1 /*
2 * Copyright (c) 2005, 2024, 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 "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/c2compiler.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/phaseX.hpp"
42 #include "opto/movenode.hpp"
43 #include "opto/narrowptrnode.hpp"
44 #include "opto/castnode.hpp"
45 #include "opto/rootnode.hpp"
46 #include "utilities/macros.hpp"
47
48 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
49 // If ReduceAllocationMerges is enabled we might call split_through_phi during
50 // split_unique_types and that will create additional nodes that need to be
51 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
52 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
53 // the array will be reallocated.
54 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
55 _in_worklist(C->comp_arena()),
56 _next_pidx(0),
57 _collecting(true),
58 _verify(false),
59 _compile(C),
60 _igvn(igvn),
61 _invocation(invocation),
62 _build_iterations(0),
63 _build_time(0.),
64 _node_map(C->comp_arena()) {
65 // Add unknown java object.
66 add_java_object(C->top(), PointsToNode::GlobalEscape);
67 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
68 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
69 // Add ConP and ConN null oop nodes
70 Node* oop_null = igvn->zerocon(T_OBJECT);
71 assert(oop_null->_idx < nodes_size(), "should be created already");
72 add_java_object(oop_null, PointsToNode::NoEscape);
73 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
74 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
75 if (UseCompressedOops) {
76 Node* noop_null = igvn->zerocon(T_NARROWOOP);
77 assert(noop_null->_idx < nodes_size(), "should be created already");
78 map_ideal_node(noop_null, null_obj);
79 }
80 }
81
82 bool ConnectionGraph::has_candidates(Compile *C) {
83 // EA brings benefits only when the code has allocations and/or locks which
84 // are represented by ideal Macro nodes.
85 int cnt = C->macro_count();
86 for (int i = 0; i < cnt; i++) {
87 Node *n = C->macro_node(i);
88 if (n->is_Allocate()) {
89 return true;
90 }
91 if (n->is_Lock()) {
92 Node* obj = n->as_Lock()->obj_node()->uncast();
93 if (!(obj->is_Parm() || obj->is_Con())) {
94 return true;
95 }
96 }
97 if (n->is_CallStaticJava() &&
98 n->as_CallStaticJava()->is_boxing_method()) {
99 return true;
100 }
101 }
102 return false;
103 }
104
105 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
106 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]);
107 ResourceMark rm;
108
109 // Add ConP and ConN null oop nodes before ConnectionGraph construction
110 // to create space for them in ConnectionGraph::_nodes[].
111 Node* oop_null = igvn->zerocon(T_OBJECT);
112 Node* noop_null = igvn->zerocon(T_NARROWOOP);
113 int invocation = 0;
114 if (C->congraph() != nullptr) {
115 invocation = C->congraph()->_invocation + 1;
116 }
117 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
118 // Perform escape analysis
119 if (congraph->compute_escape()) {
120 // There are non escaping objects.
121 C->set_congraph(congraph);
122 }
123 // Cleanup.
124 if (oop_null->outcnt() == 0) {
125 igvn->hash_delete(oop_null);
126 }
127 if (noop_null->outcnt() == 0) {
128 igvn->hash_delete(noop_null);
129 }
130 }
131
132 bool ConnectionGraph::compute_escape() {
133 Compile* C = _compile;
134 PhaseGVN* igvn = _igvn;
135
136 // Worklists used by EA.
137 Unique_Node_List delayed_worklist;
138 Unique_Node_List reducible_merges;
139 GrowableArray<Node*> alloc_worklist;
140 GrowableArray<Node*> ptr_cmp_worklist;
141 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
142 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
143 GrowableArray<PointsToNode*> ptnodes_worklist;
144 GrowableArray<JavaObjectNode*> java_objects_worklist;
145 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
146 GrowableArray<FieldNode*> oop_fields_worklist;
147 GrowableArray<SafePointNode*> sfn_worklist;
148 GrowableArray<MergeMemNode*> mergemem_worklist;
149 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
150
151 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]);
152
153 // 1. Populate Connection Graph (CG) with PointsTo nodes.
154 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
155 // Initialize worklist
156 if (C->root() != nullptr) {
157 ideal_nodes.push(C->root());
158 }
159 // Processed ideal nodes are unique on ideal_nodes list
160 // but several ideal nodes are mapped to the phantom_obj.
161 // To avoid duplicated entries on the following worklists
162 // add the phantom_obj only once to them.
163 ptnodes_worklist.append(phantom_obj);
164 java_objects_worklist.append(phantom_obj);
165 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
166 Node* n = ideal_nodes.at(next);
167 // Create PointsTo nodes and add them to Connection Graph. Called
168 // only once per ideal node since ideal_nodes is Unique_Node list.
169 add_node_to_connection_graph(n, &delayed_worklist);
170 PointsToNode* ptn = ptnode_adr(n->_idx);
171 if (ptn != nullptr && ptn != phantom_obj) {
172 ptnodes_worklist.append(ptn);
173 if (ptn->is_JavaObject()) {
174 java_objects_worklist.append(ptn->as_JavaObject());
175 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
176 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
177 // Only allocations and java static calls results are interesting.
178 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
179 }
180 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
181 oop_fields_worklist.append(ptn->as_Field());
182 }
183 }
184 // Collect some interesting nodes for further use.
185 switch (n->Opcode()) {
186 case Op_MergeMem:
187 // Collect all MergeMem nodes to add memory slices for
188 // scalar replaceable objects in split_unique_types().
189 mergemem_worklist.append(n->as_MergeMem());
190 break;
191 case Op_CmpP:
192 case Op_CmpN:
193 // Collect compare pointers nodes.
194 if (OptimizePtrCompare) {
195 ptr_cmp_worklist.append(n);
196 }
197 break;
198 case Op_MemBarStoreStore:
199 // Collect all MemBarStoreStore nodes so that depending on the
200 // escape status of the associated Allocate node some of them
201 // may be eliminated.
202 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
203 storestore_worklist.append(n->as_MemBarStoreStore());
204 }
205 break;
206 case Op_MemBarRelease:
207 if (n->req() > MemBarNode::Precedent) {
208 record_for_optimizer(n);
209 }
210 break;
211 #ifdef ASSERT
212 case Op_AddP:
213 // Collect address nodes for graph verification.
214 addp_worklist.append(n);
215 break;
216 #endif
217 case Op_ArrayCopy:
218 // Keep a list of ArrayCopy nodes so if one of its input is non
219 // escaping, we can record a unique type
220 arraycopy_worklist.append(n->as_ArrayCopy());
221 break;
222 default:
223 // not interested now, ignore...
224 break;
225 }
226 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
227 Node* m = n->fast_out(i); // Get user
228 ideal_nodes.push(m);
229 }
230 if (n->is_SafePoint()) {
231 sfn_worklist.append(n->as_SafePoint());
232 }
233 }
234
235 #ifndef PRODUCT
236 if (_compile->directive()->TraceEscapeAnalysisOption) {
237 tty->print("+++++ Initial worklist for ");
238 _compile->method()->print_name();
239 tty->print_cr(" (ea_inv=%d)", _invocation);
240 for (int i = 0; i < ptnodes_worklist.length(); i++) {
241 PointsToNode* ptn = ptnodes_worklist.at(i);
242 ptn->dump();
243 }
244 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
245 }
246 #endif
247
248 if (non_escaped_allocs_worklist.length() == 0) {
249 _collecting = false;
250 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
251 return false; // Nothing to do.
252 }
253 // Add final simple edges to graph.
254 while(delayed_worklist.size() > 0) {
255 Node* n = delayed_worklist.pop();
256 add_final_edges(n);
257 }
258
259 #ifdef ASSERT
260 if (VerifyConnectionGraph) {
261 // Verify that no new simple edges could be created and all
262 // local vars has edges.
263 _verify = true;
264 int ptnodes_length = ptnodes_worklist.length();
265 for (int next = 0; next < ptnodes_length; ++next) {
266 PointsToNode* ptn = ptnodes_worklist.at(next);
267 add_final_edges(ptn->ideal_node());
268 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
269 ptn->dump();
270 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
271 }
272 }
273 _verify = false;
274 }
275 #endif
276 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
277 // processing, calls to CI to resolve symbols (types, fields, methods)
278 // referenced in bytecode. During symbol resolution VM may throw
279 // an exception which CI cleans and converts to compilation failure.
280 if (C->failing()) {
281 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
282 return false;
283 }
284
285 // 2. Finish Graph construction by propagating references to all
286 // java objects through graph.
287 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
288 java_objects_worklist, oop_fields_worklist)) {
289 // All objects escaped or hit time or iterations limits.
290 _collecting = false;
291 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
292 return false;
293 }
294
295 // 3. Adjust scalar_replaceable state of nonescaping objects and push
296 // scalar replaceable allocations on alloc_worklist for processing
297 // in split_unique_types().
298 GrowableArray<JavaObjectNode*> jobj_worklist;
299 int non_escaped_length = non_escaped_allocs_worklist.length();
300 bool found_nsr_alloc = false;
301 for (int next = 0; next < non_escaped_length; next++) {
302 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
303 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
304 Node* n = ptn->ideal_node();
305 if (n->is_Allocate()) {
306 n->as_Allocate()->_is_non_escaping = noescape;
307 }
308 if (noescape && ptn->scalar_replaceable()) {
309 adjust_scalar_replaceable_state(ptn, reducible_merges);
310 if (ptn->scalar_replaceable()) {
311 jobj_worklist.push(ptn);
312 } else {
313 found_nsr_alloc = true;
314 }
315 }
316 }
317
318 // Propagate NSR (Not Scalar Replaceable) state.
319 if (found_nsr_alloc) {
320 find_scalar_replaceable_allocs(jobj_worklist);
321 }
322
323 // alloc_worklist will be processed in reverse push order.
324 // Therefore the reducible Phis will be processed for last and that's what we
325 // want because by then the scalarizable inputs of the merge will already have
326 // an unique instance type.
327 for (uint i = 0; i < reducible_merges.size(); i++ ) {
328 Node* n = reducible_merges.at(i);
329 alloc_worklist.append(n);
330 }
331
332 for (int next = 0; next < jobj_worklist.length(); ++next) {
333 JavaObjectNode* jobj = jobj_worklist.at(next);
334 if (jobj->scalar_replaceable()) {
335 alloc_worklist.append(jobj->ideal_node());
336 }
337 }
338
339 #ifdef ASSERT
340 if (VerifyConnectionGraph) {
341 // Verify that graph is complete - no new edges could be added or needed.
342 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
343 java_objects_worklist, addp_worklist);
344 }
345 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
346 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
347 null_obj->edge_count() == 0 &&
348 !null_obj->arraycopy_src() &&
349 !null_obj->arraycopy_dst(), "sanity");
350 #endif
351
352 _collecting = false;
353
354 } // TracePhase t3("connectionGraph")
355
356 // 4. Optimize ideal graph based on EA information.
357 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
358 if (has_non_escaping_obj) {
359 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
360 }
361
362 #ifndef PRODUCT
363 if (PrintEscapeAnalysis) {
364 dump(ptnodes_worklist); // Dump ConnectionGraph
365 }
366 #endif
367
368 #ifdef ASSERT
369 if (VerifyConnectionGraph) {
370 int alloc_length = alloc_worklist.length();
371 for (int next = 0; next < alloc_length; ++next) {
372 Node* n = alloc_worklist.at(next);
373 PointsToNode* ptn = ptnode_adr(n->_idx);
374 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
375 }
376 }
377
378 if (VerifyReduceAllocationMerges) {
379 for (uint i = 0; i < reducible_merges.size(); i++ ) {
380 Node* n = reducible_merges.at(i);
381 if (!can_reduce_phi(n->as_Phi())) {
382 TraceReduceAllocationMerges = true;
383 n->dump(2);
384 n->dump(-2);
385 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
386 }
387 }
388 }
389 #endif
390
391 // 5. Separate memory graph for scalar replaceable allcations.
392 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
393 if (has_scalar_replaceable_candidates && EliminateAllocations) {
394 assert(C->do_aliasing(), "Aliasing should be enabled");
395 // Now use the escape information to create unique types for
396 // scalar replaceable objects.
397 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
398 if (C->failing()) {
399 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
400 return false;
401 }
402 C->print_method(PHASE_AFTER_EA, 2);
403
404 #ifdef ASSERT
405 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
406 tty->print("=== No allocations eliminated for ");
407 C->method()->print_short_name();
408 if (!EliminateAllocations) {
409 tty->print(" since EliminateAllocations is off ===");
410 } else if(!has_scalar_replaceable_candidates) {
411 tty->print(" since there are no scalar replaceable candidates ===");
412 }
413 tty->cr();
414 #endif
415 }
416
417 // 6. Reduce allocation merges used as debug information. This is done after
418 // split_unique_types because the methods used to create SafePointScalarObject
419 // need to traverse the memory graph to find values for object fields. We also
420 // set to null the scalarized inputs of reducible Phis so that the Allocate
421 // that they point can be later scalar replaced.
422 bool delay = _igvn->delay_transform();
423 _igvn->set_delay_transform(true);
424 for (uint i = 0; i < reducible_merges.size(); i++) {
425 Node* n = reducible_merges.at(i);
426 if (n->outcnt() > 0) {
427 if (!reduce_phi_on_safepoints(n->as_Phi())) {
428 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
429 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
430 return false;
431 }
432
433 // Now we set the scalar replaceable inputs of ophi to null, which is
434 // the last piece that would prevent it from being scalar replaceable.
435 reset_scalar_replaceable_entries(n->as_Phi());
436 }
437 }
438 _igvn->set_delay_transform(delay);
439
440 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
441 // java calls if they pass ArgEscape objects as parameters.
442 if (has_non_escaping_obj &&
443 (C->env()->should_retain_local_variables() ||
444 C->env()->jvmti_can_get_owned_monitor_info() ||
445 C->env()->jvmti_can_walk_any_space() ||
446 DeoptimizeObjectsALot)) {
447 int sfn_length = sfn_worklist.length();
448 for (int next = 0; next < sfn_length; next++) {
449 SafePointNode* sfn = sfn_worklist.at(next);
450 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
451 if (sfn->is_CallJava()) {
452 CallJavaNode* call = sfn->as_CallJava();
453 call->set_arg_escape(has_arg_escape(call));
454 }
455 }
456 }
457
458 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
459 return has_non_escaping_obj;
460 }
461
462 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
463 // if at least one scalar replaceable allocation participates in the merge.
464 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
465 bool found_sr_allocate = false;
466
467 for (uint i = 1; i < ophi->req(); i++) {
468 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
469 if (ptn != nullptr && ptn->scalar_replaceable()) {
470 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
471
472 // Don't handle arrays.
473 if (alloc->Opcode() != Op_Allocate) {
474 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
475 continue;
476 }
477
478 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
479 found_sr_allocate = true;
480 } else {
481 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
482 ptn->set_scalar_replaceable(false);
483 }
484 }
485 }
486
487 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);)
488 return found_sr_allocate;
489 }
490
491 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
492 // I require the 'other' input to be a constant so that I can move the Cmp
493 // around safely.
494 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
495 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
496 Node* left = cmp->in(1);
497 Node* right = cmp->in(2);
498
499 return (left == n || right == n) &&
500 (left->is_Con() || right->is_Con()) &&
501 cmp->outcnt() == 1;
502 }
503
504 // We are going to check if any of the SafePointScalarMerge entries
505 // in the SafePoint reference the Phi that we are checking.
506 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
507 JVMState *jvms = sfpt->jvms();
508
509 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
510 Node* sfpt_in = sfpt->in(i);
511 if (sfpt_in->is_SafePointScalarMerge()) {
512 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
513 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
514 if (nsr_ptr == n) {
515 return true;
516 }
517 }
518 }
519
520 return false;
521 }
522
523 // Check if we are able to untangle the merge. The following patterns are
524 // supported:
525 // - Phi -> SafePoints
526 // - Phi -> CmpP/N
527 // - Phi -> AddP -> Load
528 // - Phi -> CastPP -> SafePoints
529 // - Phi -> CastPP -> AddP -> Load
530 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
531 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
532 Node* use = n->fast_out(i);
533
534 if (use->is_SafePoint()) {
535 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
536 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
537 return false;
538 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
539 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
540 return false;
541 }
542 } else if (use->is_AddP()) {
543 Node* addp = use;
544 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
545 Node* use_use = addp->fast_out(j);
546 const Type* load_type = _igvn->type(use_use);
547
548 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
549 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());)
550 return false;
551 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
552 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());)
553 return false;
554 }
555 }
556 } else if (nesting > 0) {
557 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);)
558 return false;
559 } else if (use->is_CastPP()) {
560 const Type* cast_t = _igvn->type(use);
561 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
562 #ifndef PRODUCT
563 if (TraceReduceAllocationMerges) {
564 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
565 use->dump();
566 }
567 #endif
568 return false;
569 }
570
571 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
572 if (!is_trivial_control) {
573 // If it's not a trivial control then we check if we can reduce the
574 // CmpP/N used by the If controlling the cast.
575 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
576 Node* iff = use->in(0)->in(0);
577 // We may have Opaque4 node between If and Bool nodes.
578 // Bail out in such case - we need to preserve Opaque4 for correct
579 // processing predicates after loop opts.
580 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
581 if (can_reduce) {
582 Node* iff_cmp = iff->in(1)->in(1);
583 int opc = iff_cmp->Opcode();
584 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
585 }
586 if (!can_reduce) {
587 #ifndef PRODUCT
588 if (TraceReduceAllocationMerges) {
589 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
590 n->dump(5);
591 }
592 #endif
593 return false;
594 }
595 }
596 }
597
598 if (!can_reduce_check_users(use, nesting+1)) {
599 return false;
600 }
601 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
602 if (!can_reduce_cmp(n, use)) {
603 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);)
604 return false;
605 }
606 } else {
607 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());)
608 return false;
609 }
610 }
611
612 return true;
613 }
614
615 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
616 // only used in some certain code shapes. Check comments in
617 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
618 // details.
619 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
620 // If there was an error attempting to reduce allocation merges for this
621 // method we might have disabled the compilation and be retrying with RAM
622 // disabled.
623 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
624 return false;
625 }
626
627 const Type* phi_t = _igvn->type(ophi);
628 if (phi_t == nullptr ||
629 phi_t->make_ptr() == nullptr ||
630 phi_t->make_ptr()->isa_aryptr() != nullptr) {
631 return false;
632 }
633
634 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
635 return false;
636 }
637
638 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
639 return true;
640 }
641
642 // This method will return a CmpP/N that we need to use on the If controlling a
643 // CastPP after it was split. This method is only called on bases that are
644 // nullable therefore we always need a controlling if for the splitted CastPP.
645 //
646 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
647 // If the CastPP currently doesn't have a control then the CmpP/N will be
648 // against the NULL constant, otherwise it will be against the constant input of
649 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
650 // case because we have constraints on it and because the CastPP has a control
651 // input.
652 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
653 const Type* t = base->bottom_type();
654 Node* con = nullptr;
655
656 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
657 con = _igvn->zerocon(t->basic_type());
658 } else {
659 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
660 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
661 Node* bol = curr_ctrl->in(0)->in(1);
662 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
663 Node* curr_cmp = bol->in(1);
664 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
665 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
666 }
667
668 return CmpNode::make(base, con, t->basic_type());
669 }
670
671 // This method 'specializes' the CastPP passed as parameter to the base passed
672 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
673 // means that the CastPP now will be specific for a given base instead of a Phi.
674 // An If-Then-Else-Region block is inserted to control the CastPP. The control
675 // of the CastPP is a copy of the current one (if there is one) or a check
676 // against NULL.
