1 /*
2 * Copyright (c) 2005, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/bcEscapeAnalyzer.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.hpp"
31 #include "memory/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 OpaqueConstantBool 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::DependencyType::NonFloatingNonNarrowing, 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(AddPNode::make_with_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, PhaseIterGVN::NodeOrigin::Graph);
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
1062 // If new_load is a Load but not from an AddP, it means that the load is folded into another
1063 // load. And since this load is not from a field, we cannot create a unique type for it.
1064 // For example:
1065 //
1066 // if (b) {
1067 // Holder h1 = new Holder();
1068 // Object o = ...;
1069 // h.o = o.getClass();
1070 // } else {
1071 // Holder h2 = ...;
1072 // }
1073 // Holder h = Phi(h1, h2);
1074 // Object r = h.o;
1075 //
1076 // Then, splitting r through the merge point results in:
1077 //
1078 // if (b) {
1079 // Holder h1 = new Holder();
1080 // Object o = ...;
1081 // h.o = o.getClass();
1082 // Object o1 = h.o;
1083 // } else {
1084 // Holder h2 = ...;
1085 // Object o2 = h2.o;
1086 // }
1087 // Object r = Phi(o1, o2);
1088 //
1089 // In this case, o1 is folded to o.getClass() which is a Load but not from an AddP, but from
1090 // an OopHandle that is loaded from the Klass of o.
1091 if (!new_addp->is_AddP()) {
1092 continue;
1093 }
1094 Node* base = get_addp_base(new_addp);
1095
1096 // The base might not be something that we can create an unique
1097 // type for. If that's the case we are done with that input.
1098 PointsToNode* jobj_ptn = unique_java_object(base);
1099 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1100 continue;
1101 }
1102
1103 // Push to alloc_worklist since the base has an unique_type
1104 alloc_worklist.append_if_missing(new_addp);
1105
1106 // Now let's add the node to the connection graph
1107 _nodes.at_grow(new_addp->_idx, nullptr);
1108 add_field(new_addp, fn->escape_state(), fn->offset());
1109 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1110
1111 // If the load doesn't load an object then it won't be
1112 // part of the connection graph
1113 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1114 if (curr_load_ptn != nullptr) {
1115 _nodes.at_grow(new_load->_idx, nullptr);
1116 add_local_var(new_load, curr_load_ptn->escape_state());
1117 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1118 }
1119 }
1120 }
1121 }
1122
1123 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1124 // We'll pass this to 'split_through_phi' so that it'll do the split even
1125 // though the load doesn't have an unique instance type.
1126 bool ignore_missing_instance_id = true;
1127
1128 // All AddPs are present in the connection graph
1129 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1130
1131 // Iterate over AddP looking for a Load
1132 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1133 Node* previous_load = previous_addp->raw_out(k);
1134 if (previous_load->is_Load()) {
1135 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1136
1137 // Takes care of updating CG and split_unique_types worklists due to cloned
1138 // AddP->Load.
1139 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1140
1141 _igvn->replace_node(previous_load, data_phi);
1142 }
1143 --k;
1144 k = MIN2(k, (int)previous_addp->outcnt()-1);
1145 }
1146
1147 // Remove the old AddP from the processing list because it's dead now
1148 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1149 alloc_worklist.remove_if_existing(previous_addp);
1150 }
1151
1152 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1153 // selector is:
1154 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1155 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1156 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1157 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1158 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1159 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1160 uint number_of_sr_objects = 0;
1161 for (uint i = 1; i < ophi->req(); i++) {
1162 Node* base = ophi->in(i);
1163 JavaObjectNode* ptn = unique_java_object(base);
1164
1165 if (ptn != nullptr && ptn->scalar_replaceable()) {
1166 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1167 selector->set_req(i, sr_obj_idx);
1168 number_of_sr_objects++;
1169 }
1170 }
1171
1172 return selector->as_Phi();
1173 }
1174
1175 // Returns true if the AddP node 'n' has at least one base that is a reducible
1176 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1177 // checked instead.
1178 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1179 PointsToNode* ptn = ptnode_adr(n->_idx);
1180 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1181 return false;
1182 }
1183
1184 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1185 Node* base = i.get()->ideal_node();
1186
1187 if (reducible_merges.member(base)) {
1188 return true;
1189 }
1190
1191 if (base->is_CastPP() || base->is_CheckCastPP()) {
1192 base = base->in(1);
1193 if (reducible_merges.member(base)) {
1194 return true;
1195 }
1196 }
1197 }
1198
1199 return false;
1200 }
1201
1202 // This method will call its helper method to reduce SafePoint nodes that use
1203 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1204 // "version" of Phi use the same debug information (regarding the Phi).
1205 // Therefore, I collect all safepoints and patch them all at once.
1206 //
1207 // The safepoints using the Phi node have to be processed before safepoints of
1208 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1209 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1210 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1211 // algorithm that process Phi's safepoints will think that the added Phi
1212 // reference is a regular reference.
1213 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1214 PhiNode* selector = create_selector(ophi);
1215 Unique_Node_List safepoints;
1216 Unique_Node_List casts;
1217
1218 // Just collect the users of the Phis for later processing
1219 // in the needed order.
1220 for (uint i = 0; i < ophi->outcnt(); i++) {
1221 Node* use = ophi->raw_out(i);
1222 if (use->is_SafePoint()) {
1223 safepoints.push(use);
1224 } else if (use->is_CastPP()) {
1225 casts.push(use);
1226 } else {
1227 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1228 }
1229 }
1230
1231 // Need to process safepoints using the Phi first
1232 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1233 return false;
1234 }
1235
1236 // Now process CastPP->safepoints
1237 for (uint i = 0; i < casts.size(); i++) {
1238 Node* cast = casts.at(i);
1239 Unique_Node_List cast_sfpts;
1240
1241 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1242 Node* use_use = cast->fast_out(j);
1243 if (use_use->is_SafePoint()) {
1244 cast_sfpts.push(use_use);
1245 } else {
1246 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1247 }
1248 }
1249
1250 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1251 return false;
1252 }
1253 }
1254
1255 return true;
1256 }
1257
1258 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1259 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1260 // SafePointScalarObjectNode for each scalar replaceable input. Each
1261 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1262 // check detailed description in SafePointScalarMergeNode class header.
