1 /*
2 * Copyright (c) 2005, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/bcEscapeAnalyzer.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/c2compiler.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "utilities/macros.hpp"
46
47 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
48 // If ReduceAllocationMerges is enabled we might call split_through_phi during
49 // split_unique_types and that will create additional nodes that need to be
50 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
51 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
52 // the array will be reallocated.
53 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
54 _in_worklist(C->comp_arena()),
55 _next_pidx(0),
56 _collecting(true),
57 _verify(false),
58 _compile(C),
59 _igvn(igvn),
60 _invocation(invocation),
61 _build_iterations(0),
62 _build_time(0.),
63 _node_map(C->comp_arena()) {
64 // Add unknown java object.
65 add_java_object(C->top(), PointsToNode::GlobalEscape);
66 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
67 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
68 // Add ConP and ConN null oop nodes
69 Node* oop_null = igvn->zerocon(T_OBJECT);
70 assert(oop_null->_idx < nodes_size(), "should be created already");
71 add_java_object(oop_null, PointsToNode::NoEscape);
72 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
73 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
74 if (UseCompressedOops) {
75 Node* noop_null = igvn->zerocon(T_NARROWOOP);
76 assert(noop_null->_idx < nodes_size(), "should be created already");
77 map_ideal_node(noop_null, null_obj);
78 }
79 }
80
81 bool ConnectionGraph::has_candidates(Compile *C) {
82 // EA brings benefits only when the code has allocations and/or locks which
83 // are represented by ideal Macro nodes.
84 int cnt = C->macro_count();
85 for (int i = 0; i < cnt; i++) {
86 Node *n = C->macro_node(i);
87 if (n->is_Allocate()) {
88 return true;
89 }
90 if (n->is_Lock()) {
91 Node* obj = n->as_Lock()->obj_node()->uncast();
92 if (!(obj->is_Parm() || obj->is_Con())) {
93 return true;
94 }
95 }
96 if (n->is_CallStaticJava() &&
97 n->as_CallStaticJava()->is_boxing_method()) {
98 return true;
99 }
100 }
101 return false;
102 }
103
104 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
105 Compile::TracePhase tp(Phase::_t_escapeAnalysis);
106 ResourceMark rm;
107
108 // Add ConP and ConN null oop nodes before ConnectionGraph construction
109 // to create space for them in ConnectionGraph::_nodes[].
110 Node* oop_null = igvn->zerocon(T_OBJECT);
111 Node* noop_null = igvn->zerocon(T_NARROWOOP);
112 int invocation = 0;
113 if (C->congraph() != nullptr) {
114 invocation = C->congraph()->_invocation + 1;
115 }
116 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
117 NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(congraph);)
118 // Perform escape analysis
119 if (congraph->compute_escape()) {
120 // There are non escaping objects.
121 C->set_congraph(congraph);
122 }
123 NOT_PRODUCT(if (C->should_print_igv(/* Any level */ 1)) C->igv_printer()->set_congraph(nullptr);)
124 // Cleanup.
125 if (oop_null->outcnt() == 0) {
126 igvn->hash_delete(oop_null);
127 }
128 if (noop_null->outcnt() == 0) {
129 igvn->hash_delete(noop_null);
130 }
131
132 C->print_method(PHASE_AFTER_EA, 2);
133 }
134
135 bool ConnectionGraph::compute_escape() {
136 Compile* C = _compile;
137 PhaseGVN* igvn = _igvn;
138
139 // Worklists used by EA.
140 Unique_Node_List delayed_worklist;
141 Unique_Node_List reducible_merges;
142 GrowableArray<Node*> alloc_worklist;
143 GrowableArray<Node*> ptr_cmp_worklist;
144 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
145 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
146 GrowableArray<PointsToNode*> ptnodes_worklist;
147 GrowableArray<JavaObjectNode*> java_objects_worklist;
148 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
149 GrowableArray<FieldNode*> oop_fields_worklist;
150 GrowableArray<SafePointNode*> sfn_worklist;
151 GrowableArray<MergeMemNode*> mergemem_worklist;
152 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
153
154 { Compile::TracePhase tp(Phase::_t_connectionGraph);
155
156 // 1. Populate Connection Graph (CG) with PointsTo nodes.
157 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
158 // Initialize worklist
159 if (C->root() != nullptr) {
160 ideal_nodes.push(C->root());
161 }
162 // Processed ideal nodes are unique on ideal_nodes list
163 // but several ideal nodes are mapped to the phantom_obj.
164 // To avoid duplicated entries on the following worklists
165 // add the phantom_obj only once to them.
166 ptnodes_worklist.append(phantom_obj);
167 java_objects_worklist.append(phantom_obj);
168 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
169 Node* n = ideal_nodes.at(next);
170 // Create PointsTo nodes and add them to Connection Graph. Called
171 // only once per ideal node since ideal_nodes is Unique_Node list.
172 add_node_to_connection_graph(n, &delayed_worklist);
173 PointsToNode* ptn = ptnode_adr(n->_idx);
174 if (ptn != nullptr && ptn != phantom_obj) {
175 ptnodes_worklist.append(ptn);
176 if (ptn->is_JavaObject()) {
177 java_objects_worklist.append(ptn->as_JavaObject());
178 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
179 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
180 // Only allocations and java static calls results are interesting.
181 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
182 }
183 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
184 oop_fields_worklist.append(ptn->as_Field());
185 }
186 }
187 // Collect some interesting nodes for further use.
188 switch (n->Opcode()) {
189 case Op_MergeMem:
190 // Collect all MergeMem nodes to add memory slices for
191 // scalar replaceable objects in split_unique_types().
192 mergemem_worklist.append(n->as_MergeMem());
193 break;
194 case Op_CmpP:
195 case Op_CmpN:
196 // Collect compare pointers nodes.
197 if (OptimizePtrCompare) {
198 ptr_cmp_worklist.append(n);
199 }
200 break;
201 case Op_MemBarStoreStore:
202 // Collect all MemBarStoreStore nodes so that depending on the
203 // escape status of the associated Allocate node some of them
204 // may be eliminated.
205 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
206 storestore_worklist.append(n->as_MemBarStoreStore());
207 }
208 // If MemBarStoreStore has a precedent edge add it to the worklist (like MemBarRelease)
209 case Op_MemBarRelease:
210 if (n->req() > MemBarNode::Precedent) {
211 record_for_optimizer(n);
212 }
213 break;
214 #ifdef ASSERT
215 case Op_AddP:
216 // Collect address nodes for graph verification.
217 addp_worklist.append(n);
218 break;
219 #endif
220 case Op_ArrayCopy:
221 // Keep a list of ArrayCopy nodes so if one of its input is non
222 // escaping, we can record a unique type
223 arraycopy_worklist.append(n->as_ArrayCopy());
224 break;
225 default:
226 // not interested now, ignore...
227 break;
228 }
229 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
230 Node* m = n->fast_out(i); // Get user
231 ideal_nodes.push(m);
232 }
233 if (n->is_SafePoint()) {
234 sfn_worklist.append(n->as_SafePoint());
235 }
236 }
237
238 #ifndef PRODUCT
239 if (_compile->directive()->TraceEscapeAnalysisOption) {
240 tty->print("+++++ Initial worklist for ");
241 _compile->method()->print_name();
242 tty->print_cr(" (ea_inv=%d)", _invocation);
243 for (int i = 0; i < ptnodes_worklist.length(); i++) {
244 PointsToNode* ptn = ptnodes_worklist.at(i);
245 ptn->dump();
246 }
247 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
248 }
249 #endif
250
251 if (non_escaped_allocs_worklist.length() == 0) {
252 _collecting = false;
253 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
254 return false; // Nothing to do.
255 }
256 // Add final simple edges to graph.
257 while(delayed_worklist.size() > 0) {
258 Node* n = delayed_worklist.pop();
259 add_final_edges(n);
260 }
261
262 #ifdef ASSERT
263 if (VerifyConnectionGraph) {
264 // Verify that no new simple edges could be created and all
265 // local vars has edges.
266 _verify = true;
267 int ptnodes_length = ptnodes_worklist.length();
268 for (int next = 0; next < ptnodes_length; ++next) {
269 PointsToNode* ptn = ptnodes_worklist.at(next);
270 add_final_edges(ptn->ideal_node());
271 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
272 ptn->dump();
273 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
274 }
275 }
276 _verify = false;
277 }
278 #endif
279 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
280 // processing, calls to CI to resolve symbols (types, fields, methods)
281 // referenced in bytecode. During symbol resolution VM may throw
282 // an exception which CI cleans and converts to compilation failure.
283 if (C->failing()) {
284 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
285 return false;
286 }
287
288 _compile->print_method(PHASE_EA_AFTER_INITIAL_CONGRAPH, 4);
289
290 // 2. Finish Graph construction by propagating references to all
291 // java objects through graph.
292 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
293 java_objects_worklist, oop_fields_worklist)) {
294 // All objects escaped or hit time or iterations limits.
295 _collecting = false;
296 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
297 return false;
298 }
299
300 _compile->print_method(PHASE_EA_AFTER_COMPLETE_CONGRAPH, 4);
301
302 // 3. Adjust scalar_replaceable state of nonescaping objects and push
303 // scalar replaceable allocations on alloc_worklist for processing
304 // in split_unique_types().
305 GrowableArray<JavaObjectNode*> jobj_worklist;
306 int non_escaped_length = non_escaped_allocs_worklist.length();
307 bool found_nsr_alloc = false;
308 for (int next = 0; next < non_escaped_length; next++) {
309 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
310 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
311 Node* n = ptn->ideal_node();
312 if (n->is_Allocate()) {
313 n->as_Allocate()->_is_non_escaping = noescape;
314 }
315 if (noescape && ptn->scalar_replaceable()) {
316 adjust_scalar_replaceable_state(ptn, reducible_merges);
317 if (ptn->scalar_replaceable()) {
318 jobj_worklist.push(ptn);
319 } else {
320 found_nsr_alloc = true;
321 }
322 }
323 _compile->print_method(PHASE_EA_ADJUST_SCALAR_REPLACEABLE_ITER, 6, n);
324 }
325
326 // Propagate NSR (Not Scalar Replaceable) state.
327 if (found_nsr_alloc) {
328 find_scalar_replaceable_allocs(jobj_worklist, reducible_merges);
329 }
330
331 // alloc_worklist will be processed in reverse push order.
332 // Therefore the reducible Phis will be processed for last and that's what we
333 // want because by then the scalarizable inputs of the merge will already have
334 // an unique instance type.
335 for (uint i = 0; i < reducible_merges.size(); i++ ) {
336 Node* n = reducible_merges.at(i);
337 alloc_worklist.append(n);
338 }
339
340 for (int next = 0; next < jobj_worklist.length(); ++next) {
341 JavaObjectNode* jobj = jobj_worklist.at(next);
342 if (jobj->scalar_replaceable()) {
343 alloc_worklist.append(jobj->ideal_node());
344 }
345 }
346
347 #ifdef ASSERT
348 if (VerifyConnectionGraph) {
349 // Verify that graph is complete - no new edges could be added or needed.
350 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
351 java_objects_worklist, addp_worklist);
352 }
353 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
354 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
355 null_obj->edge_count() == 0 &&
356 !null_obj->arraycopy_src() &&
357 !null_obj->arraycopy_dst(), "sanity");
358 #endif
359
360 _collecting = false;
361
362 _compile->print_method(PHASE_EA_AFTER_PROPAGATE_NSR, 4);
363 } // TracePhase t3("connectionGraph")
364
365 // 4. Optimize ideal graph based on EA information.
366 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
367 if (has_non_escaping_obj) {
368 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
369 }
370
371 #ifndef PRODUCT
372 if (PrintEscapeAnalysis) {
373 dump(ptnodes_worklist); // Dump ConnectionGraph
374 }
375 #endif
376
377 #ifdef ASSERT
378 if (VerifyConnectionGraph) {
379 int alloc_length = alloc_worklist.length();
380 for (int next = 0; next < alloc_length; ++next) {
381 Node* n = alloc_worklist.at(next);
382 PointsToNode* ptn = ptnode_adr(n->_idx);
383 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
384 }
385 }
386
387 if (VerifyReduceAllocationMerges) {
388 for (uint i = 0; i < reducible_merges.size(); i++ ) {
389 Node* n = reducible_merges.at(i);
390 if (!can_reduce_phi(n->as_Phi())) {
391 TraceReduceAllocationMerges = true;
392 n->dump(2);
393 n->dump(-2);
394 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
395 }
396 }
397 }
398 #endif
399
400 _compile->print_method(PHASE_EA_AFTER_GRAPH_OPTIMIZATION, 4);
401
402 // 5. Separate memory graph for scalar replaceable allcations.
403 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
404 if (has_scalar_replaceable_candidates && EliminateAllocations) {
405 assert(C->do_aliasing(), "Aliasing should be enabled");
406 // Now use the escape information to create unique types for
407 // scalar replaceable objects.
408 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
409 if (C->failing()) {
410 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
411 return false;
412 }
413
414 #ifdef ASSERT
415 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
416 tty->print("=== No allocations eliminated for ");
417 C->method()->print_short_name();
418 if (!EliminateAllocations) {
419 tty->print(" since EliminateAllocations is off ===");
420 } else if(!has_scalar_replaceable_candidates) {
421 tty->print(" since there are no scalar replaceable candidates ===");
422 }
423 tty->cr();
424 #endif
425 }
426
427 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES, 4);
428
429 // 6. Reduce allocation merges used as debug information. This is done after
430 // split_unique_types because the methods used to create SafePointScalarObject
431 // need to traverse the memory graph to find values for object fields. We also
432 // set to null the scalarized inputs of reducible Phis so that the Allocate
433 // that they point can be later scalar replaced.
434 bool delay = _igvn->delay_transform();
435 _igvn->set_delay_transform(true);
436 for (uint i = 0; i < reducible_merges.size(); i++) {
437 Node* n = reducible_merges.at(i);
438 if (n->outcnt() > 0) {
439 if (!reduce_phi_on_safepoints(n->as_Phi())) {
440 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
441 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
442 return false;
443 }
444
445 // Now we set the scalar replaceable inputs of ophi to null, which is
446 // the last piece that would prevent it from being scalar replaceable.
447 reset_scalar_replaceable_entries(n->as_Phi());
448 }
449 }
450 _igvn->set_delay_transform(delay);
451
452 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
453 // java calls if they pass ArgEscape objects as parameters.
454 if (has_non_escaping_obj &&
455 (C->env()->should_retain_local_variables() ||
456 C->env()->jvmti_can_get_owned_monitor_info() ||
457 C->env()->jvmti_can_walk_any_space() ||
458 DeoptimizeObjectsALot)) {
459 int sfn_length = sfn_worklist.length();
460 for (int next = 0; next < sfn_length; next++) {
461 SafePointNode* sfn = sfn_worklist.at(next);
462 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
463 if (sfn->is_CallJava()) {
464 CallJavaNode* call = sfn->as_CallJava();
465 call->set_arg_escape(has_arg_escape(call));
466 }
467 }
468 }
469
470 _compile->print_method(PHASE_EA_AFTER_REDUCE_PHI_ON_SAFEPOINTS, 4);
471
472 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
473 return has_non_escaping_obj;
474 }
475
476 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
477 // if at least one scalar replaceable allocation participates in the merge.
478 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
479 bool found_sr_allocate = false;
480
481 for (uint i = 1; i < ophi->req(); i++) {
482 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
483 if (ptn != nullptr && ptn->scalar_replaceable()) {
484 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
485
486 // Don't handle arrays.
487 if (alloc->Opcode() != Op_Allocate) {
488 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
489 continue;
490 }
491
492 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
493 found_sr_allocate = true;
494 } else {
495 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
496 ptn->set_scalar_replaceable(false);
497 }
498 }
499 }
500
501 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);)
502 return found_sr_allocate;
503 }
504
505 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
506 // I require the 'other' input to be a constant so that I can move the Cmp
507 // around safely.
508 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
509 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
510 Node* left = cmp->in(1);
511 Node* right = cmp->in(2);
512
513 return (left == n || right == n) &&
514 (left->is_Con() || right->is_Con()) &&
515 cmp->outcnt() == 1;
516 }
517
518 // We are going to check if any of the SafePointScalarMerge entries
519 // in the SafePoint reference the Phi that we are checking.
520 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
521 JVMState *jvms = sfpt->jvms();
522
523 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
524 Node* sfpt_in = sfpt->in(i);
525 if (sfpt_in->is_SafePointScalarMerge()) {
526 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
527 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
528 if (nsr_ptr == n) {
529 return true;
530 }
531 }
532 }
533
534 return false;
535 }
536
537 // Check if we are able to untangle the merge. The following patterns are
538 // supported:
539 // - Phi -> SafePoints
540 // - Phi -> CmpP/N
541 // - Phi -> AddP -> Load
542 // - Phi -> CastPP -> SafePoints
543 // - Phi -> CastPP -> AddP -> Load
544 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
545 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
546 Node* use = n->fast_out(i);
547
548 if (use->is_SafePoint()) {
549 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
550 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
551 return false;
552 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
553 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
554 return false;
555 }
556 } else if (use->is_AddP()) {
557 Node* addp = use;
558 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
559 Node* use_use = addp->fast_out(j);
560 const Type* load_type = _igvn->type(use_use);
561
562 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
563 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());)
564 return false;
565 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
566 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());)
567 return false;
568 }
569 }
570 } else if (nesting > 0) {
571 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);)
572 return false;
573 } else if (use->is_CastPP()) {
574 const Type* cast_t = _igvn->type(use);
575 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
576 #ifndef PRODUCT
577 if (TraceReduceAllocationMerges) {
578 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
579 use->dump();
580 }
581 #endif
582 return false;
583 }
584
585 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
586 if (!is_trivial_control) {
587 // If it's not a trivial control then we check if we can reduce the
588 // CmpP/N used by the If controlling the cast.
589 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
590 Node* iff = use->in(0)->in(0);
591 // We may have an OpaqueNotNull node between If and Bool nodes. But we could also have a sub class of IfNode,
592 // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
593 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
594 if (can_reduce) {
595 Node* iff_cmp = iff->in(1)->in(1);
596 int opc = iff_cmp->Opcode();
597 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
598 }
599 if (!can_reduce) {
600 #ifndef PRODUCT
601 if (TraceReduceAllocationMerges) {
602 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
603 n->dump(5);
604 }
605 #endif
606 return false;
607 }
608 }
609 }
610
611 if (!can_reduce_check_users(use, nesting+1)) {
612 return false;
613 }
614 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
615 if (!can_reduce_cmp(n, use)) {
616 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);)
617 return false;
618 }
619 } else {
620 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());)
621 return false;
622 }
623 }
624
625 return true;
626 }
627
628 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
629 // only used in some certain code shapes. Check comments in
630 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
631 // details.
632 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
633 // If there was an error attempting to reduce allocation merges for this
634 // method we might have disabled the compilation and be retrying with RAM
635 // disabled.
636 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
637 return false;
638 }
639
640 const Type* phi_t = _igvn->type(ophi);
641 if (phi_t == nullptr ||
642 phi_t->make_ptr() == nullptr ||
643 phi_t->make_ptr()->isa_aryptr() != nullptr) {
644 return false;
645 }
646
647 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
648 return false;
649 }
650
651 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
652 return true;
653 }
654
655 // This method will return a CmpP/N that we need to use on the If controlling a
656 // CastPP after it was split. This method is only called on bases that are
657 // nullable therefore we always need a controlling if for the splitted CastPP.
658 //
659 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
660 // If the CastPP currently doesn't have a control then the CmpP/N will be
661 // against the null constant, otherwise it will be against the constant input of
662 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
663 // case because we have constraints on it and because the CastPP has a control
664 // input.
665 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
666 const Type* t = base->bottom_type();
667 Node* con = nullptr;
668
669 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
670 con = _igvn->zerocon(t->basic_type());
671 } else {
672 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
673 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
674 Node* bol = curr_ctrl->in(0)->in(1);
675 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
676 Node* curr_cmp = bol->in(1);
677 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
678 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
679 }
680
681 return CmpNode::make(base, con, t->basic_type());
682 }
683
684 // This method 'specializes' the CastPP passed as parameter to the base passed
685 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
686 // means that the CastPP now will be specific for a given base instead of a Phi.
687 // An If-Then-Else-Region block is inserted to control the CastPP. The control
688 // of the CastPP is a copy of the current one (if there is one) or a check
689 // against null.