677 //
678 // Before:
679 //
680 // C1 C2 ... Cn
681 // \ | /
682 // \ | /
683 // \ | /
684 // \ | /
685 // \ | /
686 // \ | /
687 // \|/
688 // Region B1 B2 ... Bn
689 // | \ | /
690 // | \ | /
691 // | \ | /
692 // | \ | /
693 // | \ | /
694 // | \ | /
695 // ---------------> Phi
696 // |
697 // X |
698 // | |
699 // | |
700 // ------> CastPP
701 //
702 // After (only partial illustration; base = B2, current_control = C2):
703 //
704 // C2
705 // |
706 // If
707 // / \
708 // / \
709 // T F
710 // /\ /
711 // / \ /
712 // / \ /
713 // C1 CastPP Reg Cn
714 // | | |
715 // | | |
716 // | | |
717 // -------------- | ----------
718 // | | |
719 // Region
720 //
721 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
722 Node* control_successor = current_control->unique_ctrl_out();
723 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
724 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
725 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
726 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
727 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
728 Node* end_region = _igvn->transform(new RegionNode(3));
729
730 // Insert the new if-else-region block into the graph
731 end_region->set_req(1, not_eq_control);
732 end_region->set_req(2, yes_eq_control);
733 control_successor->replace_edge(current_control, end_region, _igvn);
734
735 _igvn->_worklist.push(current_control);
736 _igvn->_worklist.push(control_successor);
737
738 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::UnconditionalDependency, nullptr));
739 }
740
741 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
742 const Type* load_type = _igvn->type(curr_load);
743 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
744 Node* memory = curr_load->in(MemNode::Memory);
745
746 // The data_phi merging the loads needs to be nullable if
747 // we are loading pointers.
748 if (load_type->make_ptr() != nullptr) {
749 if (load_type->isa_narrowoop()) {
750 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
751 } else if (load_type->isa_ptr()) {
752 load_type = load_type->meet(TypePtr::NULL_PTR);
753 } else {
754 assert(false, "Unexpected load ptr type.");
755 }
756 }
757
758 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
759
760 for (int i = 1; i < bases_for_loads->length(); i++) {
761 Node* base = bases_for_loads->at(i);
762 Node* cmp_region = nullptr;
763 if (base != nullptr) {
764 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
765 cmp_region = base->unique_ctrl_out_or_null();
766 assert(cmp_region != nullptr, "There should be.");
767 base = base->find_out_with(Op_CastPP);
768 }
769
770 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
771 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
772 Node* load = curr_load->clone();
773 load->set_req(0, nullptr);
774 load->set_req(1, mem);
775 load->set_req(2, addr);
776
777 if (cmp_region != nullptr) { // see comment on previous if
778 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
779 intermediate_phi->set_req(1, _igvn->transform(load));
780 load = intermediate_phi;
781 }
782
783 data_phi->set_req(i, _igvn->transform(load));
784 } else {
785 // Just use the default, which is already in phi
786 }
787 }
788
789 // Takes care of updating CG and split_unique_types worklists due
790 // to cloned AddP->Load.
791 updates_after_load_split(data_phi, curr_load, alloc_worklist);
792
793 return _igvn->transform(data_phi);
794 }
795
796 // This method only reduces CastPP fields loads; SafePoints are handled
797 // separately. The idea here is basically to clone the CastPP and place copies
798 // on each input of the Phi, including non-scalar replaceable inputs.
799 // Experimentation shows that the resulting IR graph is simpler that way than if
800 // we just split the cast through scalar-replaceable inputs.
801 //
802 // The reduction process requires that CastPP's control be one of:
803 // 1) no control,
804 // 2) the same region as Ophi, or
805 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
806 //
807 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
808 // flow. If the CastPP originally had an IfTrue/False control input then we'll
809 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
810 // juse use a CmpP/N against the NULL constant.
811 //
812 // The If-Then-Else-Region isn't always needed. For instance, if input to
813 // splitted cast was not nullable (or if it was the NULL constant) then we don't
814 // need (shouldn't) use a CastPP at all.
815 //
816 // After the casts are splitted we'll split the AddP->Loads through the Phi and
817 // connect them to the just split CastPPs.
818 //
819 // Before (CastPP control is same as Phi):
820 //
821 // Region Allocate Null Call
822 // | \ | /
823 // | \ | /
824 // | \ | /
825 // | \ | /
826 // | \ | /
827 // | \ | /
828 // ------------------> Phi # Oop Phi
829 // | |
830 // | |
831 // | |
832 // | |
833 // ----------------> CastPP
834 // |
835 // AddP
836 // |
837 // Load
838 //
839 // After (Very much simplified):
840 //
841 // Call NULL
842 // \ /
843 // CmpP
844 // |
845 // Bool#NE
846 // |
847 // If
848 // / \
849 // T F
850 // / \ /
851 // / R
852 // CastPP |
853 // | |
854 // AddP |
855 // | |
856 // Load |
857 // \ | 0
858 // Allocate \ | /
859 // \ \ | /
860 // AddP Phi
861 // \ /
862 // Load /
863 // \ 0 /
864 // \ | /
865 // \|/
866 // Phi # "Field" Phi
867 //
868 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
869 Node* ophi = curr_castpp->in(1);
870 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
871
872 // Identify which base should be used for AddP->Load later when spliting the
873 // CastPP->Loads through ophi. Three kind of values may be stored in this
874 // array, depending on the nullability status of the corresponding input in
875 // ophi.
876 //
877 // - nullptr: Meaning that the base is actually the NULL constant and therefore
878 // we won't try to load from it.
879 //
880 // - CFG Node: Meaning that the base is a CastPP that was specialized for
881 // this input of Ophi. I.e., we added an If->Then->Else-Region
882 // that will 'activate' the CastPp only when the input is not Null.
883 //
884 // - Other Node: Meaning that the base is not nullable and therefore we'll try
885 // to load directly from it.
886 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
887
888 for (uint i = 1; i < ophi->req(); i++) {
889 Node* base = ophi->in(i);
890 const Type* base_t = _igvn->type(base);
891
892 if (base_t->maybe_null()) {
893 if (base->is_Con()) {
894 // Nothing todo as bases_for_loads[i] is already nullptr
895 } else {
896 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
897 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
898 }
899 } else {
900 bases_for_loads.at_put(i, base);
901 }
902 }
903
904 // Now let's split the CastPP->Loads through the Phi
905 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
906 Node* use = curr_castpp->raw_out(i);
907 if (use->is_AddP()) {
908 for (int j = use->outcnt()-1; j >= 0;) {
909 Node* use_use = use->raw_out(j);
910 assert(use_use->is_Load(), "Expected this to be a Load node.");
911
912 // We can't make an unconditional load from a nullable input. The
913 // 'split_castpp_load_through_phi` method will add an
914 // 'If-Then-Else-Region` around nullable bases and only load from them
915 // when the input is not null.
916 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
917 _igvn->replace_node(use_use, phi);
918
919 --j;
920 j = MIN2(j, (int)use->outcnt()-1);
921 }
922
923 _igvn->remove_dead_node(use);
924 }
925 --i;
926 i = MIN2(i, (int)curr_castpp->outcnt()-1);
927 }
928 }
929
930 // This method split a given CmpP/N through the Phi used in one of its inputs.
931 // As a result we convert a comparison with a pointer to a comparison with an
932 // integer.
933 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
934 // while the other must be a constant.
935 // The splitting process is basically just cloning the CmpP/N above the input
936 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
937 // can prove at compile time the result of the comparison.
938 //
939 // Before:
940 //
941 // in1 in2 ... inN
942 // \ | /
943 // \ | /
944 // \ | /
945 // \ | /
946 // \ | /
947 // \ | /
948 // Phi
949 // | Other
950 // | /
951 // | /
952 // | /
953 // CmpP/N
954 //
955 // After:
956 //
957 // in1 Other in2 Other inN Other
958 // | | | | | |
959 // \ | | | | |
960 // \ / | / | /
961 // CmpP/N CmpP/N CmpP/N
962 // Bool Bool Bool
963 // \ | /
964 // \ | /
965 // \ | /
966 // \ | /
967 // \ | /
968 // \ | /
969 // \ | /
970 // \ | /
971 // Phi
972 // |
973 // | Zero
974 // | /
975 // | /
976 // | /
977 // CmpI
978 //
979 //
980 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
981 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
982 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
983
984 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
985 Node* zero = _igvn->intcon(0);
986 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
987
988 // This Phi will merge the result of the Cmps split through the Phi
989 Node* res_phi = _igvn->transform(PhiNode::make(ophi->in(0), zero, TypeInt::INT));
990
991 for (uint i=1; i<ophi->req(); i++) {
992 Node* ophi_input = ophi->in(i);
993 Node* res_phi_input = nullptr;
994
995 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
996 if (tcmp->singleton()) {
997 res_phi_input = _igvn->makecon(tcmp);
998 } else {
999 Node* ncmp = _igvn->transform(cmp->clone());
1000 ncmp->set_req(1, ophi_input);
1001 ncmp->set_req(2, other);
1002 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1003 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1004 }
1005
1006 res_phi->set_req(i, res_phi_input);
1007 }
1008
1009 Node* new_cmp = _igvn->transform(new CmpINode(res_phi, zero));
1010 _igvn->replace_node(cmp, new_cmp);
1011 }
1012
1013 // Push the newly created AddP on alloc_worklist and patch
1014 // the connection graph. Note that the changes in the CG below
1015 // won't affect the ES of objects since the new nodes have the
1016 // same status as the old ones.
1017 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1018 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1019 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1020 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1021
1022 if (data_phi == nullptr || !data_phi->is_Phi()) {
1023 // Make this a retry?
1024 return ;
1025 }
1026
1027 Node* previous_addp = previous_load->in(MemNode::Address);
1028 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1029 for (uint i = 1; i < data_phi->req(); i++) {
1030 Node* new_load = data_phi->in(i);
1031
1032 if (new_load->is_Phi()) {
1033 // new_load is currently the "intermediate_phi" from an specialized
1034 // CastPP.
1035 new_load = new_load->in(1);
1036 }
1037
1038 // "new_load" might actually be a constant, parameter, etc.
1039 if (new_load->is_Load()) {
1040 Node* new_addp = new_load->in(MemNode::Address);
1041 Node* base = get_addp_base(new_addp);
1042
1043 // The base might not be something that we can create an unique
1044 // type for. If that's the case we are done with that input.
1045 PointsToNode* jobj_ptn = unique_java_object(base);
1046 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1047 continue;
1048 }
1049
1050 // Push to alloc_worklist since the base has an unique_type
1051 alloc_worklist.append_if_missing(new_addp);
1052
1053 // Now let's add the node to the connection graph
1054 _nodes.at_grow(new_addp->_idx, nullptr);
1055 add_field(new_addp, fn->escape_state(), fn->offset());
1056 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1057
1058 // If the load doesn't load an object then it won't be
1059 // part of the connection graph
1060 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1061 if (curr_load_ptn != nullptr) {
1062 _nodes.at_grow(new_load->_idx, nullptr);
1063 add_local_var(new_load, curr_load_ptn->escape_state());
1064 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1065 }
1066 }
1067 }
1068 }
1069
1070 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1071 // We'll pass this to 'split_through_phi' so that it'll do the split even
1072 // though the load doesn't have an unique instance type.
1073 bool ignore_missing_instance_id = true;
1074
1075 // All AddPs are present in the connection graph
1076 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1077
1078 // Iterate over AddP looking for a Load
1079 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1080 Node* previous_load = previous_addp->raw_out(k);
1081 if (previous_load->is_Load()) {
1082 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1083
1084 // Takes care of updating CG and split_unique_types worklists due to cloned
1085 // AddP->Load.
1086 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1087
1088 _igvn->replace_node(previous_load, data_phi);
1089 }
1090 --k;
1091 k = MIN2(k, (int)previous_addp->outcnt()-1);
1092 }
1093
1094 // Remove the old AddP from the processing list because it's dead now
1095 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1096 alloc_worklist.remove_if_existing(previous_addp);
1097 }
1098
1099 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1100 // selector is:
1101 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1102 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1103 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1104 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1105 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1106 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1107 uint number_of_sr_objects = 0;
1108 for (uint i = 1; i < ophi->req(); i++) {
1109 Node* base = ophi->in(i);
1110 JavaObjectNode* ptn = unique_java_object(base);
1111
1112 if (ptn != nullptr && ptn->scalar_replaceable()) {
1113 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1114 selector->set_req(i, sr_obj_idx);
1115 number_of_sr_objects++;
1116 }
1117 }
1118
1119 return selector->as_Phi();
1120 }
1121
1122 // Returns true if the AddP node 'n' has at least one base that is a reducible
1123 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1124 // checked instead.
1125 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1126 PointsToNode* ptn = ptnode_adr(n->_idx);
1127 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1128 return false;
1129 }
1130
1131 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1132 Node* base = i.get()->ideal_node();
1133
1134 if (reducible_merges.member(base)) {
1135 return true;
1136 }
1137
1138 if (base->is_CastPP() || base->is_CheckCastPP()) {
1139 base = base->in(1);
1140 if (reducible_merges.member(base)) {
1141 return true;
1142 }
1143 }
1144 }
1145
1146 return false;
1147 }
1148
1149 // This method will call its helper method to reduce SafePoint nodes that use
1150 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1151 // "version" of Phi use the same debug information (regarding the Phi).
1152 // Therefore, I collect all safepoints and patch them all at once.
1153 //
1154 // The safepoints using the Phi node have to be processed before safepoints of
1155 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1156 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1157 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1158 // algorithm that process Phi's safepoints will think that the added Phi
1159 // reference is a regular reference.
1160 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1161 PhiNode* selector = create_selector(ophi);
1162 Unique_Node_List safepoints;
1163 Unique_Node_List casts;
1164
1165 // Just collect the users of the Phis for later processing
1166 // in the needed order.
1167 for (uint i = 0; i < ophi->outcnt(); i++) {
1168 Node* use = ophi->raw_out(i);
1169 if (use->is_SafePoint()) {
1170 safepoints.push(use);
1171 } else if (use->is_CastPP()) {
1172 casts.push(use);
1173 } else {
1174 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1175 }
1176 }
1177
1178 // Need to process safepoints using the Phi first
1179 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1180 return false;
1181 }
1182
1183 // Now process CastPP->safepoints
1184 for (uint i = 0; i < casts.size(); i++) {
1185 Node* cast = casts.at(i);
1186 Unique_Node_List cast_sfpts;
1187
1188 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1189 Node* use_use = cast->fast_out(j);
1190 if (use_use->is_SafePoint()) {
1191 cast_sfpts.push(use_use);
1192 } else {
1193 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1194 }
1195 }
1196
1197 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1198 return false;
1199 }
1200 }
1201
1202 return true;
1203 }
1204
1205 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1206 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1207 // SafePointScalarObjectNode for each scalar replaceable input. Each
1208 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1209 // check detailed description in SafePointScalarMergeNode class header.
1210 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1211 PhaseMacroExpand mexp(*_igvn);
1212 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1213 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1214
1215 Node* nsr_merge_pointer = ophi;
1216 if (cast != nullptr) {
1217 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1218 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::RegularDependency, nullptr));
1219 }
1220
1221 for (uint spi = 0; spi < safepoints.size(); spi++) {
1222 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1223 JVMState *jvms = sfpt->jvms();
1224 uint merge_idx = (sfpt->req() - jvms->scloff());
1225 int debug_start = jvms->debug_start();
1226
1227 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1228 smerge->init_req(0, _compile->root());
1229 _igvn->register_new_node_with_optimizer(smerge);
1230
1231 // The next two inputs are:
1232 // (1) A copy of the original pointer to NSR objects.
1233 // (2) A selector, used to decide if we need to rematerialize an object
1234 // or use the pointer to a NSR object.
1235 // See more details of these fields in the declaration of SafePointScalarMergeNode
1236 sfpt->add_req(nsr_merge_pointer);
1237 sfpt->add_req(selector);
1238
1239 for (uint i = 1; i < ophi->req(); i++) {
1240 Node* base = ophi->in(i);
1241 JavaObjectNode* ptn = unique_java_object(base);
1242
1243 // If the base is not scalar replaceable we don't need to register information about
1244 // it at this time.
1245 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1246 continue;
1247 }
1248
1249 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1250 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt);
1251 if (sobj == nullptr) {
1252 return false;
1253 }
1254
1255 // Now make a pass over the debug information replacing any references
1256 // to the allocated object with "sobj"
1257 Node* ccpp = alloc->result_cast();
1258 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1259
1260 // Register the scalarized object as a candidate for reallocation
1261 smerge->add_req(sobj);
1262 }
1263
1264 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1265 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1266
1267 // The call to 'replace_edges_in_range' above might have removed the
1268 // reference to ophi that we need at _merge_pointer_idx. The line below make
1269 // sure the reference is maintained.
1270 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1271 _igvn->_worklist.push(sfpt);
1272 }
1273
1274 return true;
1275 }
1276
1277 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
1278 bool delay = _igvn->delay_transform();
1279 _igvn->set_delay_transform(true);
1280 _igvn->hash_delete(ophi);
1281
1282 // Copying all users first because some will be removed and others won't.
1283 // Ophi also may acquire some new users as part of Cast reduction.
1284 // CastPPs also need to be processed before CmpPs.
1285 Unique_Node_List castpps;
1286 Unique_Node_List others;
1287 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1288 Node* use = ophi->fast_out(i);
1289
1290 if (use->is_CastPP()) {
1291 castpps.push(use);
1292 } else if (use->is_AddP() || use->is_Cmp()) {
1293 others.push(use);
1294 } else if (use->is_SafePoint()) {
1295 // processed later
1296 } else {
1297 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1298 }
1299 }
1300
1301 // CastPPs need to be processed before Cmps because during the process of
1302 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1303 // by the If controlling the CastPP.
1304 for (uint i = 0; i < castpps.size(); i++) {
1305 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist, memnode_worklist);
1306 }
1307
1308 for (uint i = 0; i < others.size(); i++) {
1309 Node* use = others.at(i);
1310
1311 if (use->is_AddP()) {
1312 reduce_phi_on_field_access(use, alloc_worklist);
1313 } else if(use->is_Cmp()) {
1314 reduce_phi_on_cmp(use);
1315 }
1316 }
1317
1318 _igvn->set_delay_transform(delay);
1319 }
1320
1321 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1322 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1323 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1324 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1325 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1326
1327 for (uint i = 1; i < ophi->req(); i++) {
1328 Node* base = ophi->in(i);
1329 JavaObjectNode* ptn = unique_java_object(base);
1330
1331 if (ptn != nullptr && ptn->scalar_replaceable()) {
1332 new_phi->set_req(i, null_ptr);
1333 } else {
1334 new_phi->set_req(i, ophi->in(i));
1335 }
1336 }
1337
1338 for (int i = ophi->outcnt()-1; i >= 0;) {
1339 Node* out = ophi->raw_out(i);
1340
1341 if (out->is_ConstraintCast()) {
1342 const Type* out_t = _igvn->type(out)->make_ptr();
1343 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1344 bool change = out_new_t != out_t;
1345
1346 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1347 Node* out2 = out->raw_out(j);
1348 if (!out2->is_SafePoint()) {
1349 change = false;
1350 break;
1351 }
1352 }
1353
1354 if (change) {
1355 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::StrongDependency, nullptr);
1356 _igvn->replace_node(out, new_cast);
1357 _igvn->register_new_node_with_optimizer(new_cast);
1358 }
1359 }
1360
1361 --i;
1362 i = MIN2(i, (int)ophi->outcnt()-1);
1363 }
1364
1365 _igvn->replace_node(ophi, new_phi);
1366 }
1367
1368 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1369 if (!C->do_reduce_allocation_merges()) return;
1370
1371 Unique_Node_List ideal_nodes;
1372 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1373 ideal_nodes.push(root);
1374
1375 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1376 Node* n = ideal_nodes.at(next);
1377
1378 if (n->is_SafePointScalarMerge()) {
1379 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1380
1381 // Validate inputs of merge
1382 for (uint i = 1; i < merge->req(); i++) {
1383 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1384 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1385 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1386 }
1387 }
1388
1389 // Validate users of merge
1390 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1391 Node* sfpt = merge->fast_out(i);
1392 if (sfpt->is_SafePoint()) {
1393 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1394
1395 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1396 assert(false, "SafePointScalarMerge nodes can't be nested.");
1397 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1398 }
1399 } else {
1400 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1401 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1402 }
1403 }
1404 }
1405
1406 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1407 Node* m = n->fast_out(i);
1408 ideal_nodes.push(m);
1409 }
1410 }
1411 }
1412
1413 // Returns true if there is an object in the scope of sfn that does not escape globally.