1263 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1264 PhaseMacroExpand mexp(*_igvn);
1265 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1266 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1267
1268 Node* nsr_merge_pointer = ophi;
1269 if (cast != nullptr) {
1270 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1271 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::DependencyType::FloatingNarrowing, nullptr));
1272 }
1273
1274 for (uint spi = 0; spi < safepoints.size(); spi++) {
1275 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1276
1277 SafePointNode::NodeEdgeTempStorage non_debug_edges_worklist(*_igvn);
1278
1279 // All sfpt inputs are implicitly included into debug info during the scalarization process below.
1280 // Keep non-debug inputs separately, so they stay non-debug.
1281 sfpt->remove_non_debug_edges(non_debug_edges_worklist);
1282
1283 JVMState* jvms = sfpt->jvms();
1284 uint merge_idx = (sfpt->req() - jvms->scloff());
1285 int debug_start = jvms->debug_start();
1286
1287 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1288 smerge->init_req(0, _compile->root());
1289 _igvn->register_new_node_with_optimizer(smerge);
1290
1291 assert(sfpt->jvms()->endoff() == sfpt->req(), "no extra edges past debug info allowed");
1292
1293 // The next two inputs are:
1294 // (1) A copy of the original pointer to NSR objects.
1295 // (2) A selector, used to decide if we need to rematerialize an object
1296 // or use the pointer to a NSR object.
1297 // See more details of these fields in the declaration of SafePointScalarMergeNode.
1298 // It is safe to include them into debug info straight away since create_scalarized_object_description()
1299 // will include all newly added inputs into debug info anyway.
1300 sfpt->add_req(nsr_merge_pointer);
1301 sfpt->add_req(selector);
1302 sfpt->jvms()->set_endoff(sfpt->req());
1303
1304 for (uint i = 1; i < ophi->req(); i++) {
1305 Node* base = ophi->in(i);
1306 JavaObjectNode* ptn = unique_java_object(base);
1307
1308 // If the base is not scalar replaceable we don't need to register information about
1309 // it at this time.
1310 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1311 continue;
1312 }
1313
1314 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1315 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt);
1316 if (sobj == nullptr) {
1317 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1318 return false; // non-recoverable failure; recompile
1319 }
1320
1321 // Now make a pass over the debug information replacing any references
1322 // to the allocated object with "sobj"
1323 Node* ccpp = alloc->result_cast();
1324 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1325 non_debug_edges_worklist.remove_edge_if_present(ccpp); // drop scalarized input from non-debug info
1326
1327 // Register the scalarized object as a candidate for reallocation
1328 smerge->add_req(sobj);
1329 }
1330
1331 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1332 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1333 non_debug_edges_worklist.remove_edge_if_present(original_sfpt_parent); // drop scalarized input from non-debug info
1334
1335 // The call to 'replace_edges_in_range' above might have removed the
1336 // reference to ophi that we need at _merge_pointer_idx. The line below make
1337 // sure the reference is maintained.
1338 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1339
1340 sfpt->restore_non_debug_edges(non_debug_edges_worklist);
1341
1342 _igvn->_worklist.push(sfpt);
1343 }
1344
1345 return true;
1346 }
1347
1348 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1349 bool delay = _igvn->delay_transform();
1350 _igvn->set_delay_transform(true);
1351 _igvn->hash_delete(ophi);
1352
1353 // Copying all users first because some will be removed and others won't.
1354 // Ophi also may acquire some new users as part of Cast reduction.
1355 // CastPPs also need to be processed before CmpPs.
1356 Unique_Node_List castpps;
1357 Unique_Node_List others;
1358 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1359 Node* use = ophi->fast_out(i);
1360
1361 if (use->is_CastPP()) {
1362 castpps.push(use);
1363 } else if (use->is_AddP() || use->is_Cmp()) {
1364 others.push(use);
1365 } else {
1366 // Safepoints to be processed later; other users aren't expected here
1367 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1368 }
1369 }
1370
1371 _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1372
1373 // CastPPs need to be processed before Cmps because during the process of
1374 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1375 // by the If controlling the CastPP.
1376 for (uint i = 0; i < castpps.size(); i++) {
1377 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1378 _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1379 }
1380
1381 for (uint i = 0; i < others.size(); i++) {
1382 Node* use = others.at(i);
1383
1384 if (use->is_AddP()) {
1385 reduce_phi_on_field_access(use, alloc_worklist);
1386 _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1387 } else if(use->is_Cmp()) {
1388 reduce_phi_on_cmp(use);
1389 _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1390 }
1391 }
1392
1393 _igvn->set_delay_transform(delay);
1394 }
1395
1396 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1397 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1398 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1399 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1400 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1401
1402 for (uint i = 1; i < ophi->req(); i++) {
1403 Node* base = ophi->in(i);
1404 JavaObjectNode* ptn = unique_java_object(base);
1405
1406 if (ptn != nullptr && ptn->scalar_replaceable()) {
1407 new_phi->set_req(i, null_ptr);
1408 } else {
1409 new_phi->set_req(i, ophi->in(i));
1410 }
1411 }
1412
1413 for (int i = ophi->outcnt()-1; i >= 0;) {
1414 Node* out = ophi->raw_out(i);
1415
1416 if (out->is_ConstraintCast()) {
1417 const Type* out_t = _igvn->type(out)->make_ptr();
1418 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1419 bool change = out_new_t != out_t;
1420
1421 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1422 Node* out2 = out->raw_out(j);
1423 if (!out2->is_SafePoint()) {
1424 change = false;
1425 break;
1426 }
1427 }
1428
1429 if (change) {
1430 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::DependencyType::NonFloatingNarrowing, nullptr);
1431 _igvn->replace_node(out, new_cast);
1432 _igvn->register_new_node_with_optimizer(new_cast);
1433 }
1434 }
1435
1436 --i;
1437 i = MIN2(i, (int)ophi->outcnt()-1);
1438 }
1439
1440 _igvn->replace_node(ophi, new_phi);
1441 }
1442
1443 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1444 if (!C->do_reduce_allocation_merges()) return;
1445
1446 Unique_Node_List ideal_nodes;
1447 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1448 ideal_nodes.push(root);
1449
1450 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1451 Node* n = ideal_nodes.at(next);
1452
1453 if (n->is_SafePointScalarMerge()) {
1454 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1455
1456 // Validate inputs of merge
1457 for (uint i = 1; i < merge->req(); i++) {
1458 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1459 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1460 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1461 }
1462 }
1463
1464 // Validate users of merge
1465 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1466 Node* sfpt = merge->fast_out(i);
1467 if (sfpt->is_SafePoint()) {
1468 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1469
1470 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1471 assert(false, "SafePointScalarMerge nodes can't be nested.");
1472 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1473 }
1474 } else {
1475 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1476 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1477 }
1478 }
1479 }
1480
1481 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1482 Node* m = n->fast_out(i);
1483 ideal_nodes.push(m);
1484 }
1485 }
1486 }
1487
1488 // Returns true if there is an object in the scope of sfn that does not escape globally.
