994 }
3995 return result;
3996 }
3997
3998 //
3999 // The next methods are derived from methods in MemNode.
4000 //
4001 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4002 Node *mem = mmem;
4003 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4004 // means an array I have not precisely typed yet. Do not do any
4005 // alias stuff with it any time soon.
4006 if (toop->base() != Type::AnyPtr &&
4007 !(toop->isa_instptr() &&
4008 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4009 toop->offset() == Type::OffsetBot)) {
4010 mem = mmem->memory_at(alias_idx);
4011 // Update input if it is progress over what we have now
4012 }
4013 return mem;
4014 }
4015
4016 //
4017 // Move memory users to their memory slices.
4018 //
4019 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4020 Compile* C = _compile;
4021 PhaseGVN* igvn = _igvn;
4022 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4023 assert(tp != nullptr, "ptr type");
4024 int alias_idx = C->get_alias_index(tp);
4025 int general_idx = C->get_general_index(alias_idx);
4026
4027 // Move users first
4028 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4029 Node* use = n->fast_out(i);
4030 if (use->is_MergeMem()) {
4031 MergeMemNode* mmem = use->as_MergeMem();
4032 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4033 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4034 continue; // Nothing to do
4035 }
4036 // Replace previous general reference to mem node.
4037 uint orig_uniq = C->unique();
4038 Node* m = find_inst_mem(n, general_idx, orig_phis);
4039 assert(orig_uniq == C->unique(), "no new nodes");
4040 mmem->set_memory_at(general_idx, m);
4041 --imax;
4042 --i;
4043 } else if (use->is_MemBar()) {
4044 assert(!use->is_Initialize(), "initializing stores should not be moved");
4045 if (use->req() > MemBarNode::Precedent &&
4046 use->in(MemBarNode::Precedent) == n) {
4047 // Don't move related membars.
4048 record_for_optimizer(use);
4049 continue;
4050 }
4051 tp = use->as_MemBar()->adr_type()->isa_ptr();
4052 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4053 alias_idx == general_idx) {
4054 continue; // Nothing to do
4055 }
4056 // Move to general memory slice.
4057 uint orig_uniq = C->unique();
4058 Node* m = find_inst_mem(n, general_idx, orig_phis);
4059 assert(orig_uniq == C->unique(), "no new nodes");
4060 igvn->hash_delete(use);
4061 imax -= use->replace_edge(n, m, igvn);
4062 igvn->hash_insert(use);
4063 record_for_optimizer(use);
4064 --i;
4065 #ifdef ASSERT
4066 } else if (use->is_Mem()) {
4067 // Memory nodes should have new memory input.
4068 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4069 assert(tp != nullptr, "ptr type");
4070 int idx = C->get_alias_index(tp);
4071 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4072 "Following memory nodes should have new memory input or be on the same memory slice");
4073 } else if (use->is_Phi()) {
4074 // Phi nodes should be split and moved already.
4075 tp = use->as_Phi()->adr_type()->isa_ptr();
4076 assert(tp != nullptr, "ptr type");
4077 int idx = C->get_alias_index(tp);
4078 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4079 } else {
4080 use->dump();
4081 assert(false, "should not be here");
4082 #endif
4083 }
4084 }
4085 }
4086
4087 //
4088 // Search memory chain of "mem" to find a MemNode whose address
4089 // is the specified alias index.
4090 //
4091 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4092 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4093 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4094 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4095 return nullptr;
4096 }
4097 if (orig_mem == nullptr) {
4098 return orig_mem;
4099 }
4100 Compile* C = _compile;
4101 PhaseGVN* igvn = _igvn;
4102 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4103 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4104 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4105 Node *prev = nullptr;
4106 Node *result = orig_mem;
4107 while (prev != result) {
4108 prev = result;
4109 if (result == start_mem) {
4110 break; // hit one of our sentinels
4111 }
4112 if (result->is_Mem()) {
4113 const Type *at = igvn->type(result->in(MemNode::Address));
4114 if (at == Type::TOP) {
4115 break; // Dead
4116 }
4117 assert (at->isa_ptr() != nullptr, "pointer type required.");
4118 int idx = C->get_alias_index(at->is_ptr());
4119 if (idx == alias_idx) {
4120 break; // Found
4121 }
4122 if (!is_instance && (at->isa_oopptr() == nullptr ||
4123 !at->is_oopptr()->is_known_instance())) {
4124 break; // Do not skip store to general memory slice.
4125 }
4126 result = result->in(MemNode::Memory);
4127 }
4128 if (!is_instance) {
4129 continue; // don't search further for non-instance types
4130 }
4131 // skip over a call which does not affect this memory slice
4132 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4133 Node *proj_in = result->in(0);
4134 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4135 break; // hit one of our sentinels
4136 } else if (proj_in->is_Call()) {
4137 // ArrayCopy node processed here as well
4138 CallNode *call = proj_in->as_Call();
4139 if (!call->may_modify(toop, igvn)) {
4140 result = call->in(TypeFunc::Memory);
4141 }
4142 } else if (proj_in->is_Initialize()) {
4143 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4144 // Stop if this is the initialization for the object instance which
4145 // which contains this memory slice, otherwise skip over it.