1414 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1415 Compile* C = _compile;
1416 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1417 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1418 DeoptimizeObjectsALot) {
1419 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1420 int num_locs = jvms->loc_size();
1421 for (int idx = 0; idx < num_locs; idx++) {
1422 Node* l = sfn->local(jvms, idx);
1423 if (not_global_escape(l)) {
1424 return true;
1425 }
1426 }
1427 }
1428 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1429 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1430 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1431 int num_mon = jvms->nof_monitors();
1432 for (int idx = 0; idx < num_mon; idx++) {
1433 Node* m = sfn->monitor_obj(jvms, idx);
1434 if (m != nullptr && not_global_escape(m)) {
1435 return true;
1436 }
1437 }
1438 }
1439 }
1440 return false;
1441 }
1442
1443 // Returns true if at least one of the arguments to the call is an object
1444 // that does not escape globally.
1445 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1446 if (call->method() != nullptr) {
1447 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1448 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1449 Node* p = call->in(idx);
1450 if (not_global_escape(p)) {
1451 return true;
1452 }
1453 }
1454 } else {
1455 const char* name = call->as_CallStaticJava()->_name;
1456 assert(name != nullptr, "no name");
1457 // no arg escapes through uncommon traps
1458 if (strcmp(name, "uncommon_trap") != 0) {
1459 // process_call_arguments() assumes that all arguments escape globally
1460 const TypeTuple* d = call->tf()->domain();
1461 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1462 const Type* at = d->field_at(i);
1463 if (at->isa_oopptr() != nullptr) {
1464 return true;
1465 }
1466 }
1467 }
1468 }
1469 return false;
1470 }
1471
1472
1473
1474 // Utility function for nodes that load an object
1475 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1476 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1477 // ThreadLocal has RawPtr type.
1478 const Type* t = _igvn->type(n);
1479 if (t->make_ptr() != nullptr) {
1480 Node* adr = n->in(MemNode::Address);
1481 #ifdef ASSERT
1482 if (!adr->is_AddP()) {
1483 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1484 } else {
1485 assert((ptnode_adr(adr->_idx) == nullptr ||
1486 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1487 }
1488 #endif
1489 add_local_var_and_edge(n, PointsToNode::NoEscape,
1490 adr, delayed_worklist);
1491 }
1492 }
1493
1494 // Populate Connection Graph with PointsTo nodes and create simple
1495 // connection graph edges.
1496 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1497 assert(!_verify, "this method should not be called for verification");
1498 PhaseGVN* igvn = _igvn;
1499 uint n_idx = n->_idx;
1500 PointsToNode* n_ptn = ptnode_adr(n_idx);
1501 if (n_ptn != nullptr) {
1502 return; // No need to redefine PointsTo node during first iteration.
1503 }
1504 int opcode = n->Opcode();
1505 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1506 if (gc_handled) {
1507 return; // Ignore node if already handled by GC.
1508 }
1509
1510 if (n->is_Call()) {
1511 // Arguments to allocation and locking don't escape.
1512 if (n->is_AbstractLock()) {
1513 // Put Lock and Unlock nodes on IGVN worklist to process them during
1514 // first IGVN optimization when escape information is still available.
1515 record_for_optimizer(n);
1516 } else if (n->is_Allocate()) {
1517 add_call_node(n->as_Call());
1518 record_for_optimizer(n);
1519 } else {
1520 if (n->is_CallStaticJava()) {
1521 const char* name = n->as_CallStaticJava()->_name;
1522 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1523 return; // Skip uncommon traps
1524 }
1525 }
1526 // Don't mark as processed since call's arguments have to be processed.
1527 delayed_worklist->push(n);
1528 // Check if a call returns an object.
1529 if ((n->as_Call()->returns_pointer() &&
1530 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1531 (n->is_CallStaticJava() &&
1532 n->as_CallStaticJava()->is_boxing_method())) {
1533 add_call_node(n->as_Call());
1534 }
1535 }
1536 return;
1537 }
1538 // Put this check here to process call arguments since some call nodes
1539 // point to phantom_obj.
1540 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1541 return; // Skip predefined nodes.
1542 }
1543 switch (opcode) {
1544 case Op_AddP: {
1545 Node* base = get_addp_base(n);
1546 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1547 // Field nodes are created for all field types. They are used in
1548 // adjust_scalar_replaceable_state() and split_unique_types().
1549 // Note, non-oop fields will have only base edges in Connection
1550 // Graph because such fields are not used for oop loads and stores.
1551 int offset = address_offset(n, igvn);
1552 add_field(n, PointsToNode::NoEscape, offset);
1553 if (ptn_base == nullptr) {
1554 delayed_worklist->push(n); // Process it later.
1555 } else {
1556 n_ptn = ptnode_adr(n_idx);
1557 add_base(n_ptn->as_Field(), ptn_base);
1558 }
1559 break;
1560 }
1561 case Op_CastX2P: {
1562 map_ideal_node(n, phantom_obj);
1563 break;
1564 }
1565 case Op_CastPP:
1566 case Op_CheckCastPP:
1567 case Op_EncodeP:
1568 case Op_DecodeN:
1569 case Op_EncodePKlass:
1570 case Op_DecodeNKlass: {
1571 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1572 break;
1573 }
1574 case Op_CMoveP: {
1575 add_local_var(n, PointsToNode::NoEscape);
1576 // Do not add edges during first iteration because some could be
1577 // not defined yet.
1578 delayed_worklist->push(n);
1579 break;
1580 }
1581 case Op_ConP:
1582 case Op_ConN:
1583 case Op_ConNKlass: {
1584 // assume all oop constants globally escape except for null
1585 PointsToNode::EscapeState es;
1586 const Type* t = igvn->type(n);
1587 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1588 es = PointsToNode::NoEscape;
1589 } else {
1590 es = PointsToNode::GlobalEscape;
1591 }
1592 PointsToNode* ptn_con = add_java_object(n, es);
1593 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1594 break;
1595 }
1596 case Op_CreateEx: {
1597 // assume that all exception objects globally escape
1598 map_ideal_node(n, phantom_obj);
1599 break;
1600 }
1601 case Op_LoadKlass:
1602 case Op_LoadNKlass: {
1603 // Unknown class is loaded
1604 map_ideal_node(n, phantom_obj);
1605 break;
1606 }
1607 case Op_LoadP:
1608 case Op_LoadN: {
1609 add_objload_to_connection_graph(n, delayed_worklist);
1610 break;
1611 }
1612 case Op_Parm: {
1613 map_ideal_node(n, phantom_obj);
1614 break;
1615 }
1616 case Op_PartialSubtypeCheck: {
1617 // Produces Null or notNull and is used in only in CmpP so
1618 // phantom_obj could be used.
1619 map_ideal_node(n, phantom_obj); // Result is unknown
1620 break;
1621 }
1622 case Op_Phi: {
1623 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1624 // ThreadLocal has RawPtr type.
1625 const Type* t = n->as_Phi()->type();
1626 if (t->make_ptr() != nullptr) {
1627 add_local_var(n, PointsToNode::NoEscape);
1628 // Do not add edges during first iteration because some could be
1629 // not defined yet.
1630 delayed_worklist->push(n);
1631 }
1632 break;
1633 }
1634 case Op_Proj: {
1635 // we are only interested in the oop result projection from a call
1636 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1637 n->in(0)->as_Call()->returns_pointer()) {
1638 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1639 }
1640 break;
1641 }
1642 case Op_Rethrow: // Exception object escapes
1643 case Op_Return: {
1644 if (n->req() > TypeFunc::Parms &&
1645 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1646 // Treat Return value as LocalVar with GlobalEscape escape state.
1647 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1648 }
1649 break;
1650 }
1651 case Op_CompareAndExchangeP:
1652 case Op_CompareAndExchangeN:
1653 case Op_GetAndSetP:
1654 case Op_GetAndSetN: {
1655 add_objload_to_connection_graph(n, delayed_worklist);
1656 // fall-through
1657 }
1658 case Op_StoreP:
1659 case Op_StoreN:
1660 case Op_StoreNKlass:
1661 case Op_WeakCompareAndSwapP:
1662 case Op_WeakCompareAndSwapN:
1663 case Op_CompareAndSwapP:
1664 case Op_CompareAndSwapN: {
1665 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1666 break;
1667 }
1668 case Op_AryEq:
1669 case Op_CountPositives:
1670 case Op_StrComp:
1671 case Op_StrEquals:
1672 case Op_StrIndexOf:
1673 case Op_StrIndexOfChar:
1674 case Op_StrInflatedCopy:
1675 case Op_StrCompressedCopy:
1676 case Op_VectorizedHashCode:
1677 case Op_EncodeISOArray: {
1678 add_local_var(n, PointsToNode::ArgEscape);
1679 delayed_worklist->push(n); // Process it later.
1680 break;
1681 }
1682 case Op_ThreadLocal: {
1683 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1684 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1685 break;
1686 }
1687 case Op_Blackhole: {
1688 // All blackhole pointer arguments are globally escaping.
1689 // Only do this if there is at least one pointer argument.
1690 // Do not add edges during first iteration because some could be
1691 // not defined yet, defer to final step.
1692 for (uint i = 0; i < n->req(); i++) {
1693 Node* in = n->in(i);
1694 if (in != nullptr) {
1695 const Type* at = _igvn->type(in);
1696 if (!at->isa_ptr()) continue;
1697
1698 add_local_var(n, PointsToNode::GlobalEscape);
1699 delayed_worklist->push(n);
1700 break;
1701 }
1702 }
1703 break;
1704 }
1705 default:
1706 ; // Do nothing for nodes not related to EA.
1707 }
1708 return;
1709 }
1710
1711 // Add final simple edges to graph.
1712 void ConnectionGraph::add_final_edges(Node *n) {
1713 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1714 #ifdef ASSERT
1715 if (_verify && n_ptn->is_JavaObject())
1716 return; // This method does not change graph for JavaObject.
1717 #endif
1718
1719 if (n->is_Call()) {
1720 process_call_arguments(n->as_Call());
1721 return;
1722 }
1723 assert(n->is_Store() || n->is_LoadStore() ||
1724 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1725 "node should be registered already");
1726 int opcode = n->Opcode();
1727 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1728 if (gc_handled) {
1729 return; // Ignore node if already handled by GC.
1730 }
1731 switch (opcode) {
1732 case Op_AddP: {
1733 Node* base = get_addp_base(n);
1734 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1735 assert(ptn_base != nullptr, "field's base should be registered");
1736 add_base(n_ptn->as_Field(), ptn_base);
1737 break;
1738 }
1739 case Op_CastPP:
1740 case Op_CheckCastPP:
1741 case Op_EncodeP:
1742 case Op_DecodeN:
1743 case Op_EncodePKlass:
1744 case Op_DecodeNKlass: {
1745 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1746 break;
1747 }
1748 case Op_CMoveP: {
1749 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1750 Node* in = n->in(i);
1751 if (in == nullptr) {
1752 continue; // ignore null
1753 }
1754 Node* uncast_in = in->uncast();
1755 if (uncast_in->is_top() || uncast_in == n) {
1756 continue; // ignore top or inputs which go back this node
1757 }
1758 PointsToNode* ptn = ptnode_adr(in->_idx);
1759 assert(ptn != nullptr, "node should be registered");
1760 add_edge(n_ptn, ptn);
1761 }
1762 break;
1763 }
1764 case Op_LoadP:
1765 case Op_LoadN: {
1766 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1767 // ThreadLocal has RawPtr type.
1768 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1769 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1770 break;
1771 }
1772 case Op_Phi: {
1773 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1774 // ThreadLocal has RawPtr type.
1775 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1776 for (uint i = 1; i < n->req(); i++) {
1777 Node* in = n->in(i);
1778 if (in == nullptr) {
1779 continue; // ignore null
1780 }
1781 Node* uncast_in = in->uncast();
1782 if (uncast_in->is_top() || uncast_in == n) {
1783 continue; // ignore top or inputs which go back this node
1784 }
1785 PointsToNode* ptn = ptnode_adr(in->_idx);
1786 assert(ptn != nullptr, "node should be registered");
1787 add_edge(n_ptn, ptn);
1788 }
1789 break;
1790 }
1791 case Op_Proj: {
1792 // we are only interested in the oop result projection from a call
1793 assert(n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1794 n->in(0)->as_Call()->returns_pointer(), "Unexpected node type");
1795 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1796 break;
1797 }
1798 case Op_Rethrow: // Exception object escapes
1799 case Op_Return: {
1800 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1801 "Unexpected node type");
1802 // Treat Return value as LocalVar with GlobalEscape escape state.
1803 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1804 break;
1805 }
1806 case Op_CompareAndExchangeP:
1807 case Op_CompareAndExchangeN:
1808 case Op_GetAndSetP:
1809 case Op_GetAndSetN:{
1810 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1811 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1812 // fall-through
1813 }
1814 case Op_CompareAndSwapP:
1815 case Op_CompareAndSwapN:
1816 case Op_WeakCompareAndSwapP:
1817 case Op_WeakCompareAndSwapN:
1818 case Op_StoreP:
1819 case Op_StoreN:
1820 case Op_StoreNKlass:{
1821 add_final_edges_unsafe_access(n, opcode);
1822 break;
1823 }
1824 case Op_VectorizedHashCode:
1825 case Op_AryEq:
1826 case Op_CountPositives:
1827 case Op_StrComp:
1828 case Op_StrEquals:
1829 case Op_StrIndexOf:
1830 case Op_StrIndexOfChar:
1831 case Op_StrInflatedCopy:
1832 case Op_StrCompressedCopy:
1833 case Op_EncodeISOArray: {
1834 // char[]/byte[] arrays passed to string intrinsic do not escape but
1835 // they are not scalar replaceable. Adjust escape state for them.
1836 // Start from in(2) edge since in(1) is memory edge.
1837 for (uint i = 2; i < n->req(); i++) {
1838 Node* adr = n->in(i);
1839 const Type* at = _igvn->type(adr);
1840 if (!adr->is_top() && at->isa_ptr()) {
1841 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1842 at->isa_ptr() != nullptr, "expecting a pointer");
1843 if (adr->is_AddP()) {
1844 adr = get_addp_base(adr);
1845 }
1846 PointsToNode* ptn = ptnode_adr(adr->_idx);
1847 assert(ptn != nullptr, "node should be registered");
1848 add_edge(n_ptn, ptn);
1849 }
1850 }
1851 break;
1852 }
1853 case Op_Blackhole: {
1854 // All blackhole pointer arguments are globally escaping.
1855 for (uint i = 0; i < n->req(); i++) {
1856 Node* in = n->in(i);
1857 if (in != nullptr) {
1858 const Type* at = _igvn->type(in);
1859 if (!at->isa_ptr()) continue;
1860
1861 if (in->is_AddP()) {
1862 in = get_addp_base(in);
1863 }
1864
1865 PointsToNode* ptn = ptnode_adr(in->_idx);
1866 assert(ptn != nullptr, "should be defined already");
1867 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1868 add_edge(n_ptn, ptn);
1869 }
1870 }
1871 break;
1872 }
1873 default: {
1874 // This method should be called only for EA specific nodes which may
1875 // miss some edges when they were created.
1876 #ifdef ASSERT
1877 n->dump(1);
1878 #endif
1879 guarantee(false, "unknown node");
1880 }
1881 }
1882 return;
1883 }
1884
1885 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1886 Node* adr = n->in(MemNode::Address);
1887 const Type* adr_type = _igvn->type(adr);
1888 adr_type = adr_type->make_ptr();
1889 if (adr_type == nullptr) {
1890 return; // skip dead nodes
1891 }
1892 if (adr_type->isa_oopptr()
1893 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1894 && adr_type == TypeRawPtr::NOTNULL
1895 && is_captured_store_address(adr))) {
1896 delayed_worklist->push(n); // Process it later.
1897 #ifdef ASSERT
1898 assert (adr->is_AddP(), "expecting an AddP");
1899 if (adr_type == TypeRawPtr::NOTNULL) {
1900 // Verify a raw address for a store captured by Initialize node.
1901 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1902 assert(offs != Type::OffsetBot, "offset must be a constant");
1903 }
1904 #endif
1905 } else {
1906 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1907 if (adr->is_BoxLock()) {
1908 return;
1909 }
1910 // Stored value escapes in unsafe access.
1911 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1912 delayed_worklist->push(n); // Process unsafe access later.
1913 return;
1914 }
1915 #ifdef ASSERT
1916 n->dump(1);
1917 assert(false, "not unsafe");
1918 #endif
1919 }
1920 }
1921
1922 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1923 Node* adr = n->in(MemNode::Address);
1924 const Type *adr_type = _igvn->type(adr);
1925 adr_type = adr_type->make_ptr();
1926 #ifdef ASSERT
1927 if (adr_type == nullptr) {
1928 n->dump(1);
1929 assert(adr_type != nullptr, "dead node should not be on list");
1930 return true;
1931 }
1932 #endif
1933
1934 if (adr_type->isa_oopptr()
1935 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1936 && adr_type == TypeRawPtr::NOTNULL
1937 && is_captured_store_address(adr))) {
1938 // Point Address to Value
1939 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1940 assert(adr_ptn != nullptr &&
1941 adr_ptn->as_Field()->is_oop(), "node should be registered");
1942 Node* val = n->in(MemNode::ValueIn);
1943 PointsToNode* ptn = ptnode_adr(val->_idx);
1944 assert(ptn != nullptr, "node should be registered");
1945 add_edge(adr_ptn, ptn);
1946 return true;
1947 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1948 // Stored value escapes in unsafe access.
1949 Node* val = n->in(MemNode::ValueIn);
1950 PointsToNode* ptn = ptnode_adr(val->_idx);
1951 assert(ptn != nullptr, "node should be registered");
1952 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
1953 // Add edge to object for unsafe access with offset.
1954 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1955 assert(adr_ptn != nullptr, "node should be registered");
1956 if (adr_ptn->is_Field()) {
1957 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
1958 add_edge(adr_ptn, ptn);
1959 }
1960 return true;
1961 }
1962 #ifdef ASSERT
1963 n->dump(1);
1964 assert(false, "not unsafe");
1965 #endif
1966 return false;
1967 }
1968
1969 void ConnectionGraph::add_call_node(CallNode* call) {
1970 assert(call->returns_pointer(), "only for call which returns pointer");
1971 uint call_idx = call->_idx;
1972 if (call->is_Allocate()) {
1973 Node* k = call->in(AllocateNode::KlassNode);
1974 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
1975 assert(kt != nullptr, "TypeKlassPtr required.");
1976 PointsToNode::EscapeState es = PointsToNode::NoEscape;
1977 bool scalar_replaceable = true;
1978 NOT_PRODUCT(const char* nsr_reason = "");
1979 if (call->is_AllocateArray()) {
1980 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
1981 es = PointsToNode::GlobalEscape;
1982 } else {
1983 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
1984 if (length < 0) {
1985 // Not scalar replaceable if the length is not constant.
1986 scalar_replaceable = false;
1987 NOT_PRODUCT(nsr_reason = "has a non-constant length");
1988 } else if (length > EliminateAllocationArraySizeLimit) {
1989 // Not scalar replaceable if the length is too big.
1990 scalar_replaceable = false;
1991 NOT_PRODUCT(nsr_reason = "has a length that is too big");
1992 }
1993 }
1994 } else { // Allocate instance
1995 if (!kt->isa_instklassptr()) { // StressReflectiveCode
1996 es = PointsToNode::GlobalEscape;
1997 } else {
1998 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
1999 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2000 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2001 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2002 !ik->can_be_instantiated() ||
2003 ik->has_finalizer()) {
2004 es = PointsToNode::GlobalEscape;
2005 } else {
2006 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2007 if (nfields > EliminateAllocationFieldsLimit) {
2008 // Not scalar replaceable if there are too many fields.