1489 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1490 Compile* C = _compile;
1491 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1492 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1493 DeoptimizeObjectsALot) {
1494 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1495 int num_locs = jvms->loc_size();
1496 for (int idx = 0; idx < num_locs; idx++) {
1497 Node* l = sfn->local(jvms, idx);
1498 if (not_global_escape(l)) {
1499 return true;
1500 }
1501 }
1502 }
1503 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1504 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1505 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1506 int num_mon = jvms->nof_monitors();
1507 for (int idx = 0; idx < num_mon; idx++) {
1508 Node* m = sfn->monitor_obj(jvms, idx);
1509 if (m != nullptr && not_global_escape(m)) {
1510 return true;
1511 }
1512 }
1513 }
1514 }
1515 return false;
1516 }
1517
1518 // Returns true if at least one of the arguments to the call is an object
1519 // that does not escape globally.
1520 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1521 if (call->method() != nullptr) {
1522 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1523 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1524 Node* p = call->in(idx);
1525 if (not_global_escape(p)) {
1526 return true;
1527 }
1528 }
1529 } else {
1530 const char* name = call->as_CallStaticJava()->_name;
1531 assert(name != nullptr, "no name");
1532 // no arg escapes through uncommon traps
1533 if (strcmp(name, "uncommon_trap") != 0) {
1534 // process_call_arguments() assumes that all arguments escape globally
1535 const TypeTuple* d = call->tf()->domain();
1536 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1537 const Type* at = d->field_at(i);
1538 if (at->isa_oopptr() != nullptr) {
1539 return true;
1540 }
1541 }
1542 }
1543 }
1544 return false;
1545 }
1546
1547
1548
1549 // Utility function for nodes that load an object
1550 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1551 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1552 // ThreadLocal has RawPtr type.
1553 const Type* t = _igvn->type(n);
1554 if (t->make_ptr() != nullptr) {
1555 Node* adr = n->in(MemNode::Address);
1556 #ifdef ASSERT
1557 if (!adr->is_AddP()) {
1558 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1559 } else {
1560 assert((ptnode_adr(adr->_idx) == nullptr ||
1561 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1562 }
1563 #endif
1564 add_local_var_and_edge(n, PointsToNode::NoEscape,
1565 adr, delayed_worklist);
1566 }
1567 }
1568
1569 // Populate Connection Graph with PointsTo nodes and create simple
1570 // connection graph edges.
1571 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1572 assert(!_verify, "this method should not be called for verification");
1573 PhaseGVN* igvn = _igvn;
1574 uint n_idx = n->_idx;
1575 PointsToNode* n_ptn = ptnode_adr(n_idx);
1576 if (n_ptn != nullptr) {
1577 return; // No need to redefine PointsTo node during first iteration.
1578 }
1579 int opcode = n->Opcode();
1580 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1581 if (gc_handled) {
1582 return; // Ignore node if already handled by GC.
1583 }
1584
1585 if (n->is_Call()) {
1586 // Arguments to allocation and locking don't escape.
1587 if (n->is_AbstractLock()) {
1588 // Put Lock and Unlock nodes on IGVN worklist to process them during
1589 // first IGVN optimization when escape information is still available.
1590 record_for_optimizer(n);
1591 } else if (n->is_Allocate()) {
1592 add_call_node(n->as_Call());
1593 record_for_optimizer(n);
1594 } else {
1595 if (n->is_CallStaticJava()) {
1596 const char* name = n->as_CallStaticJava()->_name;
1597 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1598 return; // Skip uncommon traps
1599 }
1600 }
1601 // Don't mark as processed since call's arguments have to be processed.
1602 delayed_worklist->push(n);
1603 // Check if a call returns an object.
1604 if ((n->as_Call()->returns_pointer() &&
1605 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1606 (n->is_CallStaticJava() &&
1607 n->as_CallStaticJava()->is_boxing_method())) {
1608 add_call_node(n->as_Call());
1609 }
1610 }
1611 return;
1612 }
1613 // Put this check here to process call arguments since some call nodes
1614 // point to phantom_obj.
1615 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1616 return; // Skip predefined nodes.
1617 }
1618 switch (opcode) {
1619 case Op_AddP: {
1620 Node* base = get_addp_base(n);
1621 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1622 // Field nodes are created for all field types. They are used in
1623 // adjust_scalar_replaceable_state() and split_unique_types().
1624 // Note, non-oop fields will have only base edges in Connection
1625 // Graph because such fields are not used for oop loads and stores.
1626 int offset = address_offset(n, igvn);
1627 add_field(n, PointsToNode::NoEscape, offset);
1628 if (ptn_base == nullptr) {
1629 delayed_worklist->push(n); // Process it later.
1630 } else {
1631 n_ptn = ptnode_adr(n_idx);
1632 add_base(n_ptn->as_Field(), ptn_base);
1633 }
1634 break;
1635 }
1636 case Op_CastX2P: {
1637 map_ideal_node(n, phantom_obj);
1638 break;
1639 }
1640 case Op_CastPP:
1641 case Op_CheckCastPP:
1642 case Op_EncodeP:
1643 case Op_DecodeN:
1644 case Op_EncodePKlass:
1645 case Op_DecodeNKlass: {
1646 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1647 break;
1648 }
1649 case Op_CMoveP: {
1650 add_local_var(n, PointsToNode::NoEscape);
1651 // Do not add edges during first iteration because some could be
1652 // not defined yet.
1653 delayed_worklist->push(n);
1654 break;
1655 }
1656 case Op_ConP:
1657 case Op_ConN:
1658 case Op_ConNKlass: {
1659 // assume all oop constants globally escape except for null
1660 PointsToNode::EscapeState es;
1661 const Type* t = igvn->type(n);
1662 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1663 es = PointsToNode::NoEscape;
1664 } else {
1665 es = PointsToNode::GlobalEscape;
1666 }
1667 PointsToNode* ptn_con = add_java_object(n, es);
1668 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1669 break;
1670 }
1671 case Op_CreateEx: {
1672 // assume that all exception objects globally escape
1673 map_ideal_node(n, phantom_obj);
1674 break;
1675 }
1676 case Op_LoadKlass:
1677 case Op_LoadNKlass: {
1678 // Unknown class is loaded
1679 map_ideal_node(n, phantom_obj);
1680 break;
1681 }
1682 case Op_LoadP:
1683 case Op_LoadN: {
1684 add_objload_to_connection_graph(n, delayed_worklist);
1685 break;
1686 }
1687 case Op_Parm: {
1688 map_ideal_node(n, phantom_obj);
1689 break;
1690 }
1691 case Op_PartialSubtypeCheck: {
1692 // Produces Null or notNull and is used in only in CmpP so
1693 // phantom_obj could be used.