4146 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4147 result = proj_in->in(TypeFunc::Memory);
4148 }
4149 } else if (proj_in->is_MemBar()) {
4150 // Check if there is an array copy for a clone
4151 // Step over GC barrier when ReduceInitialCardMarks is disabled
4152 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4153 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4154
4155 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4156 // Stop if it is a clone
4157 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4158 if (ac->may_modify(toop, igvn)) {
4159 break;
4160 }
4161 }
4162 result = proj_in->in(TypeFunc::Memory);
4163 }
4164 } else if (result->is_MergeMem()) {
4165 MergeMemNode *mmem = result->as_MergeMem();
4166 result = step_through_mergemem(mmem, alias_idx, toop);
4167 if (result == mmem->base_memory()) {
4168 // Didn't find instance memory, search through general slice recursively.
4169 result = mmem->memory_at(C->get_general_index(alias_idx));
4170 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4171 if (C->failing()) {
4172 return nullptr;
4173 }
4174 mmem->set_memory_at(alias_idx, result);
4175 }
4176 } else if (result->is_Phi() &&
4177 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4178 Node *un = result->as_Phi()->unique_input(igvn);
4179 if (un != nullptr) {
4180 orig_phis.append_if_missing(result->as_Phi());
4181 result = un;
4182 } else {
4183 break;
4184 }
4185 } else if (result->is_ClearArray()) {
4186 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4187 // Can not bypass initialization of the instance
4188 // we are looking for.
4189 break;
4190 }
4191 // Otherwise skip it (the call updated 'result' value).
4192 } else if (result->Opcode() == Op_SCMemProj) {
4193 Node* mem = result->in(0);
4194 Node* adr = nullptr;
4195 if (mem->is_LoadStore()) {
4196 adr = mem->in(MemNode::Address);
4197 } else {
4198 assert(mem->Opcode() == Op_EncodeISOArray ||
4199 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4200 adr = mem->in(3); // Memory edge corresponds to destination array
4201 }
4202 const Type *at = igvn->type(adr);
4203 if (at != Type::TOP) {
4204 assert(at->isa_ptr() != nullptr, "pointer type required.");
4205 int idx = C->get_alias_index(at->is_ptr());
4206 if (idx == alias_idx) {
4207 // Assert in debug mode
4208 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4209 break; // In product mode return SCMemProj node
4210 }
4211 }
4212 result = mem->in(MemNode::Memory);
4213 } else if (result->Opcode() == Op_StrInflatedCopy) {
4214 Node* adr = result->in(3); // Memory edge corresponds to destination array
4215 const Type *at = igvn->type(adr);
4216 if (at != Type::TOP) {
4217 assert(at->isa_ptr() != nullptr, "pointer type required.");
4218 int idx = C->get_alias_index(at->is_ptr());
4219 if (idx == alias_idx) {
4220 // Assert in debug mode
4221 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4222 break; // In product mode return SCMemProj node
4223 }
4224 }
4225 result = result->in(MemNode::Memory);
4226 }
4227 }
4228 if (result->is_Phi()) {
4229 PhiNode *mphi = result->as_Phi();
4230 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4231 const TypePtr *t = mphi->adr_type();
4232 if (!is_instance) {
4233 // Push all non-instance Phis on the orig_phis worklist to update inputs
4234 // during Phase 4 if needed.
4235 orig_phis.append_if_missing(mphi);
4236 } else if (C->get_alias_index(t) != alias_idx) {
4237 // Create a new Phi with the specified alias index type.
4238 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4239 }
4240 }
4241 // the result is either MemNode, PhiNode, InitializeNode.
4242 return result;
4243 }
4244
4245 //
4246 // Convert the types of non-escaped object to instance types where possible,
4247 // propagate the new type information through the graph, and update memory
4248 // edges and MergeMem inputs to reflect the new type.
4249 //
4250 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4251 // The processing is done in 4 phases:
4252 //
4253 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4254 // types for the CheckCastPP for allocations where possible.
4255 // Propagate the new types through users as follows:
4256 // casts and Phi: push users on alloc_worklist
4257 // AddP: cast Base and Address inputs to the instance type
4258 // push any AddP users on alloc_worklist and push any memnode
4259 // users onto memnode_worklist.
4260 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4261 // search the Memory chain for a store with the appropriate type
4262 // address type. If a Phi is found, create a new version with
4263 // the appropriate memory slices from each of the Phi inputs.
4264 // For stores, process the users as follows:
4265 // MemNode: push on memnode_worklist
4266 // MergeMem: push on mergemem_worklist
4267 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4268 // moving the first node encountered of each instance type to the
4269 // the input corresponding to its alias index.
4270 // appropriate memory slice.
4271 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4272 //
4273 // In the following example, the CheckCastPP nodes are the cast of allocation
4274 // results and the allocation of node 29 is non-escaped and eligible to be an
4275 // instance type.