2009 scalar_replaceable = false;
2010 NOT_PRODUCT(nsr_reason = "has too many fields");
2011 }
2012 }
2013 }
2014 }
2015 add_java_object(call, es);
2016 PointsToNode* ptn = ptnode_adr(call_idx);
2017 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2018 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2019 }
2020 } else if (call->is_CallStaticJava()) {
2021 // Call nodes could be different types:
2022 //
2023 // 1. CallDynamicJavaNode (what happened during call is unknown):
2024 //
2025 // - mapped to GlobalEscape JavaObject node if oop is returned;
2026 //
2027 // - all oop arguments are escaping globally;
2028 //
2029 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2030 //
2031 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2032 //
2033 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2034 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2035 // during call is returned;
2036 // - mapped to ArgEscape LocalVar node pointed to object arguments
2037 // which are returned and does not escape during call;
2038 //
2039 // - oop arguments escaping status is defined by bytecode analysis;
2040 //
2041 // For a static call, we know exactly what method is being called.
2042 // Use bytecode estimator to record whether the call's return value escapes.
2043 ciMethod* meth = call->as_CallJava()->method();
2044 if (meth == nullptr) {
2045 const char* name = call->as_CallStaticJava()->_name;
2046 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
2047 // Returns a newly allocated non-escaped object.
2048 add_java_object(call, PointsToNode::NoEscape);
2049 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2050 } else if (meth->is_boxing_method()) {
2051 // Returns boxing object
2052 PointsToNode::EscapeState es;
2053 vmIntrinsics::ID intr = meth->intrinsic_id();
2054 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2055 // It does not escape if object is always allocated.
2056 es = PointsToNode::NoEscape;
2057 } else {
2058 // It escapes globally if object could be loaded from cache.
2059 es = PointsToNode::GlobalEscape;
2060 }
2061 add_java_object(call, es);
2062 if (es == PointsToNode::GlobalEscape) {
2063 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2064 }
2065 } else {
2066 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2067 call_analyzer->copy_dependencies(_compile->dependencies());
2068 if (call_analyzer->is_return_allocated()) {
2069 // Returns a newly allocated non-escaped object, simply
2070 // update dependency information.
2071 // Mark it as NoEscape so that objects referenced by
2072 // it's fields will be marked as NoEscape at least.
2073 add_java_object(call, PointsToNode::NoEscape);
2074 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2075 } else {
2076 // Determine whether any arguments are returned.
2077 const TypeTuple* d = call->tf()->domain();
2078 bool ret_arg = false;
2079 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2080 if (d->field_at(i)->isa_ptr() != nullptr &&
2081 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2082 ret_arg = true;
2083 break;
2084 }
2085 }
2086 if (ret_arg) {
2087 add_local_var(call, PointsToNode::ArgEscape);
2088 } else {
2089 // Returns unknown object.
2090 map_ideal_node(call, phantom_obj);
2091 }
2092 }
2093 }
2094 } else {
2095 // An other type of call, assume the worst case:
2096 // returned value is unknown and globally escapes.
2097 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2098 map_ideal_node(call, phantom_obj);
2099 }
2100 }
2101
2102 void ConnectionGraph::process_call_arguments(CallNode *call) {
2103 bool is_arraycopy = false;
2104 switch (call->Opcode()) {
2105 #ifdef ASSERT
2106 case Op_Allocate:
2107 case Op_AllocateArray:
2108 case Op_Lock:
2109 case Op_Unlock:
2110 assert(false, "should be done already");
2111 break;
2112 #endif
2113 case Op_ArrayCopy:
2114 case Op_CallLeafNoFP:
2115 // Most array copies are ArrayCopy nodes at this point but there
2116 // are still a few direct calls to the copy subroutines (See
2117 // PhaseStringOpts::copy_string())
2118 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2119 call->as_CallLeaf()->is_call_to_arraycopystub();
2120 // fall through
2121 case Op_CallLeafVector:
2122 case Op_CallLeaf: {
2123 // Stub calls, objects do not escape but they are not scale replaceable.
2124 // Adjust escape state for outgoing arguments.
2125 const TypeTuple * d = call->tf()->domain();
2126 bool src_has_oops = false;
2127 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2128 const Type* at = d->field_at(i);
2129 Node *arg = call->in(i);
2130 if (arg == nullptr) {
2131 continue;
2132 }
2133 const Type *aat = _igvn->type(arg);
2134 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2135 continue;
2136 }
2137 if (arg->is_AddP()) {
2138 //
2139 // The inline_native_clone() case when the arraycopy stub is called
2140 // after the allocation before Initialize and CheckCastPP nodes.
2141 // Or normal arraycopy for object arrays case.
2142 //
2143 // Set AddP's base (Allocate) as not scalar replaceable since
2144 // pointer to the base (with offset) is passed as argument.
2145 //
2146 arg = get_addp_base(arg);
2147 }
2148 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2149 assert(arg_ptn != nullptr, "should be registered");
2150 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2151 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2152 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2153 aat->isa_ptr() != nullptr, "expecting an Ptr");
2154 bool arg_has_oops = aat->isa_oopptr() &&
2155 (aat->isa_instptr() ||
2156 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)));
2157 if (i == TypeFunc::Parms) {
2158 src_has_oops = arg_has_oops;
2159 }
2160 //
2161 // src or dst could be j.l.Object when other is basic type array:
2162 //
2163 // arraycopy(char[],0,Object*,0,size);
2164 // arraycopy(Object*,0,char[],0,size);
2165 //
2166 // Don't add edges in such cases.
2167 //
2168 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2169 arg_has_oops && (i > TypeFunc::Parms);
2170 #ifdef ASSERT
2171 if (!(is_arraycopy ||
2172 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2173 (call->as_CallLeaf()->_name != nullptr &&
2174 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2175 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2176 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2177 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2178 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2179 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2180 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2181 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2182 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2183 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2184 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2185 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2186 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2187 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2188 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2189 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2190 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2191 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2192 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2193 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2194 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2195 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2196 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2197 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2198 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2199 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2200 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2201 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2202 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2203 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2204 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2205 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2206 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2207 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2208 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2209 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2210 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2211 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2212 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2213 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2214 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2215 ))) {
2216 call->dump();
2217 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2218 }
2219 #endif
2220 // Always process arraycopy's destination object since
2221 // we need to add all possible edges to references in
2222 // source object.
2223 if (arg_esc >= PointsToNode::ArgEscape &&
2224 !arg_is_arraycopy_dest) {
2225 continue;
2226 }
2227 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2228 if (call->is_ArrayCopy()) {
2229 ArrayCopyNode* ac = call->as_ArrayCopy();
2230 if (ac->is_clonebasic() ||
2231 ac->is_arraycopy_validated() ||
2232 ac->is_copyof_validated() ||
2233 ac->is_copyofrange_validated()) {
2234 es = PointsToNode::NoEscape;
2235 }
2236 }
2237 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2238 if (arg_is_arraycopy_dest) {
2239 Node* src = call->in(TypeFunc::Parms);
2240 if (src->is_AddP()) {
2241 src = get_addp_base(src);
2242 }
2243 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2244 assert(src_ptn != nullptr, "should be registered");
2245 if (arg_ptn != src_ptn) {
2246 // Special arraycopy edge:
2247 // A destination object's field can't have the source object
2248 // as base since objects escape states are not related.
2249 // Only escape state of destination object's fields affects
2250 // escape state of fields in source object.
2251 add_arraycopy(call, es, src_ptn, arg_ptn);
2252 }
2253 }
2254 }
2255 }
2256 break;
2257 }
2258 case Op_CallStaticJava: {
2259 // For a static call, we know exactly what method is being called.
2260 // Use bytecode estimator to record the call's escape affects
2261 #ifdef ASSERT
2262 const char* name = call->as_CallStaticJava()->_name;
2263 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2264 #endif
2265 ciMethod* meth = call->as_CallJava()->method();
2266 if ((meth != nullptr) && meth->is_boxing_method()) {
2267 break; // Boxing methods do not modify any oops.
2268 }
2269 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2270 // fall-through if not a Java method or no analyzer information
2271 if (call_analyzer != nullptr) {
2272 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2273 const TypeTuple* d = call->tf()->domain();
2274 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2275 const Type* at = d->field_at(i);
2276 int k = i - TypeFunc::Parms;
2277 Node* arg = call->in(i);
2278 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2279 if (at->isa_ptr() != nullptr &&
2280 call_analyzer->is_arg_returned(k)) {
2281 // The call returns arguments.
2282 if (call_ptn != nullptr) { // Is call's result used?
2283 assert(call_ptn->is_LocalVar(), "node should be registered");
2284 assert(arg_ptn != nullptr, "node should be registered");
2285 add_edge(call_ptn, arg_ptn);
2286 }
2287 }
2288 if (at->isa_oopptr() != nullptr &&
2289 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2290 if (!call_analyzer->is_arg_stack(k)) {
2291 // The argument global escapes
2292 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2293 } else {
2294 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2295 if (!call_analyzer->is_arg_local(k)) {
2296 // The argument itself doesn't escape, but any fields might
2297 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2298 }
2299 }
2300 }
2301 }
2302 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2303 // The call returns arguments.
2304 assert(call_ptn->edge_count() > 0, "sanity");
2305 if (!call_analyzer->is_return_local()) {
2306 // Returns also unknown object.
2307 add_edge(call_ptn, phantom_obj);
2308 }
2309 }
2310 break;
2311 }
2312 }
2313 default: {
2314 // Fall-through here if not a Java method or no analyzer information
2315 // or some other type of call, assume the worst case: all arguments
2316 // globally escape.
2317 const TypeTuple* d = call->tf()->domain();
2318 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2319 const Type* at = d->field_at(i);
2320 if (at->isa_oopptr() != nullptr) {
2321 Node* arg = call->in(i);
2322 if (arg->is_AddP()) {
2323 arg = get_addp_base(arg);
2324 }
2325 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2326 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2327 }
2328 }
2329 }
2330 }
2331 }
2332
2333
2334 // Finish Graph construction.
2335 bool ConnectionGraph::complete_connection_graph(
2336 GrowableArray<PointsToNode*>& ptnodes_worklist,
2337 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2338 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2339 GrowableArray<FieldNode*>& oop_fields_worklist) {
2340 // Normally only 1-3 passes needed to build Connection Graph depending
2341 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2342 // Set limit to 20 to catch situation when something did go wrong and
2343 // bailout Escape Analysis.
2344 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2345 #define GRAPH_BUILD_ITER_LIMIT 20
2346
2347 // Propagate GlobalEscape and ArgEscape escape states and check that
2348 // we still have non-escaping objects. The method pushs on _worklist
2349 // Field nodes which reference phantom_object.
2350 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2351 return false; // Nothing to do.
2352 }
2353 // Now propagate references to all JavaObject nodes.
2354 int java_objects_length = java_objects_worklist.length();
2355 elapsedTimer build_time;
2356 build_time.start();
2357 elapsedTimer time;
2358 bool timeout = false;
2359 int new_edges = 1;
2360 int iterations = 0;
2361 do {
2362 while ((new_edges > 0) &&
2363 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2364 double start_time = time.seconds();
2365 time.start();
2366 new_edges = 0;
2367 // Propagate references to phantom_object for nodes pushed on _worklist
2368 // by find_non_escaped_objects() and find_field_value().
2369 new_edges += add_java_object_edges(phantom_obj, false);
2370 for (int next = 0; next < java_objects_length; ++next) {
2371 JavaObjectNode* ptn = java_objects_worklist.at(next);
2372 new_edges += add_java_object_edges(ptn, true);
2373
2374 #define SAMPLE_SIZE 4
2375 if ((next % SAMPLE_SIZE) == 0) {
2376 // Each 4 iterations calculate how much time it will take
2377 // to complete graph construction.
2378 time.stop();
2379 // Poll for requests from shutdown mechanism to quiesce compiler
2380 // because Connection graph construction may take long time.
2381 CompileBroker::maybe_block();
2382 double stop_time = time.seconds();
2383 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2384 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2385 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2386 timeout = true;
2387 break; // Timeout
2388 }
2389 start_time = stop_time;
2390 time.start();
2391 }
2392 #undef SAMPLE_SIZE
2393
2394 }
2395 if (timeout) break;
2396 if (new_edges > 0) {
2397 // Update escape states on each iteration if graph was updated.
2398 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2399 return false; // Nothing to do.
2400 }
2401 }
2402 time.stop();
2403 if (time.seconds() >= EscapeAnalysisTimeout) {
2404 timeout = true;
2405 break;
2406 }
2407 }
2408 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2409 time.start();
2410 // Find fields which have unknown value.
2411 int fields_length = oop_fields_worklist.length();
2412 for (int next = 0; next < fields_length; next++) {
2413 FieldNode* field = oop_fields_worklist.at(next);
2414 if (field->edge_count() == 0) {
2415 new_edges += find_field_value(field);
2416 // This code may added new edges to phantom_object.
2417 // Need an other cycle to propagate references to phantom_object.
2418 }
2419 }
2420 time.stop();
2421 if (time.seconds() >= EscapeAnalysisTimeout) {
2422 timeout = true;
2423 break;
2424 }
2425 } else {
2426 new_edges = 0; // Bailout
2427 }
2428 } while (new_edges > 0);
2429
2430 build_time.stop();
2431 _build_time = build_time.seconds();
2432 _build_iterations = iterations;
2433
2434 // Bailout if passed limits.
2435 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2436 Compile* C = _compile;
2437 if (C->log() != nullptr) {
2438 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2439 C->log()->text("%s", timeout ? "time" : "iterations");
2440 C->log()->end_elem(" limit'");
2441 }
2442 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2443 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2444 // Possible infinite build_connection_graph loop,
2445 // bailout (no changes to ideal graph were made).
2446 return false;
2447 }
2448
2449 #undef GRAPH_BUILD_ITER_LIMIT
2450
2451 // Find fields initialized by null for non-escaping Allocations.
2452 int non_escaped_length = non_escaped_allocs_worklist.length();
2453 for (int next = 0; next < non_escaped_length; next++) {
2454 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2455 PointsToNode::EscapeState es = ptn->escape_state();
2456 assert(es <= PointsToNode::ArgEscape, "sanity");
2457 if (es == PointsToNode::NoEscape) {
2458 if (find_init_values_null(ptn, _igvn) > 0) {
2459 // Adding references to null object does not change escape states
2460 // since it does not escape. Also no fields are added to null object.
2461 add_java_object_edges(null_obj, false);
2462 }
2463 }
2464 Node* n = ptn->ideal_node();
2465 if (n->is_Allocate()) {
2466 // The object allocated by this Allocate node will never be
2467 // seen by an other thread. Mark it so that when it is
2468 // expanded no MemBarStoreStore is added.
2469 InitializeNode* ini = n->as_Allocate()->initialization();
2470 if (ini != nullptr)
2471 ini->set_does_not_escape();
2472 }
2473 }
2474 return true; // Finished graph construction.
2475 }
2476
2477 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2478 // and check that we still have non-escaping java objects.
2479 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2480 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist) {
2481 GrowableArray<PointsToNode*> escape_worklist;
2482 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2483 int ptnodes_length = ptnodes_worklist.length();
2484 for (int next = 0; next < ptnodes_length; ++next) {
2485 PointsToNode* ptn = ptnodes_worklist.at(next);
2486 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2487 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2488 escape_worklist.push(ptn);
2489 }
2490 }
2491 // Set escape states to referenced nodes (edges list).
2492 while (escape_worklist.length() > 0) {
2493 PointsToNode* ptn = escape_worklist.pop();
2494 PointsToNode::EscapeState es = ptn->escape_state();
2495 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2496 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2497 es >= PointsToNode::ArgEscape) {
2498 // GlobalEscape or ArgEscape state of field means it has unknown value.
2499 if (add_edge(ptn, phantom_obj)) {
2500 // New edge was added
2501 add_field_uses_to_worklist(ptn->as_Field());
2502 }
2503 }
2504 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2505 PointsToNode* e = i.get();
2506 if (e->is_Arraycopy()) {
2507 assert(ptn->arraycopy_dst(), "sanity");
2508 // Propagate only fields escape state through arraycopy edge.
2509 if (e->fields_escape_state() < field_es) {
2510 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2511 escape_worklist.push(e);
2512 }
2513 } else if (es >= field_es) {
2514 // fields_escape_state is also set to 'es' if it is less than 'es'.
2515 if (e->escape_state() < es) {
2516 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2517 escape_worklist.push(e);
2518 }
2519 } else {
2520 // Propagate field escape state.
2521 bool es_changed = false;
2522 if (e->fields_escape_state() < field_es) {
2523 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2524 es_changed = true;
2525 }
2526 if ((e->escape_state() < field_es) &&
2527 e->is_Field() && ptn->is_JavaObject() &&
2528 e->as_Field()->is_oop()) {
2529 // Change escape state of referenced fields.
2530 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2531 es_changed = true;
2532 } else if (e->escape_state() < es) {
2533 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2534 es_changed = true;
2535 }
2536 if (es_changed) {
2537 escape_worklist.push(e);
2538 }
2539 }
2540 }
2541 }
2542 // Remove escaped objects from non_escaped list.
2543 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2544 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2545 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2546 non_escaped_allocs_worklist.delete_at(next);
2547 }
2548 if (ptn->escape_state() == PointsToNode::NoEscape) {
2549 // Find fields in non-escaped allocations which have unknown value.
2550 find_init_values_phantom(ptn);
2551 }
2552 }
2553 return (non_escaped_allocs_worklist.length() > 0);
2554 }
2555
2556 // Add all references to JavaObject node by walking over all uses.
2557 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2558 int new_edges = 0;
2559 if (populate_worklist) {
2560 // Populate _worklist by uses of jobj's uses.
2561 for (UseIterator i(jobj); i.has_next(); i.next()) {
2562 PointsToNode* use = i.get();
2563 if (use->is_Arraycopy()) {
2564 continue;
2565 }
2566 add_uses_to_worklist(use);
2567 if (use->is_Field() && use->as_Field()->is_oop()) {
2568 // Put on worklist all field's uses (loads) and
2569 // related field nodes (same base and offset).
2570 add_field_uses_to_worklist(use->as_Field());
2571 }
2572 }
2573 }
2574 for (int l = 0; l < _worklist.length(); l++) {
2575 PointsToNode* use = _worklist.at(l);
2576 if (PointsToNode::is_base_use(use)) {
2577 // Add reference from jobj to field and from field to jobj (field's base).
2578 use = PointsToNode::get_use_node(use)->as_Field();
2579 if (add_base(use->as_Field(), jobj)) {
2580 new_edges++;
2581 }
2582 continue;
2583 }
2584 assert(!use->is_JavaObject(), "sanity");
2585 if (use->is_Arraycopy()) {
2586 if (jobj == null_obj) { // null object does not have field edges
2587 continue;
2588 }
2589 // Added edge from Arraycopy node to arraycopy's source java object
2590 if (add_edge(use, jobj)) {
2591 jobj->set_arraycopy_src();
2592 new_edges++;
2593 }
2594 // and stop here.
2595 continue;
2596 }
2597 if (!add_edge(use, jobj)) {
2598 continue; // No new edge added, there was such edge already.
2599 }
2600 new_edges++;
2601 if (use->is_LocalVar()) {
2602 add_uses_to_worklist(use);
2603 if (use->arraycopy_dst()) {
2604 for (EdgeIterator i(use); i.has_next(); i.next()) {
2605 PointsToNode* e = i.get();
2606 if (e->is_Arraycopy()) {
2607 if (jobj == null_obj) { // null object does not have field edges
2608 continue;
2609 }
2610 // Add edge from arraycopy's destination java object to Arraycopy node.
2611 if (add_edge(jobj, e)) {
2612 new_edges++;
2613 jobj->set_arraycopy_dst();
2614 }
2615 }
2616 }
2617 }
2618 } else {
2619 // Added new edge to stored in field values.
2620 // Put on worklist all field's uses (loads) and
2621 // related field nodes (same base and offset).
2622 add_field_uses_to_worklist(use->as_Field());
2623 }
2624 }
2625 _worklist.clear();
2626 _in_worklist.reset();
2627 return new_edges;
2628 }
2629
2630 // Put on worklist all related field nodes.
2631 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2632 assert(field->is_oop(), "sanity");
2633 int offset = field->offset();
2634 add_uses_to_worklist(field);
2635 // Loop over all bases of this field and push on worklist Field nodes
2636 // with the same offset and base (since they may reference the same field).