1694 map_ideal_node(n, phantom_obj); // Result is unknown
1695 break;
1696 }
1697 case Op_Phi: {
1698 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1699 // ThreadLocal has RawPtr type.
1700 const Type* t = n->as_Phi()->type();
1701 if (t->make_ptr() != nullptr) {
1702 add_local_var(n, PointsToNode::NoEscape);
1703 // Do not add edges during first iteration because some could be
1704 // not defined yet.
1705 delayed_worklist->push(n);
1706 }
1707 break;
1708 }
1709 case Op_Proj: {
1710 // we are only interested in the oop result projection from a call
1711 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&1712 n->in(0)->as_Call()->returns_pointer()) {1713 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);1714 }
1715 break;
1716 }
1717 case Op_Rethrow: // Exception object escapes
1718 case Op_Return: {
1719 if (n->req() > TypeFunc::Parms &&
1720 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1721 // Treat Return value as LocalVar with GlobalEscape escape state.
1722 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1723 }
1724 break;
1725 }
1726 case Op_CompareAndExchangeP:
1727 case Op_CompareAndExchangeN:
1728 case Op_GetAndSetP:
1729 case Op_GetAndSetN: {
1730 add_objload_to_connection_graph(n, delayed_worklist);
1731 // fall-through
1732 }
1733 case Op_StoreP:
1734 case Op_StoreN:
1735 case Op_StoreNKlass:
1736 case Op_WeakCompareAndSwapP:
1737 case Op_WeakCompareAndSwapN:
1738 case Op_CompareAndSwapP:
1739 case Op_CompareAndSwapN: {
1740 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1741 break;
1742 }
1743 case Op_AryEq:
1744 case Op_CountPositives:
1745 case Op_StrComp:
1746 case Op_StrEquals:
1747 case Op_StrIndexOf:
1748 case Op_StrIndexOfChar:
1749 case Op_StrInflatedCopy:
1750 case Op_StrCompressedCopy:
1751 case Op_VectorizedHashCode:
1752 case Op_EncodeISOArray: {
1753 add_local_var(n, PointsToNode::ArgEscape);
1754 delayed_worklist->push(n); // Process it later.
1755 break;
1756 }
1757 case Op_ThreadLocal: {
1758 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1759 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1760 break;
1761 }
1762 case Op_Blackhole: {
1763 // All blackhole pointer arguments are globally escaping.
1764 // Only do this if there is at least one pointer argument.
1765 // Do not add edges during first iteration because some could be
1766 // not defined yet, defer to final step.
1767 for (uint i = 0; i < n->req(); i++) {
1768 Node* in = n->in(i);
1769 if (in != nullptr) {
1770 const Type* at = _igvn->type(in);
1771 if (!at->isa_ptr()) continue;
1772
1773 add_local_var(n, PointsToNode::GlobalEscape);
1774 delayed_worklist->push(n);
1775 break;
1776 }
1777 }
1778 break;
1779 }
1780 default:
1781 ; // Do nothing for nodes not related to EA.
1782 }
1783 return;
1784 }
1785
1786 // Add final simple edges to graph.
1787 void ConnectionGraph::add_final_edges(Node *n) {
1788 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1789 #ifdef ASSERT
1790 if (_verify && n_ptn->is_JavaObject())
1791 return; // This method does not change graph for JavaObject.
1792 #endif
1793
1794 if (n->is_Call()) {
1795 process_call_arguments(n->as_Call());
1796 return;
1797 }
1798 assert(n->is_Store() || n->is_LoadStore() ||
1799 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1800 "node should be registered already");
1801 int opcode = n->Opcode();
1802 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1803 if (gc_handled) {
1804 return; // Ignore node if already handled by GC.
1805 }
1806 switch (opcode) {
1807 case Op_AddP: {
1808 Node* base = get_addp_base(n);
1809 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1810 assert(ptn_base != nullptr, "field's base should be registered");
1811 add_base(n_ptn->as_Field(), ptn_base);
1812 break;
1813 }
1814 case Op_CastPP:
1815 case Op_CheckCastPP:
1816 case Op_EncodeP:
1817 case Op_DecodeN:
1818 case Op_EncodePKlass:
1819 case Op_DecodeNKlass: {
1820 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1821 break;
1822 }
1823 case Op_CMoveP: {
1824 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1825 Node* in = n->in(i);
1826 if (in == nullptr) {
1827 continue; // ignore null
1828 }
1829 Node* uncast_in = in->uncast();
1830 if (uncast_in->is_top() || uncast_in == n) {
1831 continue; // ignore top or inputs which go back this node
1832 }
1833 PointsToNode* ptn = ptnode_adr(in->_idx);
1834 assert(ptn != nullptr, "node should be registered");
1835 add_edge(n_ptn, ptn);
1836 }
1837 break;
1838 }
1839 case Op_LoadP:
1840 case Op_LoadN: {
1841 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1842 // ThreadLocal has RawPtr type.
1843 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1844 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1845 break;
1846 }
1847 case Op_Phi: {
1848 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1849 // ThreadLocal has RawPtr type.
1850 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1851 for (uint i = 1; i < n->req(); i++) {
1852 Node* in = n->in(i);
1853 if (in == nullptr) {
1854 continue; // ignore null
1855 }
1856 Node* uncast_in = in->uncast();
1857 if (uncast_in->is_top() || uncast_in == n) {
1858 continue; // ignore top or inputs which go back this node
1859 }
1860 PointsToNode* ptn = ptnode_adr(in->_idx);
1861 assert(ptn != nullptr, "node should be registered");
1862 add_edge(n_ptn, ptn);
1863 }
1864 break;
1865 }
1866 case Op_Proj: {
1867 // we are only interested in the oop result projection from a call1868 assert(n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&1869 n->in(0)->as_Call()->returns_pointer(), "Unexpected node type");1870 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1871 break;
1872 }
1873 case Op_Rethrow: // Exception object escapes
1874 case Op_Return: {
1875 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1876 "Unexpected node type");
1877 // Treat Return value as LocalVar with GlobalEscape escape state.