4276 //
4277 // We start with:
4278 //
4279 // 7 Parm #memory
4280 // 10 ConI "12"
4281 // 19 CheckCastPP "Foo"
4282 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4283 // 29 CheckCastPP "Foo"
4284 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4285 //
4286 // 40 StoreP 25 7 20 ... alias_index=4
4287 // 50 StoreP 35 40 30 ... alias_index=4
4288 // 60 StoreP 45 50 20 ... alias_index=4
4289 // 70 LoadP _ 60 30 ... alias_index=4
4290 // 80 Phi 75 50 60 Memory alias_index=4
4291 // 90 LoadP _ 80 30 ... alias_index=4
4292 // 100 LoadP _ 80 20 ... alias_index=4
4293 //
4294 //
4295 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4296 // and creating a new alias index for node 30. This gives:
4297 //
4298 // 7 Parm #memory
4299 // 10 ConI "12"
4300 // 19 CheckCastPP "Foo"
4301 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4302 // 29 CheckCastPP "Foo" iid=24
4303 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4304 //
4305 // 40 StoreP 25 7 20 ... alias_index=4
4306 // 50 StoreP 35 40 30 ... alias_index=6
4307 // 60 StoreP 45 50 20 ... alias_index=4
4308 // 70 LoadP _ 60 30 ... alias_index=6
4309 // 80 Phi 75 50 60 Memory alias_index=4
4310 // 90 LoadP _ 80 30 ... alias_index=6
4311 // 100 LoadP _ 80 20 ... alias_index=4
4312 //
4313 // In phase 2, new memory inputs are computed for the loads and stores,
4314 // And a new version of the phi is created. In phase 4, the inputs to
4315 // node 80 are updated and then the memory nodes are updated with the
4316 // values computed in phase 2. This results in:
4317 //
4318 // 7 Parm #memory
4319 // 10 ConI "12"
4320 // 19 CheckCastPP "Foo"
4321 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4322 // 29 CheckCastPP "Foo" iid=24
4323 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4324 //
4325 // 40 StoreP 25 7 20 ... alias_index=4
4326 // 50 StoreP 35 7 30 ... alias_index=6
4327 // 60 StoreP 45 40 20 ... alias_index=4
4328 // 70 LoadP _ 50 30 ... alias_index=6
4329 // 80 Phi 75 40 60 Memory alias_index=4
4330 // 120 Phi 75 50 50 Memory alias_index=6
4331 // 90 LoadP _ 120 30 ... alias_index=6
4332 // 100 LoadP _ 80 20 ... alias_index=4
4333 //
4334 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4335 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4336 GrowableArray<MergeMemNode*> &mergemem_worklist,
4337 Unique_Node_List &reducible_merges) {
4338 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4339 GrowableArray<Node *> memnode_worklist;
4340 GrowableArray<PhiNode *> orig_phis;
4341 PhaseIterGVN *igvn = _igvn;
4342 uint new_index_start = (uint) _compile->num_alias_types();
4343 VectorSet visited;
4344 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4345 uint unique_old = _compile->unique();
4346
4347 // Phase 1: Process possible allocations from alloc_worklist.
4348 // Create instance types for the CheckCastPP for allocations where possible.
4349 //
4350 // (Note: don't forget to change the order of the second AddP node on
4351 // the alloc_worklist if the order of the worklist processing is changed,
4352 // see the comment in find_second_addp().)
4353 //
4354 while (alloc_worklist.length() != 0) {
4355 Node *n = alloc_worklist.pop();
4356 uint ni = n->_idx;
4357 if (n->is_Call()) {
4358 CallNode *alloc = n->as_Call();
4359 // copy escape information to call node
4360 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4361 PointsToNode::EscapeState es = ptn->escape_state();
4362 // We have an allocation or call which returns a Java object,
4363 // see if it is non-escaped.
4364 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4365 continue;
4366 }
4367 // Find CheckCastPP for the allocate or for the return value of a call
4368 n = alloc->result_cast();
4369 if (n == nullptr) { // No uses except Initialize node
4370 if (alloc->is_Allocate()) {
4371 // Set the scalar_replaceable flag for allocation
4372 // so it could be eliminated if it has no uses.
4373 alloc->as_Allocate()->_is_scalar_replaceable = true;
4374 }
4375 continue;
4376 }
4377 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4378 // we could reach here for allocate case if one init is associated with many allocs.
4379 if (alloc->is_Allocate()) {
4380 alloc->as_Allocate()->_is_scalar_replaceable = false;
4381 }
4382 continue;
4383 }
4384
4385 // The inline code for Object.clone() casts the allocation result to
4386 // java.lang.Object and then to the actual type of the allocated
4387 // object. Detect this case and use the second cast.
4388 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4389 // the allocation result is cast to java.lang.Object and then
4390 // to the actual Array type.
4391 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4392 && (alloc->is_AllocateArray() ||
4393 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4394 Node *cast2 = nullptr;
4395 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4396 Node *use = n->fast_out(i);
4397 if (use->is_CheckCastPP()) {
4398 cast2 = use;
4399 break;
4400 }
4401 }
4402 if (cast2 != nullptr) {
4403 n = cast2;
4404 } else {
4405 // Non-scalar replaceable if the allocation type is unknown statically
4406 // (reflection allocation), the object can't be restored during
4407 // deoptimization without precise type.