2637 for (BaseIterator i(field); i.has_next(); i.next()) {
2638 PointsToNode* base = i.get();
2639 add_fields_to_worklist(field, base);
2640 // Check if the base was source object of arraycopy and go over arraycopy's
2641 // destination objects since values stored to a field of source object are
2642 // accessible by uses (loads) of fields of destination objects.
2643 if (base->arraycopy_src()) {
2644 for (UseIterator j(base); j.has_next(); j.next()) {
2645 PointsToNode* arycp = j.get();
2646 if (arycp->is_Arraycopy()) {
2647 for (UseIterator k(arycp); k.has_next(); k.next()) {
2648 PointsToNode* abase = k.get();
2649 if (abase->arraycopy_dst() && abase != base) {
2650 // Look for the same arraycopy reference.
2651 add_fields_to_worklist(field, abase);
2652 }
2653 }
2654 }
2655 }
2656 }
2657 }
2658 }
2659
2660 // Put on worklist all related field nodes.
2661 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2662 int offset = field->offset();
2663 if (base->is_LocalVar()) {
2664 for (UseIterator j(base); j.has_next(); j.next()) {
2665 PointsToNode* f = j.get();
2666 if (PointsToNode::is_base_use(f)) { // Field
2667 f = PointsToNode::get_use_node(f);
2668 if (f == field || !f->as_Field()->is_oop()) {
2669 continue;
2670 }
2671 int offs = f->as_Field()->offset();
2672 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2673 add_to_worklist(f);
2674 }
2675 }
2676 }
2677 } else {
2678 assert(base->is_JavaObject(), "sanity");
2679 if (// Skip phantom_object since it is only used to indicate that
2680 // this field's content globally escapes.
2681 (base != phantom_obj) &&
2682 // null object node does not have fields.
2683 (base != null_obj)) {
2684 for (EdgeIterator i(base); i.has_next(); i.next()) {
2685 PointsToNode* f = i.get();
2686 // Skip arraycopy edge since store to destination object field
2687 // does not update value in source object field.
2688 if (f->is_Arraycopy()) {
2689 assert(base->arraycopy_dst(), "sanity");
2690 continue;
2691 }
2692 if (f == field || !f->as_Field()->is_oop()) {
2693 continue;
2694 }
2695 int offs = f->as_Field()->offset();
2696 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2697 add_to_worklist(f);
2698 }
2699 }
2700 }
2701 }
2702 }
2703
2704 // Find fields which have unknown value.
2705 int ConnectionGraph::find_field_value(FieldNode* field) {
2706 // Escaped fields should have init value already.
2707 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2708 int new_edges = 0;
2709 for (BaseIterator i(field); i.has_next(); i.next()) {
2710 PointsToNode* base = i.get();
2711 if (base->is_JavaObject()) {
2712 // Skip Allocate's fields which will be processed later.
2713 if (base->ideal_node()->is_Allocate()) {
2714 return 0;
2715 }
2716 assert(base == null_obj, "only null ptr base expected here");
2717 }
2718 }
2719 if (add_edge(field, phantom_obj)) {
2720 // New edge was added
2721 new_edges++;
2722 add_field_uses_to_worklist(field);
2723 }
2724 return new_edges;
2725 }
2726
2727 // Find fields initializing values for allocations.
2728 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2729 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2730 Node* alloc = pta->ideal_node();
2731
2732 // Do nothing for Allocate nodes since its fields values are
2733 // "known" unless they are initialized by arraycopy/clone.
2734 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2735 return 0;
2736 }
2737 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
2738 #ifdef ASSERT
2739 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == nullptr) {
2740 const char* name = alloc->as_CallStaticJava()->_name;
2741 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
2742 }
2743 #endif
2744 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2745 int new_edges = 0;
2746 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2747 PointsToNode* field = i.get();
2748 if (field->is_Field() && field->as_Field()->is_oop()) {
2749 if (add_edge(field, phantom_obj)) {
2750 // New edge was added
2751 new_edges++;
2752 add_field_uses_to_worklist(field->as_Field());
2753 }
2754 }
2755 }
2756 return new_edges;
2757 }
2758
2759 // Find fields initializing values for allocations.
2760 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2761 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2762 Node* alloc = pta->ideal_node();
2763 // Do nothing for Call nodes since its fields values are unknown.
2764 if (!alloc->is_Allocate()) {
2765 return 0;
2766 }
2767 InitializeNode* ini = alloc->as_Allocate()->initialization();
2768 bool visited_bottom_offset = false;
2769 GrowableArray<int> offsets_worklist;
2770 int new_edges = 0;
2771
2772 // Check if an oop field's initializing value is recorded and add
2773 // a corresponding null if field's value if it is not recorded.
2774 // Connection Graph does not record a default initialization by null
2775 // captured by Initialize node.
2776 //
2777 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2778 PointsToNode* field = i.get(); // Field (AddP)
2779 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2780 continue; // Not oop field
2781 }
2782 int offset = field->as_Field()->offset();
2783 if (offset == Type::OffsetBot) {
2784 if (!visited_bottom_offset) {
2785 // OffsetBot is used to reference array's element,
2786 // always add reference to null to all Field nodes since we don't
2787 // known which element is referenced.
2788 if (add_edge(field, null_obj)) {
2789 // New edge was added
2790 new_edges++;
2791 add_field_uses_to_worklist(field->as_Field());
2792 visited_bottom_offset = true;
2793 }
2794 }
2795 } else {
2796 // Check only oop fields.
2797 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2798 if (adr_type->isa_rawptr()) {
2799 #ifdef ASSERT
2800 // Raw pointers are used for initializing stores so skip it
2801 // since it should be recorded already
2802 Node* base = get_addp_base(field->ideal_node());
2803 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2804 #endif
2805 continue;
2806 }
2807 if (!offsets_worklist.contains(offset)) {
2808 offsets_worklist.append(offset);
2809 Node* value = nullptr;
2810 if (ini != nullptr) {
2811 // StoreP::memory_type() == T_ADDRESS
2812 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2813 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2814 // Make sure initializing store has the same type as this AddP.
2815 // This AddP may reference non existing field because it is on a
2816 // dead branch of bimorphic call which is not eliminated yet.
2817 if (store != nullptr && store->is_Store() &&
2818 store->as_Store()->memory_type() == ft) {
2819 value = store->in(MemNode::ValueIn);
2820 #ifdef ASSERT
2821 if (VerifyConnectionGraph) {
2822 // Verify that AddP already points to all objects the value points to.
2823 PointsToNode* val = ptnode_adr(value->_idx);
2824 assert((val != nullptr), "should be processed already");
2825 PointsToNode* missed_obj = nullptr;
2826 if (val->is_JavaObject()) {
2827 if (!field->points_to(val->as_JavaObject())) {
2828 missed_obj = val;
2829 }
2830 } else {
2831 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2832 tty->print_cr("----------init store has invalid value -----");
2833 store->dump();
2834 val->dump();
2835 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2836 }
2837 for (EdgeIterator j(val); j.has_next(); j.next()) {
2838 PointsToNode* obj = j.get();
2839 if (obj->is_JavaObject()) {
2840 if (!field->points_to(obj->as_JavaObject())) {
2841 missed_obj = obj;
2842 break;
2843 }
2844 }
2845 }
2846 }
2847 if (missed_obj != nullptr) {
2848 tty->print_cr("----------field---------------------------------");
2849 field->dump();
2850 tty->print_cr("----------missed referernce to object-----------");
2851 missed_obj->dump();
2852 tty->print_cr("----------object referernced by init store -----");
2853 store->dump();
2854 val->dump();
2855 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2856 }
2857 }
2858 #endif
2859 } else {
2860 // There could be initializing stores which follow allocation.
2861 // For example, a volatile field store is not collected
2862 // by Initialize node.
2863 //
2864 // Need to check for dependent loads to separate such stores from
2865 // stores which follow loads. For now, add initial value null so
2866 // that compare pointers optimization works correctly.
2867 }
2868 }
2869 if (value == nullptr) {
2870 // A field's initializing value was not recorded. Add null.
2871 if (add_edge(field, null_obj)) {
2872 // New edge was added
2873 new_edges++;
2874 add_field_uses_to_worklist(field->as_Field());
2875 }
2876 }
2877 }
2878 }
2879 }
2880 return new_edges;
2881 }
2882
2883 // Adjust scalar_replaceable state after Connection Graph is built.
2884 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2885 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2886 // returns true. If one of the constraints in this method set 'jobj' to NSR
2887 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2888 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2889 // that other object and the Phi will become a reducible Phi.
2890 // There could be multiple merges involving the same jobj.
2891 Unique_Node_List candidates;
2892
2893 // Search for non-escaping objects which are not scalar replaceable
2894 // and mark them to propagate the state to referenced objects.
2895
2896 for (UseIterator i(jobj); i.has_next(); i.next()) {
2897 PointsToNode* use = i.get();
2898 if (use->is_Arraycopy()) {
2899 continue;
2900 }
2901 if (use->is_Field()) {
2902 FieldNode* field = use->as_Field();
2903 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2904 // 1. An object is not scalar replaceable if the field into which it is
2905 // stored has unknown offset (stored into unknown element of an array).
2906 if (field->offset() == Type::OffsetBot) {
2907 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2908 return;
2909 }
2910 for (BaseIterator i(field); i.has_next(); i.next()) {
2911 PointsToNode* base = i.get();
2912 // 2. An object is not scalar replaceable if the field into which it is
2913 // stored has multiple bases one of which is null.
2914 if ((base == null_obj) && (field->base_count() > 1)) {
2915 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2916 return;
2917 }
2918 // 2.5. An object is not scalar replaceable if the field into which it is
2919 // stored has NSR base.
2920 if (!base->scalar_replaceable()) {
2921 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2922 return;
2923 }
2924 }
2925 }
2926 assert(use->is_Field() || use->is_LocalVar(), "sanity");
2927 // 3. An object is not scalar replaceable if it is merged with other objects
2928 // and we can't remove the merge
2929 for (EdgeIterator j(use); j.has_next(); j.next()) {
2930 PointsToNode* ptn = j.get();
2931 if (ptn->is_JavaObject() && ptn != jobj) {
2932 Node* use_n = use->ideal_node();
2933
2934 // These other local vars may point to multiple objects through a Phi
2935 // In this case we skip them and see if we can reduce the Phi.
2936 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
2937 use_n = use_n->in(1);
2938 }
2939
2940 // If it's already a candidate or confirmed reducible merge we can skip verification
2941 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
2942 continue;
2943 }
2944
2945 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
2946 candidates.push(use_n);
2947 } else {
2948 // Mark all objects as NSR if we can't remove the merge
2949 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
2950 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
2951 }
2952 }
2953 }
2954 if (!jobj->scalar_replaceable()) {
2955 return;
2956 }
2957 }
2958
2959 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
2960 if (j.get()->is_Arraycopy()) {
2961 continue;
2962 }
2963
2964 // Non-escaping object node should point only to field nodes.
2965 FieldNode* field = j.get()->as_Field();
2966 int offset = field->as_Field()->offset();
2967
2968 // 4. An object is not scalar replaceable if it has a field with unknown
2969 // offset (array's element is accessed in loop).
2970 if (offset == Type::OffsetBot) {
2971 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
2972 return;
2973 }
2974 // 5. Currently an object is not scalar replaceable if a LoadStore node
2975 // access its field since the field value is unknown after it.
2976 //
2977 Node* n = field->ideal_node();
2978
2979 // Test for an unsafe access that was parsed as maybe off heap
2980 // (with a CheckCastPP to raw memory).
2981 assert(n->is_AddP(), "expect an address computation");
2982 if (n->in(AddPNode::Base)->is_top() &&
2983 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
2984 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
2985 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
2986 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
2987 return;
2988 }
2989
2990 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2991 Node* u = n->fast_out(i);
2992 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
2993 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
2994 return;
2995 }
2996 }
2997
2998 // 6. Or the address may point to more then one object. This may produce
2999 // the false positive result (set not scalar replaceable)
3000 // since the flow-insensitive escape analysis can't separate
3001 // the case when stores overwrite the field's value from the case
3002 // when stores happened on different control branches.
3003 //
3004 // Note: it will disable scalar replacement in some cases:
3005 //
3006 // Point p[] = new Point[1];
3007 // p[0] = new Point(); // Will be not scalar replaced
3008 //
3009 // but it will save us from incorrect optimizations in next cases:
3010 //
3011 // Point p[] = new Point[1];
3012 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3013 //
3014 if (field->base_count() > 1 && candidates.size() == 0) {
3015 if (has_non_reducible_merge(field, reducible_merges)) {
3016 for (BaseIterator i(field); i.has_next(); i.next()) {
3017 PointsToNode* base = i.get();
3018 // Don't take into account LocalVar nodes which
3019 // may point to only one object which should be also
3020 // this field's base by now.
3021 if (base->is_JavaObject() && base != jobj) {
3022 // Mark all bases.
3023 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3024 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3025 }
3026 }
3027
3028 if (!jobj->scalar_replaceable()) {
3029 return;
3030 }
3031 }
3032 }
3033 }
3034
3035 // The candidate is truly a reducible merge only if none of the other
3036 // constraints ruled it as NSR. There could be multiple merges involving the
3037 // same jobj.
3038 assert(jobj->scalar_replaceable(), "sanity");
3039 for (uint i = 0; i < candidates.size(); i++ ) {
3040 Node* candidate = candidates.at(i);
3041 reducible_merges.push(candidate);
3042 }
3043 }
3044
3045 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3046 for (BaseIterator i(field); i.has_next(); i.next()) {
3047 Node* base = i.get()->ideal_node();
3048 if (base->is_Phi() && !reducible_merges.member(base)) {
3049 return true;
3050 }
3051 }
3052 return false;
3053 }
3054
3055 // Propagate NSR (Not scalar replaceable) state.
3056 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist) {
3057 int jobj_length = jobj_worklist.length();
3058 bool found_nsr_alloc = true;
3059 while (found_nsr_alloc) {
3060 found_nsr_alloc = false;
3061 for (int next = 0; next < jobj_length; ++next) {
3062 JavaObjectNode* jobj = jobj_worklist.at(next);
3063 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3064 PointsToNode* use = i.get();
3065 if (use->is_Field()) {
3066 FieldNode* field = use->as_Field();
3067 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3068 assert(field->offset() != Type::OffsetBot, "sanity");
3069 for (BaseIterator i(field); i.has_next(); i.next()) {
3070 PointsToNode* base = i.get();
3071 // An object is not scalar replaceable if the field into which
3072 // it is stored has NSR base.
3073 if ((base != null_obj) && !base->scalar_replaceable()) {
3074 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3075 found_nsr_alloc = true;
3076 break;
3077 }
3078 }
3079 }
3080 }
3081 }
3082 }
3083 }
3084
3085 #ifdef ASSERT
3086 void ConnectionGraph::verify_connection_graph(
3087 GrowableArray<PointsToNode*>& ptnodes_worklist,
3088 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3089 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3090 GrowableArray<Node*>& addp_worklist) {
3091 // Verify that graph is complete - no new edges could be added.
3092 int java_objects_length = java_objects_worklist.length();
3093 int non_escaped_length = non_escaped_allocs_worklist.length();
3094 int new_edges = 0;
3095 for (int next = 0; next < java_objects_length; ++next) {
3096 JavaObjectNode* ptn = java_objects_worklist.at(next);
3097 new_edges += add_java_object_edges(ptn, true);
3098 }
3099 assert(new_edges == 0, "graph was not complete");
3100 // Verify that escape state is final.
3101 int length = non_escaped_allocs_worklist.length();
3102 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist);
3103 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3104 (non_escaped_length == length) &&
3105 (_worklist.length() == 0), "escape state was not final");
3106
3107 // Verify fields information.
3108 int addp_length = addp_worklist.length();
3109 for (int next = 0; next < addp_length; ++next ) {
3110 Node* n = addp_worklist.at(next);
3111 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3112 if (field->is_oop()) {
3113 // Verify that field has all bases
3114 Node* base = get_addp_base(n);
3115 PointsToNode* ptn = ptnode_adr(base->_idx);
3116 if (ptn->is_JavaObject()) {
3117 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3118 } else {
3119 assert(ptn->is_LocalVar(), "sanity");
3120 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3121 PointsToNode* e = i.get();
3122 if (e->is_JavaObject()) {
3123 assert(field->has_base(e->as_JavaObject()), "sanity");
3124 }
3125 }
3126 }
3127 // Verify that all fields have initializing values.
3128 if (field->edge_count() == 0) {
3129 tty->print_cr("----------field does not have references----------");
3130 field->dump();
3131 for (BaseIterator i(field); i.has_next(); i.next()) {
3132 PointsToNode* base = i.get();
3133 tty->print_cr("----------field has next base---------------------");
3134 base->dump();
3135 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3136 tty->print_cr("----------base has fields-------------------------");
3137 for (EdgeIterator j(base); j.has_next(); j.next()) {
3138 j.get()->dump();
3139 }
3140 tty->print_cr("----------base has references---------------------");
3141 for (UseIterator j(base); j.has_next(); j.next()) {
3142 j.get()->dump();
3143 }
3144 }
3145 }
3146 for (UseIterator i(field); i.has_next(); i.next()) {
3147 i.get()->dump();
3148 }
3149 assert(field->edge_count() > 0, "sanity");
3150 }
3151 }
3152 }
3153 }
3154 #endif
3155
3156 // Optimize ideal graph.
3157 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3158 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3159 Compile* C = _compile;
3160 PhaseIterGVN* igvn = _igvn;
3161 if (EliminateLocks) {
3162 // Mark locks before changing ideal graph.
3163 int cnt = C->macro_count();
3164 for (int i = 0; i < cnt; i++) {
3165 Node *n = C->macro_node(i);
3166 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3167 AbstractLockNode* alock = n->as_AbstractLock();
3168 if (!alock->is_non_esc_obj()) {
3169 if (can_eliminate_lock(alock)) {
3170 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3171 // The lock could be marked eliminated by lock coarsening
3172 // code during first IGVN before EA. Replace coarsened flag
3173 // to eliminate all associated locks/unlocks.
3174 #ifdef ASSERT
3175 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3176 #endif
3177 alock->set_non_esc_obj();
3178 }
3179 }
3180 }
3181 }
3182 }
3183
3184 if (OptimizePtrCompare) {
3185 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3186 Node *n = ptr_cmp_worklist.at(i);
3187 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3188 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3189 if (tcmp->singleton()) {
3190 Node* cmp = igvn->makecon(tcmp);
3191 #ifndef PRODUCT
3192 if (PrintOptimizePtrCompare) {
3193 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"));
3194 if (Verbose) {
3195 n->dump(1);
3196 }
3197 }
3198 #endif
3199 igvn->replace_node(n, cmp);
3200 }
3201 }
3202 }
3203
3204 // For MemBarStoreStore nodes added in library_call.cpp, check
3205 // escape status of associated AllocateNode and optimize out
3206 // MemBarStoreStore node if the allocated object never escapes.
3207 for (int i = 0; i < storestore_worklist.length(); i++) {
3208 Node* storestore = storestore_worklist.at(i);
3209 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3210 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3211 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3212 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3213 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3214 igvn->register_new_node_with_optimizer(mb);
3215 igvn->replace_node(storestore, mb);
3216 }
3217 }
3218 }
3219
3220 // Optimize objects compare.
3221 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3222 assert(OptimizePtrCompare, "sanity");
3223 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3224 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3225 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3226
3227 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3228 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3229 JavaObjectNode* jobj1 = unique_java_object(left);
3230 JavaObjectNode* jobj2 = unique_java_object(right);
3231
3232 // The use of this method during allocation merge reduction may cause 'left'
3233 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3234 // that doesn't reference an unique java object.
3235 if (ptn1 == nullptr || ptn2 == nullptr ||
3236 jobj1 == nullptr || jobj2 == nullptr) {
3237 return UNKNOWN;
3238 }
3239
3240 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3241 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3242
3243 // Check simple cases first.
3244 if (jobj1 != nullptr) {
3245 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3246 if (jobj1 == jobj2) {
3247 // Comparing the same not escaping object.
3248 return EQ;
3249 }
3250 Node* obj = jobj1->ideal_node();
3251 // Comparing not escaping allocation.
3252 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3253 !ptn2->points_to(jobj1)) {
3254 return NE; // This includes nullness check.