1878 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1879 break;
1880 }
1881 case Op_CompareAndExchangeP:
1882 case Op_CompareAndExchangeN:
1883 case Op_GetAndSetP:
1884 case Op_GetAndSetN:{
1885 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1886 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1887 // fall-through
1888 }
1889 case Op_CompareAndSwapP:
1890 case Op_CompareAndSwapN:
1891 case Op_WeakCompareAndSwapP:
1892 case Op_WeakCompareAndSwapN:
1893 case Op_StoreP:
1894 case Op_StoreN:
1895 case Op_StoreNKlass:{
1896 add_final_edges_unsafe_access(n, opcode);
1897 break;
1898 }
1899 case Op_VectorizedHashCode:
1900 case Op_AryEq:
1901 case Op_CountPositives:
1902 case Op_StrComp:
1903 case Op_StrEquals:
1904 case Op_StrIndexOf:
1905 case Op_StrIndexOfChar:
1906 case Op_StrInflatedCopy:
1907 case Op_StrCompressedCopy:
1908 case Op_EncodeISOArray: {
1909 // char[]/byte[] arrays passed to string intrinsic do not escape but
1910 // they are not scalar replaceable. Adjust escape state for them.
1911 // Start from in(2) edge since in(1) is memory edge.
1912 for (uint i = 2; i < n->req(); i++) {
1913 Node* adr = n->in(i);
1914 const Type* at = _igvn->type(adr);
1915 if (!adr->is_top() && at->isa_ptr()) {
1916 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1917 at->isa_ptr() != nullptr, "expecting a pointer");
1918 if (adr->is_AddP()) {
1919 adr = get_addp_base(adr);
1920 }
1921 PointsToNode* ptn = ptnode_adr(adr->_idx);
1922 assert(ptn != nullptr, "node should be registered");
1923 add_edge(n_ptn, ptn);
1924 }
1925 }
1926 break;
1927 }
1928 case Op_Blackhole: {
1929 // All blackhole pointer arguments are globally escaping.
1930 for (uint i = 0; i < n->req(); i++) {
1931 Node* in = n->in(i);
1932 if (in != nullptr) {
1933 const Type* at = _igvn->type(in);
1934 if (!at->isa_ptr()) continue;
1935
1936 if (in->is_AddP()) {
1937 in = get_addp_base(in);
1938 }
1939
1940 PointsToNode* ptn = ptnode_adr(in->_idx);
1941 assert(ptn != nullptr, "should be defined already");
1942 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1943 add_edge(n_ptn, ptn);
1944 }
1945 }
1946 break;
1947 }
1948 default: {
1949 // This method should be called only for EA specific nodes which may
1950 // miss some edges when they were created.
1951 #ifdef ASSERT
1952 n->dump(1);
1953 #endif
1954 guarantee(false, "unknown node");
1955 }
1956 }
1957 return;
1958 }
1959
1960 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1961 Node* adr = n->in(MemNode::Address);
1962 const Type* adr_type = _igvn->type(adr);
1963 adr_type = adr_type->make_ptr();
1964 if (adr_type == nullptr) {
1965 return; // skip dead nodes
1966 }
1967 if (adr_type->isa_oopptr()
1968 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1969 && adr_type == TypeRawPtr::NOTNULL
1970 && is_captured_store_address(adr))) {
1971 delayed_worklist->push(n); // Process it later.
1972 #ifdef ASSERT
1973 assert (adr->is_AddP(), "expecting an AddP");
1974 if (adr_type == TypeRawPtr::NOTNULL) {
1975 // Verify a raw address for a store captured by Initialize node.
1976 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1977 assert(offs != Type::OffsetBot, "offset must be a constant");
1978 }
1979 #endif
1980 } else {
1981 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1982 if (adr->is_BoxLock()) {
1983 return;
1984 }
1985 // Stored value escapes in unsafe access.
1986 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1987 delayed_worklist->push(n); // Process unsafe access later.
1988 return;
1989 }
1990 #ifdef ASSERT
1991 n->dump(1);
1992 assert(false, "not unsafe");
1993 #endif
1994 }
1995 }
1996
1997 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1998 Node* adr = n->in(MemNode::Address);
1999 const Type *adr_type = _igvn->type(adr);
2000 adr_type = adr_type->make_ptr();
2001 #ifdef ASSERT
2002 if (adr_type == nullptr) {
2003 n->dump(1);
2004 assert(adr_type != nullptr, "dead node should not be on list");
2005 return true;
2006 }
2007 #endif
2008
2009 if (adr_type->isa_oopptr()
2010 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
2011 && adr_type == TypeRawPtr::NOTNULL
2012 && is_captured_store_address(adr))) {
2013 // Point Address to Value
2014 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2015 assert(adr_ptn != nullptr &&
2016 adr_ptn->as_Field()->is_oop(), "node should be registered");
2017 Node* val = n->in(MemNode::ValueIn);
2018 PointsToNode* ptn = ptnode_adr(val->_idx);
2019 assert(ptn != nullptr, "node should be registered");
2020 add_edge(adr_ptn, ptn);
2021 return true;
2022 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2023 // Stored value escapes in unsafe access.
2024 Node* val = n->in(MemNode::ValueIn);
2025 PointsToNode* ptn = ptnode_adr(val->_idx);
2026 assert(ptn != nullptr, "node should be registered");
2027 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2028 // Add edge to object for unsafe access with offset.
2029 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2030 assert(adr_ptn != nullptr, "node should be registered");
2031 if (adr_ptn->is_Field()) {
2032 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2033 add_edge(adr_ptn, ptn);
2034 }
2035 return true;
2036 }
2037 #ifdef ASSERT
2038 n->dump(1);
2039 assert(false, "not unsafe");
2040 #endif
2041 return false;
2042 }
2043
2044 void ConnectionGraph::add_call_node(CallNode* call) {
2045 assert(call->returns_pointer(), "only for call which returns pointer");
2046 uint call_idx = call->_idx;
2047 if (call->is_Allocate()) {
2048 Node* k = call->in(AllocateNode::KlassNode);
2049 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2050 assert(kt != nullptr, "TypeKlassPtr required.");
2051 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2052 bool scalar_replaceable = true;
2053 NOT_PRODUCT(const char* nsr_reason = "");
2054 if (call->is_AllocateArray()) {
2055 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2056 es = PointsToNode::GlobalEscape;
2057 } else {
2058 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2059 if (length < 0) {
2060 // Not scalar replaceable if the length is not constant.
2061 scalar_replaceable = false;
2062 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2063 } else if (length > EliminateAllocationArraySizeLimit) {
2064 // Not scalar replaceable if the length is too big.