4408 continue;
4409 }
4410 }
4411
4412 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4413 if (t == nullptr) {
4414 continue; // not a TypeOopPtr
4415 }
4416 if (!t->klass_is_exact()) {
4417 continue; // not an unique type
4418 }
4419 if (alloc->is_Allocate()) {
4420 // Set the scalar_replaceable flag for allocation
4421 // so it could be eliminated.
4422 alloc->as_Allocate()->_is_scalar_replaceable = true;
4423 }
4424 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4425 // in order for an object to be scalar-replaceable, it must be:
4426 // - a direct allocation (not a call returning an object)
4427 // - non-escaping
4428 // - eligible to be a unique type
4429 // - not determined to be ineligible by escape analysis
4430 set_map(alloc, n);
4431 set_map(n, alloc);
4432 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4433 igvn->hash_delete(n);
4434 igvn->set_type(n, tinst);
4435 n->raise_bottom_type(tinst);
4436 igvn->hash_insert(n);
4437 record_for_optimizer(n);
4438 // Allocate an alias index for the header fields. Accesses to
4439 // the header emitted during macro expansion wouldn't have
4440 // correct memory state otherwise.
4441 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4442 _compile->get_alias_index(tinst->add_offset(Type::klass_offset()));
4443 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4444
4445 // First, put on the worklist all Field edges from Connection Graph
4446 // which is more accurate than putting immediate users from Ideal Graph.
4447 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4448 PointsToNode* tgt = e.get();
4449 if (tgt->is_Arraycopy()) {
4450 continue;
4451 }
4452 Node* use = tgt->ideal_node();
4453 assert(tgt->is_Field() && use->is_AddP(),
4454 "only AddP nodes are Field edges in CG");
4455 if (use->outcnt() > 0) { // Don't process dead nodes
4456 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4457 if (addp2 != nullptr) {
4458 assert(alloc->is_AllocateArray(),"array allocation was expected");
4459 alloc_worklist.append_if_missing(addp2);
4460 }
4461 alloc_worklist.append_if_missing(use);
4462 }
4463 }
4464
4465 // An allocation may have an Initialize which has raw stores. Scan
4466 // the users of the raw allocation result and push AddP users
4467 // on alloc_worklist.
4468 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4469 assert (raw_result != nullptr, "must have an allocation result");
4470 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4471 Node *use = raw_result->fast_out(i);
4472 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4473 Node* addp2 = find_second_addp(use, raw_result);
4474 if (addp2 != nullptr) {
4475 assert(alloc->is_AllocateArray(),"array allocation was expected");
4476 alloc_worklist.append_if_missing(addp2);
4477 }
4478 alloc_worklist.append_if_missing(use);
4479 } else if (use->is_MemBar()) {
4480 memnode_worklist.append_if_missing(use);
4481 }
4482 }
4483 }
4484 } else if (n->is_AddP()) {
4485 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4486 // This AddP will go away when we reduce the Phi
4487 continue;
4488 }
4489 Node* addp_base = get_addp_base(n);
4490 JavaObjectNode* jobj = unique_java_object(addp_base);
4491 if (jobj == nullptr || jobj == phantom_obj) {
4492 #ifdef ASSERT
4493 ptnode_adr(get_addp_base(n)->_idx)->dump();
4494 ptnode_adr(n->_idx)->dump();
4495 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4496 #endif
4497 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4498 return;
4499 }
4500 Node *base = get_map(jobj->idx()); // CheckCastPP node
4501 if (!split_AddP(n, base)) continue; // wrong type from dead path
4502 } else if (n->is_Phi() ||
4503 n->is_CheckCastPP() ||
4504 n->is_EncodeP() ||
4505 n->is_DecodeN() ||
4506 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4507 if (visited.test_set(n->_idx)) {
4508 assert(n->is_Phi(), "loops only through Phi's");
4509 continue; // already processed
4510 }
4511 // Reducible Phi's will be removed from the graph after split_unique_types
4512 // finishes. For now we just try to split out the SR inputs of the merge.
4513 Node* parent = n->in(1);
4514 if (reducible_merges.member(n)) {
4515 reduce_phi(n->as_Phi(), alloc_worklist, memnode_worklist);
4516 #ifdef ASSERT
4517 if (VerifyReduceAllocationMerges) {
4518 reduced_merges.push(n);
4519 }
4520 #endif
4521 continue;
4522 } else if (reducible_merges.member(parent)) {
4523 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4524 // part of reduce_merge.