3255 }
3256 }
3257 }
3258 if (jobj2 != nullptr) {
3259 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3260 Node* obj = jobj2->ideal_node();
3261 // Comparing not escaping allocation.
3262 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3263 !ptn1->points_to(jobj2)) {
3264 return NE; // This includes nullness check.
3265 }
3266 }
3267 }
3268 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3269 jobj2 != nullptr && jobj2 != phantom_obj &&
3270 jobj1->ideal_node()->is_Con() &&
3271 jobj2->ideal_node()->is_Con()) {
3272 // Klass or String constants compare. Need to be careful with
3273 // compressed pointers - compare types of ConN and ConP instead of nodes.
3274 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3275 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3276 if (t1->make_ptr() == t2->make_ptr()) {
3277 return EQ;
3278 } else {
3279 return NE;
3280 }
3281 }
3282 if (ptn1->meet(ptn2)) {
3283 return UNKNOWN; // Sets are not disjoint
3284 }
3285
3286 // Sets are disjoint.
3287 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3288 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3289 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3290 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3291 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3292 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3293 // Check nullness of unknown object.
3294 return UNKNOWN;
3295 }
3296
3297 // Disjointness by itself is not sufficient since
3298 // alias analysis is not complete for escaped objects.
3299 // Disjoint sets are definitely unrelated only when
3300 // at least one set has only not escaping allocations.
3301 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3302 if (ptn1->non_escaping_allocation()) {
3303 return NE;
3304 }
3305 }
3306 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3307 if (ptn2->non_escaping_allocation()) {
3308 return NE;
3309 }
3310 }
3311 return UNKNOWN;
3312 }
3313
3314 // Connection Graph construction functions.
3315
3316 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3317 PointsToNode* ptadr = _nodes.at(n->_idx);
3318 if (ptadr != nullptr) {
3319 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3320 return;
3321 }
3322 Compile* C = _compile;
3323 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3324 map_ideal_node(n, ptadr);
3325 }
3326
3327 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3328 PointsToNode* ptadr = _nodes.at(n->_idx);
3329 if (ptadr != nullptr) {
3330 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3331 return ptadr;
3332 }
3333 Compile* C = _compile;
3334 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3335 map_ideal_node(n, ptadr);
3336 return ptadr;
3337 }
3338
3339 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3340 PointsToNode* ptadr = _nodes.at(n->_idx);
3341 if (ptadr != nullptr) {
3342 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3343 return;
3344 }
3345 bool unsafe = false;
3346 bool is_oop = is_oop_field(n, offset, &unsafe);
3347 if (unsafe) {
3348 es = PointsToNode::GlobalEscape;
3349 }
3350 Compile* C = _compile;
3351 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3352 map_ideal_node(n, field);
3353 }
3354
3355 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3356 PointsToNode* src, PointsToNode* dst) {
3357 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3358 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3359 PointsToNode* ptadr = _nodes.at(n->_idx);
3360 if (ptadr != nullptr) {
3361 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3362 return;
3363 }
3364 Compile* C = _compile;
3365 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3366 map_ideal_node(n, ptadr);
3367 // Add edge from arraycopy node to source object.
3368 (void)add_edge(ptadr, src);
3369 src->set_arraycopy_src();
3370 // Add edge from destination object to arraycopy node.
3371 (void)add_edge(dst, ptadr);
3372 dst->set_arraycopy_dst();
3373 }
3374
3375 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3376 const Type* adr_type = n->as_AddP()->bottom_type();
3377 BasicType bt = T_INT;
3378 if (offset == Type::OffsetBot) {
3379 // Check only oop fields.
3380 if (!adr_type->isa_aryptr() ||
3381 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3382 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3383 // OffsetBot is used to reference array's element. Ignore first AddP.
3384 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3385 bt = T_OBJECT;
3386 }
3387 }
3388 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3389 if (adr_type->isa_instptr()) {
3390 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
3391 if (field != nullptr) {
3392 bt = field->layout_type();
3393 } else {
3394 // Check for unsafe oop field access
3395 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3396 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3397 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3398 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3399 bt = T_OBJECT;
3400 (*unsafe) = true;
3401 }
3402 }
3403 } else if (adr_type->isa_aryptr()) {
3404 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3405 // Ignore array length load.
3406 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3407 // Ignore first AddP.
3408 } else {
3409 const Type* elemtype = adr_type->isa_aryptr()->elem();
3410 bt = elemtype->array_element_basic_type();
3411 }
3412 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3413 // Allocation initialization, ThreadLocal field access, unsafe access
3414 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3415 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3416 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3417 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3418 bt = T_OBJECT;
3419 }
3420 }
3421 }
3422 // Note: T_NARROWOOP is not classed as a real reference type
3423 return (is_reference_type(bt) || bt == T_NARROWOOP);
3424 }
3425
3426 // Returns unique pointed java object or null.
3427 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3428 // If the node was created after the escape computation we can't answer.
3429 uint idx = n->_idx;
3430 if (idx >= nodes_size()) {
3431 return nullptr;
3432 }
3433 PointsToNode* ptn = ptnode_adr(idx);
3434 if (ptn == nullptr) {
3435 return nullptr;
3436 }
3437 if (ptn->is_JavaObject()) {
3438 return ptn->as_JavaObject();
3439 }
3440 assert(ptn->is_LocalVar(), "sanity");
3441 // Check all java objects it points to.
3442 JavaObjectNode* jobj = nullptr;
3443 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3444 PointsToNode* e = i.get();
3445 if (e->is_JavaObject()) {
3446 if (jobj == nullptr) {
3447 jobj = e->as_JavaObject();
3448 } else if (jobj != e) {
3449 return nullptr;
3450 }
3451 }
3452 }
3453 return jobj;
3454 }
3455
3456 // Return true if this node points only to non-escaping allocations.
3457 bool PointsToNode::non_escaping_allocation() {
3458 if (is_JavaObject()) {
3459 Node* n = ideal_node();
3460 if (n->is_Allocate() || n->is_CallStaticJava()) {
3461 return (escape_state() == PointsToNode::NoEscape);
3462 } else {
3463 return false;
3464 }
3465 }
3466 assert(is_LocalVar(), "sanity");
3467 // Check all java objects it points to.
3468 for (EdgeIterator i(this); i.has_next(); i.next()) {
3469 PointsToNode* e = i.get();
3470 if (e->is_JavaObject()) {
3471 Node* n = e->ideal_node();
3472 if ((e->escape_state() != PointsToNode::NoEscape) ||
3473 !(n->is_Allocate() || n->is_CallStaticJava())) {
3474 return false;
3475 }
3476 }
3477 }
3478 return true;
3479 }
3480
3481 // Return true if we know the node does not escape globally.
3482 bool ConnectionGraph::not_global_escape(Node *n) {
3483 assert(!_collecting, "should not call during graph construction");
3484 // If the node was created after the escape computation we can't answer.
3485 uint idx = n->_idx;
3486 if (idx >= nodes_size()) {
3487 return false;
3488 }
3489 PointsToNode* ptn = ptnode_adr(idx);
3490 if (ptn == nullptr) {
3491 return false; // not in congraph (e.g. ConI)
3492 }
3493 PointsToNode::EscapeState es = ptn->escape_state();
3494 // If we have already computed a value, return it.
3495 if (es >= PointsToNode::GlobalEscape) {
3496 return false;
3497 }
3498 if (ptn->is_JavaObject()) {
3499 return true; // (es < PointsToNode::GlobalEscape);
3500 }
3501 assert(ptn->is_LocalVar(), "sanity");
3502 // Check all java objects it points to.
3503 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3504 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3505 return false;
3506 }
3507 }
3508 return true;
3509 }
3510
3511 // Return true if locked object does not escape globally
3512 // and locked code region (identified by BoxLockNode) is balanced:
3513 // all compiled code paths have corresponding Lock/Unlock pairs.
3514 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3515 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3516 if (EliminateNestedLocks) {
3517 // We can mark whole locking region as Local only when only
3518 // one object is used for locking.
3519 alock->box_node()->as_BoxLock()->set_local();
3520 }
3521 return true;
3522 }
3523 return false;
3524 }
3525
3526 // Helper functions
3527
3528 // Return true if this node points to specified node or nodes it points to.
3529 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3530 if (is_JavaObject()) {
3531 return (this == ptn);
3532 }
3533 assert(is_LocalVar() || is_Field(), "sanity");
3534 for (EdgeIterator i(this); i.has_next(); i.next()) {
3535 if (i.get() == ptn) {
3536 return true;
3537 }
3538 }
3539 return false;
3540 }
3541
3542 // Return true if one node points to an other.
3543 bool PointsToNode::meet(PointsToNode* ptn) {
3544 if (this == ptn) {
3545 return true;
3546 } else if (ptn->is_JavaObject()) {
3547 return this->points_to(ptn->as_JavaObject());
3548 } else if (this->is_JavaObject()) {
3549 return ptn->points_to(this->as_JavaObject());
3550 }
3551 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3552 int ptn_count = ptn->edge_count();
3553 for (EdgeIterator i(this); i.has_next(); i.next()) {
3554 PointsToNode* this_e = i.get();
3555 for (int j = 0; j < ptn_count; j++) {
3556 if (this_e == ptn->edge(j)) {
3557 return true;
3558 }
3559 }
3560 }
3561 return false;
3562 }
3563
3564 #ifdef ASSERT
3565 // Return true if bases point to this java object.
3566 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3567 for (BaseIterator i(this); i.has_next(); i.next()) {
3568 if (i.get() == jobj) {
3569 return true;
3570 }
3571 }
3572 return false;
3573 }
3574 #endif
3575
3576 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3577 // Handle simple case first.
3578 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3579 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3580 return true;
3581 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3582 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3583 Node* addp_use = addp->fast_out(i);
3584 if (addp_use->is_Store()) {
3585 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3586 if (addp_use->fast_out(j)->is_Initialize()) {
3587 return true;
3588 }
3589 }
3590 }
3591 }
3592 }
3593 return false;
3594 }
3595
3596 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3597 const Type *adr_type = phase->type(adr);
3598 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3599 // We are computing a raw address for a store captured by an Initialize
3600 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3601 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3602 assert(offs != Type::OffsetBot ||
3603 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3604 "offset must be a constant or it is initialization of array");
3605 return offs;
3606 }
3607 const TypePtr *t_ptr = adr_type->isa_ptr();
3608 assert(t_ptr != nullptr, "must be a pointer type");
3609 return t_ptr->offset();
3610 }
3611
3612 Node* ConnectionGraph::get_addp_base(Node *addp) {
3613 assert(addp->is_AddP(), "must be AddP");
3614 //
3615 // AddP cases for Base and Address inputs:
3616 // case #1. Direct object's field reference:
3617 // Allocate
3618 // |
3619 // Proj #5 ( oop result )
3620 // |
3621 // CheckCastPP (cast to instance type)
3622 // | |
3623 // AddP ( base == address )
3624 //
3625 // case #2. Indirect object's field reference:
3626 // Phi
3627 // |
3628 // CastPP (cast to instance type)
3629 // | |
3630 // AddP ( base == address )
3631 //
3632 // case #3. Raw object's field reference for Initialize node:
3633 // Allocate
3634 // |
3635 // Proj #5 ( oop result )
3636 // top |
3637 // \ |
3638 // AddP ( base == top )
3639 //
3640 // case #4. Array's element reference:
3641 // {CheckCastPP | CastPP}
3642 // | | |
3643 // | AddP ( array's element offset )
3644 // | |
3645 // AddP ( array's offset )
3646 //
3647 // case #5. Raw object's field reference for arraycopy stub call:
3648 // The inline_native_clone() case when the arraycopy stub is called
3649 // after the allocation before Initialize and CheckCastPP nodes.
3650 // Allocate
3651 // |
3652 // Proj #5 ( oop result )
3653 // | |
3654 // AddP ( base == address )
3655 //
3656 // case #6. Constant Pool, ThreadLocal, CastX2P or
3657 // Raw object's field reference:
3658 // {ConP, ThreadLocal, CastX2P, raw Load}
3659 // top |
3660 // \ |
3661 // AddP ( base == top )
3662 //
3663 // case #7. Klass's field reference.
3664 // LoadKlass
3665 // | |
3666 // AddP ( base == address )
3667 //
3668 // case #8. narrow Klass's field reference.
3669 // LoadNKlass
3670 // |
3671 // DecodeN
3672 // | |
3673 // AddP ( base == address )
3674 //
3675 // case #9. Mixed unsafe access
3676 // {instance}
3677 // |
3678 // CheckCastPP (raw)
3679 // top |
3680 // \ |
3681 // AddP ( base == top )
3682 //
3683 Node *base = addp->in(AddPNode::Base);
3684 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3685 base = addp->in(AddPNode::Address);
3686 while (base->is_AddP()) {
3687 // Case #6 (unsafe access) may have several chained AddP nodes.
3688 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3689 base = base->in(AddPNode::Address);
3690 }
3691 if (base->Opcode() == Op_CheckCastPP &&
3692 base->bottom_type()->isa_rawptr() &&
3693 _igvn->type(base->in(1))->isa_oopptr()) {
3694 base = base->in(1); // Case #9
3695 } else {
3696 Node* uncast_base = base->uncast();
3697 int opcode = uncast_base->Opcode();
3698 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3699 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3700 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3701 is_captured_store_address(addp), "sanity");
3702 }
3703 }
3704 return base;
3705 }
3706
3707 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3708 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3709 Node* addp2 = addp->raw_out(0);
3710 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3711 addp2->in(AddPNode::Base) == n &&
3712 addp2->in(AddPNode::Address) == addp) {
3713 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3714 //
3715 // Find array's offset to push it on worklist first and
3716 // as result process an array's element offset first (pushed second)
3717 // to avoid CastPP for the array's offset.
3718 // Otherwise the inserted CastPP (LocalVar) will point to what
3719 // the AddP (Field) points to. Which would be wrong since
3720 // the algorithm expects the CastPP has the same point as
3721 // as AddP's base CheckCastPP (LocalVar).
3722 //
3723 // ArrayAllocation
3724 // |
3725 // CheckCastPP
3726 // |
3727 // memProj (from ArrayAllocation CheckCastPP)
3728 // | ||
3729 // | || Int (element index)
3730 // | || | ConI (log(element size))
3731 // | || | /
3732 // | || LShift
3733 // | || /
3734 // | AddP (array's element offset)
3735 // | |
3736 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3737 // | / /
3738 // AddP (array's offset)
3739 // |
3740 // Load/Store (memory operation on array's element)
3741 //
3742 return addp2;
3743 }
3744 return nullptr;
3745 }
3746
3747 //
3748 // Adjust the type and inputs of an AddP which computes the
3749 // address of a field of an instance
3750 //
3751 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3752 PhaseGVN* igvn = _igvn;
3753 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3754 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3755 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3756 if (t == nullptr) {
3757 // We are computing a raw address for a store captured by an Initialize
3758 // compute an appropriate address type (cases #3 and #5).
3759 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3760 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3761 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3762 assert(offs != Type::OffsetBot, "offset must be a constant");
3763 t = base_t->add_offset(offs)->is_oopptr();
3764 }
3765 int inst_id = base_t->instance_id();
3766 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3767 "old type must be non-instance or match new type");
3768
3769 // The type 't' could be subclass of 'base_t'.
3770 // As result t->offset() could be large then base_t's size and it will
3771 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3772 // constructor verifies correctness of the offset.
3773 //
3774 // It could happened on subclass's branch (from the type profiling
3775 // inlining) which was not eliminated during parsing since the exactness
3776 // of the allocation type was not propagated to the subclass type check.
3777 //
3778 // Or the type 't' could be not related to 'base_t' at all.
3779 // It could happened when CHA type is different from MDO type on a dead path
3780 // (for example, from instanceof check) which is not collapsed during parsing.
3781 //
3782 // Do nothing for such AddP node and don't process its users since
3783 // this code branch will go away.
3784 //
3785 if (!t->is_known_instance() &&
3786 !base_t->maybe_java_subtype_of(t)) {
3787 return false; // bail out
3788 }
3789 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
3790 // Do NOT remove the next line: ensure a new alias index is allocated
3791 // for the instance type. Note: C++ will not remove it since the call
3792 // has side effect.
3793 int alias_idx = _compile->get_alias_index(tinst);
3794 igvn->set_type(addp, tinst);
3795 // record the allocation in the node map
3796 set_map(addp, get_map(base->_idx));
3797 // Set addp's Base and Address to 'base'.
3798 Node *abase = addp->in(AddPNode::Base);
3799 Node *adr = addp->in(AddPNode::Address);
3800 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3801 adr->in(0)->_idx == (uint)inst_id) {
3802 // Skip AddP cases #3 and #5.
3803 } else {
3804 assert(!abase->is_top(), "sanity"); // AddP case #3
3805 if (abase != base) {
3806 igvn->hash_delete(addp);
3807 addp->set_req(AddPNode::Base, base);
3808 if (abase == adr) {
3809 addp->set_req(AddPNode::Address, base);
3810 } else {
3811 // AddP case #4 (adr is array's element offset AddP node)
3812 #ifdef ASSERT
3813 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3814 assert(adr->is_AddP() && atype != nullptr &&
3815 atype->instance_id() == inst_id, "array's element offset should be processed first");
3816 #endif
3817 }
3818 igvn->hash_insert(addp);
3819 }
3820 }
3821 // Put on IGVN worklist since at least addp's type was changed above.
3822 record_for_optimizer(addp);
3823 return true;
3824 }
3825
3826 //
3827 // Create a new version of orig_phi if necessary. Returns either the newly
3828 // created phi or an existing phi. Sets create_new to indicate whether a new
3829 // phi was created. Cache the last newly created phi in the node map.
3830 //
3831 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3832 Compile *C = _compile;
3833 PhaseGVN* igvn = _igvn;
3834 new_created = false;
3835 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3836 // nothing to do if orig_phi is bottom memory or matches alias_idx
3837 if (phi_alias_idx == alias_idx) {
3838 return orig_phi;
3839 }
3840 // Have we recently created a Phi for this alias index?
3841 PhiNode *result = get_map_phi(orig_phi->_idx);
3842 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3843 return result;
3844 }
3845 // Previous check may fail when the same wide memory Phi was split into Phis
3846 // for different memory slices. Search all Phis for this region.
3847 if (result != nullptr) {
3848 Node* region = orig_phi->in(0);
3849 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3850 Node* phi = region->fast_out(i);
3851 if (phi->is_Phi() &&
3852 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3853 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3854 return phi->as_Phi();
3855 }
3856 }
3857 }
3858 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
3859 if (C->do_escape_analysis() == true && !C->failing()) {
3860 // Retry compilation without escape analysis.
3861 // If this is the first failure, the sentinel string will "stick"
3862 // to the Compile object, and the C2Compiler will see it and retry.
3863 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
3864 }
3865 return nullptr;
3866 }
3867 orig_phi_worklist.append_if_missing(orig_phi);
3868 const TypePtr *atype = C->get_adr_type(alias_idx);
3869 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
3870 C->copy_node_notes_to(result, orig_phi);
3871 igvn->set_type(result, result->bottom_type());
3872 record_for_optimizer(result);
3873 set_map(orig_phi, result);
3874 new_created = true;
3875 return result;
3876 }
3877
3878 //
3879 // Return a new version of Memory Phi "orig_phi" with the inputs having the
3880 // specified alias index.
3881 //
3882 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
3883 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
3884 Compile *C = _compile;
3885 PhaseGVN* igvn = _igvn;
3886 bool new_phi_created;
3887 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
3888 if (!new_phi_created) {
3889 return result;
3890 }
3891 GrowableArray<PhiNode *> phi_list;
3892 GrowableArray<uint> cur_input;
3893 PhiNode *phi = orig_phi;
3894 uint idx = 1;
3895 bool finished = false;
3896 while(!finished) {
3897 while (idx < phi->req()) {
3898 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
3899 if (mem != nullptr && mem->is_Phi()) {
3900 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
3901 if (new_phi_created) {
3902 // found an phi for which we created a new split, push current one on worklist and begin
3903 // processing new one
3904 phi_list.push(phi);
3905 cur_input.push(idx);
3906 phi = mem->as_Phi();
3907 result = newphi;
3908 idx = 1;
3909 continue;
3910 } else {
3911 mem = newphi;
3912 }
3913 }
3914 if (C->failing()) {
3915 return nullptr;
3916 }
3917 result->set_req(idx++, mem);
3918 }
3919 #ifdef ASSERT
3920 // verify that the new Phi has an input for each input of the original
3921 assert( phi->req() == result->req(), "must have same number of inputs.");
3922 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
3923 #endif
3924 // Check if all new phi's inputs have specified alias index.