2065 scalar_replaceable = false;
2066 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2067 }
2068 }
2069 } else { // Allocate instance
2070 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2071 es = PointsToNode::GlobalEscape;
2072 } else {
2073 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2074 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2075 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2076 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2077 !ik->can_be_instantiated() ||
2078 ik->has_finalizer()) {
2079 es = PointsToNode::GlobalEscape;
2080 } else {
2081 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2082 if (nfields > EliminateAllocationFieldsLimit) {
2083 // Not scalar replaceable if there are too many fields.
2084 scalar_replaceable = false;
2085 NOT_PRODUCT(nsr_reason = "has too many fields");
2086 }
2087 }
2088 }
2089 }
2090 add_java_object(call, es);
2091 PointsToNode* ptn = ptnode_adr(call_idx);
2092 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2093 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2094 }
2095 } else if (call->is_CallStaticJava()) {
2096 // Call nodes could be different types:
2097 //
2098 // 1. CallDynamicJavaNode (what happened during call is unknown):
2099 //
2100 // - mapped to GlobalEscape JavaObject node if oop is returned;
2101 //
2102 // - all oop arguments are escaping globally;
2103 //
2104 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2105 //
2106 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2107 //
2108 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2109 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2110 // during call is returned;
2111 // - mapped to ArgEscape LocalVar node pointed to object arguments
2112 // which are returned and does not escape during call;
2113 //
2114 // - oop arguments escaping status is defined by bytecode analysis;
2115 //
2116 // For a static call, we know exactly what method is being called.
2117 // Use bytecode estimator to record whether the call's return value escapes.
2118 ciMethod* meth = call->as_CallJava()->method();
2119 if (meth == nullptr) {
2120 assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub(), "TODO: add failed case check");
2121 // Returns a newly allocated non-escaped object.
2122 add_java_object(call, PointsToNode::NoEscape);
2123 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2124 } else if (meth->is_boxing_method()) {
2125 // Returns boxing object
2126 PointsToNode::EscapeState es;
2127 vmIntrinsics::ID intr = meth->intrinsic_id();
2128 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2129 // It does not escape if object is always allocated.
2130 es = PointsToNode::NoEscape;
2131 } else {
2132 // It escapes globally if object could be loaded from cache.
2133 es = PointsToNode::GlobalEscape;
2134 }
2135 add_java_object(call, es);
2136 if (es == PointsToNode::GlobalEscape) {
2137 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2138 }
2139 } else {
2140 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2141 call_analyzer->copy_dependencies(_compile->dependencies());
2142 if (call_analyzer->is_return_allocated()) {
2143 // Returns a newly allocated non-escaped object, simply
2144 // update dependency information.
2145 // Mark it as NoEscape so that objects referenced by
2146 // it's fields will be marked as NoEscape at least.
2147 add_java_object(call, PointsToNode::NoEscape);
2148 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2149 } else {
2150 // Determine whether any arguments are returned.2151 const TypeTuple* d = call->tf()->domain();2152 bool ret_arg = false;2153 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {2154 if (d->field_at(i)->isa_ptr() != nullptr &&2155 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {2156 ret_arg = true;2157 break;2158 }2159 }2160 if (ret_arg) {
2161 add_local_var(call, PointsToNode::ArgEscape);
2162 } else {
2163 // Returns unknown object.
2164 map_ideal_node(call, phantom_obj);
2165 }
2166 }
2167 }
2168 } else {
2169 // An other type of call, assume the worst case:
2170 // returned value is unknown and globally escapes.
2171 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2172 map_ideal_node(call, phantom_obj);
2173 }
2174 }
2175
2176 void ConnectionGraph::process_call_arguments(CallNode *call) {
2177 bool is_arraycopy = false;
2178 switch (call->Opcode()) {
2179 #ifdef ASSERT
2180 case Op_Allocate:
2181 case Op_AllocateArray:
2182 case Op_Lock:
2183 case Op_Unlock:
2184 assert(false, "should be done already");
2185 break;
2186 #endif
2187 case Op_ArrayCopy:
2188 case Op_CallLeafNoFP:
2189 // Most array copies are ArrayCopy nodes at this point but there
2190 // are still a few direct calls to the copy subroutines (See
2191 // PhaseStringOpts::copy_string())
2192 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2193 call->as_CallLeaf()->is_call_to_arraycopystub();
2194 // fall through
2195 case Op_CallLeafVector:
2196 case Op_CallLeaf: {
2197 // Stub calls, objects do not escape but they are not scale replaceable.
2198 // Adjust escape state for outgoing arguments.
2199 const TypeTuple * d = call->tf()->domain();
2200 bool src_has_oops = false;
2201 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2202 const Type* at = d->field_at(i);
2203 Node *arg = call->in(i);
2204 if (arg == nullptr) {
2205 continue;
2206 }
2207 const Type *aat = _igvn->type(arg);
2208 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2209 continue;
2210 }
2211 if (arg->is_AddP()) {
2212 //
2213 // The inline_native_clone() case when the arraycopy stub is called
2214 // after the allocation before Initialize and CheckCastPP nodes.
2215 // Or normal arraycopy for object arrays case.
2216 //
2217 // Set AddP's base (Allocate) as not scalar replaceable since
2218 // pointer to the base (with offset) is passed as argument.
2219 //
2220 arg = get_addp_base(arg);
2221 }
2222 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2223 assert(arg_ptn != nullptr, "should be registered");
2224 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2225 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2226 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2227 aat->isa_ptr() != nullptr, "expecting an Ptr");
2228 bool arg_has_oops = aat->isa_oopptr() &&
2229 (aat->isa_instptr() ||
2230 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)));
2231 if (i == TypeFunc::Parms) {
2232 src_has_oops = arg_has_oops;
2233 }
2234 //
2235 // src or dst could be j.l.Object when other is basic type array:
2236 //
2237 // arraycopy(char[],0,Object*,0,size);
2238 // arraycopy(Object*,0,char[],0,size);
2239 //
2240 // Don't add edges in such cases.
2241 //
2242 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2243 arg_has_oops && (i > TypeFunc::Parms);
2244 #ifdef ASSERT
2245 if (!(is_arraycopy ||
2246 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2247 (call->as_CallLeaf()->_name != nullptr &&
2248 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2249 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2250 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2251 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2252 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2253 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2254 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2255 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2256 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2257 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2258 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2259 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2260 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2261 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2262 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2263 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2264 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2265 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2266 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2267 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2268 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2269 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2270 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2271 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2272 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2273 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2274 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2275 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2276 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2277 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2278 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2279 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2280 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2281 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2282 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2283 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2284 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2285 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2286 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2287 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2288 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2289 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2290 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2291 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2292 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2293 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2294 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2295 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2296 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2297 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2298 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2299 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2300 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2301 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2302 ))) {
2303 call->dump();
2304 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2305 }
2306 #endif
2307 // Always process arraycopy's destination object since
2308 // we need to add all possible edges to references in
2309 // source object.