4525 continue;
4526 }
4527 JavaObjectNode* jobj = unique_java_object(n);
4528 if (jobj == nullptr || jobj == phantom_obj) {
4529 #ifdef ASSERT
4530 ptnode_adr(n->_idx)->dump();
4531 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4532 #endif
4533 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4534 return;
4535 } else {
4536 Node *val = get_map(jobj->idx()); // CheckCastPP node
4537 TypeNode *tn = n->as_Type();
4538 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4539 assert(tinst != nullptr && tinst->is_known_instance() &&
4540 tinst->instance_id() == jobj->idx() , "instance type expected.");
4541
4542 const Type *tn_type = igvn->type(tn);
4543 const TypeOopPtr *tn_t;
4544 if (tn_type->isa_narrowoop()) {
4545 tn_t = tn_type->make_ptr()->isa_oopptr();
4546 } else {
4547 tn_t = tn_type->isa_oopptr();
4548 }
4549 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4550 if (tn_type->isa_narrowoop()) {
4551 tn_type = tinst->make_narrowoop();
4552 } else {
4553 tn_type = tinst;
4554 }
4555 igvn->hash_delete(tn);
4556 igvn->set_type(tn, tn_type);
4557 tn->set_type(tn_type);
4558 igvn->hash_insert(tn);
4559 record_for_optimizer(n);
4560 } else {
4561 assert(tn_type == TypePtr::NULL_PTR ||
4562 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4563 "unexpected type");
4564 continue; // Skip dead path with different type
4565 }
4566 }
4567 } else {
4568 DEBUG_ONLY(n->dump();)
4569 assert(false, "EA: unexpected node");
4570 continue;
4571 }
4572 // push allocation's users on appropriate worklist
4573 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4574 Node *use = n->fast_out(i);
4575 if(use->is_Mem() && use->in(MemNode::Address) == n) {
4576 // Load/store to instance's field
4577 memnode_worklist.append_if_missing(use);
4578 } else if (use->is_MemBar()) {
4579 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4580 memnode_worklist.append_if_missing(use);
4581 }
4582 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4583 Node* addp2 = find_second_addp(use, n);
4584 if (addp2 != nullptr) {
4585 alloc_worklist.append_if_missing(addp2);
4586 }
4587 alloc_worklist.append_if_missing(use);
4588 } else if (use->is_Phi() ||
4589 use->is_CheckCastPP() ||
4590 use->is_EncodeNarrowPtr() ||
4591 use->is_DecodeNarrowPtr() ||
4592 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4593 alloc_worklist.append_if_missing(use);
4594 #ifdef ASSERT
4595 } else if (use->is_Mem()) {
4596 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4597 } else if (use->is_MergeMem()) {
4598 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4599 } else if (use->is_SafePoint()) {
4600 // Look for MergeMem nodes for calls which reference unique allocation
4601 // (through CheckCastPP nodes) even for debug info.
4602 Node* m = use->in(TypeFunc::Memory);
4603 if (m->is_MergeMem()) {
4604 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4605 }
4606 } else if (use->Opcode() == Op_EncodeISOArray) {
4607 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4608 // EncodeISOArray overwrites destination array
4609 memnode_worklist.append_if_missing(use);
4610 }
4611 } else {
4612 uint op = use->Opcode();
4613 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4614 (use->in(MemNode::Memory) == n)) {
4615 // They overwrite memory edge corresponding to destination array,
4616 memnode_worklist.append_if_missing(use);
4617 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4618 op == Op_CastP2X ||
4619 op == Op_FastLock || op == Op_AryEq ||
4620 op == Op_StrComp || op == Op_CountPositives ||
4621 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4622 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4623 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4624 op == Op_SubTypeCheck ||
4625 op == Op_ReinterpretS2HF ||
4626 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4627 n->dump();
4628 use->dump();
4629 assert(false, "EA: missing allocation reference path");
4630 }
4631 #endif
4632 }
4633 }
4634
4635 }
4636
4637 #ifdef ASSERT
4638 if (VerifyReduceAllocationMerges) {
4639 for (uint i = 0; i < reducible_merges.size(); i++) {
4640 Node* phi = reducible_merges.at(i);
4641
4642 if (!reduced_merges.member(phi)) {
4643 phi->dump(2);
4644 phi->dump(-2);
4645 assert(false, "This reducible merge wasn't reduced.");
4646 }
4647
4648 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4649 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4650 Node* use = phi->fast_out(j);
4651 if (!use->is_SafePoint() && !use->is_CastPP()) {
4652 phi->dump(2);
4653 phi->dump(-2);
4654 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4655 }
4656 }
4657 }
4658 }
4659 #endif
4660
4661 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4662 // type, record it in the ArrayCopy node so we know what memory this
4663 // node uses/modified.
4664 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4665 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4666 Node* dest = ac->in(ArrayCopyNode::Dest);
4667 if (dest->is_AddP()) {
4668 dest = get_addp_base(dest);
4669 }
4670 JavaObjectNode* jobj = unique_java_object(dest);
4671 if (jobj != nullptr) {
4672 Node *base = get_map(jobj->idx());
4673 if (base != nullptr) {
4674 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4675 ac->_dest_type = base_t;
4676 }
4677 }
4678 Node* src = ac->in(ArrayCopyNode::Src);
4679 if (src->is_AddP()) {
4680 src = get_addp_base(src);
4681 }
4682 jobj = unique_java_object(src);
4683 if (jobj != nullptr) {
4684 Node* base = get_map(jobj->idx());
4685 if (base != nullptr) {
4686 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4687 ac->_src_type = base_t;
4688 }
4689 }
4690 }
4691
4692 // New alias types were created in split_AddP().