3925 // Otherwise use old phi.
3926 for (uint i = 1; i < phi->req(); i++) {
3927 Node* in = result->in(i);
3928 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
3929 }
3930 // we have finished processing a Phi, see if there are any more to do
3931 finished = (phi_list.length() == 0 );
3932 if (!finished) {
3933 phi = phi_list.pop();
3934 idx = cur_input.pop();
3935 PhiNode *prev_result = get_map_phi(phi->_idx);
3936 prev_result->set_req(idx++, result);
3937 result = prev_result;
3938 }
3939 }
3940 return result;
3941 }
3942
3943 //
3944 // The next methods are derived from methods in MemNode.
3945 //
3946 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
3947 Node *mem = mmem;
3948 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
3949 // means an array I have not precisely typed yet. Do not do any
3950 // alias stuff with it any time soon.
3951 if (toop->base() != Type::AnyPtr &&
3952 !(toop->isa_instptr() &&
3953 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
3954 toop->offset() == Type::OffsetBot)) {
3955 mem = mmem->memory_at(alias_idx);
3956 // Update input if it is progress over what we have now
3957 }
3958 return mem;
3959 }
3960
3961 //
3962 // Move memory users to their memory slices.
3963 //
3964 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
3965 Compile* C = _compile;
3966 PhaseGVN* igvn = _igvn;
3967 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
3968 assert(tp != nullptr, "ptr type");
3969 int alias_idx = C->get_alias_index(tp);
3970 int general_idx = C->get_general_index(alias_idx);
3971
3972 // Move users first
3973 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3974 Node* use = n->fast_out(i);
3975 if (use->is_MergeMem()) {
3976 MergeMemNode* mmem = use->as_MergeMem();
3977 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
3978 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
3979 continue; // Nothing to do
3980 }
3981 // Replace previous general reference to mem node.
3982 uint orig_uniq = C->unique();
3983 Node* m = find_inst_mem(n, general_idx, orig_phis);
3984 assert(orig_uniq == C->unique(), "no new nodes");
3985 mmem->set_memory_at(general_idx, m);
3986 --imax;
3987 --i;
3988 } else if (use->is_MemBar()) {
3989 assert(!use->is_Initialize(), "initializing stores should not be moved");
3990 if (use->req() > MemBarNode::Precedent &&
3991 use->in(MemBarNode::Precedent) == n) {
3992 // Don't move related membars.
3993 record_for_optimizer(use);
3994 continue;
3995 }
3996 tp = use->as_MemBar()->adr_type()->isa_ptr();
3997 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
3998 alias_idx == general_idx) {
3999 continue; // Nothing to do
4000 }
4001 // Move to general memory slice.
4002 uint orig_uniq = C->unique();
4003 Node* m = find_inst_mem(n, general_idx, orig_phis);
4004 assert(orig_uniq == C->unique(), "no new nodes");
4005 igvn->hash_delete(use);
4006 imax -= use->replace_edge(n, m, igvn);
4007 igvn->hash_insert(use);
4008 record_for_optimizer(use);
4009 --i;
4010 #ifdef ASSERT
4011 } else if (use->is_Mem()) {
4012 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
4013 // Don't move related cardmark.
4014 continue;
4015 }
4016 // Memory nodes should have new memory input.
4017 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4018 assert(tp != nullptr, "ptr type");
4019 int idx = C->get_alias_index(tp);
4020 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4021 "Following memory nodes should have new memory input or be on the same memory slice");
4022 } else if (use->is_Phi()) {
4023 // Phi nodes should be split and moved already.
4024 tp = use->as_Phi()->adr_type()->isa_ptr();
4025 assert(tp != nullptr, "ptr type");
4026 int idx = C->get_alias_index(tp);
4027 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4028 } else {
4029 use->dump();
4030 assert(false, "should not be here");
4031 #endif
4032 }
4033 }
4034 }
4035
4036 //
4037 // Search memory chain of "mem" to find a MemNode whose address
4038 // is the specified alias index.
4039 //
4040 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4041 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4042 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4043 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4044 return nullptr;
4045 }
4046 if (orig_mem == nullptr) {
4047 return orig_mem;
4048 }
4049 Compile* C = _compile;
4050 PhaseGVN* igvn = _igvn;
4051 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4052 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4053 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4054 Node *prev = nullptr;
4055 Node *result = orig_mem;
4056 while (prev != result) {
4057 prev = result;
4058 if (result == start_mem) {
4059 break; // hit one of our sentinels
4060 }
4061 if (result->is_Mem()) {
4062 const Type *at = igvn->type(result->in(MemNode::Address));
4063 if (at == Type::TOP) {
4064 break; // Dead
4065 }
4066 assert (at->isa_ptr() != nullptr, "pointer type required.");
4067 int idx = C->get_alias_index(at->is_ptr());
4068 if (idx == alias_idx) {
4069 break; // Found
4070 }
4071 if (!is_instance && (at->isa_oopptr() == nullptr ||
4072 !at->is_oopptr()->is_known_instance())) {
4073 break; // Do not skip store to general memory slice.
4074 }
4075 result = result->in(MemNode::Memory);
4076 }
4077 if (!is_instance) {
4078 continue; // don't search further for non-instance types
4079 }
4080 // skip over a call which does not affect this memory slice
4081 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4082 Node *proj_in = result->in(0);
4083 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4084 break; // hit one of our sentinels
4085 } else if (proj_in->is_Call()) {
4086 // ArrayCopy node processed here as well
4087 CallNode *call = proj_in->as_Call();
4088 if (!call->may_modify(toop, igvn)) {
4089 result = call->in(TypeFunc::Memory);
4090 }
4091 } else if (proj_in->is_Initialize()) {
4092 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4093 // Stop if this is the initialization for the object instance which
4094 // which contains this memory slice, otherwise skip over it.
4095 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4096 result = proj_in->in(TypeFunc::Memory);
4097 }
4098 } else if (proj_in->is_MemBar()) {
4099 // Check if there is an array copy for a clone
4100 // Step over GC barrier when ReduceInitialCardMarks is disabled
4101 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4102 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4103
4104 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4105 // Stop if it is a clone
4106 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4107 if (ac->may_modify(toop, igvn)) {
4108 break;
4109 }
4110 }
4111 result = proj_in->in(TypeFunc::Memory);
4112 }
4113 } else if (result->is_MergeMem()) {
4114 MergeMemNode *mmem = result->as_MergeMem();
4115 result = step_through_mergemem(mmem, alias_idx, toop);
4116 if (result == mmem->base_memory()) {
4117 // Didn't find instance memory, search through general slice recursively.
4118 result = mmem->memory_at(C->get_general_index(alias_idx));
4119 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4120 if (C->failing()) {
4121 return nullptr;
4122 }
4123 mmem->set_memory_at(alias_idx, result);
4124 }
4125 } else if (result->is_Phi() &&
4126 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4127 Node *un = result->as_Phi()->unique_input(igvn);
4128 if (un != nullptr) {
4129 orig_phis.append_if_missing(result->as_Phi());
4130 result = un;
4131 } else {
4132 break;
4133 }
4134 } else if (result->is_ClearArray()) {
4135 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4136 // Can not bypass initialization of the instance
4137 // we are looking for.
4138 break;
4139 }
4140 // Otherwise skip it (the call updated 'result' value).
4141 } else if (result->Opcode() == Op_SCMemProj) {
4142 Node* mem = result->in(0);
4143 Node* adr = nullptr;
4144 if (mem->is_LoadStore()) {
4145 adr = mem->in(MemNode::Address);
4146 } else {
4147 assert(mem->Opcode() == Op_EncodeISOArray ||
4148 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4149 adr = mem->in(3); // Memory edge corresponds to destination array
4150 }
4151 const Type *at = igvn->type(adr);
4152 if (at != Type::TOP) {
4153 assert(at->isa_ptr() != nullptr, "pointer type required.");
4154 int idx = C->get_alias_index(at->is_ptr());
4155 if (idx == alias_idx) {
4156 // Assert in debug mode
4157 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4158 break; // In product mode return SCMemProj node
4159 }
4160 }
4161 result = mem->in(MemNode::Memory);
4162 } else if (result->Opcode() == Op_StrInflatedCopy) {
4163 Node* adr = result->in(3); // Memory edge corresponds to destination array
4164 const Type *at = igvn->type(adr);
4165 if (at != Type::TOP) {
4166 assert(at->isa_ptr() != nullptr, "pointer type required.");
4167 int idx = C->get_alias_index(at->is_ptr());
4168 if (idx == alias_idx) {
4169 // Assert in debug mode
4170 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4171 break; // In product mode return SCMemProj node
4172 }
4173 }
4174 result = result->in(MemNode::Memory);
4175 }
4176 }
4177 if (result->is_Phi()) {
4178 PhiNode *mphi = result->as_Phi();
4179 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4180 const TypePtr *t = mphi->adr_type();
4181 if (!is_instance) {
4182 // Push all non-instance Phis on the orig_phis worklist to update inputs
4183 // during Phase 4 if needed.
4184 orig_phis.append_if_missing(mphi);
4185 } else if (C->get_alias_index(t) != alias_idx) {
4186 // Create a new Phi with the specified alias index type.
4187 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4188 }
4189 }
4190 // the result is either MemNode, PhiNode, InitializeNode.
4191 return result;
4192 }
4193
4194 //
4195 // Convert the types of non-escaped object to instance types where possible,
4196 // propagate the new type information through the graph, and update memory
4197 // edges and MergeMem inputs to reflect the new type.
4198 //
4199 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4200 // The processing is done in 4 phases:
4201 //
4202 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4203 // types for the CheckCastPP for allocations where possible.
4204 // Propagate the new types through users as follows:
4205 // casts and Phi: push users on alloc_worklist
4206 // AddP: cast Base and Address inputs to the instance type
4207 // push any AddP users on alloc_worklist and push any memnode
4208 // users onto memnode_worklist.
4209 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4210 // search the Memory chain for a store with the appropriate type
4211 // address type. If a Phi is found, create a new version with
4212 // the appropriate memory slices from each of the Phi inputs.
4213 // For stores, process the users as follows:
4214 // MemNode: push on memnode_worklist
4215 // MergeMem: push on mergemem_worklist
4216 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4217 // moving the first node encountered of each instance type to the
4218 // the input corresponding to its alias index.
4219 // appropriate memory slice.
4220 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4221 //
4222 // In the following example, the CheckCastPP nodes are the cast of allocation
4223 // results and the allocation of node 29 is non-escaped and eligible to be an
4224 // instance type.
4225 //
4226 // We start with:
4227 //
4228 // 7 Parm #memory
4229 // 10 ConI "12"
4230 // 19 CheckCastPP "Foo"
4231 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4232 // 29 CheckCastPP "Foo"
4233 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4234 //
4235 // 40 StoreP 25 7 20 ... alias_index=4
4236 // 50 StoreP 35 40 30 ... alias_index=4
4237 // 60 StoreP 45 50 20 ... alias_index=4
4238 // 70 LoadP _ 60 30 ... alias_index=4
4239 // 80 Phi 75 50 60 Memory alias_index=4
4240 // 90 LoadP _ 80 30 ... alias_index=4
4241 // 100 LoadP _ 80 20 ... alias_index=4
4242 //
4243 //
4244 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4245 // and creating a new alias index for node 30. This gives:
4246 //
4247 // 7 Parm #memory
4248 // 10 ConI "12"
4249 // 19 CheckCastPP "Foo"
4250 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4251 // 29 CheckCastPP "Foo" iid=24
4252 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4253 //
4254 // 40 StoreP 25 7 20 ... alias_index=4
4255 // 50 StoreP 35 40 30 ... alias_index=6
4256 // 60 StoreP 45 50 20 ... alias_index=4
4257 // 70 LoadP _ 60 30 ... alias_index=6
4258 // 80 Phi 75 50 60 Memory alias_index=4
4259 // 90 LoadP _ 80 30 ... alias_index=6
4260 // 100 LoadP _ 80 20 ... alias_index=4
4261 //
4262 // In phase 2, new memory inputs are computed for the loads and stores,
4263 // And a new version of the phi is created. In phase 4, the inputs to
4264 // node 80 are updated and then the memory nodes are updated with the
4265 // values computed in phase 2. This results in:
4266 //
4267 // 7 Parm #memory
4268 // 10 ConI "12"
4269 // 19 CheckCastPP "Foo"
4270 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4271 // 29 CheckCastPP "Foo" iid=24
4272 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4273 //
4274 // 40 StoreP 25 7 20 ... alias_index=4
4275 // 50 StoreP 35 7 30 ... alias_index=6
4276 // 60 StoreP 45 40 20 ... alias_index=4
4277 // 70 LoadP _ 50 30 ... alias_index=6
4278 // 80 Phi 75 40 60 Memory alias_index=4
4279 // 120 Phi 75 50 50 Memory alias_index=6
4280 // 90 LoadP _ 120 30 ... alias_index=6
4281 // 100 LoadP _ 80 20 ... alias_index=4
4282 //
4283 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4284 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4285 GrowableArray<MergeMemNode*> &mergemem_worklist,
4286 Unique_Node_List &reducible_merges) {
4287 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4288 GrowableArray<Node *> memnode_worklist;
4289 GrowableArray<PhiNode *> orig_phis;
4290 PhaseIterGVN *igvn = _igvn;
4291 uint new_index_start = (uint) _compile->num_alias_types();
4292 VectorSet visited;
4293 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4294 uint unique_old = _compile->unique();
4295
4296 // Phase 1: Process possible allocations from alloc_worklist.
4297 // Create instance types for the CheckCastPP for allocations where possible.
4298 //
4299 // (Note: don't forget to change the order of the second AddP node on
4300 // the alloc_worklist if the order of the worklist processing is changed,
4301 // see the comment in find_second_addp().)
4302 //
4303 while (alloc_worklist.length() != 0) {
4304 Node *n = alloc_worklist.pop();
4305 uint ni = n->_idx;
4306 if (n->is_Call()) {
4307 CallNode *alloc = n->as_Call();
4308 // copy escape information to call node
4309 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4310 PointsToNode::EscapeState es = ptn->escape_state();
4311 // We have an allocation or call which returns a Java object,
4312 // see if it is non-escaped.
4313 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4314 continue;
4315 }
4316 // Find CheckCastPP for the allocate or for the return value of a call
4317 n = alloc->result_cast();
4318 if (n == nullptr) { // No uses except Initialize node
4319 if (alloc->is_Allocate()) {
4320 // Set the scalar_replaceable flag for allocation
4321 // so it could be eliminated if it has no uses.
4322 alloc->as_Allocate()->_is_scalar_replaceable = true;
4323 }
4324 continue;
4325 }
4326 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4327 // we could reach here for allocate case if one init is associated with many allocs.
4328 if (alloc->is_Allocate()) {
4329 alloc->as_Allocate()->_is_scalar_replaceable = false;
4330 }
4331 continue;
4332 }
4333
4334 // The inline code for Object.clone() casts the allocation result to
4335 // java.lang.Object and then to the actual type of the allocated
4336 // object. Detect this case and use the second cast.
4337 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4338 // the allocation result is cast to java.lang.Object and then
4339 // to the actual Array type.
4340 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4341 && (alloc->is_AllocateArray() ||
4342 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4343 Node *cast2 = nullptr;
4344 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4345 Node *use = n->fast_out(i);
4346 if (use->is_CheckCastPP()) {
4347 cast2 = use;
4348 break;
4349 }
4350 }
4351 if (cast2 != nullptr) {
4352 n = cast2;
4353 } else {
4354 // Non-scalar replaceable if the allocation type is unknown statically
4355 // (reflection allocation), the object can't be restored during
4356 // deoptimization without precise type.
4357 continue;
4358 }
4359 }
4360
4361 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4362 if (t == nullptr) {
4363 continue; // not a TypeOopPtr
4364 }
4365 if (!t->klass_is_exact()) {
4366 continue; // not an unique type
4367 }
4368 if (alloc->is_Allocate()) {
4369 // Set the scalar_replaceable flag for allocation
4370 // so it could be eliminated.
4371 alloc->as_Allocate()->_is_scalar_replaceable = true;
4372 }
4373 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4374 // in order for an object to be scalar-replaceable, it must be:
4375 // - a direct allocation (not a call returning an object)
4376 // - non-escaping
4377 // - eligible to be a unique type
4378 // - not determined to be ineligible by escape analysis
4379 set_map(alloc, n);
4380 set_map(n, alloc);
4381 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4382 igvn->hash_delete(n);
4383 igvn->set_type(n, tinst);
4384 n->raise_bottom_type(tinst);
4385 igvn->hash_insert(n);
4386 record_for_optimizer(n);
4387 // Allocate an alias index for the header fields. Accesses to
4388 // the header emitted during macro expansion wouldn't have
4389 // correct memory state otherwise.
4390 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4391 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4392 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4393
4394 // First, put on the worklist all Field edges from Connection Graph
4395 // which is more accurate than putting immediate users from Ideal Graph.
4396 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4397 PointsToNode* tgt = e.get();
4398 if (tgt->is_Arraycopy()) {
4399 continue;
4400 }
4401 Node* use = tgt->ideal_node();
4402 assert(tgt->is_Field() && use->is_AddP(),
4403 "only AddP nodes are Field edges in CG");
4404 if (use->outcnt() > 0) { // Don't process dead nodes
4405 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4406 if (addp2 != nullptr) {
4407 assert(alloc->is_AllocateArray(),"array allocation was expected");
4408 alloc_worklist.append_if_missing(addp2);
4409 }
4410 alloc_worklist.append_if_missing(use);
4411 }
4412 }
4413
4414 // An allocation may have an Initialize which has raw stores. Scan
4415 // the users of the raw allocation result and push AddP users
4416 // on alloc_worklist.
4417 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4418 assert (raw_result != nullptr, "must have an allocation result");
4419 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4420 Node *use = raw_result->fast_out(i);
4421 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4422 Node* addp2 = find_second_addp(use, raw_result);
4423 if (addp2 != nullptr) {
4424 assert(alloc->is_AllocateArray(),"array allocation was expected");
4425 alloc_worklist.append_if_missing(addp2);
4426 }
4427 alloc_worklist.append_if_missing(use);
4428 } else if (use->is_MemBar()) {
4429 memnode_worklist.append_if_missing(use);
4430 }
4431 }
4432 }
4433 } else if (n->is_AddP()) {
4434 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4435 // This AddP will go away when we reduce the the Phi
4436 continue;
4437 }
4438 Node* addp_base = get_addp_base(n);
4439 JavaObjectNode* jobj = unique_java_object(addp_base);
4440 if (jobj == nullptr || jobj == phantom_obj) {
4441 #ifdef ASSERT
4442 ptnode_adr(get_addp_base(n)->_idx)->dump();
4443 ptnode_adr(n->_idx)->dump();
4444 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4445 #endif
4446 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4447 return;
4448 }
4449 Node *base = get_map(jobj->idx()); // CheckCastPP node
4450 if (!split_AddP(n, base)) continue; // wrong type from dead path
4451 } else if (n->is_Phi() ||
4452 n->is_CheckCastPP() ||
4453 n->is_EncodeP() ||
4454 n->is_DecodeN() ||
4455 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4456 if (visited.test_set(n->_idx)) {
4457 assert(n->is_Phi(), "loops only through Phi's");
4458 continue; // already processed
4459 }
4460 // Reducible Phi's will be removed from the graph after split_unique_types
4461 // finishes. For now we just try to split out the SR inputs of the merge.
4462 Node* parent = n->in(1);
4463 if (reducible_merges.member(n)) {
4464 reduce_phi(n->as_Phi(), alloc_worklist, memnode_worklist);
4465 #ifdef ASSERT
4466 if (VerifyReduceAllocationMerges) {
4467 reduced_merges.push(n);
4468 }
4469 #endif
4470 continue;
4471 } else if (reducible_merges.member(parent)) {
4472 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4473 // part of reduce_merge.