690 //
691 // Before:
692 //
693 // C1 C2 ... Cn
694 // \ | /
695 // \ | /
696 // \ | /
697 // \ | /
698 // \ | /
699 // \ | /
700 // \|/
701 // Region B1 B2 ... Bn
702 // | \ | /
703 // | \ | /
704 // | \ | /
705 // | \ | /
706 // | \ | /
707 // | \ | /
708 // ---------------> Phi
709 // |
710 // X |
711 // | |
712 // | |
713 // ------> CastPP
714 //
715 // After (only partial illustration; base = B2, current_control = C2):
716 //
717 // C2
718 // |
719 // If
720 // / \
721 // / \
722 // T F
723 // /\ /
724 // / \ /
725 // / \ /
726 // C1 CastPP Reg Cn
727 // | | |
728 // | | |
729 // | | |
730 // -------------- | ----------
731 // | | |
732 // Region
733 //
734 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
735 Node* control_successor = current_control->unique_ctrl_out();
736 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
737 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
738 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
739 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
740 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
741 Node* end_region = _igvn->transform(new RegionNode(3));
742
743 // Insert the new if-else-region block into the graph
744 end_region->set_req(1, not_eq_control);
745 end_region->set_req(2, yes_eq_control);
746 control_successor->replace_edge(current_control, end_region, _igvn);
747
748 _igvn->_worklist.push(current_control);
749 _igvn->_worklist.push(control_successor);
750
751 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::UnconditionalDependency, nullptr));
752 }
753
754 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
755 const Type* load_type = _igvn->type(curr_load);
756 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
757 Node* memory = curr_load->in(MemNode::Memory);
758
759 // The data_phi merging the loads needs to be nullable if
760 // we are loading pointers.
761 if (load_type->make_ptr() != nullptr) {
762 if (load_type->isa_narrowoop()) {
763 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
764 } else if (load_type->isa_ptr()) {
765 load_type = load_type->meet(TypePtr::NULL_PTR);
766 } else {
767 assert(false, "Unexpected load ptr type.");
768 }
769 }
770
771 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
772
773 for (int i = 1; i < bases_for_loads->length(); i++) {
774 Node* base = bases_for_loads->at(i);
775 Node* cmp_region = nullptr;
776 if (base != nullptr) {
777 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
778 cmp_region = base->unique_ctrl_out_or_null();
779 assert(cmp_region != nullptr, "There should be.");
780 base = base->find_out_with(Op_CastPP);
781 }
782
783 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
784 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
785 Node* load = curr_load->clone();
786 load->set_req(0, nullptr);
787 load->set_req(1, mem);
788 load->set_req(2, addr);
789
790 if (cmp_region != nullptr) { // see comment on previous if
791 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
792 intermediate_phi->set_req(1, _igvn->transform(load));
793 load = intermediate_phi;
794 }
795
796 data_phi->set_req(i, _igvn->transform(load));
797 } else {
798 // Just use the default, which is already in phi
799 }
800 }
801
802 // Takes care of updating CG and split_unique_types worklists due
803 // to cloned AddP->Load.
804 updates_after_load_split(data_phi, curr_load, alloc_worklist);
805
806 return _igvn->transform(data_phi);
807 }
808
809 // This method only reduces CastPP fields loads; SafePoints are handled
810 // separately. The idea here is basically to clone the CastPP and place copies
811 // on each input of the Phi, including non-scalar replaceable inputs.
812 // Experimentation shows that the resulting IR graph is simpler that way than if
813 // we just split the cast through scalar-replaceable inputs.
814 //
815 // The reduction process requires that CastPP's control be one of:
816 // 1) no control,
817 // 2) the same region as Ophi, or
818 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
819 //
820 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
821 // flow. If the CastPP originally had an IfTrue/False control input then we'll
822 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
823 // juse use a CmpP/N against the null constant.
824 //
825 // The If-Then-Else-Region isn't always needed. For instance, if input to
826 // splitted cast was not nullable (or if it was the null constant) then we don't
827 // need (shouldn't) use a CastPP at all.
828 //
829 // After the casts are splitted we'll split the AddP->Loads through the Phi and
830 // connect them to the just split CastPPs.
831 //
832 // Before (CastPP control is same as Phi):
833 //
834 // Region Allocate Null Call
835 // | \ | /
836 // | \ | /
837 // | \ | /
838 // | \ | /
839 // | \ | /
840 // | \ | /
841 // ------------------> Phi # Oop Phi
842 // | |
843 // | |
844 // | |
845 // | |
846 // ----------------> CastPP
847 // |
848 // AddP
849 // |
850 // Load
851 //
852 // After (Very much simplified):
853 //
854 // Call Null
855 // \ /
856 // CmpP
857 // |
858 // Bool#NE
859 // |
860 // If
861 // / \
862 // T F
863 // / \ /
864 // / R
865 // CastPP |
866 // | |
867 // AddP |
868 // | |
869 // Load |
870 // \ | 0
871 // Allocate \ | /
872 // \ \ | /
873 // AddP Phi
874 // \ /
875 // Load /
876 // \ 0 /
877 // \ | /
878 // \|/
879 // Phi # "Field" Phi
880 //
881 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node*> &alloc_worklist) {
882 Node* ophi = curr_castpp->in(1);
883 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
884
885 // Identify which base should be used for AddP->Load later when spliting the
886 // CastPP->Loads through ophi. Three kind of values may be stored in this
887 // array, depending on the nullability status of the corresponding input in
888 // ophi.
889 //
890 // - nullptr: Meaning that the base is actually the null constant and therefore
891 // we won't try to load from it.
892 //
893 // - CFG Node: Meaning that the base is a CastPP that was specialized for
894 // this input of Ophi. I.e., we added an If->Then->Else-Region
895 // that will 'activate' the CastPp only when the input is not Null.
896 //
897 // - Other Node: Meaning that the base is not nullable and therefore we'll try
898 // to load directly from it.
899 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
900
901 for (uint i = 1; i < ophi->req(); i++) {
902 Node* base = ophi->in(i);
903 const Type* base_t = _igvn->type(base);
904
905 if (base_t->maybe_null()) {
906 if (base->is_Con()) {
907 // Nothing todo as bases_for_loads[i] is already null
908 } else {
909 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
910 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
911 }
912 } else {
913 bases_for_loads.at_put(i, base);
914 }
915 }
916
917 // Now let's split the CastPP->Loads through the Phi
918 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
919 Node* use = curr_castpp->raw_out(i);
920 if (use->is_AddP()) {
921 for (int j = use->outcnt()-1; j >= 0;) {
922 Node* use_use = use->raw_out(j);
923 assert(use_use->is_Load(), "Expected this to be a Load node.");
924
925 // We can't make an unconditional load from a nullable input. The
926 // 'split_castpp_load_through_phi` method will add an
927 // 'If-Then-Else-Region` around nullable bases and only load from them
928 // when the input is not null.
929 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
930 _igvn->replace_node(use_use, phi);
931
932 --j;
933 j = MIN2(j, (int)use->outcnt()-1);
934 }
935
936 _igvn->remove_dead_node(use);
937 }
938 --i;
939 i = MIN2(i, (int)curr_castpp->outcnt()-1);
940 }
941 }
942
943 // This method split a given CmpP/N through the Phi used in one of its inputs.
944 // As a result we convert a comparison with a pointer to a comparison with an
945 // integer.
946 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
947 // while the other must be a constant.
948 // The splitting process is basically just cloning the CmpP/N above the input
949 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
950 // can prove at compile time the result of the comparison.
951 //
952 // Before:
953 //
954 // in1 in2 ... inN
955 // \ | /
956 // \ | /
957 // \ | /
958 // \ | /
959 // \ | /
960 // \ | /
961 // Phi
962 // | Other
963 // | /
964 // | /
965 // | /
966 // CmpP/N
967 //
968 // After:
969 //
970 // in1 Other in2 Other inN Other
971 // | | | | | |
972 // \ | | | | |
973 // \ / | / | /
974 // CmpP/N CmpP/N CmpP/N
975 // Bool Bool Bool
976 // \ | /
977 // \ | /
978 // \ | /
979 // \ | /
980 // \ | /
981 // \ | /
982 // \ | /
983 // \ | /
984 // Phi
985 // |
986 // | Zero
987 // | /
988 // | /
989 // | /
990 // CmpI
991 //
992 //
993 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
994 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
995 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
996
997 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
998 Node* zero = _igvn->intcon(0);
999 Node* one = _igvn->intcon(1);
1000 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
1001
1002 // This Phi will merge the result of the Cmps split through the Phi
1003 Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
1004
1005 for (uint i=1; i<ophi->req(); i++) {
1006 Node* ophi_input = ophi->in(i);
1007 Node* res_phi_input = nullptr;
1008
1009 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1010 if (tcmp->singleton()) {
1011 if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
1012 (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
1013 res_phi_input = one;
1014 } else {
1015 res_phi_input = zero;
1016 }
1017 } else {
1018 Node* ncmp = _igvn->transform(cmp->clone());
1019 ncmp->set_req(1, ophi_input);
1020 ncmp->set_req(2, other);
1021 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1022 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1023 }
1024
1025 res_phi->set_req(i, res_phi_input);
1026 }
1027
1028 // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1029 Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1030 _igvn->replace_node(cmp, new_cmp);
1031 }
1032
1033 // Push the newly created AddP on alloc_worklist and patch
1034 // the connection graph. Note that the changes in the CG below
1035 // won't affect the ES of objects since the new nodes have the
1036 // same status as the old ones.
1037 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1038 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1039 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1040 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1041
1042 if (data_phi == nullptr || !data_phi->is_Phi()) {
1043 // Make this a retry?
1044 return ;
1045 }
1046
1047 Node* previous_addp = previous_load->in(MemNode::Address);
1048 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1049 for (uint i = 1; i < data_phi->req(); i++) {
1050 Node* new_load = data_phi->in(i);
1051
1052 if (new_load->is_Phi()) {
1053 // new_load is currently the "intermediate_phi" from an specialized
1054 // CastPP.
1055 new_load = new_load->in(1);
1056 }
1057
1058 // "new_load" might actually be a constant, parameter, etc.
1059 if (new_load->is_Load()) {
1060 Node* new_addp = new_load->in(MemNode::Address);
1061 Node* base = get_addp_base(new_addp);
1062
1063 // The base might not be something that we can create an unique
1064 // type for. If that's the case we are done with that input.
1065 PointsToNode* jobj_ptn = unique_java_object(base);
1066 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1067 continue;
1068 }
1069
1070 // Push to alloc_worklist since the base has an unique_type
1071 alloc_worklist.append_if_missing(new_addp);
1072
1073 // Now let's add the node to the connection graph
1074 _nodes.at_grow(new_addp->_idx, nullptr);
1075 add_field(new_addp, fn->escape_state(), fn->offset());
1076 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1077
1078 // If the load doesn't load an object then it won't be
1079 // part of the connection graph
1080 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1081 if (curr_load_ptn != nullptr) {
1082 _nodes.at_grow(new_load->_idx, nullptr);
1083 add_local_var(new_load, curr_load_ptn->escape_state());
1084 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1085 }
1086 }
1087 }
1088 }
1089
1090 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1091 // We'll pass this to 'split_through_phi' so that it'll do the split even
1092 // though the load doesn't have an unique instance type.
1093 bool ignore_missing_instance_id = true;
1094
1095 // All AddPs are present in the connection graph
1096 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1097
1098 // Iterate over AddP looking for a Load
1099 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1100 Node* previous_load = previous_addp->raw_out(k);
1101 if (previous_load->is_Load()) {
1102 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1103
1104 // Takes care of updating CG and split_unique_types worklists due to cloned
1105 // AddP->Load.
1106 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1107
1108 _igvn->replace_node(previous_load, data_phi);
1109 }
1110 --k;
1111 k = MIN2(k, (int)previous_addp->outcnt()-1);
1112 }
1113
1114 // Remove the old AddP from the processing list because it's dead now
1115 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1116 alloc_worklist.remove_if_existing(previous_addp);
1117 }
1118
1119 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1120 // selector is:
1121 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1122 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1123 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1124 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1125 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1126 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1127 uint number_of_sr_objects = 0;
1128 for (uint i = 1; i < ophi->req(); i++) {
1129 Node* base = ophi->in(i);
1130 JavaObjectNode* ptn = unique_java_object(base);
1131
1132 if (ptn != nullptr && ptn->scalar_replaceable()) {
1133 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1134 selector->set_req(i, sr_obj_idx);
1135 number_of_sr_objects++;
1136 }
1137 }
1138
1139 return selector->as_Phi();
1140 }
1141
1142 // Returns true if the AddP node 'n' has at least one base that is a reducible
1143 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1144 // checked instead.
1145 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1146 PointsToNode* ptn = ptnode_adr(n->_idx);
1147 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1148 return false;
1149 }
1150
1151 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1152 Node* base = i.get()->ideal_node();
1153
1154 if (reducible_merges.member(base)) {
1155 return true;
1156 }
1157
1158 if (base->is_CastPP() || base->is_CheckCastPP()) {
1159 base = base->in(1);
1160 if (reducible_merges.member(base)) {
1161 return true;
1162 }
1163 }
1164 }
1165
1166 return false;
1167 }
1168
1169 // This method will call its helper method to reduce SafePoint nodes that use
1170 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1171 // "version" of Phi use the same debug information (regarding the Phi).
1172 // Therefore, I collect all safepoints and patch them all at once.
1173 //
1174 // The safepoints using the Phi node have to be processed before safepoints of
1175 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1176 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1177 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1178 // algorithm that process Phi's safepoints will think that the added Phi
1179 // reference is a regular reference.
1180 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1181 PhiNode* selector = create_selector(ophi);
1182 Unique_Node_List safepoints;
1183 Unique_Node_List casts;
1184
1185 // Just collect the users of the Phis for later processing
1186 // in the needed order.
1187 for (uint i = 0; i < ophi->outcnt(); i++) {
1188 Node* use = ophi->raw_out(i);
1189 if (use->is_SafePoint()) {
1190 safepoints.push(use);
1191 } else if (use->is_CastPP()) {
1192 casts.push(use);
1193 } else {
1194 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1195 }
1196 }
1197
1198 // Need to process safepoints using the Phi first
1199 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1200 return false;
1201 }
1202
1203 // Now process CastPP->safepoints
1204 for (uint i = 0; i < casts.size(); i++) {
1205 Node* cast = casts.at(i);
1206 Unique_Node_List cast_sfpts;
1207
1208 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1209 Node* use_use = cast->fast_out(j);
1210 if (use_use->is_SafePoint()) {
1211 cast_sfpts.push(use_use);
1212 } else {
1213 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1214 }
1215 }
1216
1217 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1218 return false;
1219 }
1220 }
1221
1222 return true;
1223 }
1224
1225 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1226 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1227 // SafePointScalarObjectNode for each scalar replaceable input. Each
1228 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1229 // check detailed description in SafePointScalarMergeNode class header.
1230 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1231 PhaseMacroExpand mexp(*_igvn);
1232 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1233 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1234
1235 Node* nsr_merge_pointer = ophi;
1236 if (cast != nullptr) {
1237 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1238 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::RegularDependency, nullptr));
1239 }
1240
1241 for (uint spi = 0; spi < safepoints.size(); spi++) {
1242 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1243 JVMState *jvms = sfpt->jvms();
1244 uint merge_idx = (sfpt->req() - jvms->scloff());
1245 int debug_start = jvms->debug_start();
1246
1247 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1248 smerge->init_req(0, _compile->root());
1249 _igvn->register_new_node_with_optimizer(smerge);
1250
1251 // The next two inputs are:
1252 // (1) A copy of the original pointer to NSR objects.
1253 // (2) A selector, used to decide if we need to rematerialize an object
1254 // or use the pointer to a NSR object.
1255 // See more details of these fields in the declaration of SafePointScalarMergeNode
1256 sfpt->add_req(nsr_merge_pointer);
1257 sfpt->add_req(selector);
1258
1259 for (uint i = 1; i < ophi->req(); i++) {
1260 Node* base = ophi->in(i);
1261 JavaObjectNode* ptn = unique_java_object(base);
1262
1263 // If the base is not scalar replaceable we don't need to register information about
1264 // it at this time.
1265 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1266 continue;
1267 }
1268
1269 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1270 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt);
1271 if (sobj == nullptr) {
1272 return false;
1273 }
1274
1275 // Now make a pass over the debug information replacing any references
1276 // to the allocated object with "sobj"
1277 Node* ccpp = alloc->result_cast();
1278 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1279
1280 // Register the scalarized object as a candidate for reallocation
1281 smerge->add_req(sobj);
1282 }
1283
1284 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1285 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1286
1287 // The call to 'replace_edges_in_range' above might have removed the
1288 // reference to ophi that we need at _merge_pointer_idx. The line below make
1289 // sure the reference is maintained.
1290 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1291 _igvn->_worklist.push(sfpt);
1292 }
1293
1294 return true;
1295 }
1296
1297 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1298 bool delay = _igvn->delay_transform();
1299 _igvn->set_delay_transform(true);
1300 _igvn->hash_delete(ophi);
1301
1302 // Copying all users first because some will be removed and others won't.
1303 // Ophi also may acquire some new users as part of Cast reduction.
1304 // CastPPs also need to be processed before CmpPs.
1305 Unique_Node_List castpps;
1306 Unique_Node_List others;
1307 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1308 Node* use = ophi->fast_out(i);
1309
1310 if (use->is_CastPP()) {
1311 castpps.push(use);
1312 } else if (use->is_AddP() || use->is_Cmp()) {
1313 others.push(use);
1314 } else {
1315 // Safepoints to be processed later; other users aren't expected here
1316 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1317 }
1318 }
1319
1320 _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1321
1322 // CastPPs need to be processed before Cmps because during the process of
1323 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1324 // by the If controlling the CastPP.
1325 for (uint i = 0; i < castpps.size(); i++) {
1326 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1327 _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1328 }
1329
1330 for (uint i = 0; i < others.size(); i++) {
1331 Node* use = others.at(i);
1332
1333 if (use->is_AddP()) {
1334 reduce_phi_on_field_access(use, alloc_worklist);
1335 _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1336 } else if(use->is_Cmp()) {
1337 reduce_phi_on_cmp(use);
1338 _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1339 }
1340 }
1341
1342 _igvn->set_delay_transform(delay);
1343 }
1344
1345 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1346 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1347 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1348 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1349 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1350
1351 for (uint i = 1; i < ophi->req(); i++) {
1352 Node* base = ophi->in(i);
1353 JavaObjectNode* ptn = unique_java_object(base);
1354
1355 if (ptn != nullptr && ptn->scalar_replaceable()) {
1356 new_phi->set_req(i, null_ptr);
1357 } else {
1358 new_phi->set_req(i, ophi->in(i));
1359 }
1360 }
1361
1362 for (int i = ophi->outcnt()-1; i >= 0;) {
1363 Node* out = ophi->raw_out(i);
1364
1365 if (out->is_ConstraintCast()) {
1366 const Type* out_t = _igvn->type(out)->make_ptr();
1367 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1368 bool change = out_new_t != out_t;
1369
1370 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1371 Node* out2 = out->raw_out(j);
1372 if (!out2->is_SafePoint()) {
1373 change = false;
1374 break;
1375 }
1376 }
1377
1378 if (change) {
1379 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::StrongDependency, nullptr);
1380 _igvn->replace_node(out, new_cast);
1381 _igvn->register_new_node_with_optimizer(new_cast);
1382 }
1383 }
1384
1385 --i;
1386 i = MIN2(i, (int)ophi->outcnt()-1);
1387 }
1388
1389 _igvn->replace_node(ophi, new_phi);
1390 }
1391
1392 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1393 if (!C->do_reduce_allocation_merges()) return;
1394
1395 Unique_Node_List ideal_nodes;
1396 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1397 ideal_nodes.push(root);
1398
1399 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1400 Node* n = ideal_nodes.at(next);
1401
1402 if (n->is_SafePointScalarMerge()) {
1403 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1404
1405 // Validate inputs of merge
1406 for (uint i = 1; i < merge->req(); i++) {
1407 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1408 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1409 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1410 }
1411 }
1412
1413 // Validate users of merge
1414 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1415 Node* sfpt = merge->fast_out(i);
1416 if (sfpt->is_SafePoint()) {
1417 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1418
1419 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1420 assert(false, "SafePointScalarMerge nodes can't be nested.");
1421 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1422 }
1423 } else {
1424 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1425 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1426 }
1427 }
1428 }
1429
1430 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1431 Node* m = n->fast_out(i);
1432 ideal_nodes.push(m);
1433 }
1434 }
1435 }
1436
1437 // Returns true if there is an object in the scope of sfn that does not escape globally.