4693 uint new_index_end = (uint) _compile->num_alias_types();
4694
4695 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4696 // compute new values for Memory inputs (the Memory inputs are not
4697 // actually updated until phase 4.)
4698 if (memnode_worklist.length() == 0)
4699 return; // nothing to do
4700 while (memnode_worklist.length() != 0) {
4701 Node *n = memnode_worklist.pop();
4702 if (visited.test_set(n->_idx)) {
4703 continue;
4704 }
4705 if (n->is_Phi() || n->is_ClearArray()) {
4706 // we don't need to do anything, but the users must be pushed
4707 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
4708 // we don't need to do anything, but the users must be pushed
4709 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4710 if (n == nullptr) {
4711 continue;
4712 }
4713 } else if (n->is_CallLeaf()) {
4714 // Runtime calls with narrow memory input (no MergeMem node)
4715 // get the memory projection
4716 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4717 if (n == nullptr) {
4718 continue;
4719 }
4720 } else if (n->Opcode() == Op_StrInflatedCopy) {
4721 // Check direct uses of StrInflatedCopy.
4722 // It is memory type Node - no special SCMemProj node.
4723 } else if (n->Opcode() == Op_StrCompressedCopy ||
4724 n->Opcode() == Op_EncodeISOArray) {
4725 // get the memory projection
4726 n = n->find_out_with(Op_SCMemProj);
4727 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4728 } else {
4729 #ifdef ASSERT
4730 if (!n->is_Mem()) {
4731 n->dump();
4732 }
4733 assert(n->is_Mem(), "memory node required.");
4734 #endif
4735 Node *addr = n->in(MemNode::Address);
4736 const Type *addr_t = igvn->type(addr);
4737 if (addr_t == Type::TOP) {
4738 continue;
4739 }
4740 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4741 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4742 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4743 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4744 if (_compile->failing()) {
4745 return;
4746 }
4747 if (mem != n->in(MemNode::Memory)) {
4748 // We delay the memory edge update since we need old one in
4749 // MergeMem code below when instances memory slices are separated.
4750 set_map(n, mem);
4751 }
4752 if (n->is_Load()) {
4753 continue; // don't push users
4754 } else if (n->is_LoadStore()) {
4755 // get the memory projection
4756 n = n->find_out_with(Op_SCMemProj);
4757 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4758 }
4759 }
4760 // push user on appropriate worklist
4761 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4762 Node *use = n->fast_out(i);
4763 if (use->is_Phi() || use->is_ClearArray()) {
4764 memnode_worklist.append_if_missing(use);
4765 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4766 memnode_worklist.append_if_missing(use);
4767 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4768 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4769 memnode_worklist.append_if_missing(use);
4770 }
4771 #ifdef ASSERT
4772 } else if(use->is_Mem()) {
4773 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4774 } else if (use->is_MergeMem()) {
4775 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4776 } else if (use->Opcode() == Op_EncodeISOArray) {
4777 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4778 // EncodeISOArray overwrites destination array
4779 memnode_worklist.append_if_missing(use);
4780 }
4781 } else {
4782 uint op = use->Opcode();
4783 if ((use->in(MemNode::Memory) == n) &&
4784 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4785 // They overwrite memory edge corresponding to destination array,
4786 memnode_worklist.append_if_missing(use);
4787 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4788 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4789 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4790 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
4791 n->dump();
4792 use->dump();
4793 assert(false, "EA: missing memory path");
4794 }
4795 #endif
4796 }
4797 }
4798 }
4799
4800 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4801 // Walk each memory slice moving the first node encountered of each
4802 // instance type to the input corresponding to its alias index.
4803 uint length = mergemem_worklist.length();
4804 for( uint next = 0; next < length; ++next ) {
4805 MergeMemNode* nmm = mergemem_worklist.at(next);
4806 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4807 // Note: we don't want to use MergeMemStream here because we only want to
4808 // scan inputs which exist at the start, not ones we add during processing.
4809 // Note 2: MergeMem may already contains instance memory slices added
4810 // during find_inst_mem() call when memory nodes were processed above.
4811 igvn->hash_delete(nmm);
4812 uint nslices = MIN2(nmm->req(), new_index_start);
4813 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4814 Node* mem = nmm->in(i);
4815 Node* cur = nullptr;
4816 if (mem == nullptr || mem->is_top()) {
4817 continue;
4818 }
4819 // First, update mergemem by moving memory nodes to corresponding slices
4820 // if their type became more precise since this mergemem was created.
4821 while (mem->is_Mem()) {
4822 const Type *at = igvn->type(mem->in(MemNode::Address));
4823 if (at != Type::TOP) {
4824 assert (at->isa_ptr() != nullptr, "pointer type required.");
4825 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4826 if (idx == i) {
4827 if (cur == nullptr) {
4828 cur = mem;
4829 }
4830 } else {
4831 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4832 nmm->set_memory_at(idx, mem);
4833 }
4834 }
4835 }
4836 mem = mem->in(MemNode::Memory);
4837 }
4838 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4839 // Find any instance of the current type if we haven't encountered
4840 // already a memory slice of the instance along the memory chain.