2310 if (arg_esc >= PointsToNode::ArgEscape &&
2311 !arg_is_arraycopy_dest) {
2312 continue;
2313 }
2314 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2315 if (call->is_ArrayCopy()) {
2316 ArrayCopyNode* ac = call->as_ArrayCopy();
2317 if (ac->is_clonebasic() ||
2318 ac->is_arraycopy_validated() ||
2319 ac->is_copyof_validated() ||
2320 ac->is_copyofrange_validated()) {
2321 es = PointsToNode::NoEscape;
2322 }
2323 }
2324 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2325 if (arg_is_arraycopy_dest) {
2326 Node* src = call->in(TypeFunc::Parms);
2327 if (src->is_AddP()) {
2328 src = get_addp_base(src);
2329 }
2330 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2331 assert(src_ptn != nullptr, "should be registered");
2332 // Special arraycopy edge:
2333 // Only escape state of destination object's fields affects
2334 // escape state of fields in source object.
2335 add_arraycopy(call, es, src_ptn, arg_ptn);
2336 }
2337 }
2338 }
2339 break;
2340 }
2341 case Op_CallStaticJava: {
2342 // For a static call, we know exactly what method is being called.
2343 // Use bytecode estimator to record the call's escape affects
2344 #ifdef ASSERT
2345 const char* name = call->as_CallStaticJava()->_name;
2346 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2347 #endif
2348 ciMethod* meth = call->as_CallJava()->method();
2349 if ((meth != nullptr) && meth->is_boxing_method()) {
2350 break; // Boxing methods do not modify any oops.
2351 }
2352 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2353 // fall-through if not a Java method or no analyzer information
2354 if (call_analyzer != nullptr) {
2355 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2356 const TypeTuple* d = call->tf()->domain();2357 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {2358 const Type* at = d->field_at(i);2359 int k = i - TypeFunc::Parms;2360 Node* arg = call->in(i);
2361 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2362 if (at->isa_ptr() != nullptr &&2363 call_analyzer->is_arg_returned(k)) {
2364 // The call returns arguments.
2365 if (call_ptn != nullptr) { // Is call's result used?
2366 assert(call_ptn->is_LocalVar(), "node should be registered");
2367 assert(arg_ptn != nullptr, "node should be registered");
2368 add_edge(call_ptn, arg_ptn);
2369 }
2370 }
2371 if (at->isa_oopptr() != nullptr &&
2372 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2373 if (!call_analyzer->is_arg_stack(k)) {
2374 // The argument global escapes
2375 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2376 } else {
2377 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2378 if (!call_analyzer->is_arg_local(k)) {
2379 // The argument itself doesn't escape, but any fields might
2380 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2381 }
2382 }
2383 }
2384 }
2385 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2386 // The call returns arguments.
2387 assert(call_ptn->edge_count() > 0, "sanity");
2388 if (!call_analyzer->is_return_local()) {
2389 // Returns also unknown object.
2390 add_edge(call_ptn, phantom_obj);
2391 }
2392 }
2393 break;
2394 }
2395 }
2396 default: {
2397 // Fall-through here if not a Java method or no analyzer information
2398 // or some other type of call, assume the worst case: all arguments
2399 // globally escape.
2400 const TypeTuple* d = call->tf()->domain();
2401 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2402 const Type* at = d->field_at(i);
2403 if (at->isa_oopptr() != nullptr) {
2404 Node* arg = call->in(i);
2405 if (arg->is_AddP()) {
2406 arg = get_addp_base(arg);
2407 }
2408 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2409 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2410 }
2411 }
2412 }
2413 }
2414 }
2415
2416
2417 // Finish Graph construction.
2418 bool ConnectionGraph::complete_connection_graph(
2419 GrowableArray<PointsToNode*>& ptnodes_worklist,
2420 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2421 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2422 GrowableArray<FieldNode*>& oop_fields_worklist) {
2423 // Normally only 1-3 passes needed to build Connection Graph depending
2424 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2425 // Set limit to 20 to catch situation when something did go wrong and
2426 // bailout Escape Analysis.
2427 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2428 #define GRAPH_BUILD_ITER_LIMIT 20
2429
2430 // Propagate GlobalEscape and ArgEscape escape states and check that
2431 // we still have non-escaping objects. The method pushs on _worklist
2432 // Field nodes which reference phantom_object.
2433 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2434 return false; // Nothing to do.
2435 }
2436 // Now propagate references to all JavaObject nodes.
2437 int java_objects_length = java_objects_worklist.length();
2438 elapsedTimer build_time;
2439 build_time.start();
2440 elapsedTimer time;
2441 bool timeout = false;
2442 int new_edges = 1;
2443 int iterations = 0;
2444 do {
2445 while ((new_edges > 0) &&
2446 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2447 double start_time = time.seconds();
2448 time.start();
2449 new_edges = 0;
2450 // Propagate references to phantom_object for nodes pushed on _worklist
2451 // by find_non_escaped_objects() and find_field_value().
2452 new_edges += add_java_object_edges(phantom_obj, false);
2453 for (int next = 0; next < java_objects_length; ++next) {
2454 JavaObjectNode* ptn = java_objects_worklist.at(next);
2455 new_edges += add_java_object_edges(ptn, true);
2456
2457 #define SAMPLE_SIZE 4
2458 if ((next % SAMPLE_SIZE) == 0) {
2459 // Each 4 iterations calculate how much time it will take
2460 // to complete graph construction.
2461 time.stop();
2462 // Poll for requests from shutdown mechanism to quiesce compiler
2463 // because Connection graph construction may take long time.
2464 CompileBroker::maybe_block();
2465 double stop_time = time.seconds();
2466 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2467 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2468 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2469 timeout = true;
2470 break; // Timeout
2471 }
2472 start_time = stop_time;
2473 time.start();
2474 }
2475 #undef SAMPLE_SIZE
2476
2477 }
2478 if (timeout) break;
2479 if (new_edges > 0) {
2480 // Update escape states on each iteration if graph was updated.
2481 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2482 return false; // Nothing to do.
2483 }
2484 }
2485 time.stop();
2486 if (time.seconds() >= EscapeAnalysisTimeout) {
2487 timeout = true;
2488 break;
2489 }
2490 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2491 }
2492 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2493 time.start();
2494 // Find fields which have unknown value.