1438 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1439 Compile* C = _compile;
1440 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1441 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1442 DeoptimizeObjectsALot) {
1443 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1444 int num_locs = jvms->loc_size();
1445 for (int idx = 0; idx < num_locs; idx++) {
1446 Node* l = sfn->local(jvms, idx);
1447 if (not_global_escape(l)) {
1448 return true;
1449 }
1450 }
1451 }
1452 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1453 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1454 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1455 int num_mon = jvms->nof_monitors();
1456 for (int idx = 0; idx < num_mon; idx++) {
1457 Node* m = sfn->monitor_obj(jvms, idx);
1458 if (m != nullptr && not_global_escape(m)) {
1459 return true;
1460 }
1461 }
1462 }
1463 }
1464 return false;
1465 }
1466
1467 // Returns true if at least one of the arguments to the call is an object
1468 // that does not escape globally.
1469 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1470 if (call->method() != nullptr) {
1471 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1472 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1473 Node* p = call->in(idx);
1474 if (not_global_escape(p)) {
1475 return true;
1476 }
1477 }
1478 } else {
1479 const char* name = call->as_CallStaticJava()->_name;
1480 assert(name != nullptr, "no name");
1481 // no arg escapes through uncommon traps
1482 if (strcmp(name, "uncommon_trap") != 0) {
1483 // process_call_arguments() assumes that all arguments escape globally
1484 const TypeTuple* d = call->tf()->domain();
1485 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1486 const Type* at = d->field_at(i);
1487 if (at->isa_oopptr() != nullptr) {
1488 return true;
1489 }
1490 }
1491 }
1492 }
1493 return false;
1494 }
1495
1496
1497
1498 // Utility function for nodes that load an object
1499 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1500 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1501 // ThreadLocal has RawPtr type.
1502 const Type* t = _igvn->type(n);
1503 if (t->make_ptr() != nullptr) {
1504 Node* adr = n->in(MemNode::Address);
1505 #ifdef ASSERT
1506 if (!adr->is_AddP()) {
1507 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1508 } else {
1509 assert((ptnode_adr(adr->_idx) == nullptr ||
1510 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1511 }
1512 #endif
1513 add_local_var_and_edge(n, PointsToNode::NoEscape,
1514 adr, delayed_worklist);
1515 }
1516 }
1517
1518 // Populate Connection Graph with PointsTo nodes and create simple
1519 // connection graph edges.
1520 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1521 assert(!_verify, "this method should not be called for verification");
1522 PhaseGVN* igvn = _igvn;
1523 uint n_idx = n->_idx;
1524 PointsToNode* n_ptn = ptnode_adr(n_idx);
1525 if (n_ptn != nullptr) {
1526 return; // No need to redefine PointsTo node during first iteration.
1527 }
1528 int opcode = n->Opcode();
1529 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1530 if (gc_handled) {
1531 return; // Ignore node if already handled by GC.
1532 }
1533
1534 if (n->is_Call()) {
1535 // Arguments to allocation and locking don't escape.
1536 if (n->is_AbstractLock()) {
1537 // Put Lock and Unlock nodes on IGVN worklist to process them during
1538 // first IGVN optimization when escape information is still available.
1539 record_for_optimizer(n);
1540 } else if (n->is_Allocate()) {
1541 add_call_node(n->as_Call());
1542 record_for_optimizer(n);
1543 } else {
1544 if (n->is_CallStaticJava()) {
1545 const char* name = n->as_CallStaticJava()->_name;
1546 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1547 return; // Skip uncommon traps
1548 }
1549 }
1550 // Don't mark as processed since call's arguments have to be processed.
1551 delayed_worklist->push(n);
1552 // Check if a call returns an object.
1553 if ((n->as_Call()->returns_pointer() &&
1554 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1555 (n->is_CallStaticJava() &&
1556 n->as_CallStaticJava()->is_boxing_method())) {
1557 add_call_node(n->as_Call());
1558 }
1559 }
1560 return;
1561 }
1562 // Put this check here to process call arguments since some call nodes
1563 // point to phantom_obj.
1564 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1565 return; // Skip predefined nodes.
1566 }
1567 switch (opcode) {
1568 case Op_AddP: {
1569 Node* base = get_addp_base(n);
1570 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1571 // Field nodes are created for all field types. They are used in
1572 // adjust_scalar_replaceable_state() and split_unique_types().
1573 // Note, non-oop fields will have only base edges in Connection
1574 // Graph because such fields are not used for oop loads and stores.
1575 int offset = address_offset(n, igvn);
1576 add_field(n, PointsToNode::NoEscape, offset);
1577 if (ptn_base == nullptr) {
1578 delayed_worklist->push(n); // Process it later.
1579 } else {
1580 n_ptn = ptnode_adr(n_idx);
1581 add_base(n_ptn->as_Field(), ptn_base);
1582 }
1583 break;
1584 }
1585 case Op_CastX2P: {
1586 map_ideal_node(n, phantom_obj);
1587 break;
1588 }
1589 case Op_CastPP:
1590 case Op_CheckCastPP:
1591 case Op_EncodeP:
1592 case Op_DecodeN:
1593 case Op_EncodePKlass:
1594 case Op_DecodeNKlass: {
1595 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1596 break;
1597 }
1598 case Op_CMoveP: {
1599 add_local_var(n, PointsToNode::NoEscape);
1600 // Do not add edges during first iteration because some could be
1601 // not defined yet.
1602 delayed_worklist->push(n);
1603 break;
1604 }
1605 case Op_ConP:
1606 case Op_ConN:
1607 case Op_ConNKlass: {
1608 // assume all oop constants globally escape except for null
1609 PointsToNode::EscapeState es;
1610 const Type* t = igvn->type(n);
1611 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1612 es = PointsToNode::NoEscape;
1613 } else {
1614 es = PointsToNode::GlobalEscape;
1615 }
1616 PointsToNode* ptn_con = add_java_object(n, es);
1617 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1618 break;
1619 }
1620 case Op_CreateEx: {
1621 // assume that all exception objects globally escape
1622 map_ideal_node(n, phantom_obj);
1623 break;
1624 }
1625 case Op_LoadKlass:
1626 case Op_LoadNKlass: {
1627 // Unknown class is loaded
1628 map_ideal_node(n, phantom_obj);
1629 break;
1630 }
1631 case Op_LoadP:
1632 case Op_LoadN: {
1633 add_objload_to_connection_graph(n, delayed_worklist);
1634 break;
1635 }
1636 case Op_Parm: {
1637 map_ideal_node(n, phantom_obj);
1638 break;
1639 }
1640 case Op_PartialSubtypeCheck: {
1641 // Produces Null or notNull and is used in only in CmpP so
1642 // phantom_obj could be used.
1643 map_ideal_node(n, phantom_obj); // Result is unknown
1644 break;
1645 }
1646 case Op_Phi: {
1647 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1648 // ThreadLocal has RawPtr type.
1649 const Type* t = n->as_Phi()->type();
1650 if (t->make_ptr() != nullptr) {
1651 add_local_var(n, PointsToNode::NoEscape);
1652 // Do not add edges during first iteration because some could be
1653 // not defined yet.
1654 delayed_worklist->push(n);
1655 }
1656 break;
1657 }
1658 case Op_Proj: {
1659 // we are only interested in the oop result projection from a call
1660 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1661 n->in(0)->as_Call()->returns_pointer()) {
1662 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1663 }
1664 break;
1665 }
1666 case Op_Rethrow: // Exception object escapes
1667 case Op_Return: {
1668 if (n->req() > TypeFunc::Parms &&
1669 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1670 // Treat Return value as LocalVar with GlobalEscape escape state.
1671 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1672 }
1673 break;
1674 }
1675 case Op_CompareAndExchangeP:
1676 case Op_CompareAndExchangeN:
1677 case Op_GetAndSetP:
1678 case Op_GetAndSetN: {
1679 add_objload_to_connection_graph(n, delayed_worklist);
1680 // fall-through
1681 }
1682 case Op_StoreP:
1683 case Op_StoreN:
1684 case Op_StoreNKlass:
1685 case Op_WeakCompareAndSwapP:
1686 case Op_WeakCompareAndSwapN:
1687 case Op_CompareAndSwapP:
1688 case Op_CompareAndSwapN: {
1689 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1690 break;
1691 }
1692 case Op_AryEq:
1693 case Op_CountPositives:
1694 case Op_StrComp:
1695 case Op_StrEquals:
1696 case Op_StrIndexOf:
1697 case Op_StrIndexOfChar:
1698 case Op_StrInflatedCopy:
1699 case Op_StrCompressedCopy:
1700 case Op_VectorizedHashCode:
1701 case Op_EncodeISOArray: {
1702 add_local_var(n, PointsToNode::ArgEscape);
1703 delayed_worklist->push(n); // Process it later.
1704 break;
1705 }
1706 case Op_ThreadLocal: {
1707 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1708 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1709 break;
1710 }
1711 case Op_Blackhole: {
1712 // All blackhole pointer arguments are globally escaping.
1713 // Only do this if there is at least one pointer argument.
1714 // Do not add edges during first iteration because some could be
1715 // not defined yet, defer to final step.
1716 for (uint i = 0; i < n->req(); i++) {
1717 Node* in = n->in(i);
1718 if (in != nullptr) {
1719 const Type* at = _igvn->type(in);
1720 if (!at->isa_ptr()) continue;
1721
1722 add_local_var(n, PointsToNode::GlobalEscape);
1723 delayed_worklist->push(n);
1724 break;
1725 }
1726 }
1727 break;
1728 }
1729 default:
1730 ; // Do nothing for nodes not related to EA.
1731 }
1732 return;
1733 }
1734
1735 // Add final simple edges to graph.
1736 void ConnectionGraph::add_final_edges(Node *n) {
1737 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1738 #ifdef ASSERT
1739 if (_verify && n_ptn->is_JavaObject())
1740 return; // This method does not change graph for JavaObject.
1741 #endif
1742
1743 if (n->is_Call()) {
1744 process_call_arguments(n->as_Call());
1745 return;
1746 }
1747 assert(n->is_Store() || n->is_LoadStore() ||
1748 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1749 "node should be registered already");
1750 int opcode = n->Opcode();
1751 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1752 if (gc_handled) {
1753 return; // Ignore node if already handled by GC.
1754 }
1755 switch (opcode) {
1756 case Op_AddP: {
1757 Node* base = get_addp_base(n);
1758 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1759 assert(ptn_base != nullptr, "field's base should be registered");
1760 add_base(n_ptn->as_Field(), ptn_base);
1761 break;
1762 }
1763 case Op_CastPP:
1764 case Op_CheckCastPP:
1765 case Op_EncodeP:
1766 case Op_DecodeN:
1767 case Op_EncodePKlass:
1768 case Op_DecodeNKlass: {
1769 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1770 break;
1771 }
1772 case Op_CMoveP: {
1773 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1774 Node* in = n->in(i);
1775 if (in == nullptr) {
1776 continue; // ignore null
1777 }
1778 Node* uncast_in = in->uncast();
1779 if (uncast_in->is_top() || uncast_in == n) {
1780 continue; // ignore top or inputs which go back this node
1781 }
1782 PointsToNode* ptn = ptnode_adr(in->_idx);
1783 assert(ptn != nullptr, "node should be registered");
1784 add_edge(n_ptn, ptn);
1785 }
1786 break;
1787 }
1788 case Op_LoadP:
1789 case Op_LoadN: {
1790 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1791 // ThreadLocal has RawPtr type.
1792 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1793 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1794 break;
1795 }
1796 case Op_Phi: {
1797 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1798 // ThreadLocal has RawPtr type.
1799 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1800 for (uint i = 1; i < n->req(); i++) {
1801 Node* in = n->in(i);
1802 if (in == nullptr) {
1803 continue; // ignore null
1804 }
1805 Node* uncast_in = in->uncast();
1806 if (uncast_in->is_top() || uncast_in == n) {
1807 continue; // ignore top or inputs which go back this node
1808 }
1809 PointsToNode* ptn = ptnode_adr(in->_idx);
1810 assert(ptn != nullptr, "node should be registered");
1811 add_edge(n_ptn, ptn);
1812 }
1813 break;
1814 }
1815 case Op_Proj: {
1816 // we are only interested in the oop result projection from a call
1817 assert(n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
1818 n->in(0)->as_Call()->returns_pointer(), "Unexpected node type");
1819 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1820 break;
1821 }
1822 case Op_Rethrow: // Exception object escapes
1823 case Op_Return: {
1824 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1825 "Unexpected node type");
1826 // Treat Return value as LocalVar with GlobalEscape escape state.
1827 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1828 break;
1829 }
1830 case Op_CompareAndExchangeP:
1831 case Op_CompareAndExchangeN:
1832 case Op_GetAndSetP:
1833 case Op_GetAndSetN:{
1834 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1835 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1836 // fall-through
1837 }
1838 case Op_CompareAndSwapP:
1839 case Op_CompareAndSwapN:
1840 case Op_WeakCompareAndSwapP:
1841 case Op_WeakCompareAndSwapN:
1842 case Op_StoreP:
1843 case Op_StoreN:
1844 case Op_StoreNKlass:{
1845 add_final_edges_unsafe_access(n, opcode);
1846 break;
1847 }
1848 case Op_VectorizedHashCode:
1849 case Op_AryEq:
1850 case Op_CountPositives:
1851 case Op_StrComp:
1852 case Op_StrEquals:
1853 case Op_StrIndexOf:
1854 case Op_StrIndexOfChar:
1855 case Op_StrInflatedCopy:
1856 case Op_StrCompressedCopy:
1857 case Op_EncodeISOArray: {
1858 // char[]/byte[] arrays passed to string intrinsic do not escape but
1859 // they are not scalar replaceable. Adjust escape state for them.
1860 // Start from in(2) edge since in(1) is memory edge.
1861 for (uint i = 2; i < n->req(); i++) {
1862 Node* adr = n->in(i);
1863 const Type* at = _igvn->type(adr);
1864 if (!adr->is_top() && at->isa_ptr()) {
1865 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1866 at->isa_ptr() != nullptr, "expecting a pointer");
1867 if (adr->is_AddP()) {
1868 adr = get_addp_base(adr);
1869 }
1870 PointsToNode* ptn = ptnode_adr(adr->_idx);
1871 assert(ptn != nullptr, "node should be registered");
1872 add_edge(n_ptn, ptn);
1873 }
1874 }
1875 break;
1876 }
1877 case Op_Blackhole: {
1878 // All blackhole pointer arguments are globally escaping.
1879 for (uint i = 0; i < n->req(); i++) {
1880 Node* in = n->in(i);
1881 if (in != nullptr) {
1882 const Type* at = _igvn->type(in);
1883 if (!at->isa_ptr()) continue;
1884
1885 if (in->is_AddP()) {
1886 in = get_addp_base(in);
1887 }
1888
1889 PointsToNode* ptn = ptnode_adr(in->_idx);
1890 assert(ptn != nullptr, "should be defined already");
1891 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1892 add_edge(n_ptn, ptn);
1893 }
1894 }
1895 break;
1896 }
1897 default: {
1898 // This method should be called only for EA specific nodes which may
1899 // miss some edges when they were created.
1900 #ifdef ASSERT
1901 n->dump(1);
1902 #endif
1903 guarantee(false, "unknown node");
1904 }
1905 }
1906 return;
1907 }
1908
1909 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1910 Node* adr = n->in(MemNode::Address);
1911 const Type* adr_type = _igvn->type(adr);
1912 adr_type = adr_type->make_ptr();
1913 if (adr_type == nullptr) {
1914 return; // skip dead nodes
1915 }
1916 if (adr_type->isa_oopptr()
1917 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1918 && adr_type == TypeRawPtr::NOTNULL
1919 && is_captured_store_address(adr))) {
1920 delayed_worklist->push(n); // Process it later.
1921 #ifdef ASSERT
1922 assert (adr->is_AddP(), "expecting an AddP");
1923 if (adr_type == TypeRawPtr::NOTNULL) {
1924 // Verify a raw address for a store captured by Initialize node.
1925 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1926 assert(offs != Type::OffsetBot, "offset must be a constant");
1927 }
1928 #endif
1929 } else {
1930 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1931 if (adr->is_BoxLock()) {
1932 return;
1933 }
1934 // Stored value escapes in unsafe access.
1935 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1936 delayed_worklist->push(n); // Process unsafe access later.
1937 return;
1938 }
1939 #ifdef ASSERT
1940 n->dump(1);
1941 assert(false, "not unsafe");
1942 #endif
1943 }
1944 }
1945
1946 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1947 Node* adr = n->in(MemNode::Address);
1948 const Type *adr_type = _igvn->type(adr);
1949 adr_type = adr_type->make_ptr();
1950 #ifdef ASSERT
1951 if (adr_type == nullptr) {
1952 n->dump(1);
1953 assert(adr_type != nullptr, "dead node should not be on list");
1954 return true;
1955 }
1956 #endif
1957
1958 if (adr_type->isa_oopptr()
1959 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1960 && adr_type == TypeRawPtr::NOTNULL
1961 && is_captured_store_address(adr))) {
1962 // Point Address to Value
1963 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1964 assert(adr_ptn != nullptr &&
1965 adr_ptn->as_Field()->is_oop(), "node should be registered");
1966 Node* val = n->in(MemNode::ValueIn);
1967 PointsToNode* ptn = ptnode_adr(val->_idx);
1968 assert(ptn != nullptr, "node should be registered");
1969 add_edge(adr_ptn, ptn);
1970 return true;
1971 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1972 // Stored value escapes in unsafe access.
1973 Node* val = n->in(MemNode::ValueIn);
1974 PointsToNode* ptn = ptnode_adr(val->_idx);
1975 assert(ptn != nullptr, "node should be registered");
1976 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
1977 // Add edge to object for unsafe access with offset.
1978 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1979 assert(adr_ptn != nullptr, "node should be registered");
1980 if (adr_ptn->is_Field()) {
1981 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
1982 add_edge(adr_ptn, ptn);
1983 }
1984 return true;
1985 }
1986 #ifdef ASSERT
1987 n->dump(1);
1988 assert(false, "not unsafe");
1989 #endif
1990 return false;
1991 }
1992
1993 void ConnectionGraph::add_call_node(CallNode* call) {
1994 assert(call->returns_pointer(), "only for call which returns pointer");
1995 uint call_idx = call->_idx;
1996 if (call->is_Allocate()) {
1997 Node* k = call->in(AllocateNode::KlassNode);
1998 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
1999 assert(kt != nullptr, "TypeKlassPtr required.");
2000 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2001 bool scalar_replaceable = true;
2002 NOT_PRODUCT(const char* nsr_reason = "");
2003 if (call->is_AllocateArray()) {
2004 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2005 es = PointsToNode::GlobalEscape;
2006 } else {
2007 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2008 if (length < 0) {
2009 // Not scalar replaceable if the length is not constant.
2010 scalar_replaceable = false;
2011 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2012 } else if (length > EliminateAllocationArraySizeLimit) {
2013 // Not scalar replaceable if the length is too big.
2014 scalar_replaceable = false;
2015 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2016 }
2017 }
2018 } else { // Allocate instance
2019 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2020 es = PointsToNode::GlobalEscape;
2021 } else {
2022 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2023 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2024 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2025 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2026 !ik->can_be_instantiated() ||
2027 ik->has_finalizer()) {
2028 es = PointsToNode::GlobalEscape;
2029 } else {
2030 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2031 if (nfields > EliminateAllocationFieldsLimit) {
2032 // Not scalar replaceable if there are too many fields.
2033 scalar_replaceable = false;
2034 NOT_PRODUCT(nsr_reason = "has too many fields");
2035 }
2036 }
2037 }
2038 }
2039 add_java_object(call, es);
2040 PointsToNode* ptn = ptnode_adr(call_idx);
2041 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2042 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2043 }
2044 } else if (call->is_CallStaticJava()) {
2045 // Call nodes could be different types:
2046 //
2047 // 1. CallDynamicJavaNode (what happened during call is unknown):
2048 //
2049 // - mapped to GlobalEscape JavaObject node if oop is returned;
2050 //
2051 // - all oop arguments are escaping globally;
2052 //
2053 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2054 //
2055 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2056 //
2057 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2058 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2059 // during call is returned;
2060 // - mapped to ArgEscape LocalVar node pointed to object arguments
2061 // which are returned and does not escape during call;
2062 //
2063 // - oop arguments escaping status is defined by bytecode analysis;
2064 //
2065 // For a static call, we know exactly what method is being called.
2066 // Use bytecode estimator to record whether the call's return value escapes.