4841 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4842 if((uint)_compile->get_general_index(ni) == i) {
4843 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4844 if (nmm->is_empty_memory(m)) {
4845 Node* result = find_inst_mem(mem, ni, orig_phis);
4846 if (_compile->failing()) {
4847 return;
4848 }
4849 nmm->set_memory_at(ni, result);
4850 }
4851 }
4852 }
4853 }
4854 // Find the rest of instances values
4855 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4856 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4857 Node* result = step_through_mergemem(nmm, ni, tinst);
4858 if (result == nmm->base_memory()) {
4859 // Didn't find instance memory, search through general slice recursively.
4860 result = nmm->memory_at(_compile->get_general_index(ni));
4861 result = find_inst_mem(result, ni, orig_phis);
4862 if (_compile->failing()) {
4863 return;
4864 }
4865 nmm->set_memory_at(ni, result);
4866 }
4867 }
4868
4869 // If we have crossed the 3/4 point of max node limit it's too risky
4870 // to continue with EA/SR because we might hit the max node limit.
4871 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4872 if (_compile->do_reduce_allocation_merges()) {
4873 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4874 } else if (_invocation > 0) {
4875 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4876 } else {
4877 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4878 }
4879 return;
4880 }
4881
4882 igvn->hash_insert(nmm);
4883 record_for_optimizer(nmm);
4884 }
4885
4886 // Phase 4: Update the inputs of non-instance memory Phis and
4887 // the Memory input of memnodes
4888 // First update the inputs of any non-instance Phi's from
4889 // which we split out an instance Phi. Note we don't have
4890 // to recursively process Phi's encountered on the input memory
4891 // chains as is done in split_memory_phi() since they will
4892 // also be processed here.
4893 for (int j = 0; j < orig_phis.length(); j++) {
4894 PhiNode *phi = orig_phis.at(j);
4895 int alias_idx = _compile->get_alias_index(phi->adr_type());
4896 igvn->hash_delete(phi);
4897 for (uint i = 1; i < phi->req(); i++) {
4898 Node *mem = phi->in(i);
4899 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
4900 if (_compile->failing()) {
4901 return;
4902 }
4903 if (mem != new_mem) {
4904 phi->set_req(i, new_mem);
4905 }
4906 }
4907 igvn->hash_insert(phi);
4908 record_for_optimizer(phi);
4909 }
4910
4911 // Update the memory inputs of MemNodes with the value we computed
4912 // in Phase 2 and move stores memory users to corresponding memory slices.
4913 // Disable memory split verification code until the fix for 6984348.
4914 // Currently it produces false negative results since it does not cover all cases.
4915 #if 0 // ifdef ASSERT
4916 visited.Reset();
4917 Node_Stack old_mems(arena, _compile->unique() >> 2);
4918 #endif
4919 for (uint i = 0; i < ideal_nodes.size(); i++) {
4920 Node* n = ideal_nodes.at(i);
4921 Node* nmem = get_map(n->_idx);
4922 assert(nmem != nullptr, "sanity");
4923 if (n->is_Mem()) {
4924 #if 0 // ifdef ASSERT
4925 Node* old_mem = n->in(MemNode::Memory);
4926 if (!visited.test_set(old_mem->_idx)) {
4927 old_mems.push(old_mem, old_mem->outcnt());
4928 }
4929 #endif
4930 assert(n->in(MemNode::Memory) != nmem, "sanity");
4931 if (!n->is_Load()) {
4932 // Move memory users of a store first.
