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