2067 ciMethod* meth = call->as_CallJava()->method();
2068 if (meth == nullptr) {
2069 const char* name = call->as_CallStaticJava()->_name;
2070 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0, "TODO: add failed case check");
2071 // Returns a newly allocated non-escaped object.
2072 add_java_object(call, PointsToNode::NoEscape);
2073 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2074 } else if (meth->is_boxing_method()) {
2075 // Returns boxing object
2076 PointsToNode::EscapeState es;
2077 vmIntrinsics::ID intr = meth->intrinsic_id();
2078 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2079 // It does not escape if object is always allocated.
2080 es = PointsToNode::NoEscape;
2081 } else {
2082 // It escapes globally if object could be loaded from cache.
2083 es = PointsToNode::GlobalEscape;
2084 }
2085 add_java_object(call, es);
2086 if (es == PointsToNode::GlobalEscape) {
2087 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2088 }
2089 } else {
2090 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2091 call_analyzer->copy_dependencies(_compile->dependencies());
2092 if (call_analyzer->is_return_allocated()) {
2093 // Returns a newly allocated non-escaped object, simply
2094 // update dependency information.
2095 // Mark it as NoEscape so that objects referenced by
2096 // it's fields will be marked as NoEscape at least.
2097 add_java_object(call, PointsToNode::NoEscape);
2098 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2099 } else {
2100 // Determine whether any arguments are returned.
2101 const TypeTuple* d = call->tf()->domain();
2102 bool ret_arg = false;
2103 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2104 if (d->field_at(i)->isa_ptr() != nullptr &&
2105 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2106 ret_arg = true;
2107 break;
2108 }
2109 }
2110 if (ret_arg) {
2111 add_local_var(call, PointsToNode::ArgEscape);
2112 } else {
2113 // Returns unknown object.
2114 map_ideal_node(call, phantom_obj);
2115 }
2116 }
2117 }
2118 } else {
2119 // An other type of call, assume the worst case:
2120 // returned value is unknown and globally escapes.
2121 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2122 map_ideal_node(call, phantom_obj);
2123 }
2124 }
2125
2126 void ConnectionGraph::process_call_arguments(CallNode *call) {
2127 bool is_arraycopy = false;
2128 switch (call->Opcode()) {
2129 #ifdef ASSERT
2130 case Op_Allocate:
2131 case Op_AllocateArray:
2132 case Op_Lock:
2133 case Op_Unlock:
2134 assert(false, "should be done already");
2135 break;
2136 #endif
2137 case Op_ArrayCopy:
2138 case Op_CallLeafNoFP:
2139 // Most array copies are ArrayCopy nodes at this point but there
2140 // are still a few direct calls to the copy subroutines (See
2141 // PhaseStringOpts::copy_string())
2142 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2143 call->as_CallLeaf()->is_call_to_arraycopystub();
2144 // fall through
2145 case Op_CallLeafVector:
2146 case Op_CallLeaf: {
2147 // Stub calls, objects do not escape but they are not scale replaceable.
2148 // Adjust escape state for outgoing arguments.
2149 const TypeTuple * d = call->tf()->domain();
2150 bool src_has_oops = false;
2151 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2152 const Type* at = d->field_at(i);
2153 Node *arg = call->in(i);
2154 if (arg == nullptr) {
2155 continue;
2156 }
2157 const Type *aat = _igvn->type(arg);
2158 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2159 continue;
2160 }
2161 if (arg->is_AddP()) {
2162 //
2163 // The inline_native_clone() case when the arraycopy stub is called
2164 // after the allocation before Initialize and CheckCastPP nodes.
2165 // Or normal arraycopy for object arrays case.
2166 //
2167 // Set AddP's base (Allocate) as not scalar replaceable since
2168 // pointer to the base (with offset) is passed as argument.
2169 //
2170 arg = get_addp_base(arg);
2171 }
2172 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2173 assert(arg_ptn != nullptr, "should be registered");
2174 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2175 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2176 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2177 aat->isa_ptr() != nullptr, "expecting an Ptr");
2178 bool arg_has_oops = aat->isa_oopptr() &&
2179 (aat->isa_instptr() ||
2180 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)));
2181 if (i == TypeFunc::Parms) {
2182 src_has_oops = arg_has_oops;
2183 }
2184 //
2185 // src or dst could be j.l.Object when other is basic type array:
2186 //
2187 // arraycopy(char[],0,Object*,0,size);
2188 // arraycopy(Object*,0,char[],0,size);
2189 //
2190 // Don't add edges in such cases.
2191 //
2192 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2193 arg_has_oops && (i > TypeFunc::Parms);
2194 #ifdef ASSERT
2195 if (!(is_arraycopy ||
2196 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2197 (call->as_CallLeaf()->_name != nullptr &&
2198 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2199 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2200 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2201 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2202 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2203 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2204 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2205 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2206 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2207 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2208 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2209 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2210 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2211 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2212 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2213 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2214 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2215 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2216 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2217 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2218 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2219 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2220 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2221 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2222 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2223 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2224 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2225 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2226 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2227 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2228 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2229 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2230 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2231 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2232 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2233 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2234 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2235 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2236 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2237 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2238 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2239 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2240 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2241 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2242 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2243 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2244 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2245 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2246 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2247 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2248 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2249 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2250 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2251 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2252 ))) {
2253 call->dump();
2254 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2255 }
2256 #endif
2257 // Always process arraycopy's destination object since
2258 // we need to add all possible edges to references in
2259 // source object.
2260 if (arg_esc >= PointsToNode::ArgEscape &&
2261 !arg_is_arraycopy_dest) {
2262 continue;
2263 }
2264 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2265 if (call->is_ArrayCopy()) {
2266 ArrayCopyNode* ac = call->as_ArrayCopy();
2267 if (ac->is_clonebasic() ||
2268 ac->is_arraycopy_validated() ||
2269 ac->is_copyof_validated() ||
2270 ac->is_copyofrange_validated()) {
2271 es = PointsToNode::NoEscape;
2272 }
2273 }
2274 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2275 if (arg_is_arraycopy_dest) {
2276 Node* src = call->in(TypeFunc::Parms);
2277 if (src->is_AddP()) {
2278 src = get_addp_base(src);
2279 }
2280 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2281 assert(src_ptn != nullptr, "should be registered");
2282 // Special arraycopy edge:
2283 // Only escape state of destination object's fields affects
2284 // escape state of fields in source object.
2285 add_arraycopy(call, es, src_ptn, arg_ptn);
2286 }
2287 }
2288 }
2289 break;
2290 }
2291 case Op_CallStaticJava: {
2292 // For a static call, we know exactly what method is being called.
2293 // Use bytecode estimator to record the call's escape affects
2294 #ifdef ASSERT
2295 const char* name = call->as_CallStaticJava()->_name;
2296 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2297 #endif
2298 ciMethod* meth = call->as_CallJava()->method();
2299 if ((meth != nullptr) && meth->is_boxing_method()) {
2300 break; // Boxing methods do not modify any oops.
2301 }
2302 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2303 // fall-through if not a Java method or no analyzer information
2304 if (call_analyzer != nullptr) {
2305 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2306 const TypeTuple* d = call->tf()->domain();
2307 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2308 const Type* at = d->field_at(i);
2309 int k = i - TypeFunc::Parms;
2310 Node* arg = call->in(i);
2311 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2312 if (at->isa_ptr() != nullptr &&
2313 call_analyzer->is_arg_returned(k)) {
2314 // The call returns arguments.
2315 if (call_ptn != nullptr) { // Is call's result used?
2316 assert(call_ptn->is_LocalVar(), "node should be registered");
2317 assert(arg_ptn != nullptr, "node should be registered");
2318 add_edge(call_ptn, arg_ptn);
2319 }
2320 }
2321 if (at->isa_oopptr() != nullptr &&
2322 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2323 if (!call_analyzer->is_arg_stack(k)) {
2324 // The argument global escapes
2325 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2326 } else {
2327 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2328 if (!call_analyzer->is_arg_local(k)) {
2329 // The argument itself doesn't escape, but any fields might
2330 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2331 }
2332 }
2333 }
2334 }
2335 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2336 // The call returns arguments.
2337 assert(call_ptn->edge_count() > 0, "sanity");
2338 if (!call_analyzer->is_return_local()) {
2339 // Returns also unknown object.
2340 add_edge(call_ptn, phantom_obj);
2341 }
2342 }
2343 break;
2344 }
2345 }
2346 default: {
2347 // Fall-through here if not a Java method or no analyzer information
2348 // or some other type of call, assume the worst case: all arguments
2349 // globally escape.
2350 const TypeTuple* d = call->tf()->domain();
2351 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2352 const Type* at = d->field_at(i);
2353 if (at->isa_oopptr() != nullptr) {
2354 Node* arg = call->in(i);
2355 if (arg->is_AddP()) {
2356 arg = get_addp_base(arg);
2357 }
2358 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2359 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2360 }
2361 }
2362 }
2363 }
2364 }
2365
2366
2367 // Finish Graph construction.
2368 bool ConnectionGraph::complete_connection_graph(
2369 GrowableArray<PointsToNode*>& ptnodes_worklist,
2370 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2371 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2372 GrowableArray<FieldNode*>& oop_fields_worklist) {
2373 // Normally only 1-3 passes needed to build Connection Graph depending
2374 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2375 // Set limit to 20 to catch situation when something did go wrong and
2376 // bailout Escape Analysis.
2377 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2378 #define GRAPH_BUILD_ITER_LIMIT 20
2379
2380 // Propagate GlobalEscape and ArgEscape escape states and check that
2381 // we still have non-escaping objects. The method pushs on _worklist
2382 // Field nodes which reference phantom_object.
2383 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2384 return false; // Nothing to do.
2385 }
2386 // Now propagate references to all JavaObject nodes.
2387 int java_objects_length = java_objects_worklist.length();
2388 elapsedTimer build_time;
2389 build_time.start();
2390 elapsedTimer time;
2391 bool timeout = false;
2392 int new_edges = 1;
2393 int iterations = 0;
2394 do {
2395 while ((new_edges > 0) &&
2396 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2397 double start_time = time.seconds();
2398 time.start();
2399 new_edges = 0;
2400 // Propagate references to phantom_object for nodes pushed on _worklist
2401 // by find_non_escaped_objects() and find_field_value().
2402 new_edges += add_java_object_edges(phantom_obj, false);
2403 for (int next = 0; next < java_objects_length; ++next) {
2404 JavaObjectNode* ptn = java_objects_worklist.at(next);
2405 new_edges += add_java_object_edges(ptn, true);
2406
2407 #define SAMPLE_SIZE 4
2408 if ((next % SAMPLE_SIZE) == 0) {
2409 // Each 4 iterations calculate how much time it will take
2410 // to complete graph construction.
2411 time.stop();
2412 // Poll for requests from shutdown mechanism to quiesce compiler
2413 // because Connection graph construction may take long time.
2414 CompileBroker::maybe_block();
2415 double stop_time = time.seconds();
2416 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2417 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2418 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2419 timeout = true;
2420 break; // Timeout
2421 }
2422 start_time = stop_time;
2423 time.start();
2424 }
2425 #undef SAMPLE_SIZE
2426
2427 }
2428 if (timeout) break;
2429 if (new_edges > 0) {
2430 // Update escape states on each iteration if graph was updated.
2431 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2432 return false; // Nothing to do.
2433 }
2434 }
2435 time.stop();
2436 if (time.seconds() >= EscapeAnalysisTimeout) {
2437 timeout = true;
2438 break;
2439 }
2440 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2441 }
2442 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2443 time.start();
2444 // Find fields which have unknown value.
2445 int fields_length = oop_fields_worklist.length();
2446 for (int next = 0; next < fields_length; next++) {
2447 FieldNode* field = oop_fields_worklist.at(next);
2448 if (field->edge_count() == 0) {
2449 new_edges += find_field_value(field);
2450 // This code may added new edges to phantom_object.
2451 // Need an other cycle to propagate references to phantom_object.
2452 }
2453 }
2454 time.stop();
2455 if (time.seconds() >= EscapeAnalysisTimeout) {
2456 timeout = true;
2457 break;
2458 }
2459 } else {
2460 new_edges = 0; // Bailout
2461 }
2462 } while (new_edges > 0);
2463
2464 build_time.stop();
2465 _build_time = build_time.seconds();
2466 _build_iterations = iterations;
2467
2468 // Bailout if passed limits.
2469 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2470 Compile* C = _compile;
2471 if (C->log() != nullptr) {
2472 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2473 C->log()->text("%s", timeout ? "time" : "iterations");
2474 C->log()->end_elem(" limit'");
2475 }
2476 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2477 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2478 // Possible infinite build_connection_graph loop,
2479 // bailout (no changes to ideal graph were made).
2480 return false;
2481 }
2482
2483 #undef GRAPH_BUILD_ITER_LIMIT
2484
2485 // Find fields initialized by null for non-escaping Allocations.
2486 int non_escaped_length = non_escaped_allocs_worklist.length();
2487 for (int next = 0; next < non_escaped_length; next++) {
2488 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2489 PointsToNode::EscapeState es = ptn->escape_state();
2490 assert(es <= PointsToNode::ArgEscape, "sanity");
2491 if (es == PointsToNode::NoEscape) {
2492 if (find_init_values_null(ptn, _igvn) > 0) {
2493 // Adding references to null object does not change escape states
2494 // since it does not escape. Also no fields are added to null object.
2495 add_java_object_edges(null_obj, false);
2496 }
2497 }
2498 Node* n = ptn->ideal_node();
2499 if (n->is_Allocate()) {
2500 // The object allocated by this Allocate node will never be
2501 // seen by an other thread. Mark it so that when it is
2502 // expanded no MemBarStoreStore is added.
2503 InitializeNode* ini = n->as_Allocate()->initialization();
2504 if (ini != nullptr)
2505 ini->set_does_not_escape();
2506 }
2507 }
2508 return true; // Finished graph construction.
2509 }
2510
2511 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2512 // and check that we still have non-escaping java objects.
2513 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2514 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2515 bool print_method) {
2516 GrowableArray<PointsToNode*> escape_worklist;
2517 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2518 int ptnodes_length = ptnodes_worklist.length();
2519 for (int next = 0; next < ptnodes_length; ++next) {
2520 PointsToNode* ptn = ptnodes_worklist.at(next);
2521 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2522 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2523 escape_worklist.push(ptn);
2524 }
2525 }
2526 // Set escape states to referenced nodes (edges list).
2527 while (escape_worklist.length() > 0) {
2528 PointsToNode* ptn = escape_worklist.pop();
2529 PointsToNode::EscapeState es = ptn->escape_state();
2530 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2531 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2532 es >= PointsToNode::ArgEscape) {
2533 // GlobalEscape or ArgEscape state of field means it has unknown value.
2534 if (add_edge(ptn, phantom_obj)) {
2535 // New edge was added
2536 add_field_uses_to_worklist(ptn->as_Field());
2537 }
2538 }
2539 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2540 PointsToNode* e = i.get();
2541 if (e->is_Arraycopy()) {
2542 assert(ptn->arraycopy_dst(), "sanity");
2543 // Propagate only fields escape state through arraycopy edge.
2544 if (e->fields_escape_state() < field_es) {
2545 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2546 escape_worklist.push(e);
2547 }
2548 } else if (es >= field_es) {
2549 // fields_escape_state is also set to 'es' if it is less than 'es'.
2550 if (e->escape_state() < es) {
2551 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2552 escape_worklist.push(e);
2553 }
2554 } else {
2555 // Propagate field escape state.
2556 bool es_changed = false;
2557 if (e->fields_escape_state() < field_es) {
2558 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2559 es_changed = true;
2560 }
2561 if ((e->escape_state() < field_es) &&
2562 e->is_Field() && ptn->is_JavaObject() &&
2563 e->as_Field()->is_oop()) {
2564 // Change escape state of referenced fields.
2565 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2566 es_changed = true;
2567 } else if (e->escape_state() < es) {
2568 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2569 es_changed = true;
2570 }
2571 if (es_changed) {
2572 escape_worklist.push(e);
2573 }
2574 }
2575 if (print_method) {
2576 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2577 }
2578 }
2579 }
2580 // Remove escaped objects from non_escaped list.
2581 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2582 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2583 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2584 non_escaped_allocs_worklist.delete_at(next);
2585 }
2586 if (ptn->escape_state() == PointsToNode::NoEscape) {
2587 // Find fields in non-escaped allocations which have unknown value.
2588 find_init_values_phantom(ptn);
2589 }
2590 }
2591 return (non_escaped_allocs_worklist.length() > 0);
2592 }
2593
2594 // Add all references to JavaObject node by walking over all uses.
2595 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2596 int new_edges = 0;
2597 if (populate_worklist) {
2598 // Populate _worklist by uses of jobj's uses.
2599 for (UseIterator i(jobj); i.has_next(); i.next()) {
2600 PointsToNode* use = i.get();
2601 if (use->is_Arraycopy()) {
2602 continue;
2603 }
2604 add_uses_to_worklist(use);
2605 if (use->is_Field() && use->as_Field()->is_oop()) {
2606 // Put on worklist all field's uses (loads) and
2607 // related field nodes (same base and offset).
2608 add_field_uses_to_worklist(use->as_Field());
2609 }
2610 }
2611 }
2612 for (int l = 0; l < _worklist.length(); l++) {
2613 PointsToNode* use = _worklist.at(l);
2614 if (PointsToNode::is_base_use(use)) {
2615 // Add reference from jobj to field and from field to jobj (field's base).
2616 use = PointsToNode::get_use_node(use)->as_Field();
2617 if (add_base(use->as_Field(), jobj)) {
2618 new_edges++;
2619 }
2620 continue;
2621 }
2622 assert(!use->is_JavaObject(), "sanity");
2623 if (use->is_Arraycopy()) {
2624 if (jobj == null_obj) { // null object does not have field edges
2625 continue;
2626 }
2627 // Added edge from Arraycopy node to arraycopy's source java object
2628 if (add_edge(use, jobj)) {
2629 jobj->set_arraycopy_src();
2630 new_edges++;
2631 }
2632 // and stop here.
2633 continue;
2634 }
2635 if (!add_edge(use, jobj)) {
2636 continue; // No new edge added, there was such edge already.
2637 }
2638 new_edges++;
2639 if (use->is_LocalVar()) {
2640 add_uses_to_worklist(use);
2641 if (use->arraycopy_dst()) {
2642 for (EdgeIterator i(use); i.has_next(); i.next()) {
2643 PointsToNode* e = i.get();
2644 if (e->is_Arraycopy()) {
2645 if (jobj == null_obj) { // null object does not have field edges
2646 continue;
2647 }
2648 // Add edge from arraycopy's destination java object to Arraycopy node.
2649 if (add_edge(jobj, e)) {
2650 new_edges++;
2651 jobj->set_arraycopy_dst();
2652 }
2653 }
2654 }
2655 }
2656 } else {
2657 // Added new edge to stored in field values.
2658 // Put on worklist all field's uses (loads) and
2659 // related field nodes (same base and offset).
2660 add_field_uses_to_worklist(use->as_Field());
2661 }
2662 }
2663 _worklist.clear();
2664 _in_worklist.reset();
2665 return new_edges;
2666 }
2667
2668 // Put on worklist all related field nodes.
2669 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2670 assert(field->is_oop(), "sanity");
2671 int offset = field->offset();
2672 add_uses_to_worklist(field);
2673 // Loop over all bases of this field and push on worklist Field nodes
2674 // with the same offset and base (since they may reference the same field).
2675 for (BaseIterator i(field); i.has_next(); i.next()) {
2676 PointsToNode* base = i.get();
2677 add_fields_to_worklist(field, base);
2678 // Check if the base was source object of arraycopy and go over arraycopy's
2679 // destination objects since values stored to a field of source object are
2680 // accessible by uses (loads) of fields of destination objects.
2681 if (base->arraycopy_src()) {
2682 for (UseIterator j(base); j.has_next(); j.next()) {
2683 PointsToNode* arycp = j.get();
2684 if (arycp->is_Arraycopy()) {
2685 for (UseIterator k(arycp); k.has_next(); k.next()) {
2686 PointsToNode* abase = k.get();
2687 if (abase->arraycopy_dst() && abase != base) {
2688 // Look for the same arraycopy reference.
2689 add_fields_to_worklist(field, abase);
2690 }
2691 }
2692 }
2693 }
2694 }
2695 }
2696 }
2697
2698 // Put on worklist all related field nodes.