4933 move_inst_mem(n, orig_phis);
4934 }
4935 // Now update memory input
4936 igvn->hash_delete(n);
4937 n->set_req(MemNode::Memory, nmem);
4938 igvn->hash_insert(n);
4939 record_for_optimizer(n);
4940 } else {
4941 assert(n->is_Allocate() || n->is_CheckCastPP() ||
4942 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
4943 }
4944 }
4945 #if 0 // ifdef ASSERT
4946 // Verify that memory was split correctly
4947 while (old_mems.is_nonempty()) {
4948 Node* old_mem = old_mems.node();
4949 uint old_cnt = old_mems.index();
4950 old_mems.pop();
4951 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
4952 }
4953 #endif
4954 }
4955
4956 #ifndef PRODUCT
4957 int ConnectionGraph::_no_escape_counter = 0;
4958 int ConnectionGraph::_arg_escape_counter = 0;
4959 int ConnectionGraph::_global_escape_counter = 0;
4960
4961 static const char *node_type_names[] = {
4962 "UnknownType",
4963 "JavaObject",
4964 "LocalVar",
4965 "Field",
4966 "Arraycopy"
4967 };
4968
4969 static const char *esc_names[] = {
4970 "UnknownEscape",
4971 "NoEscape",
4972 "ArgEscape",
4973 "GlobalEscape"
4974 };
4975
4976 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
4977 NodeType nt = node_type();
4978 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
4979 if (print_state) {
4980 EscapeState es = escape_state();
4981 EscapeState fields_es = fields_escape_state();
4982 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
4983 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
4984 out->print("NSR ");
4985 }
4986 }
4987 }
4988
4989 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
4990 dump_header(print_state, out);
4991 if (is_Field()) {
4992 FieldNode* f = (FieldNode*)this;
4993 if (f->is_oop()) {
4994 out->print("oop ");
4995 }
4996 if (f->offset() > 0) {
4997 out->print("+%d ", f->offset());
4998 }
4999 out->print("(");
5000 for (BaseIterator i(f); i.has_next(); i.next()) {
5001 PointsToNode* b = i.get();
5002 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5003 }
5004 out->print(" )");
5005 }
5006 out->print("[");
5007 for (EdgeIterator i(this); i.has_next(); i.next()) {
5008 PointsToNode* e = i.get();
5009 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5010 }
5011 out->print(" [");
5012 for (UseIterator i(this); i.has_next(); i.next()) {
5013 PointsToNode* u = i.get();
5014 bool is_base = false;
5015 if (PointsToNode::is_base_use(u)) {
5016 is_base = true;
5017 u = PointsToNode::get_use_node(u)->as_Field();
5018 }
5019 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5020 }
5021 out->print(" ]] ");
5022 if (_node == nullptr) {
5023 out->print("<null>%s", newline ? "\n" : "");
5024 } else {
5025 _node->dump(newline ? "\n" : "", false, out);
5026 }
5027 }
5028
5029 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5030 bool first = true;
5031 int ptnodes_length = ptnodes_worklist.length();
5032 for (int i = 0; i < ptnodes_length; i++) {
5033 PointsToNode *ptn = ptnodes_worklist.at(i);
5034 if (ptn == nullptr || !ptn->is_JavaObject()) {
5035 continue;
5036 }
5037 PointsToNode::EscapeState es = ptn->escape_state();
5038 if ((es != PointsToNode::NoEscape) && !Verbose) {
5039 continue;
5040 }
5041 Node* n = ptn->ideal_node();
5042 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5043 n->as_CallStaticJava()->is_boxing_method())) {
5044 if (first) {
5045 tty->cr();
5046 tty->print("======== Connection graph for ");
5047 _compile->method()->print_short_name();
5048 tty->cr();
5049 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5050 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5051 tty->cr();
5052 first = false;
5053 }
5054 ptn->dump();
5055 // Print all locals and fields which reference this allocation
5056 for (UseIterator j(ptn); j.has_next(); j.next()) {
5057 PointsToNode* use = j.get();
5058 if (use->is_LocalVar()) {
5059 use->dump(Verbose);
5060 } else if (Verbose) {
5061 use->dump();
5062 }
5063 }
5064 tty->cr();
5065 }
5066 }
5067 }
5068
5069 void ConnectionGraph::print_statistics() {
5070 tty->print_cr("No escape = %d, Arg escape = %d, Global escape = %d", Atomic::load(&_no_escape_counter), Atomic::load(&_arg_escape_counter), Atomic::load(&_global_escape_counter));
5071 }
5072
5073 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5074 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5075 return;
5076 }
5077 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5078 JavaObjectNode* ptn = java_objects_worklist.at(next);
5079 if (ptn->ideal_node()->is_Allocate()) {
5080 if (ptn->escape_state() == PointsToNode::NoEscape) {
5081 Atomic::inc(&ConnectionGraph::_no_escape_counter);
5082 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5083 Atomic::inc(&ConnectionGraph::_arg_escape_counter);
5084 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5085 Atomic::inc(&ConnectionGraph::_global_escape_counter);
5086 } else {
5087 assert(false, "Unexpected Escape State");
5088 }
5089 }
5090 }
5091 }
5092
5093 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5094 if (_compile->directive()->TraceEscapeAnalysisOption) {
5095 assert(ptn != nullptr, "should not be null");
5096 assert(reason != nullptr, "should not be null");
5097 ptn->dump_header(true);
5098 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5099 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5100 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5101 }
5102 }
5103
5104 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5105 if (_compile->directive()->TraceEscapeAnalysisOption) {
5106 stringStream ss;
5107 ss.print("propagated from: ");
5108 from->dump(true, &ss, false);
5109 return ss.as_string();
5110 } else {
5111 return nullptr;
5112 }
5113 }
5114
5115 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5116 if (_compile->directive()->TraceEscapeAnalysisOption) {
5117 stringStream ss;
5118 ss.print("escapes as arg to:");
5119 call->dump("", false, &ss);
5120 return ss.as_string();
5121 } else {
5122 return nullptr;
5123 }
5124 }
5125
5126 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5127 if (_compile->directive()->TraceEscapeAnalysisOption) {
5128 stringStream ss;
5129 ss.print("is merged with other object: ");
5130 other->dump_header(true, &ss);
5131 return ss.as_string();
5132 } else {
5133 return nullptr;
5134 }
5135 }
5136
5137 #endif
5138
5139 void ConnectionGraph::record_for_optimizer(Node *n) {
5140 _igvn->_worklist.push(n);
5141 _igvn->add_users_to_worklist(n);
5142 }