4474 continue;
4475 }
4476 JavaObjectNode* jobj = unique_java_object(n);
4477 if (jobj == nullptr || jobj == phantom_obj) {
4478 #ifdef ASSERT
4479 ptnode_adr(n->_idx)->dump();
4480 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4481 #endif
4482 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4483 return;
4484 } else {
4485 Node *val = get_map(jobj->idx()); // CheckCastPP node
4486 TypeNode *tn = n->as_Type();
4487 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4488 assert(tinst != nullptr && tinst->is_known_instance() &&
4489 tinst->instance_id() == jobj->idx() , "instance type expected.");
4490
4491 const Type *tn_type = igvn->type(tn);
4492 const TypeOopPtr *tn_t;
4493 if (tn_type->isa_narrowoop()) {
4494 tn_t = tn_type->make_ptr()->isa_oopptr();
4495 } else {
4496 tn_t = tn_type->isa_oopptr();
4497 }
4498 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4499 if (tn_type->isa_narrowoop()) {
4500 tn_type = tinst->make_narrowoop();
4501 } else {
4502 tn_type = tinst;
4503 }
4504 igvn->hash_delete(tn);
4505 igvn->set_type(tn, tn_type);
4506 tn->set_type(tn_type);
4507 igvn->hash_insert(tn);
4508 record_for_optimizer(n);
4509 } else {
4510 assert(tn_type == TypePtr::NULL_PTR ||
4511 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4512 "unexpected type");
4513 continue; // Skip dead path with different type
4514 }
4515 }
4516 } else {
4517 debug_only(n->dump();)
4518 assert(false, "EA: unexpected node");
4519 continue;
4520 }
4521 // push allocation's users on appropriate worklist
4522 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4523 Node *use = n->fast_out(i);
4524 if(use->is_Mem() && use->in(MemNode::Address) == n) {
4525 // Load/store to instance's field
4526 memnode_worklist.append_if_missing(use);
4527 } else if (use->is_MemBar()) {
4528 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4529 memnode_worklist.append_if_missing(use);
4530 }
4531 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4532 Node* addp2 = find_second_addp(use, n);
4533 if (addp2 != nullptr) {
4534 alloc_worklist.append_if_missing(addp2);
4535 }
4536 alloc_worklist.append_if_missing(use);
4537 } else if (use->is_Phi() ||
4538 use->is_CheckCastPP() ||
4539 use->is_EncodeNarrowPtr() ||
4540 use->is_DecodeNarrowPtr() ||
4541 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4542 alloc_worklist.append_if_missing(use);
4543 #ifdef ASSERT
4544 } else if (use->is_Mem()) {
4545 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4546 } else if (use->is_MergeMem()) {
4547 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4548 } else if (use->is_SafePoint()) {
4549 // Look for MergeMem nodes for calls which reference unique allocation
4550 // (through CheckCastPP nodes) even for debug info.
4551 Node* m = use->in(TypeFunc::Memory);
4552 if (m->is_MergeMem()) {
4553 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4554 }
4555 } else if (use->Opcode() == Op_EncodeISOArray) {
4556 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4557 // EncodeISOArray overwrites destination array
4558 memnode_worklist.append_if_missing(use);
4559 }
4560 } else {
4561 uint op = use->Opcode();
4562 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4563 (use->in(MemNode::Memory) == n)) {
4564 // They overwrite memory edge corresponding to destination array,
4565 memnode_worklist.append_if_missing(use);
4566 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4567 op == Op_CastP2X || op == Op_StoreCM ||
4568 op == Op_FastLock || op == Op_AryEq ||
4569 op == Op_StrComp || op == Op_CountPositives ||
4570 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4571 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4572 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4573 op == Op_SubTypeCheck ||
4574 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4575 n->dump();
4576 use->dump();
4577 assert(false, "EA: missing allocation reference path");
4578 }
4579 #endif
4580 }
4581 }
4582
4583 }
4584
4585 #ifdef ASSERT
4586 if (VerifyReduceAllocationMerges) {
4587 for (uint i = 0; i < reducible_merges.size(); i++) {
4588 Node* phi = reducible_merges.at(i);
4589
4590 if (!reduced_merges.member(phi)) {
4591 phi->dump(2);
4592 phi->dump(-2);
4593 assert(false, "This reducible merge wasn't reduced.");
4594 }
4595
4596 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4597 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4598 Node* use = phi->fast_out(j);
4599 if (!use->is_SafePoint() && !use->is_CastPP()) {
4600 phi->dump(2);
4601 phi->dump(-2);
4602 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4603 }
4604 }
4605 }
4606 }
4607 #endif
4608
4609 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4610 // type, record it in the ArrayCopy node so we know what memory this
4611 // node uses/modified.
4612 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4613 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4614 Node* dest = ac->in(ArrayCopyNode::Dest);
4615 if (dest->is_AddP()) {
4616 dest = get_addp_base(dest);
4617 }
4618 JavaObjectNode* jobj = unique_java_object(dest);
4619 if (jobj != nullptr) {
4620 Node *base = get_map(jobj->idx());
4621 if (base != nullptr) {
4622 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4623 ac->_dest_type = base_t;
4624 }
4625 }
4626 Node* src = ac->in(ArrayCopyNode::Src);
4627 if (src->is_AddP()) {
4628 src = get_addp_base(src);
4629 }
4630 jobj = unique_java_object(src);
4631 if (jobj != nullptr) {
4632 Node* base = get_map(jobj->idx());
4633 if (base != nullptr) {
4634 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4635 ac->_src_type = base_t;
4636 }
4637 }
4638 }
4639
4640 // New alias types were created in split_AddP().
4641 uint new_index_end = (uint) _compile->num_alias_types();
4642
4643 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4644 // compute new values for Memory inputs (the Memory inputs are not
4645 // actually updated until phase 4.)
4646 if (memnode_worklist.length() == 0)
4647 return; // nothing to do
4648 while (memnode_worklist.length() != 0) {
4649 Node *n = memnode_worklist.pop();
4650 if (visited.test_set(n->_idx)) {
4651 continue;
4652 }
4653 if (n->is_Phi() || n->is_ClearArray()) {
4654 // we don't need to do anything, but the users must be pushed
4655 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
4656 // we don't need to do anything, but the users must be pushed
4657 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4658 if (n == nullptr) {
4659 continue;
4660 }
4661 } else if (n->is_CallLeaf()) {
4662 // Runtime calls with narrow memory input (no MergeMem node)
4663 // get the memory projection
4664 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4665 if (n == nullptr) {
4666 continue;
4667 }
4668 } else if (n->Opcode() == Op_StrCompressedCopy ||
4669 n->Opcode() == Op_EncodeISOArray) {
4670 // get the memory projection
4671 n = n->find_out_with(Op_SCMemProj);
4672 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4673 } else {
4674 assert(n->is_Mem(), "memory node required.");
4675 Node *addr = n->in(MemNode::Address);
4676 const Type *addr_t = igvn->type(addr);
4677 if (addr_t == Type::TOP) {
4678 continue;
4679 }
4680 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4681 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4682 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4683 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4684 if (_compile->failing()) {
4685 return;
4686 }
4687 if (mem != n->in(MemNode::Memory)) {
4688 // We delay the memory edge update since we need old one in
4689 // MergeMem code below when instances memory slices are separated.
4690 set_map(n, mem);
4691 }
4692 if (n->is_Load()) {
4693 continue; // don't push users
4694 } else if (n->is_LoadStore()) {
4695 // get the memory projection
4696 n = n->find_out_with(Op_SCMemProj);
4697 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4698 }
4699 }
4700 // push user on appropriate worklist
4701 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4702 Node *use = n->fast_out(i);
4703 if (use->is_Phi() || use->is_ClearArray()) {
4704 memnode_worklist.append_if_missing(use);
4705 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4706 if (use->Opcode() == Op_StoreCM) { // Ignore cardmark stores
4707 continue;
4708 }
4709 memnode_worklist.append_if_missing(use);
4710 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4711 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4712 memnode_worklist.append_if_missing(use);
4713 }
4714 #ifdef ASSERT
4715 } else if(use->is_Mem()) {
4716 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4717 } else if (use->is_MergeMem()) {
4718 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4719 } else if (use->Opcode() == Op_EncodeISOArray) {
4720 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4721 // EncodeISOArray overwrites destination array
4722 memnode_worklist.append_if_missing(use);
4723 }
4724 } else {
4725 uint op = use->Opcode();
4726 if ((use->in(MemNode::Memory) == n) &&
4727 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4728 // They overwrite memory edge corresponding to destination array,
4729 memnode_worklist.append_if_missing(use);
4730 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4731 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4732 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4733 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
4734 n->dump();
4735 use->dump();
4736 assert(false, "EA: missing memory path");
4737 }
4738 #endif
4739 }
4740 }
4741 }
4742
4743 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4744 // Walk each memory slice moving the first node encountered of each
4745 // instance type to the input corresponding to its alias index.
4746 uint length = mergemem_worklist.length();
4747 for( uint next = 0; next < length; ++next ) {
4748 MergeMemNode* nmm = mergemem_worklist.at(next);
4749 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4750 // Note: we don't want to use MergeMemStream here because we only want to
4751 // scan inputs which exist at the start, not ones we add during processing.
4752 // Note 2: MergeMem may already contains instance memory slices added
4753 // during find_inst_mem() call when memory nodes were processed above.
4754 igvn->hash_delete(nmm);
4755 uint nslices = MIN2(nmm->req(), new_index_start);
4756 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4757 Node* mem = nmm->in(i);
4758 Node* cur = nullptr;
4759 if (mem == nullptr || mem->is_top()) {
4760 continue;
4761 }
4762 // First, update mergemem by moving memory nodes to corresponding slices
4763 // if their type became more precise since this mergemem was created.
4764 while (mem->is_Mem()) {
4765 const Type *at = igvn->type(mem->in(MemNode::Address));
4766 if (at != Type::TOP) {
4767 assert (at->isa_ptr() != nullptr, "pointer type required.");
4768 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4769 if (idx == i) {
4770 if (cur == nullptr) {
4771 cur = mem;
4772 }
4773 } else {
4774 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4775 nmm->set_memory_at(idx, mem);
4776 }
4777 }
4778 }
4779 mem = mem->in(MemNode::Memory);
4780 }
4781 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4782 // Find any instance of the current type if we haven't encountered
4783 // already a memory slice of the instance along the memory chain.
4784 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4785 if((uint)_compile->get_general_index(ni) == i) {
4786 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4787 if (nmm->is_empty_memory(m)) {
4788 Node* result = find_inst_mem(mem, ni, orig_phis);
4789 if (_compile->failing()) {
4790 return;
4791 }
4792 nmm->set_memory_at(ni, result);
4793 }
4794 }
4795 }
4796 }
4797 // Find the rest of instances values
4798 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4799 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4800 Node* result = step_through_mergemem(nmm, ni, tinst);
4801 if (result == nmm->base_memory()) {
4802 // Didn't find instance memory, search through general slice recursively.
4803 result = nmm->memory_at(_compile->get_general_index(ni));
4804 result = find_inst_mem(result, ni, orig_phis);
4805 if (_compile->failing()) {
4806 return;
4807 }
4808 nmm->set_memory_at(ni, result);
4809 }
4810 }
4811
4812 // If we have crossed the 3/4 point of max node limit it's too risky
4813 // to continue with EA/SR because we might hit the max node limit.
4814 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4815 if (_compile->do_reduce_allocation_merges()) {
4816 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4817 } else if (_invocation > 0) {
4818 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4819 } else {
4820 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4821 }
4822 return;
4823 }
4824
4825 igvn->hash_insert(nmm);
4826 record_for_optimizer(nmm);
4827 }
4828
4829 // Phase 4: Update the inputs of non-instance memory Phis and
4830 // the Memory input of memnodes
4831 // First update the inputs of any non-instance Phi's from
4832 // which we split out an instance Phi. Note we don't have
4833 // to recursively process Phi's encountered on the input memory
4834 // chains as is done in split_memory_phi() since they will
4835 // also be processed here.
4836 for (int j = 0; j < orig_phis.length(); j++) {
4837 PhiNode *phi = orig_phis.at(j);
4838 int alias_idx = _compile->get_alias_index(phi->adr_type());
4839 igvn->hash_delete(phi);
4840 for (uint i = 1; i < phi->req(); i++) {
4841 Node *mem = phi->in(i);
4842 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
4843 if (_compile->failing()) {
4844 return;
4845 }
4846 if (mem != new_mem) {
4847 phi->set_req(i, new_mem);
4848 }
4849 }
4850 igvn->hash_insert(phi);
4851 record_for_optimizer(phi);
4852 }
4853
4854 // Update the memory inputs of MemNodes with the value we computed
4855 // in Phase 2 and move stores memory users to corresponding memory slices.
4856 // Disable memory split verification code until the fix for 6984348.
4857 // Currently it produces false negative results since it does not cover all cases.
4858 #if 0 // ifdef ASSERT
4859 visited.Reset();
4860 Node_Stack old_mems(arena, _compile->unique() >> 2);
4861 #endif
4862 for (uint i = 0; i < ideal_nodes.size(); i++) {
4863 Node* n = ideal_nodes.at(i);
4864 Node* nmem = get_map(n->_idx);
4865 assert(nmem != nullptr, "sanity");
4866 if (n->is_Mem()) {
4867 #if 0 // ifdef ASSERT
4868 Node* old_mem = n->in(MemNode::Memory);
4869 if (!visited.test_set(old_mem->_idx)) {
4870 old_mems.push(old_mem, old_mem->outcnt());
4871 }
4872 #endif
4873 assert(n->in(MemNode::Memory) != nmem, "sanity");
4874 if (!n->is_Load()) {
4875 // Move memory users of a store first.
4876 move_inst_mem(n, orig_phis);
4877 }
4878 // Now update memory input
4879 igvn->hash_delete(n);
4880 n->set_req(MemNode::Memory, nmem);
4881 igvn->hash_insert(n);
4882 record_for_optimizer(n);
4883 } else {
4884 assert(n->is_Allocate() || n->is_CheckCastPP() ||
4885 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
4886 }
4887 }
4888 #if 0 // ifdef ASSERT
4889 // Verify that memory was split correctly
4890 while (old_mems.is_nonempty()) {
4891 Node* old_mem = old_mems.node();
4892 uint old_cnt = old_mems.index();
4893 old_mems.pop();
4894 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
4895 }
4896 #endif
4897 }
4898
4899 #ifndef PRODUCT
4900 int ConnectionGraph::_no_escape_counter = 0;
4901 int ConnectionGraph::_arg_escape_counter = 0;
4902 int ConnectionGraph::_global_escape_counter = 0;
4903
4904 static const char *node_type_names[] = {
4905 "UnknownType",
4906 "JavaObject",
4907 "LocalVar",
4908 "Field",
4909 "Arraycopy"
4910 };
4911
4912 static const char *esc_names[] = {
4913 "UnknownEscape",
4914 "NoEscape",
4915 "ArgEscape",
4916 "GlobalEscape"
4917 };
4918
4919 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
4920 NodeType nt = node_type();
4921 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
4922 if (print_state) {
4923 EscapeState es = escape_state();
4924 EscapeState fields_es = fields_escape_state();
4925 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
4926 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
4927 out->print("NSR ");
4928 }
4929 }
4930 }
4931
4932 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
4933 dump_header(print_state, out);
4934 if (is_Field()) {
4935 FieldNode* f = (FieldNode*)this;
4936 if (f->is_oop()) {
4937 out->print("oop ");
4938 }
4939 if (f->offset() > 0) {
4940 out->print("+%d ", f->offset());
4941 }
4942 out->print("(");
4943 for (BaseIterator i(f); i.has_next(); i.next()) {
4944 PointsToNode* b = i.get();
4945 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
4946 }
4947 out->print(" )");
4948 }
4949 out->print("[");
4950 for (EdgeIterator i(this); i.has_next(); i.next()) {
4951 PointsToNode* e = i.get();
4952 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
4953 }
4954 out->print(" [");
4955 for (UseIterator i(this); i.has_next(); i.next()) {
4956 PointsToNode* u = i.get();
4957 bool is_base = false;
4958 if (PointsToNode::is_base_use(u)) {
4959 is_base = true;
4960 u = PointsToNode::get_use_node(u)->as_Field();
4961 }
4962 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
4963 }
4964 out->print(" ]] ");
4965 if (_node == nullptr) {
4966 out->print("<null>%s", newline ? "\n" : "");
4967 } else {
4968 _node->dump(newline ? "\n" : "", false, out);
4969 }
4970 }
4971
4972 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
4973 bool first = true;
4974 int ptnodes_length = ptnodes_worklist.length();
4975 for (int i = 0; i < ptnodes_length; i++) {
4976 PointsToNode *ptn = ptnodes_worklist.at(i);
4977 if (ptn == nullptr || !ptn->is_JavaObject()) {
4978 continue;
4979 }
4980 PointsToNode::EscapeState es = ptn->escape_state();
4981 if ((es != PointsToNode::NoEscape) && !Verbose) {
4982 continue;
4983 }
4984 Node* n = ptn->ideal_node();
4985 if (n->is_Allocate() || (n->is_CallStaticJava() &&
4986 n->as_CallStaticJava()->is_boxing_method())) {
4987 if (first) {
4988 tty->cr();
4989 tty->print("======== Connection graph for ");
4990 _compile->method()->print_short_name();
4991 tty->cr();
4992 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
4993 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
4994 tty->cr();
4995 first = false;
4996 }
4997 ptn->dump();
4998 // Print all locals and fields which reference this allocation
4999 for (UseIterator j(ptn); j.has_next(); j.next()) {
5000 PointsToNode* use = j.get();
5001 if (use->is_LocalVar()) {
5002 use->dump(Verbose);
5003 } else if (Verbose) {
5004 use->dump();
5005 }
5006 }
5007 tty->cr();
5008 }
5009 }
5010 }
5011
5012 void ConnectionGraph::print_statistics() {
5013 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));
5014 }
5015
5016 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5017 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5018 return;
5019 }
5020 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5021 JavaObjectNode* ptn = java_objects_worklist.at(next);
5022 if (ptn->ideal_node()->is_Allocate()) {
5023 if (ptn->escape_state() == PointsToNode::NoEscape) {
5024 Atomic::inc(&ConnectionGraph::_no_escape_counter);
5025 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5026 Atomic::inc(&ConnectionGraph::_arg_escape_counter);
5027 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5028 Atomic::inc(&ConnectionGraph::_global_escape_counter);
5029 } else {
5030 assert(false, "Unexpected Escape State");
5031 }
5032 }
5033 }
5034 }
5035
5036 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5037 if (_compile->directive()->TraceEscapeAnalysisOption) {
5038 assert(ptn != nullptr, "should not be null");
5039 assert(reason != nullptr, "should not be null");
5040 ptn->dump_header(true);
5041 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5042 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5043 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5044 }
5045 }
5046
5047 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5048 if (_compile->directive()->TraceEscapeAnalysisOption) {
5049 stringStream ss;
5050 ss.print("propagated from: ");
5051 from->dump(true, &ss, false);
5052 return ss.as_string();
5053 } else {
5054 return nullptr;
5055 }
5056 }
5057
5058 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5059 if (_compile->directive()->TraceEscapeAnalysisOption) {
5060 stringStream ss;
5061 ss.print("escapes as arg to:");
5062 call->dump("", false, &ss);
5063 return ss.as_string();
5064 } else {
5065 return nullptr;
5066 }
5067 }
5068
5069 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5070 if (_compile->directive()->TraceEscapeAnalysisOption) {
5071 stringStream ss;
5072 ss.print("is merged with other object: ");
5073 other->dump_header(true, &ss);
5074 return ss.as_string();
5075 } else {
5076 return nullptr;
5077 }
5078 }
5079
5080 #endif
5081
5082 void ConnectionGraph::record_for_optimizer(Node *n) {
5083 _igvn->_worklist.push(n);
5084 _igvn->add_users_to_worklist(n);
5085 }