2699 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2700 int offset = field->offset();
2701 if (base->is_LocalVar()) {
2702 for (UseIterator j(base); j.has_next(); j.next()) {
2703 PointsToNode* f = j.get();
2704 if (PointsToNode::is_base_use(f)) { // Field
2705 f = PointsToNode::get_use_node(f);
2706 if (f == field || !f->as_Field()->is_oop()) {
2707 continue;
2708 }
2709 int offs = f->as_Field()->offset();
2710 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2711 add_to_worklist(f);
2712 }
2713 }
2714 }
2715 } else {
2716 assert(base->is_JavaObject(), "sanity");
2717 if (// Skip phantom_object since it is only used to indicate that
2718 // this field's content globally escapes.
2719 (base != phantom_obj) &&
2720 // null object node does not have fields.
2721 (base != null_obj)) {
2722 for (EdgeIterator i(base); i.has_next(); i.next()) {
2723 PointsToNode* f = i.get();
2724 // Skip arraycopy edge since store to destination object field
2725 // does not update value in source object field.
2726 if (f->is_Arraycopy()) {
2727 assert(base->arraycopy_dst(), "sanity");
2728 continue;
2729 }
2730 if (f == field || !f->as_Field()->is_oop()) {
2731 continue;
2732 }
2733 int offs = f->as_Field()->offset();
2734 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2735 add_to_worklist(f);
2736 }
2737 }
2738 }
2739 }
2740 }
2741
2742 // Find fields which have unknown value.
2743 int ConnectionGraph::find_field_value(FieldNode* field) {
2744 // Escaped fields should have init value already.
2745 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2746 int new_edges = 0;
2747 for (BaseIterator i(field); i.has_next(); i.next()) {
2748 PointsToNode* base = i.get();
2749 if (base->is_JavaObject()) {
2750 // Skip Allocate's fields which will be processed later.
2751 if (base->ideal_node()->is_Allocate()) {
2752 return 0;
2753 }
2754 assert(base == null_obj, "only null ptr base expected here");
2755 }
2756 }
2757 if (add_edge(field, phantom_obj)) {
2758 // New edge was added
2759 new_edges++;
2760 add_field_uses_to_worklist(field);
2761 }
2762 return new_edges;
2763 }
2764
2765 // Find fields initializing values for allocations.
2766 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2767 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2768 Node* alloc = pta->ideal_node();
2769
2770 // Do nothing for Allocate nodes since its fields values are
2771 // "known" unless they are initialized by arraycopy/clone.
2772 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2773 return 0;
2774 }
2775 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
2776 #ifdef ASSERT
2777 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == nullptr) {
2778 const char* name = alloc->as_CallStaticJava()->_name;
2779 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0, "sanity");
2780 }
2781 #endif
2782 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2783 int new_edges = 0;
2784 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2785 PointsToNode* field = i.get();
2786 if (field->is_Field() && field->as_Field()->is_oop()) {
2787 if (add_edge(field, phantom_obj)) {
2788 // New edge was added
2789 new_edges++;
2790 add_field_uses_to_worklist(field->as_Field());
2791 }
2792 }
2793 }
2794 return new_edges;
2795 }
2796
2797 // Find fields initializing values for allocations.
2798 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2799 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2800 Node* alloc = pta->ideal_node();
2801 // Do nothing for Call nodes since its fields values are unknown.
2802 if (!alloc->is_Allocate()) {
2803 return 0;
2804 }
2805 InitializeNode* ini = alloc->as_Allocate()->initialization();
2806 bool visited_bottom_offset = false;
2807 GrowableArray<int> offsets_worklist;
2808 int new_edges = 0;
2809
2810 // Check if an oop field's initializing value is recorded and add
2811 // a corresponding null if field's value if it is not recorded.
2812 // Connection Graph does not record a default initialization by null
2813 // captured by Initialize node.
2814 //
2815 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2816 PointsToNode* field = i.get(); // Field (AddP)
2817 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2818 continue; // Not oop field
2819 }
2820 int offset = field->as_Field()->offset();
2821 if (offset == Type::OffsetBot) {
2822 if (!visited_bottom_offset) {
2823 // OffsetBot is used to reference array's element,
2824 // always add reference to null to all Field nodes since we don't
2825 // known which element is referenced.
2826 if (add_edge(field, null_obj)) {
2827 // New edge was added
2828 new_edges++;
2829 add_field_uses_to_worklist(field->as_Field());
2830 visited_bottom_offset = true;
2831 }
2832 }
2833 } else {
2834 // Check only oop fields.
2835 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2836 if (adr_type->isa_rawptr()) {
2837 #ifdef ASSERT
2838 // Raw pointers are used for initializing stores so skip it
2839 // since it should be recorded already
2840 Node* base = get_addp_base(field->ideal_node());
2841 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2842 #endif
2843 continue;
2844 }
2845 if (!offsets_worklist.contains(offset)) {
2846 offsets_worklist.append(offset);
2847 Node* value = nullptr;
2848 if (ini != nullptr) {
2849 // StoreP::value_basic_type() == T_ADDRESS
2850 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2851 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2852 // Make sure initializing store has the same type as this AddP.
2853 // This AddP may reference non existing field because it is on a
2854 // dead branch of bimorphic call which is not eliminated yet.
2855 if (store != nullptr && store->is_Store() &&
2856 store->as_Store()->value_basic_type() == ft) {
2857 value = store->in(MemNode::ValueIn);
2858 #ifdef ASSERT
2859 if (VerifyConnectionGraph) {
2860 // Verify that AddP already points to all objects the value points to.
2861 PointsToNode* val = ptnode_adr(value->_idx);
2862 assert((val != nullptr), "should be processed already");
2863 PointsToNode* missed_obj = nullptr;
2864 if (val->is_JavaObject()) {
2865 if (!field->points_to(val->as_JavaObject())) {
2866 missed_obj = val;
2867 }
2868 } else {
2869 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2870 tty->print_cr("----------init store has invalid value -----");
2871 store->dump();
2872 val->dump();
2873 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2874 }
2875 for (EdgeIterator j(val); j.has_next(); j.next()) {
2876 PointsToNode* obj = j.get();
2877 if (obj->is_JavaObject()) {
2878 if (!field->points_to(obj->as_JavaObject())) {
2879 missed_obj = obj;
2880 break;
2881 }
2882 }
2883 }
2884 }
2885 if (missed_obj != nullptr) {
2886 tty->print_cr("----------field---------------------------------");
2887 field->dump();
2888 tty->print_cr("----------missed referernce to object-----------");
2889 missed_obj->dump();
2890 tty->print_cr("----------object referernced by init store -----");
2891 store->dump();
2892 val->dump();
2893 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2894 }
2895 }
2896 #endif
2897 } else {
2898 // There could be initializing stores which follow allocation.
2899 // For example, a volatile field store is not collected
2900 // by Initialize node.
2901 //
2902 // Need to check for dependent loads to separate such stores from
2903 // stores which follow loads. For now, add initial value null so
2904 // that compare pointers optimization works correctly.
2905 }
2906 }
2907 if (value == nullptr) {
2908 // A field's initializing value was not recorded. Add null.
2909 if (add_edge(field, null_obj)) {
2910 // New edge was added
2911 new_edges++;
2912 add_field_uses_to_worklist(field->as_Field());
2913 }
2914 }
2915 }
2916 }
2917 }
2918 return new_edges;
2919 }
2920
2921 // Adjust scalar_replaceable state after Connection Graph is built.
2922 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2923 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2924 // returns true. If one of the constraints in this method set 'jobj' to NSR
2925 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2926 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2927 // that other object and the Phi will become a reducible Phi.
2928 // There could be multiple merges involving the same jobj.
2929 Unique_Node_List candidates;
2930
2931 // Search for non-escaping objects which are not scalar replaceable
2932 // and mark them to propagate the state to referenced objects.
2933
2934 for (UseIterator i(jobj); i.has_next(); i.next()) {
2935 PointsToNode* use = i.get();
2936 if (use->is_Arraycopy()) {
2937 continue;
2938 }
2939 if (use->is_Field()) {
2940 FieldNode* field = use->as_Field();
2941 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2942 // 1. An object is not scalar replaceable if the field into which it is
2943 // stored has unknown offset (stored into unknown element of an array).
2944 if (field->offset() == Type::OffsetBot) {
2945 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2946 return;
2947 }
2948 for (BaseIterator i(field); i.has_next(); i.next()) {
2949 PointsToNode* base = i.get();
2950 // 2. An object is not scalar replaceable if the field into which it is
2951 // stored has multiple bases one of which is null.
2952 if ((base == null_obj) && (field->base_count() > 1)) {
2953 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2954 return;
2955 }
2956 // 2.5. An object is not scalar replaceable if the field into which it is
2957 // stored has NSR base.
2958 if (!base->scalar_replaceable()) {
2959 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2960 return;
2961 }
2962 }
2963 }
2964 assert(use->is_Field() || use->is_LocalVar(), "sanity");
2965 // 3. An object is not scalar replaceable if it is merged with other objects
2966 // and we can't remove the merge
2967 for (EdgeIterator j(use); j.has_next(); j.next()) {
2968 PointsToNode* ptn = j.get();
2969 if (ptn->is_JavaObject() && ptn != jobj) {
2970 Node* use_n = use->ideal_node();
2971
2972 // These other local vars may point to multiple objects through a Phi
2973 // In this case we skip them and see if we can reduce the Phi.
2974 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
2975 use_n = use_n->in(1);
2976 }
2977
2978 // If it's already a candidate or confirmed reducible merge we can skip verification
2979 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
2980 continue;
2981 }
2982
2983 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
2984 candidates.push(use_n);
2985 } else {
2986 // Mark all objects as NSR if we can't remove the merge
2987 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
2988 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
2989 }
2990 }
2991 }
2992 if (!jobj->scalar_replaceable()) {
2993 return;
2994 }
2995 }
2996
2997 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
2998 if (j.get()->is_Arraycopy()) {
2999 continue;
3000 }
3001
3002 // Non-escaping object node should point only to field nodes.
3003 FieldNode* field = j.get()->as_Field();
3004 int offset = field->as_Field()->offset();
3005
3006 // 4. An object is not scalar replaceable if it has a field with unknown
3007 // offset (array's element is accessed in loop).
3008 if (offset == Type::OffsetBot) {
3009 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3010 return;
3011 }
3012 // 5. Currently an object is not scalar replaceable if a LoadStore node
3013 // access its field since the field value is unknown after it.
3014 //
3015 Node* n = field->ideal_node();
3016
3017 // Test for an unsafe access that was parsed as maybe off heap
3018 // (with a CheckCastPP to raw memory).
3019 assert(n->is_AddP(), "expect an address computation");
3020 if (n->in(AddPNode::Base)->is_top() &&
3021 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3022 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3023 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3024 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3025 return;
3026 }
3027
3028 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3029 Node* u = n->fast_out(i);
3030 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3031 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3032 return;
3033 }
3034 }
3035
3036 // 6. Or the address may point to more then one object. This may produce
3037 // the false positive result (set not scalar replaceable)
3038 // since the flow-insensitive escape analysis can't separate
3039 // the case when stores overwrite the field's value from the case
3040 // when stores happened on different control branches.
3041 //
3042 // Note: it will disable scalar replacement in some cases:
3043 //
3044 // Point p[] = new Point[1];
3045 // p[0] = new Point(); // Will be not scalar replaced
3046 //
3047 // but it will save us from incorrect optimizations in next cases:
3048 //
3049 // Point p[] = new Point[1];
3050 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3051 //
3052 if (field->base_count() > 1 && candidates.size() == 0) {
3053 if (has_non_reducible_merge(field, reducible_merges)) {
3054 for (BaseIterator i(field); i.has_next(); i.next()) {
3055 PointsToNode* base = i.get();
3056 // Don't take into account LocalVar nodes which
3057 // may point to only one object which should be also
3058 // this field's base by now.
3059 if (base->is_JavaObject() && base != jobj) {
3060 // Mark all bases.
3061 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3062 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3063 }
3064 }
3065
3066 if (!jobj->scalar_replaceable()) {
3067 return;
3068 }
3069 }
3070 }
3071 }
3072
3073 // The candidate is truly a reducible merge only if none of the other
3074 // constraints ruled it as NSR. There could be multiple merges involving the
3075 // same jobj.
3076 assert(jobj->scalar_replaceable(), "sanity");
3077 for (uint i = 0; i < candidates.size(); i++ ) {
3078 Node* candidate = candidates.at(i);
3079 reducible_merges.push(candidate);
3080 }
3081 }
3082
3083 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3084 for (BaseIterator i(field); i.has_next(); i.next()) {
3085 Node* base = i.get()->ideal_node();
3086 if (base->is_Phi() && !reducible_merges.member(base)) {
3087 return true;
3088 }
3089 }
3090 return false;
3091 }
3092
3093 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3094 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3095
3096 // Look for 'phis' that refer to 'jobj' as the last
3097 // remaining scalar replaceable input.
3098 uint reducible_merges_cnt = reducible_merges.size();
3099 for (uint i = 0; i < reducible_merges_cnt; i++) {
3100 Node* phi = reducible_merges.at(i);
3101
3102 // This 'Phi' will be a 'good' if it still points to
3103 // at least one scalar replaceable object. Note that 'obj'
3104 // was/should be marked as NSR before calling this function.
3105 bool good_phi = false;
3106
3107 for (uint j = 1; j < phi->req(); j++) {
3108 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3109 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3110 good_phi = true;
3111 break;
3112 }
3113 }
3114
3115 if (!good_phi) {
3116 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3117 reducible_merges.remove(i);
3118
3119 // Decrement the index because the 'remove' call above actually
3120 // moves the last entry of the list to position 'i'.
3121 i--;
3122
3123 reducible_merges_cnt--;
3124 }
3125 }
3126 }
3127
3128 // Propagate NSR (Not scalar replaceable) state.
3129 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3130 int jobj_length = jobj_worklist.length();
3131 bool found_nsr_alloc = true;
3132 while (found_nsr_alloc) {
3133 found_nsr_alloc = false;
3134 for (int next = 0; next < jobj_length; ++next) {
3135 JavaObjectNode* jobj = jobj_worklist.at(next);
3136 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3137 PointsToNode* use = i.get();
3138 if (use->is_Field()) {
3139 FieldNode* field = use->as_Field();
3140 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3141 assert(field->offset() != Type::OffsetBot, "sanity");
3142 for (BaseIterator i(field); i.has_next(); i.next()) {
3143 PointsToNode* base = i.get();
3144 // An object is not scalar replaceable if the field into which
3145 // it is stored has NSR base.
3146 if ((base != null_obj) && !base->scalar_replaceable()) {
3147 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3148 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3149 // because there is no point in reducing a Phi that won't improve the number of SR
3150 // objects.
3151 revisit_reducible_phi_status(jobj, reducible_merges);
3152 found_nsr_alloc = true;
3153 break;
3154 }
3155 }
3156 } else if (use->is_LocalVar()) {
3157 Node* phi = use->ideal_node();
3158 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3159 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3160 reducible_merges.yank(phi);
3161 found_nsr_alloc = true;
3162 break;
3163 }
3164 }
3165 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3166 }
3167 }
3168 }
3169 }
3170
3171 #ifdef ASSERT
3172 void ConnectionGraph::verify_connection_graph(
3173 GrowableArray<PointsToNode*>& ptnodes_worklist,
3174 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3175 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3176 GrowableArray<Node*>& addp_worklist) {
3177 // Verify that graph is complete - no new edges could be added.
3178 int java_objects_length = java_objects_worklist.length();
3179 int non_escaped_length = non_escaped_allocs_worklist.length();
3180 int new_edges = 0;
3181 for (int next = 0; next < java_objects_length; ++next) {
3182 JavaObjectNode* ptn = java_objects_worklist.at(next);
3183 new_edges += add_java_object_edges(ptn, true);
3184 }
3185 assert(new_edges == 0, "graph was not complete");
3186 // Verify that escape state is final.
3187 int length = non_escaped_allocs_worklist.length();
3188 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3189 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3190 (non_escaped_length == length) &&
3191 (_worklist.length() == 0), "escape state was not final");
3192
3193 // Verify fields information.
3194 int addp_length = addp_worklist.length();
3195 for (int next = 0; next < addp_length; ++next ) {
3196 Node* n = addp_worklist.at(next);
3197 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3198 if (field->is_oop()) {
3199 // Verify that field has all bases
3200 Node* base = get_addp_base(n);
3201 PointsToNode* ptn = ptnode_adr(base->_idx);
3202 if (ptn->is_JavaObject()) {
3203 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3204 } else {
3205 assert(ptn->is_LocalVar(), "sanity");
3206 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3207 PointsToNode* e = i.get();
3208 if (e->is_JavaObject()) {
3209 assert(field->has_base(e->as_JavaObject()), "sanity");
3210 }
3211 }
3212 }
3213 // Verify that all fields have initializing values.
3214 if (field->edge_count() == 0) {
3215 tty->print_cr("----------field does not have references----------");
3216 field->dump();
3217 for (BaseIterator i(field); i.has_next(); i.next()) {
3218 PointsToNode* base = i.get();
3219 tty->print_cr("----------field has next base---------------------");
3220 base->dump();
3221 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3222 tty->print_cr("----------base has fields-------------------------");
3223 for (EdgeIterator j(base); j.has_next(); j.next()) {
3224 j.get()->dump();
3225 }
3226 tty->print_cr("----------base has references---------------------");
3227 for (UseIterator j(base); j.has_next(); j.next()) {
3228 j.get()->dump();
3229 }
3230 }
3231 }
3232 for (UseIterator i(field); i.has_next(); i.next()) {
3233 i.get()->dump();
3234 }
3235 assert(field->edge_count() > 0, "sanity");
3236 }
3237 }
3238 }
3239 }
3240 #endif
3241
3242 // Optimize ideal graph.
3243 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3244 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3245 Compile* C = _compile;
3246 PhaseIterGVN* igvn = _igvn;
3247 if (EliminateLocks) {
3248 // Mark locks before changing ideal graph.
3249 int cnt = C->macro_count();
3250 for (int i = 0; i < cnt; i++) {
3251 Node *n = C->macro_node(i);
3252 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3253 AbstractLockNode* alock = n->as_AbstractLock();
3254 if (!alock->is_non_esc_obj()) {
3255 if (can_eliminate_lock(alock)) {
3256 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3257 // The lock could be marked eliminated by lock coarsening
3258 // code during first IGVN before EA. Replace coarsened flag
3259 // to eliminate all associated locks/unlocks.
3260 #ifdef ASSERT
3261 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3262 #endif
3263 alock->set_non_esc_obj();
3264 }
3265 }
3266 }
3267 }
3268 }
3269
3270 if (OptimizePtrCompare) {
3271 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3272 Node *n = ptr_cmp_worklist.at(i);
3273 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3274 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3275 if (tcmp->singleton()) {
3276 Node* cmp = igvn->makecon(tcmp);
3277 #ifndef PRODUCT
3278 if (PrintOptimizePtrCompare) {
3279 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"));
3280 if (Verbose) {
3281 n->dump(1);
3282 }
3283 }
3284 #endif
3285 igvn->replace_node(n, cmp);
3286 }
3287 }
3288 }
3289
3290 // For MemBarStoreStore nodes added in library_call.cpp, check
3291 // escape status of associated AllocateNode and optimize out
3292 // MemBarStoreStore node if the allocated object never escapes.
3293 for (int i = 0; i < storestore_worklist.length(); i++) {
3294 Node* storestore = storestore_worklist.at(i);
3295 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3296 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3297 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3298 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3299 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3300 igvn->register_new_node_with_optimizer(mb);
3301 igvn->replace_node(storestore, mb);
3302 }
3303 }
3304 }
3305
3306 // Optimize objects compare.
3307 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3308 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3309 if (!OptimizePtrCompare) {
3310 return UNKNOWN;
3311 }
3312 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3313 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3314
3315 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3316 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3317 JavaObjectNode* jobj1 = unique_java_object(left);
3318 JavaObjectNode* jobj2 = unique_java_object(right);
3319
3320 // The use of this method during allocation merge reduction may cause 'left'
3321 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3322 // that doesn't reference an unique java object.
3323 if (ptn1 == nullptr || ptn2 == nullptr ||
3324 jobj1 == nullptr || jobj2 == nullptr) {
3325 return UNKNOWN;
3326 }
3327
3328 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3329 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3330
3331 // Check simple cases first.
3332 if (jobj1 != nullptr) {
3333 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3334 if (jobj1 == jobj2) {
3335 // Comparing the same not escaping object.
3336 return EQ;
3337 }
3338 Node* obj = jobj1->ideal_node();
3339 // Comparing not escaping allocation.
3340 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3341 !ptn2->points_to(jobj1)) {
3342 return NE; // This includes nullness check.
3343 }
3344 }
3345 }
3346 if (jobj2 != nullptr) {
3347 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3348 Node* obj = jobj2->ideal_node();
3349 // Comparing not escaping allocation.
3350 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3351 !ptn1->points_to(jobj2)) {
3352 return NE; // This includes nullness check.
3353 }
3354 }
3355 }
3356 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3357 jobj2 != nullptr && jobj2 != phantom_obj &&
3358 jobj1->ideal_node()->is_Con() &&
3359 jobj2->ideal_node()->is_Con()) {
3360 // Klass or String constants compare. Need to be careful with
3361 // compressed pointers - compare types of ConN and ConP instead of nodes.
3362 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3363 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3364 if (t1->make_ptr() == t2->make_ptr()) {
3365 return EQ;
3366 } else {
3367 return NE;
3368 }
3369 }
3370 if (ptn1->meet(ptn2)) {
3371 return UNKNOWN; // Sets are not disjoint
3372 }
3373
3374 // Sets are disjoint.
3375 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3376 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3377 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3378 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3379 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3380 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3381 // Check nullness of unknown object.
3382 return UNKNOWN;
3383 }
3384
3385 // Disjointness by itself is not sufficient since
3386 // alias analysis is not complete for escaped objects.
3387 // Disjoint sets are definitely unrelated only when
3388 // at least one set has only not escaping allocations.
3389 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3390 if (ptn1->non_escaping_allocation()) {
3391 return NE;
3392 }
3393 }
3394 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3395 if (ptn2->non_escaping_allocation()) {
3396 return NE;
3397 }
3398 }
3399 return UNKNOWN;
3400 }
3401
3402 // Connection Graph construction functions.
3403
3404 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3405 PointsToNode* ptadr = _nodes.at(n->_idx);
3406 if (ptadr != nullptr) {
3407 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3408 return;
3409 }
3410 Compile* C = _compile;
3411 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3412 map_ideal_node(n, ptadr);
3413 }
3414
3415 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3416 PointsToNode* ptadr = _nodes.at(n->_idx);
3417 if (ptadr != nullptr) {
3418 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3419 return ptadr;
3420 }
3421 Compile* C = _compile;
3422 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3423 map_ideal_node(n, ptadr);
3424 return ptadr;
3425 }
3426
3427 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3428 PointsToNode* ptadr = _nodes.at(n->_idx);
3429 if (ptadr != nullptr) {
3430 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3431 return;
3432 }
3433 bool unsafe = false;
3434 bool is_oop = is_oop_field(n, offset, &unsafe);
3435 if (unsafe) {
3436 es = PointsToNode::GlobalEscape;
3437 }
3438 Compile* C = _compile;
3439 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3440 map_ideal_node(n, field);
3441 }
3442
3443 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3444 PointsToNode* src, PointsToNode* dst) {
3445 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3446 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3447 PointsToNode* ptadr = _nodes.at(n->_idx);
3448 if (ptadr != nullptr) {
3449 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3450 return;
3451 }
3452 Compile* C = _compile;
3453 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3454 map_ideal_node(n, ptadr);
3455 // Add edge from arraycopy node to source object.
3456 (void)add_edge(ptadr, src);
3457 src->set_arraycopy_src();
3458 // Add edge from destination object to arraycopy node.
3459 (void)add_edge(dst, ptadr);
3460 dst->set_arraycopy_dst();
3461 }
3462
3463 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3464 const Type* adr_type = n->as_AddP()->bottom_type();
3465 BasicType bt = T_INT;
3466 if (offset == Type::OffsetBot) {
3467 // Check only oop fields.
3468 if (!adr_type->isa_aryptr() ||
3469 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3470 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3471 // OffsetBot is used to reference array's element. Ignore first AddP.
3472 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3473 bt = T_OBJECT;
3474 }
3475 }
3476 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3477 if (adr_type->isa_instptr()) {
3478 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
3479 if (field != nullptr) {
3480 bt = field->layout_type();
3481 } else {
3482 // Check for unsafe oop field access
3483 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3484 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3485 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3486 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3487 bt = T_OBJECT;
3488 (*unsafe) = true;
3489 }
3490 }
3491 } else if (adr_type->isa_aryptr()) {
3492 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3493 // Ignore array length load.
3494 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3495 // Ignore first AddP.
3496 } else {
3497 const Type* elemtype = adr_type->isa_aryptr()->elem();
3498 bt = elemtype->array_element_basic_type();
3499 }
3500 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3501 // Allocation initialization, ThreadLocal field access, unsafe access
3502 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3503 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3504 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3505 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3506 bt = T_OBJECT;
3507 }
3508 }
3509 }
3510 // Note: T_NARROWOOP is not classed as a real reference type
3511 return (is_reference_type(bt) || bt == T_NARROWOOP);
3512 }
3513
3514 // Returns unique pointed java object or null.
3515 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3516 // If the node was created after the escape computation we can't answer.
3517 uint idx = n->_idx;
3518 if (idx >= nodes_size()) {
3519 return nullptr;
3520 }
3521 PointsToNode* ptn = ptnode_adr(idx);
3522 if (ptn == nullptr) {
3523 return nullptr;
3524 }
3525 if (ptn->is_JavaObject()) {
3526 return ptn->as_JavaObject();
3527 }
3528 assert(ptn->is_LocalVar(), "sanity");
3529 // Check all java objects it points to.
3530 JavaObjectNode* jobj = nullptr;
3531 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3532 PointsToNode* e = i.get();
3533 if (e->is_JavaObject()) {
3534 if (jobj == nullptr) {
3535 jobj = e->as_JavaObject();
3536 } else if (jobj != e) {
3537 return nullptr;
3538 }
3539 }
3540 }
3541 return jobj;
3542 }
3543
3544 // Return true if this node points only to non-escaping allocations.
3545 bool PointsToNode::non_escaping_allocation() {
3546 if (is_JavaObject()) {
3547 Node* n = ideal_node();
3548 if (n->is_Allocate() || n->is_CallStaticJava()) {
3549 return (escape_state() == PointsToNode::NoEscape);
3550 } else {
3551 return false;
3552 }
3553 }
3554 assert(is_LocalVar(), "sanity");
3555 // Check all java objects it points to.
3556 for (EdgeIterator i(this); i.has_next(); i.next()) {
3557 PointsToNode* e = i.get();
3558 if (e->is_JavaObject()) {
3559 Node* n = e->ideal_node();
3560 if ((e->escape_state() != PointsToNode::NoEscape) ||
3561 !(n->is_Allocate() || n->is_CallStaticJava())) {
3562 return false;
3563 }
3564 }
3565 }
3566 return true;
3567 }
3568
3569 // Return true if we know the node does not escape globally.
3570 bool ConnectionGraph::not_global_escape(Node *n) {
3571 assert(!_collecting, "should not call during graph construction");
3572 // If the node was created after the escape computation we can't answer.
3573 uint idx = n->_idx;
3574 if (idx >= nodes_size()) {
3575 return false;
3576 }
3577 PointsToNode* ptn = ptnode_adr(idx);
3578 if (ptn == nullptr) {
3579 return false; // not in congraph (e.g. ConI)
3580 }
3581 PointsToNode::EscapeState es = ptn->escape_state();
3582 // If we have already computed a value, return it.
3583 if (es >= PointsToNode::GlobalEscape) {
3584 return false;
3585 }
3586 if (ptn->is_JavaObject()) {
3587 return true; // (es < PointsToNode::GlobalEscape);
3588 }
3589 assert(ptn->is_LocalVar(), "sanity");
3590 // Check all java objects it points to.
3591 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3592 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3593 return false;
3594 }
3595 }
3596 return true;
3597 }
3598
3599 // Return true if locked object does not escape globally
3600 // and locked code region (identified by BoxLockNode) is balanced:
3601 // all compiled code paths have corresponding Lock/Unlock pairs.
3602 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3603 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3604 if (EliminateNestedLocks) {
3605 // We can mark whole locking region as Local only when only
3606 // one object is used for locking.
3607 alock->box_node()->as_BoxLock()->set_local();
3608 }
3609 return true;
3610 }
3611 return false;
3612 }
3613
3614 // Helper functions
3615
3616 // Return true if this node points to specified node or nodes it points to.
3617 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3618 if (is_JavaObject()) {
3619 return (this == ptn);
3620 }
3621 assert(is_LocalVar() || is_Field(), "sanity");
3622 for (EdgeIterator i(this); i.has_next(); i.next()) {
3623 if (i.get() == ptn) {
3624 return true;
3625 }
3626 }
3627 return false;
3628 }
3629
3630 // Return true if one node points to an other.
3631 bool PointsToNode::meet(PointsToNode* ptn) {
3632 if (this == ptn) {
3633 return true;
3634 } else if (ptn->is_JavaObject()) {
3635 return this->points_to(ptn->as_JavaObject());
3636 } else if (this->is_JavaObject()) {
3637 return ptn->points_to(this->as_JavaObject());
3638 }
3639 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3640 int ptn_count = ptn->edge_count();
3641 for (EdgeIterator i(this); i.has_next(); i.next()) {
3642 PointsToNode* this_e = i.get();
3643 for (int j = 0; j < ptn_count; j++) {
3644 if (this_e == ptn->edge(j)) {
3645 return true;
3646 }
3647 }
3648 }
3649 return false;
3650 }
3651
3652 #ifdef ASSERT
3653 // Return true if bases point to this java object.
3654 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3655 for (BaseIterator i(this); i.has_next(); i.next()) {
3656 if (i.get() == jobj) {
3657 return true;
3658 }
3659 }
3660 return false;
3661 }
3662 #endif
3663
3664 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3665 // Handle simple case first.
3666 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3667 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3668 return true;
3669 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3670 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3671 Node* addp_use = addp->fast_out(i);
3672 if (addp_use->is_Store()) {
3673 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3674 if (addp_use->fast_out(j)->is_Initialize()) {
3675 return true;
3676 }
3677 }
3678 }
3679 }
3680 }
3681 return false;
3682 }
3683
3684 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3685 const Type *adr_type = phase->type(adr);
3686 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3687 // We are computing a raw address for a store captured by an Initialize
3688 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3689 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3690 assert(offs != Type::OffsetBot ||
3691 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3692 "offset must be a constant or it is initialization of array");
3693 return offs;
3694 }
3695 const TypePtr *t_ptr = adr_type->isa_ptr();
3696 assert(t_ptr != nullptr, "must be a pointer type");
3697 return t_ptr->offset();
3698 }
3699
3700 Node* ConnectionGraph::get_addp_base(Node *addp) {
3701 assert(addp->is_AddP(), "must be AddP");
3702 //
3703 // AddP cases for Base and Address inputs:
3704 // case #1. Direct object's field reference:
3705 // Allocate
3706 // |
3707 // Proj #5 ( oop result )
3708 // |
3709 // CheckCastPP (cast to instance type)
3710 // | |
3711 // AddP ( base == address )
3712 //
3713 // case #2. Indirect object's field reference:
3714 // Phi
3715 // |
3716 // CastPP (cast to instance type)
3717 // | |
3718 // AddP ( base == address )
3719 //
3720 // case #3. Raw object's field reference for Initialize node:
3721 // Allocate
3722 // |
3723 // Proj #5 ( oop result )
3724 // top |
3725 // \ |
3726 // AddP ( base == top )
3727 //
3728 // case #4. Array's element reference:
3729 // {CheckCastPP | CastPP}
3730 // | | |
3731 // | AddP ( array's element offset )
3732 // | |
3733 // AddP ( array's offset )
3734 //
3735 // case #5. Raw object's field reference for arraycopy stub call:
3736 // The inline_native_clone() case when the arraycopy stub is called
3737 // after the allocation before Initialize and CheckCastPP nodes.
3738 // Allocate
3739 // |
3740 // Proj #5 ( oop result )
3741 // | |
3742 // AddP ( base == address )
3743 //
3744 // case #6. Constant Pool, ThreadLocal, CastX2P or
3745 // Raw object's field reference:
3746 // {ConP, ThreadLocal, CastX2P, raw Load}
3747 // top |
3748 // \ |
3749 // AddP ( base == top )
3750 //
3751 // case #7. Klass's field reference.
3752 // LoadKlass
3753 // | |
3754 // AddP ( base == address )
3755 //
3756 // case #8. narrow Klass's field reference.
3757 // LoadNKlass
3758 // |
3759 // DecodeN
3760 // | |
3761 // AddP ( base == address )
3762 //
3763 // case #9. Mixed unsafe access
3764 // {instance}
3765 // |
3766 // CheckCastPP (raw)
3767 // top |
3768 // \ |
3769 // AddP ( base == top )
3770 //
3771 Node *base = addp->in(AddPNode::Base);
3772 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3773 base = addp->in(AddPNode::Address);
3774 while (base->is_AddP()) {
3775 // Case #6 (unsafe access) may have several chained AddP nodes.
3776 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3777 base = base->in(AddPNode::Address);
3778 }
3779 if (base->Opcode() == Op_CheckCastPP &&
3780 base->bottom_type()->isa_rawptr() &&
3781 _igvn->type(base->in(1))->isa_oopptr()) {
3782 base = base->in(1); // Case #9
3783 } else {
3784 Node* uncast_base = base->uncast();
3785 int opcode = uncast_base->Opcode();
3786 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3787 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3788 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3789 is_captured_store_address(addp), "sanity");
3790 }
3791 }
3792 return base;
3793 }
3794
3795 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3796 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3797 Node* addp2 = addp->raw_out(0);
3798 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3799 addp2->in(AddPNode::Base) == n &&
3800 addp2->in(AddPNode::Address) == addp) {
3801 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3802 //
3803 // Find array's offset to push it on worklist first and
3804 // as result process an array's element offset first (pushed second)
3805 // to avoid CastPP for the array's offset.
3806 // Otherwise the inserted CastPP (LocalVar) will point to what
3807 // the AddP (Field) points to. Which would be wrong since
3808 // the algorithm expects the CastPP has the same point as
3809 // as AddP's base CheckCastPP (LocalVar).
3810 //
3811 // ArrayAllocation
3812 // |
3813 // CheckCastPP
3814 // |
3815 // memProj (from ArrayAllocation CheckCastPP)
3816 // | ||
3817 // | || Int (element index)
3818 // | || | ConI (log(element size))
3819 // | || | /
3820 // | || LShift
3821 // | || /
3822 // | AddP (array's element offset)
3823 // | |
3824 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3825 // | / /
3826 // AddP (array's offset)
3827 // |
3828 // Load/Store (memory operation on array's element)
3829 //
3830 return addp2;
3831 }
3832 return nullptr;
3833 }
3834
3835 //
3836 // Adjust the type and inputs of an AddP which computes the
3837 // address of a field of an instance
3838 //
3839 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3840 PhaseGVN* igvn = _igvn;
3841 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3842 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3843 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3844 if (t == nullptr) {
3845 // We are computing a raw address for a store captured by an Initialize
3846 // compute an appropriate address type (cases #3 and #5).
3847 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3848 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3849 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3850 assert(offs != Type::OffsetBot, "offset must be a constant");
3851 t = base_t->add_offset(offs)->is_oopptr();
3852 }
3853 int inst_id = base_t->instance_id();
3854 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3855 "old type must be non-instance or match new type");
3856
3857 // The type 't' could be subclass of 'base_t'.
3858 // As result t->offset() could be large then base_t's size and it will
3859 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3860 // constructor verifies correctness of the offset.
3861 //
3862 // It could happened on subclass's branch (from the type profiling
3863 // inlining) which was not eliminated during parsing since the exactness
3864 // of the allocation type was not propagated to the subclass type check.
3865 //
3866 // Or the type 't' could be not related to 'base_t' at all.
3867 // It could happened when CHA type is different from MDO type on a dead path
3868 // (for example, from instanceof check) which is not collapsed during parsing.
3869 //
3870 // Do nothing for such AddP node and don't process its users since
3871 // this code branch will go away.
3872 //
3873 if (!t->is_known_instance() &&
3874 !base_t->maybe_java_subtype_of(t)) {
3875 return false; // bail out
3876 }
3877 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
3878 // Do NOT remove the next line: ensure a new alias index is allocated
3879 // for the instance type. Note: C++ will not remove it since the call
3880 // has side effect.
3881 int alias_idx = _compile->get_alias_index(tinst);
3882 igvn->set_type(addp, tinst);
3883 // record the allocation in the node map
3884 set_map(addp, get_map(base->_idx));
3885 // Set addp's Base and Address to 'base'.
3886 Node *abase = addp->in(AddPNode::Base);
3887 Node *adr = addp->in(AddPNode::Address);
3888 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3889 adr->in(0)->_idx == (uint)inst_id) {
3890 // Skip AddP cases #3 and #5.
3891 } else {
3892 assert(!abase->is_top(), "sanity"); // AddP case #3
3893 if (abase != base) {
3894 igvn->hash_delete(addp);
3895 addp->set_req(AddPNode::Base, base);
3896 if (abase == adr) {
3897 addp->set_req(AddPNode::Address, base);
3898 } else {
3899 // AddP case #4 (adr is array's element offset AddP node)
3900 #ifdef ASSERT
3901 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3902 assert(adr->is_AddP() && atype != nullptr &&
3903 atype->instance_id() == inst_id, "array's element offset should be processed first");
3904 #endif
3905 }
3906 igvn->hash_insert(addp);
3907 }
3908 }
3909 // Put on IGVN worklist since at least addp's type was changed above.
3910 record_for_optimizer(addp);
3911 return true;
3912 }
3913
3914 //
3915 // Create a new version of orig_phi if necessary. Returns either the newly
3916 // created phi or an existing phi. Sets create_new to indicate whether a new
3917 // phi was created. Cache the last newly created phi in the node map.
3918 //
3919 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3920 Compile *C = _compile;
3921 PhaseGVN* igvn = _igvn;
3922 new_created = false;
3923 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3924 // nothing to do if orig_phi is bottom memory or matches alias_idx
3925 if (phi_alias_idx == alias_idx) {
3926 return orig_phi;
3927 }
3928 // Have we recently created a Phi for this alias index?
3929 PhiNode *result = get_map_phi(orig_phi->_idx);
3930 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3931 return result;
3932 }
3933 // Previous check may fail when the same wide memory Phi was split into Phis
3934 // for different memory slices. Search all Phis for this region.
3935 if (result != nullptr) {
3936 Node* region = orig_phi->in(0);
3937 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3938 Node* phi = region->fast_out(i);
3939 if (phi->is_Phi() &&
3940 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3941 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3942 return phi->as_Phi();
3943 }
3944 }
3945 }
3946 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
3947 if (C->do_escape_analysis() == true && !C->failing()) {
3948 // Retry compilation without escape analysis.
3949 // If this is the first failure, the sentinel string will "stick"
3950 // to the Compile object, and the C2Compiler will see it and retry.
3951 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
3952 }
3953 return nullptr;
3954 }
3955 orig_phi_worklist.append_if_missing(orig_phi);
3956 const TypePtr *atype = C->get_adr_type(alias_idx);
3957 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
3958 C->copy_node_notes_to(result, orig_phi);
3959 igvn->set_type(result, result->bottom_type());
3960 record_for_optimizer(result);
3961 set_map(orig_phi, result);
3962 new_created = true;
3963 return result;
3964 }
3965
3966 //
3967 // Return a new version of Memory Phi "orig_phi" with the inputs having the
3968 // specified alias index.
3969 //
3970 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
3971 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
3972 Compile *C = _compile;
3973 PhaseGVN* igvn = _igvn;
3974 bool new_phi_created;
3975 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
3976 if (!new_phi_created) {
3977 return result;
3978 }
3979 GrowableArray<PhiNode *> phi_list;
3980 GrowableArray<uint> cur_input;
3981 PhiNode *phi = orig_phi;
3982 uint idx = 1;
3983 bool finished = false;
3984 while(!finished) {
3985 while (idx < phi->req()) {
3986 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
3987 if (mem != nullptr && mem->is_Phi()) {
3988 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
3989 if (new_phi_created) {
3990 // found an phi for which we created a new split, push current one on worklist and begin
3991 // processing new one
3992 phi_list.push(phi);
3993 cur_input.push(idx);
3994 phi = mem->as_Phi();
3995 result = newphi;
3996 idx = 1;
3997 continue;
3998 } else {
3999 mem = newphi;
4000 }
4001 }
4002 if (C->failing()) {
4003 return nullptr;
4004 }
4005 result->set_req(idx++, mem);
4006 }
4007 #ifdef ASSERT
4008 // verify that the new Phi has an input for each input of the original
4009 assert( phi->req() == result->req(), "must have same number of inputs.");
4010 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4011 #endif
4012 // Check if all new phi's inputs have specified alias index.
4013 // Otherwise use old phi.
4014 for (uint i = 1; i < phi->req(); i++) {
4015 Node* in = result->in(i);
4016 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4017 }
4018 // we have finished processing a Phi, see if there are any more to do
4019 finished = (phi_list.length() == 0 );
4020 if (!finished) {
4021 phi = phi_list.pop();
4022 idx = cur_input.pop();
4023 PhiNode *prev_result = get_map_phi(phi->_idx);
4024 prev_result->set_req(idx++, result);
4025 result = prev_result;
4026 }
4027 }
4028 return result;
4029 }
4030
4031 //
4032 // The next methods are derived from methods in MemNode.
4033 //
4034 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4035 Node *mem = mmem;
4036 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4037 // means an array I have not precisely typed yet. Do not do any
4038 // alias stuff with it any time soon.
4039 if (toop->base() != Type::AnyPtr &&
4040 !(toop->isa_instptr() &&
4041 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4042 toop->offset() == Type::OffsetBot)) {
4043 mem = mmem->memory_at(alias_idx);
4044 // Update input if it is progress over what we have now
4045 }
4046 return mem;
4047 }
4048
4049 //
4050 // Move memory users to their memory slices.
4051 //
4052 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4053 Compile* C = _compile;
4054 PhaseGVN* igvn = _igvn;
4055 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4056 assert(tp != nullptr, "ptr type");
4057 int alias_idx = C->get_alias_index(tp);
4058 int general_idx = C->get_general_index(alias_idx);
4059
4060 // Move users first
4061 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4062 Node* use = n->fast_out(i);
4063 if (use->is_MergeMem()) {
4064 MergeMemNode* mmem = use->as_MergeMem();
4065 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4066 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4067 continue; // Nothing to do
4068 }
4069 // Replace previous general reference to mem node.
4070 uint orig_uniq = C->unique();
4071 Node* m = find_inst_mem(n, general_idx, orig_phis);
4072 assert(orig_uniq == C->unique(), "no new nodes");
4073 mmem->set_memory_at(general_idx, m);
4074 --imax;
4075 --i;
4076 } else if (use->is_MemBar()) {
4077 assert(!use->is_Initialize(), "initializing stores should not be moved");
4078 if (use->req() > MemBarNode::Precedent &&
4079 use->in(MemBarNode::Precedent) == n) {
4080 // Don't move related membars.
4081 record_for_optimizer(use);
4082 continue;
4083 }
4084 tp = use->as_MemBar()->adr_type()->isa_ptr();
4085 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4086 alias_idx == general_idx) {
4087 continue; // Nothing to do
4088 }
4089 // Move to general memory slice.
4090 uint orig_uniq = C->unique();
4091 Node* m = find_inst_mem(n, general_idx, orig_phis);
4092 assert(orig_uniq == C->unique(), "no new nodes");
4093 igvn->hash_delete(use);
4094 imax -= use->replace_edge(n, m, igvn);
4095 igvn->hash_insert(use);
4096 record_for_optimizer(use);
4097 --i;
4098 #ifdef ASSERT
4099 } else if (use->is_Mem()) {
4100 // Memory nodes should have new memory input.
4101 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4102 assert(tp != nullptr, "ptr type");
4103 int idx = C->get_alias_index(tp);
4104 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4105 "Following memory nodes should have new memory input or be on the same memory slice");
4106 } else if (use->is_Phi()) {
4107 // Phi nodes should be split and moved already.
4108 tp = use->as_Phi()->adr_type()->isa_ptr();
4109 assert(tp != nullptr, "ptr type");
4110 int idx = C->get_alias_index(tp);
4111 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4112 } else {
4113 use->dump();
4114 assert(false, "should not be here");
4115 #endif
4116 }
4117 }
4118 }
4119
4120 //
4121 // Search memory chain of "mem" to find a MemNode whose address
4122 // is the specified alias index.
4123 //
4124 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4125 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4126 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4127 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4128 return nullptr;
4129 }
4130 if (orig_mem == nullptr) {
4131 return orig_mem;
4132 }
4133 Compile* C = _compile;
4134 PhaseGVN* igvn = _igvn;
4135 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4136 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4137 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4138 Node *prev = nullptr;
4139 Node *result = orig_mem;
4140 while (prev != result) {
4141 prev = result;
4142 if (result == start_mem) {
4143 break; // hit one of our sentinels
4144 }
4145 if (result->is_Mem()) {
4146 const Type *at = igvn->type(result->in(MemNode::Address));
4147 if (at == Type::TOP) {
4148 break; // Dead
4149 }
4150 assert (at->isa_ptr() != nullptr, "pointer type required.");
4151 int idx = C->get_alias_index(at->is_ptr());
4152 if (idx == alias_idx) {
4153 break; // Found
4154 }
4155 if (!is_instance && (at->isa_oopptr() == nullptr ||
4156 !at->is_oopptr()->is_known_instance())) {
4157 break; // Do not skip store to general memory slice.
4158 }
4159 result = result->in(MemNode::Memory);
4160 }
4161 if (!is_instance) {
4162 continue; // don't search further for non-instance types
4163 }
4164 // skip over a call which does not affect this memory slice
4165 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4166 Node *proj_in = result->in(0);
4167 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4168 break; // hit one of our sentinels
4169 } else if (proj_in->is_Call()) {
4170 // ArrayCopy node processed here as well
4171 CallNode *call = proj_in->as_Call();
4172 if (!call->may_modify(toop, igvn)) {
4173 result = call->in(TypeFunc::Memory);
4174 }
4175 } else if (proj_in->is_Initialize()) {
4176 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4177 // Stop if this is the initialization for the object instance which
4178 // which contains this memory slice, otherwise skip over it.
4179 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4180 result = proj_in->in(TypeFunc::Memory);
4181 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4182 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4183 result = get_map(result->_idx);
4184 assert(result != nullptr, "new projection should have been allocated");
4185 break;
4186 }
4187 } else if (proj_in->is_MemBar()) {
4188 // Check if there is an array copy for a clone
4189 // Step over GC barrier when ReduceInitialCardMarks is disabled
4190 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4191 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4192
4193 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4194 // Stop if it is a clone
4195 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4196 if (ac->may_modify(toop, igvn)) {
4197 break;
4198 }
4199 }
4200 result = proj_in->in(TypeFunc::Memory);
4201 }
4202 } else if (result->is_MergeMem()) {
4203 MergeMemNode *mmem = result->as_MergeMem();
4204 result = step_through_mergemem(mmem, alias_idx, toop);
4205 if (result == mmem->base_memory()) {
4206 // Didn't find instance memory, search through general slice recursively.
4207 result = mmem->memory_at(C->get_general_index(alias_idx));
4208 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4209 if (C->failing()) {
4210 return nullptr;
4211 }
4212 mmem->set_memory_at(alias_idx, result);
4213 }
4214 } else if (result->is_Phi() &&
4215 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4216 Node *un = result->as_Phi()->unique_input(igvn);
4217 if (un != nullptr) {
4218 orig_phis.append_if_missing(result->as_Phi());
4219 result = un;
4220 } else {
4221 break;
4222 }
4223 } else if (result->is_ClearArray()) {
4224 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4225 // Can not bypass initialization of the instance
4226 // we are looking for.
4227 break;
4228 }
4229 // Otherwise skip it (the call updated 'result' value).
4230 } else if (result->Opcode() == Op_SCMemProj) {
4231 Node* mem = result->in(0);
4232 Node* adr = nullptr;
4233 if (mem->is_LoadStore()) {
4234 adr = mem->in(MemNode::Address);
4235 } else {
4236 assert(mem->Opcode() == Op_EncodeISOArray ||
4237 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4238 adr = mem->in(3); // Memory edge corresponds to destination array
4239 }
4240 const Type *at = igvn->type(adr);
4241 if (at != Type::TOP) {
4242 assert(at->isa_ptr() != nullptr, "pointer type required.");
4243 int idx = C->get_alias_index(at->is_ptr());
4244 if (idx == alias_idx) {
4245 // Assert in debug mode
4246 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4247 break; // In product mode return SCMemProj node
4248 }
4249 }
4250 result = mem->in(MemNode::Memory);
4251 } else if (result->Opcode() == Op_StrInflatedCopy) {
4252 Node* adr = result->in(3); // Memory edge corresponds to destination array
4253 const Type *at = igvn->type(adr);
4254 if (at != Type::TOP) {
4255 assert(at->isa_ptr() != nullptr, "pointer type required.");
4256 int idx = C->get_alias_index(at->is_ptr());
4257 if (idx == alias_idx) {
4258 // Assert in debug mode
4259 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4260 break; // In product mode return SCMemProj node
4261 }
4262 }
4263 result = result->in(MemNode::Memory);
4264 }
4265 }
4266 if (result->is_Phi()) {
4267 PhiNode *mphi = result->as_Phi();
4268 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4269 const TypePtr *t = mphi->adr_type();
4270 if (!is_instance) {
4271 // Push all non-instance Phis on the orig_phis worklist to update inputs
4272 // during Phase 4 if needed.
4273 orig_phis.append_if_missing(mphi);
4274 } else if (C->get_alias_index(t) != alias_idx) {
4275 // Create a new Phi with the specified alias index type.
4276 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4277 }
4278 }
4279 // the result is either MemNode, PhiNode, InitializeNode.
4280 return result;
4281 }
4282
4283 //
4284 // Convert the types of non-escaped object to instance types where possible,
4285 // propagate the new type information through the graph, and update memory
4286 // edges and MergeMem inputs to reflect the new type.
4287 //
4288 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4289 // The processing is done in 4 phases:
4290 //
4291 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4292 // types for the CheckCastPP for allocations where possible.
4293 // Propagate the new types through users as follows:
4294 // casts and Phi: push users on alloc_worklist
4295 // AddP: cast Base and Address inputs to the instance type
4296 // push any AddP users on alloc_worklist and push any memnode
4297 // users onto memnode_worklist.
4298 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4299 // search the Memory chain for a store with the appropriate type
4300 // address type. If a Phi is found, create a new version with
4301 // the appropriate memory slices from each of the Phi inputs.
4302 // For stores, process the users as follows:
4303 // MemNode: push on memnode_worklist
4304 // MergeMem: push on mergemem_worklist
4305 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4306 // moving the first node encountered of each instance type to the
4307 // the input corresponding to its alias index.
4308 // appropriate memory slice.
4309 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4310 //
4311 // In the following example, the CheckCastPP nodes are the cast of allocation
4312 // results and the allocation of node 29 is non-escaped and eligible to be an
4313 // instance type.
4314 //
4315 // We start with:
4316 //
4317 // 7 Parm #memory
4318 // 10 ConI "12"
4319 // 19 CheckCastPP "Foo"
4320 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4321 // 29 CheckCastPP "Foo"
4322 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4323 //
4324 // 40 StoreP 25 7 20 ... alias_index=4
4325 // 50 StoreP 35 40 30 ... alias_index=4
4326 // 60 StoreP 45 50 20 ... alias_index=4
4327 // 70 LoadP _ 60 30 ... alias_index=4
4328 // 80 Phi 75 50 60 Memory alias_index=4
4329 // 90 LoadP _ 80 30 ... alias_index=4
4330 // 100 LoadP _ 80 20 ... alias_index=4
4331 //
4332 //
4333 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4334 // and creating a new alias index for node 30. This gives:
4335 //
4336 // 7 Parm #memory
4337 // 10 ConI "12"
4338 // 19 CheckCastPP "Foo"
4339 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4340 // 29 CheckCastPP "Foo" iid=24
4341 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4342 //
4343 // 40 StoreP 25 7 20 ... alias_index=4
4344 // 50 StoreP 35 40 30 ... alias_index=6
4345 // 60 StoreP 45 50 20 ... alias_index=4
4346 // 70 LoadP _ 60 30 ... alias_index=6
4347 // 80 Phi 75 50 60 Memory alias_index=4
4348 // 90 LoadP _ 80 30 ... alias_index=6
4349 // 100 LoadP _ 80 20 ... alias_index=4
4350 //
4351 // In phase 2, new memory inputs are computed for the loads and stores,
4352 // And a new version of the phi is created. In phase 4, the inputs to
4353 // node 80 are updated and then the memory nodes are updated with the
4354 // values computed in phase 2. This results in:
4355 //
4356 // 7 Parm #memory
4357 // 10 ConI "12"
4358 // 19 CheckCastPP "Foo"
4359 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4360 // 29 CheckCastPP "Foo" iid=24
4361 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4362 //
4363 // 40 StoreP 25 7 20 ... alias_index=4
4364 // 50 StoreP 35 7 30 ... alias_index=6
4365 // 60 StoreP 45 40 20 ... alias_index=4
4366 // 70 LoadP _ 50 30 ... alias_index=6
4367 // 80 Phi 75 40 60 Memory alias_index=4
4368 // 120 Phi 75 50 50 Memory alias_index=6
4369 // 90 LoadP _ 120 30 ... alias_index=6
4370 // 100 LoadP _ 80 20 ... alias_index=4
4371 //
4372 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4373 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4374 GrowableArray<MergeMemNode*> &mergemem_worklist,
4375 Unique_Node_List &reducible_merges) {
4376 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4377 GrowableArray<Node *> memnode_worklist;
4378 GrowableArray<PhiNode *> orig_phis;
4379 PhaseIterGVN *igvn = _igvn;
4380 uint new_index_start = (uint) _compile->num_alias_types();
4381 VectorSet visited;
4382 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4383 uint unique_old = _compile->unique();
4384
4385 // Phase 1: Process possible allocations from alloc_worklist.
4386 // Create instance types for the CheckCastPP for allocations where possible.
4387 //
4388 // (Note: don't forget to change the order of the second AddP node on
4389 // the alloc_worklist if the order of the worklist processing is changed,
4390 // see the comment in find_second_addp().)
4391 //
4392 while (alloc_worklist.length() != 0) {
4393 Node *n = alloc_worklist.pop();
4394 uint ni = n->_idx;
4395 if (n->is_Call()) {
4396 CallNode *alloc = n->as_Call();
4397 // copy escape information to call node
4398 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4399 PointsToNode::EscapeState es = ptn->escape_state();
4400 // We have an allocation or call which returns a Java object,
4401 // see if it is non-escaped.
4402 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4403 continue;
4404 }
4405 // Find CheckCastPP for the allocate or for the return value of a call
4406 n = alloc->result_cast();
4407 if (n == nullptr) { // No uses except Initialize node
4408 if (alloc->is_Allocate()) {
4409 // Set the scalar_replaceable flag for allocation
4410 // so it could be eliminated if it has no uses.
4411 alloc->as_Allocate()->_is_scalar_replaceable = true;
4412 }
4413 continue;
4414 }
4415 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4416 // we could reach here for allocate case if one init is associated with many allocs.
4417 if (alloc->is_Allocate()) {
4418 alloc->as_Allocate()->_is_scalar_replaceable = false;
4419 }
4420 continue;
4421 }
4422
4423 // The inline code for Object.clone() casts the allocation result to
4424 // java.lang.Object and then to the actual type of the allocated
4425 // object. Detect this case and use the second cast.
4426 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4427 // the allocation result is cast to java.lang.Object and then
4428 // to the actual Array type.
4429 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4430 && (alloc->is_AllocateArray() ||
4431 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4432 Node *cast2 = nullptr;
4433 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4434 Node *use = n->fast_out(i);
4435 if (use->is_CheckCastPP()) {
4436 cast2 = use;
4437 break;
4438 }
4439 }
4440 if (cast2 != nullptr) {
4441 n = cast2;
4442 } else {
4443 // Non-scalar replaceable if the allocation type is unknown statically
4444 // (reflection allocation), the object can't be restored during
4445 // deoptimization without precise type.
4446 continue;
4447 }
4448 }
4449
4450 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4451 if (t == nullptr) {
4452 continue; // not a TypeOopPtr
4453 }
4454 if (!t->klass_is_exact()) {
4455 continue; // not an unique type
4456 }
4457 if (alloc->is_Allocate()) {
4458 // Set the scalar_replaceable flag for allocation
4459 // so it could be eliminated.
4460 alloc->as_Allocate()->_is_scalar_replaceable = true;
4461 }
4462 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4463 // in order for an object to be scalar-replaceable, it must be:
4464 // - a direct allocation (not a call returning an object)
4465 // - non-escaping
4466 // - eligible to be a unique type
4467 // - not determined to be ineligible by escape analysis
4468 set_map(alloc, n);
4469 set_map(n, alloc);
4470 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4471 igvn->hash_delete(n);
4472 igvn->set_type(n, tinst);
4473 n->raise_bottom_type(tinst);
4474 igvn->hash_insert(n);
4475 record_for_optimizer(n);
4476 // Allocate an alias index for the header fields. Accesses to
4477 // the header emitted during macro expansion wouldn't have
4478 // correct memory state otherwise.
4479 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4480 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4481 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4482 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4483 InitializeNode* init = alloc->as_Allocate()->initialization();
4484 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4485 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4486 const TypePtr* adr_type = proj->adr_type();
4487 const TypePtr* new_adr_type = tinst->add_offset(adr_type->offset());
4488 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4489 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4490 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4491 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4492 igvn->set_type(new_proj, new_proj->bottom_type());
4493 record_for_optimizer(new_proj);
4494 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4495 }
4496 };
4497 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4498
4499 // First, put on the worklist all Field edges from Connection Graph
4500 // which is more accurate than putting immediate users from Ideal Graph.
4501 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4502 PointsToNode* tgt = e.get();
4503 if (tgt->is_Arraycopy()) {
4504 continue;
4505 }
4506 Node* use = tgt->ideal_node();
4507 assert(tgt->is_Field() && use->is_AddP(),
4508 "only AddP nodes are Field edges in CG");
4509 if (use->outcnt() > 0) { // Don't process dead nodes
4510 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4511 if (addp2 != nullptr) {
4512 assert(alloc->is_AllocateArray(),"array allocation was expected");
4513 alloc_worklist.append_if_missing(addp2);
4514 }
4515 alloc_worklist.append_if_missing(use);
4516 }
4517 }
4518
4519 // An allocation may have an Initialize which has raw stores. Scan
4520 // the users of the raw allocation result and push AddP users
4521 // on alloc_worklist.
4522 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4523 assert (raw_result != nullptr, "must have an allocation result");
4524 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4525 Node *use = raw_result->fast_out(i);
4526 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4527 Node* addp2 = find_second_addp(use, raw_result);
4528 if (addp2 != nullptr) {
4529 assert(alloc->is_AllocateArray(),"array allocation was expected");
4530 alloc_worklist.append_if_missing(addp2);
4531 }
4532 alloc_worklist.append_if_missing(use);
4533 } else if (use->is_MemBar()) {
4534 memnode_worklist.append_if_missing(use);
4535 }
4536 }
4537 }
4538 } else if (n->is_AddP()) {
4539 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4540 // This AddP will go away when we reduce the Phi
4541 continue;
4542 }
4543 Node* addp_base = get_addp_base(n);
4544 JavaObjectNode* jobj = unique_java_object(addp_base);
4545 if (jobj == nullptr || jobj == phantom_obj) {
4546 #ifdef ASSERT
4547 ptnode_adr(get_addp_base(n)->_idx)->dump();
4548 ptnode_adr(n->_idx)->dump();
4549 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4550 #endif
4551 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4552 return;
4553 }
4554 Node *base = get_map(jobj->idx()); // CheckCastPP node
4555 if (!split_AddP(n, base)) continue; // wrong type from dead path
4556 } else if (n->is_Phi() ||
4557 n->is_CheckCastPP() ||
4558 n->is_EncodeP() ||
4559 n->is_DecodeN() ||
4560 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4561 if (visited.test_set(n->_idx)) {
4562 assert(n->is_Phi(), "loops only through Phi's");
4563 continue; // already processed
4564 }
4565 // Reducible Phi's will be removed from the graph after split_unique_types
4566 // finishes. For now we just try to split out the SR inputs of the merge.
4567 Node* parent = n->in(1);
4568 if (reducible_merges.member(n)) {
4569 reduce_phi(n->as_Phi(), alloc_worklist);
4570 #ifdef ASSERT
4571 if (VerifyReduceAllocationMerges) {
4572 reduced_merges.push(n);
4573 }
4574 #endif
4575 continue;
4576 } else if (reducible_merges.member(parent)) {
4577 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4578 // part of reduce_merge.
4579 continue;
4580 }
4581 JavaObjectNode* jobj = unique_java_object(n);
4582 if (jobj == nullptr || jobj == phantom_obj) {
4583 #ifdef ASSERT
4584 ptnode_adr(n->_idx)->dump();
4585 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4586 #endif
4587 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4588 return;
4589 } else {
4590 Node *val = get_map(jobj->idx()); // CheckCastPP node
4591 TypeNode *tn = n->as_Type();
4592 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4593 assert(tinst != nullptr && tinst->is_known_instance() &&
4594 tinst->instance_id() == jobj->idx() , "instance type expected.");
4595
4596 const Type *tn_type = igvn->type(tn);
4597 const TypeOopPtr *tn_t;
4598 if (tn_type->isa_narrowoop()) {
4599 tn_t = tn_type->make_ptr()->isa_oopptr();
4600 } else {
4601 tn_t = tn_type->isa_oopptr();
4602 }
4603 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4604 if (tn_type->isa_narrowoop()) {
4605 tn_type = tinst->make_narrowoop();
4606 } else {
4607 tn_type = tinst;
4608 }
4609 igvn->hash_delete(tn);
4610 igvn->set_type(tn, tn_type);
4611 tn->set_type(tn_type);
4612 igvn->hash_insert(tn);
4613 record_for_optimizer(n);
4614 } else {
4615 assert(tn_type == TypePtr::NULL_PTR ||
4616 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4617 "unexpected type");
4618 continue; // Skip dead path with different type
4619 }
4620 }
4621 } else {
4622 DEBUG_ONLY(n->dump();)
4623 assert(false, "EA: unexpected node");
4624 continue;
4625 }
4626 // push allocation's users on appropriate worklist
4627 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4628 Node *use = n->fast_out(i);
4629 if(use->is_Mem() && use->in(MemNode::Address) == n) {
4630 // Load/store to instance's field
4631 memnode_worklist.append_if_missing(use);
4632 } else if (use->is_MemBar()) {
4633 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4634 memnode_worklist.append_if_missing(use);
4635 }
4636 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4637 Node* addp2 = find_second_addp(use, n);
4638 if (addp2 != nullptr) {
4639 alloc_worklist.append_if_missing(addp2);
4640 }
4641 alloc_worklist.append_if_missing(use);
4642 } else if (use->is_Phi() ||
4643 use->is_CheckCastPP() ||
4644 use->is_EncodeNarrowPtr() ||
4645 use->is_DecodeNarrowPtr() ||
4646 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4647 alloc_worklist.append_if_missing(use);
4648 #ifdef ASSERT
4649 } else if (use->is_Mem()) {
4650 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4651 } else if (use->is_MergeMem()) {
4652 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4653 } else if (use->is_SafePoint()) {
4654 // Look for MergeMem nodes for calls which reference unique allocation
4655 // (through CheckCastPP nodes) even for debug info.
4656 Node* m = use->in(TypeFunc::Memory);
4657 if (m->is_MergeMem()) {
4658 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4659 }
4660 } else if (use->Opcode() == Op_EncodeISOArray) {
4661 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4662 // EncodeISOArray overwrites destination array
4663 memnode_worklist.append_if_missing(use);
4664 }
4665 } else {
4666 uint op = use->Opcode();
4667 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4668 (use->in(MemNode::Memory) == n)) {
4669 // They overwrite memory edge corresponding to destination array,
4670 memnode_worklist.append_if_missing(use);
4671 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4672 op == Op_CastP2X ||
4673 op == Op_FastLock || op == Op_AryEq ||
4674 op == Op_StrComp || op == Op_CountPositives ||
4675 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4676 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4677 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4678 op == Op_SubTypeCheck ||
4679 op == Op_ReinterpretS2HF ||
4680 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4681 n->dump();
4682 use->dump();
4683 assert(false, "EA: missing allocation reference path");
4684 }
4685 #endif
4686 }
4687 }
4688
4689 }
4690
4691 #ifdef ASSERT
4692 if (VerifyReduceAllocationMerges) {
4693 for (uint i = 0; i < reducible_merges.size(); i++) {
4694 Node* phi = reducible_merges.at(i);
4695
4696 if (!reduced_merges.member(phi)) {
4697 phi->dump(2);
4698 phi->dump(-2);
4699 assert(false, "This reducible merge wasn't reduced.");
4700 }
4701
4702 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4703 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4704 Node* use = phi->fast_out(j);
4705 if (!use->is_SafePoint() && !use->is_CastPP()) {
4706 phi->dump(2);
4707 phi->dump(-2);
4708 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4709 }
4710 }
4711 }
4712 }
4713 #endif
4714
4715 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4716 // type, record it in the ArrayCopy node so we know what memory this
4717 // node uses/modified.
4718 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4719 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4720 Node* dest = ac->in(ArrayCopyNode::Dest);
4721 if (dest->is_AddP()) {
4722 dest = get_addp_base(dest);
4723 }
4724 JavaObjectNode* jobj = unique_java_object(dest);
4725 if (jobj != nullptr) {
4726 Node *base = get_map(jobj->idx());
4727 if (base != nullptr) {
4728 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4729 ac->_dest_type = base_t;
4730 }
4731 }
4732 Node* src = ac->in(ArrayCopyNode::Src);
4733 if (src->is_AddP()) {
4734 src = get_addp_base(src);
4735 }
4736 jobj = unique_java_object(src);
4737 if (jobj != nullptr) {
4738 Node* base = get_map(jobj->idx());
4739 if (base != nullptr) {
4740 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4741 ac->_src_type = base_t;
4742 }
4743 }
4744 }
4745
4746 // New alias types were created in split_AddP().
4747 uint new_index_end = (uint) _compile->num_alias_types();
4748
4749 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
4750
4751 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4752 // compute new values for Memory inputs (the Memory inputs are not
4753 // actually updated until phase 4.)
4754 if (memnode_worklist.length() == 0)
4755 return; // nothing to do
4756 while (memnode_worklist.length() != 0) {
4757 Node *n = memnode_worklist.pop();
4758 if (visited.test_set(n->_idx)) {
4759 continue;
4760 }
4761 if (n->is_Phi() || n->is_ClearArray()) {
4762 // we don't need to do anything, but the users must be pushed
4763 } else if (n->is_MemBar()) { // MemBar nodes
4764 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
4765 // we don't need to do anything, but the users must be pushed
4766 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4767 if (n == nullptr) {
4768 continue;
4769 }
4770 }
4771 } else if (n->is_CallLeaf()) {
4772 // Runtime calls with narrow memory input (no MergeMem node)
4773 // get the memory projection
4774 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4775 if (n == nullptr) {
4776 continue;
4777 }
4778 } else if (n->Opcode() == Op_StrInflatedCopy) {
4779 // Check direct uses of StrInflatedCopy.
4780 // It is memory type Node - no special SCMemProj node.
4781 } else if (n->Opcode() == Op_StrCompressedCopy ||
4782 n->Opcode() == Op_EncodeISOArray) {
4783 // get the memory projection
4784 n = n->find_out_with(Op_SCMemProj);
4785 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4786 } else if (n->is_Proj()) {
4787 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
4788 } else {
4789 #ifdef ASSERT
4790 if (!n->is_Mem()) {
4791 n->dump();
4792 }
4793 assert(n->is_Mem(), "memory node required.");
4794 #endif
4795 Node *addr = n->in(MemNode::Address);
4796 const Type *addr_t = igvn->type(addr);
4797 if (addr_t == Type::TOP) {
4798 continue;
4799 }
4800 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4801 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4802 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4803 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4804 if (_compile->failing()) {
4805 return;
4806 }
4807 if (mem != n->in(MemNode::Memory)) {
4808 // We delay the memory edge update since we need old one in
4809 // MergeMem code below when instances memory slices are separated.
4810 set_map(n, mem);
4811 }
4812 if (n->is_Load()) {
4813 continue; // don't push users
4814 } else if (n->is_LoadStore()) {
4815 // get the memory projection
4816 n = n->find_out_with(Op_SCMemProj);
4817 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4818 }
4819 }
4820 // push user on appropriate worklist
4821 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4822 Node *use = n->fast_out(i);
4823 if (use->is_Phi() || use->is_ClearArray()) {
4824 memnode_worklist.append_if_missing(use);
4825 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4826 memnode_worklist.append_if_missing(use);
4827 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4828 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4829 memnode_worklist.append_if_missing(use);
4830 }
4831 } else if (use->is_Proj()) {
4832 assert(n->is_Initialize(), "We only push projections of Initialize");
4833 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
4834 memnode_worklist.append_if_missing(use);
4835 }
4836 #ifdef ASSERT
4837 } else if(use->is_Mem()) {
4838 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4839 } else if (use->is_MergeMem()) {
4840 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4841 } else if (use->Opcode() == Op_EncodeISOArray) {
4842 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4843 // EncodeISOArray overwrites destination array
4844 memnode_worklist.append_if_missing(use);
4845 }
4846 } else {
4847 uint op = use->Opcode();
4848 if ((use->in(MemNode::Memory) == n) &&
4849 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4850 // They overwrite memory edge corresponding to destination array,
4851 memnode_worklist.append_if_missing(use);
4852 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4853 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4854 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4855 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
4856 n->dump();
4857 use->dump();
4858 assert(false, "EA: missing memory path");
4859 }
4860 #endif
4861 }
4862 }
4863 }
4864
4865 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4866 // Walk each memory slice moving the first node encountered of each
4867 // instance type to the input corresponding to its alias index.
4868 uint length = mergemem_worklist.length();
4869 for( uint next = 0; next < length; ++next ) {
4870 MergeMemNode* nmm = mergemem_worklist.at(next);
4871 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4872 // Note: we don't want to use MergeMemStream here because we only want to
4873 // scan inputs which exist at the start, not ones we add during processing.
4874 // Note 2: MergeMem may already contains instance memory slices added
4875 // during find_inst_mem() call when memory nodes were processed above.
4876 igvn->hash_delete(nmm);
4877 uint nslices = MIN2(nmm->req(), new_index_start);
4878 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4879 Node* mem = nmm->in(i);
4880 Node* cur = nullptr;
4881 if (mem == nullptr || mem->is_top()) {
4882 continue;
4883 }
4884 // First, update mergemem by moving memory nodes to corresponding slices
4885 // if their type became more precise since this mergemem was created.
4886 while (mem->is_Mem()) {
4887 const Type* at = igvn->type(mem->in(MemNode::Address));
4888 if (at != Type::TOP) {
4889 assert (at->isa_ptr() != nullptr, "pointer type required.");
4890 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4891 if (idx == i) {
4892 if (cur == nullptr) {
4893 cur = mem;
4894 }
4895 } else {
4896 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4897 nmm->set_memory_at(idx, mem);
4898 }
4899 }
4900 }
4901 mem = mem->in(MemNode::Memory);
4902 }
4903 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4904 // Find any instance of the current type if we haven't encountered
4905 // already a memory slice of the instance along the memory chain.
4906 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4907 if((uint)_compile->get_general_index(ni) == i) {
4908 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4909 if (nmm->is_empty_memory(m)) {
4910 Node* result = find_inst_mem(mem, ni, orig_phis);
4911 if (_compile->failing()) {
4912 return;
4913 }
4914 nmm->set_memory_at(ni, result);
4915 }
4916 }
4917 }
4918 }
4919 // Find the rest of instances values
4920 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4921 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4922 Node* result = step_through_mergemem(nmm, ni, tinst);
4923 if (result == nmm->base_memory()) {
4924 // Didn't find instance memory, search through general slice recursively.
4925 result = nmm->memory_at(_compile->get_general_index(ni));
4926 result = find_inst_mem(result, ni, orig_phis);
4927 if (_compile->failing()) {
4928 return;
4929 }
4930 nmm->set_memory_at(ni, result);
4931 }
4932 }
4933
4934 // If we have crossed the 3/4 point of max node limit it's too risky
4935 // to continue with EA/SR because we might hit the max node limit.
4936 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4937 if (_compile->do_reduce_allocation_merges()) {
4938 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4939 } else if (_invocation > 0) {
4940 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4941 } else {
4942 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4943 }
4944 return;
4945 }
4946
4947 igvn->hash_insert(nmm);
4948 record_for_optimizer(nmm);
4949 }
4950
4951 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
4952
4953 // Phase 4: Update the inputs of non-instance memory Phis and
4954 // the Memory input of memnodes
4955 // First update the inputs of any non-instance Phi's from
4956 // which we split out an instance Phi. Note we don't have
4957 // to recursively process Phi's encountered on the input memory
4958 // chains as is done in split_memory_phi() since they will
4959 // also be processed here.
4960 for (int j = 0; j < orig_phis.length(); j++) {
4961 PhiNode *phi = orig_phis.at(j);
4962 int alias_idx = _compile->get_alias_index(phi->adr_type());
4963 igvn->hash_delete(phi);
4964 for (uint i = 1; i < phi->req(); i++) {
4965 Node *mem = phi->in(i);
4966 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
4967 if (_compile->failing()) {
4968 return;
4969 }
4970 if (mem != new_mem) {
4971 phi->set_req(i, new_mem);
4972 }
4973 }
4974 igvn->hash_insert(phi);
4975 record_for_optimizer(phi);
4976 }
4977
4978 // Update the memory inputs of MemNodes with the value we computed
4979 // in Phase 2 and move stores memory users to corresponding memory slices.
4980 // Disable memory split verification code until the fix for 6984348.
4981 // Currently it produces false negative results since it does not cover all cases.
4982 #if 0 // ifdef ASSERT
4983 visited.Reset();
4984 Node_Stack old_mems(arena, _compile->unique() >> 2);
4985 #endif
4986 for (uint i = 0; i < ideal_nodes.size(); i++) {
4987 Node* n = ideal_nodes.at(i);
4988 Node* nmem = get_map(n->_idx);
4989 assert(nmem != nullptr, "sanity");
4990 if (n->is_Mem()) {
4991 #if 0 // ifdef ASSERT
4992 Node* old_mem = n->in(MemNode::Memory);
4993 if (!visited.test_set(old_mem->_idx)) {
4994 old_mems.push(old_mem, old_mem->outcnt());
4995 }
4996 #endif
4997 assert(n->in(MemNode::Memory) != nmem, "sanity");
4998 if (!n->is_Load()) {
4999 // Move memory users of a store first.
5000 move_inst_mem(n, orig_phis);
5001 }
5002 // Now update memory input
5003 igvn->hash_delete(n);
5004 n->set_req(MemNode::Memory, nmem);
5005 igvn->hash_insert(n);
5006 record_for_optimizer(n);
5007 } else {
5008 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5009 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5010 }
5011 }
5012 #if 0 // ifdef ASSERT
5013 // Verify that memory was split correctly
5014 while (old_mems.is_nonempty()) {
5015 Node* old_mem = old_mems.node();
5016 uint old_cnt = old_mems.index();
5017 old_mems.pop();
5018 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5019 }
5020 #endif
5021 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5022 }
5023
5024 #ifndef PRODUCT
5025 int ConnectionGraph::_no_escape_counter = 0;
5026 int ConnectionGraph::_arg_escape_counter = 0;
5027 int ConnectionGraph::_global_escape_counter = 0;
5028
5029 static const char *node_type_names[] = {
5030 "UnknownType",
5031 "JavaObject",
5032 "LocalVar",
5033 "Field",
5034 "Arraycopy"
5035 };
5036
5037 static const char *esc_names[] = {
5038 "UnknownEscape",
5039 "NoEscape",
5040 "ArgEscape",
5041 "GlobalEscape"
5042 };
5043
5044 const char* PointsToNode::esc_name() const {
5045 return esc_names[(int)escape_state()];
5046 }
5047
5048 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5049 NodeType nt = node_type();
5050 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5051 if (print_state) {
5052 EscapeState es = escape_state();
5053 EscapeState fields_es = fields_escape_state();
5054 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5055 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5056 out->print("NSR ");
5057 }
5058 }
5059 }
5060
5061 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5062 dump_header(print_state, out);
5063 if (is_Field()) {
5064 FieldNode* f = (FieldNode*)this;
5065 if (f->is_oop()) {
5066 out->print("oop ");
5067 }
5068 if (f->offset() > 0) {
5069 out->print("+%d ", f->offset());
5070 }
5071 out->print("(");
5072 for (BaseIterator i(f); i.has_next(); i.next()) {
5073 PointsToNode* b = i.get();
5074 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5075 }
5076 out->print(" )");
5077 }
5078 out->print("[");
5079 for (EdgeIterator i(this); i.has_next(); i.next()) {
5080 PointsToNode* e = i.get();
5081 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5082 }
5083 out->print(" [");
5084 for (UseIterator i(this); i.has_next(); i.next()) {
5085 PointsToNode* u = i.get();
5086 bool is_base = false;
5087 if (PointsToNode::is_base_use(u)) {
5088 is_base = true;
5089 u = PointsToNode::get_use_node(u)->as_Field();
5090 }
5091 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5092 }
5093 out->print(" ]] ");
5094 if (_node == nullptr) {
5095 out->print("<null>%s", newline ? "\n" : "");
5096 } else {
5097 _node->dump(newline ? "\n" : "", false, out);
5098 }
5099 }
5100
5101 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5102 bool first = true;
5103 int ptnodes_length = ptnodes_worklist.length();
5104 for (int i = 0; i < ptnodes_length; i++) {
5105 PointsToNode *ptn = ptnodes_worklist.at(i);
5106 if (ptn == nullptr || !ptn->is_JavaObject()) {
5107 continue;
5108 }
5109 PointsToNode::EscapeState es = ptn->escape_state();
5110 if ((es != PointsToNode::NoEscape) && !Verbose) {
5111 continue;
5112 }
5113 Node* n = ptn->ideal_node();
5114 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5115 n->as_CallStaticJava()->is_boxing_method())) {
5116 if (first) {
5117 tty->cr();
5118 tty->print("======== Connection graph for ");
5119 _compile->method()->print_short_name();
5120 tty->cr();
5121 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5122 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5123 tty->cr();
5124 first = false;
5125 }
5126 ptn->dump();
5127 // Print all locals and fields which reference this allocation
5128 for (UseIterator j(ptn); j.has_next(); j.next()) {
5129 PointsToNode* use = j.get();
5130 if (use->is_LocalVar()) {
5131 use->dump(Verbose);
5132 } else if (Verbose) {
5133 use->dump();
5134 }
5135 }
5136 tty->cr();
5137 }
5138 }
5139 }
5140
5141 void ConnectionGraph::print_statistics() {
5142 tty->print_cr("No escape = %d, Arg escape = %d, Global escape = %d", AtomicAccess::load(&_no_escape_counter), AtomicAccess::load(&_arg_escape_counter), AtomicAccess::load(&_global_escape_counter));
5143 }
5144
5145 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5146 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5147 return;
5148 }
5149 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5150 JavaObjectNode* ptn = java_objects_worklist.at(next);
5151 if (ptn->ideal_node()->is_Allocate()) {
5152 if (ptn->escape_state() == PointsToNode::NoEscape) {
5153 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5154 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5155 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5156 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5157 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5158 } else {
5159 assert(false, "Unexpected Escape State");
5160 }
5161 }
5162 }
5163 }
5164
5165 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5166 if (_compile->directive()->TraceEscapeAnalysisOption) {
5167 assert(ptn != nullptr, "should not be null");
5168 assert(reason != nullptr, "should not be null");
5169 ptn->dump_header(true);
5170 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5171 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5172 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5173 }
5174 }
5175
5176 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5177 if (_compile->directive()->TraceEscapeAnalysisOption) {
5178 stringStream ss;
5179 ss.print("propagated from: ");
5180 from->dump(true, &ss, false);
5181 return ss.as_string();
5182 } else {
5183 return nullptr;
5184 }
5185 }
5186
5187 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5188 if (_compile->directive()->TraceEscapeAnalysisOption) {
5189 stringStream ss;
5190 ss.print("escapes as arg to:");
5191 call->dump("", false, &ss);
5192 return ss.as_string();
5193 } else {
5194 return nullptr;
5195 }
5196 }
5197
5198 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5199 if (_compile->directive()->TraceEscapeAnalysisOption) {
5200 stringStream ss;
5201 ss.print("is merged with other object: ");
5202 other->dump_header(true, &ss);
5203 return ss.as_string();
5204 } else {
5205 return nullptr;
5206 }
5207 }
5208
5209 #endif
5210
5211 void ConnectionGraph::record_for_optimizer(Node *n) {
5212 _igvn->_worklist.push(n);
5213 _igvn->add_users_to_worklist(n);
5214 }