2495 int fields_length = oop_fields_worklist.length();
2496 for (int next = 0; next < fields_length; next++) {
2497 FieldNode* field = oop_fields_worklist.at(next);
2498 if (field->edge_count() == 0) {
2499 new_edges += find_field_value(field);
2500 // This code may added new edges to phantom_object.
2501 // Need an other cycle to propagate references to phantom_object.
2502 }
2503 }
2504 time.stop();
2505 if (time.seconds() >= EscapeAnalysisTimeout) {
2506 timeout = true;
2507 break;
2508 }
2509 } else {
2510 new_edges = 0; // Bailout
2511 }
2512 } while (new_edges > 0);
2513
2514 build_time.stop();
2515 _build_time = build_time.seconds();
2516 _build_iterations = iterations;
2517
2518 // Bailout if passed limits.
2519 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2520 Compile* C = _compile;
2521 if (C->log() != nullptr) {
2522 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2523 C->log()->text("%s", timeout ? "time" : "iterations");
2524 C->log()->end_elem(" limit'");
2525 }
2526 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2527 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2528 // Possible infinite build_connection_graph loop,
2529 // bailout (no changes to ideal graph were made).
2530 return false;
2531 }
2532
2533 #undef GRAPH_BUILD_ITER_LIMIT
2534
2535 // Find fields initialized by null for non-escaping Allocations.
2536 int non_escaped_length = non_escaped_allocs_worklist.length();
2537 for (int next = 0; next < non_escaped_length; next++) {
2538 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2539 PointsToNode::EscapeState es = ptn->escape_state();
2540 assert(es <= PointsToNode::ArgEscape, "sanity");
2541 if (es == PointsToNode::NoEscape) {
2542 if (find_init_values_null(ptn, _igvn) > 0) {
2543 // Adding references to null object does not change escape states
2544 // since it does not escape. Also no fields are added to null object.
2545 add_java_object_edges(null_obj, false);
2546 }
2547 }
2548 Node* n = ptn->ideal_node();
2549 if (n->is_Allocate()) {
2550 // The object allocated by this Allocate node will never be
2551 // seen by an other thread. Mark it so that when it is
2552 // expanded no MemBarStoreStore is added.
2553 InitializeNode* ini = n->as_Allocate()->initialization();
2554 if (ini != nullptr)
2555 ini->set_does_not_escape();
2556 }
2557 }
2558 return true; // Finished graph construction.
2559 }
2560
2561 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2562 // and check that we still have non-escaping java objects.
2563 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2564 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2565 bool print_method) {
2566 GrowableArray<PointsToNode*> escape_worklist;
2567 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2568 int ptnodes_length = ptnodes_worklist.length();
2569 for (int next = 0; next < ptnodes_length; ++next) {
2570 PointsToNode* ptn = ptnodes_worklist.at(next);
2571 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2572 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2573 escape_worklist.push(ptn);
2574 }
2575 }
2576 // Set escape states to referenced nodes (edges list).
2577 while (escape_worklist.length() > 0) {
2578 PointsToNode* ptn = escape_worklist.pop();
2579 PointsToNode::EscapeState es = ptn->escape_state();
2580 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2581 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2582 es >= PointsToNode::ArgEscape) {
2583 // GlobalEscape or ArgEscape state of field means it has unknown value.
2584 if (add_edge(ptn, phantom_obj)) {
2585 // New edge was added
2586 add_field_uses_to_worklist(ptn->as_Field());
2587 }
2588 }
2589 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2590 PointsToNode* e = i.get();
2591 if (e->is_Arraycopy()) {
2592 assert(ptn->arraycopy_dst(), "sanity");
2593 // Propagate only fields escape state through arraycopy edge.
2594 if (e->fields_escape_state() < field_es) {
2595 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2596 escape_worklist.push(e);
2597 }
2598 } else if (es >= field_es) {
2599 // fields_escape_state is also set to 'es' if it is less than 'es'.
2600 if (e->escape_state() < es) {
2601 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2602 escape_worklist.push(e);
2603 }
2604 } else {
2605 // Propagate field escape state.
2606 bool es_changed = false;
2607 if (e->fields_escape_state() < field_es) {
2608 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2609 es_changed = true;
2610 }
2611 if ((e->escape_state() < field_es) &&
2612 e->is_Field() && ptn->is_JavaObject() &&
2613 e->as_Field()->is_oop()) {
2614 // Change escape state of referenced fields.
2615 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2616 es_changed = true;
2617 } else if (e->escape_state() < es) {
2618 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2619 es_changed = true;
2620 }
2621 if (es_changed) {
2622 escape_worklist.push(e);
2623 }
2624 }
2625 if (print_method) {
2626 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2627 }
2628 }
2629 }
2630 // Remove escaped objects from non_escaped list.
2631 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2632 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2633 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2634 non_escaped_allocs_worklist.delete_at(next);
2635 }
2636 if (ptn->escape_state() == PointsToNode::NoEscape) {
2637 // Find fields in non-escaped allocations which have unknown value.
2638 find_init_values_phantom(ptn);
2639 }
2640 }
2641 return (non_escaped_allocs_worklist.length() > 0);
2642 }
2643
2644 // Add all references to JavaObject node by walking over all uses.
2645 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2646 int new_edges = 0;
2647 if (populate_worklist) {
2648 // Populate _worklist by uses of jobj's uses.
2649 for (UseIterator i(jobj); i.has_next(); i.next()) {
2650 PointsToNode* use = i.get();
2651 if (use->is_Arraycopy()) {
2652 continue;
2653 }
2654 add_uses_to_worklist(use);
2655 if (use->is_Field() && use->as_Field()->is_oop()) {
2656 // Put on worklist all field's uses (loads) and
2657 // related field nodes (same base and offset).
2658 add_field_uses_to_worklist(use->as_Field());
2659 }
2660 }
2661 }
2662 for (int l = 0; l < _worklist.length(); l++) {
2663 PointsToNode* use = _worklist.at(l);
2664 if (PointsToNode::is_base_use(use)) {
2665 // Add reference from jobj to field and from field to jobj (field's base).
2666 use = PointsToNode::get_use_node(use)->as_Field();
2667 if (add_base(use->as_Field(), jobj)) {
2668 new_edges++;
2669 }
2670 continue;
2671 }
2672 assert(!use->is_JavaObject(), "sanity");
2673 if (use->is_Arraycopy()) {
2674 if (jobj == null_obj) { // null object does not have field edges
2675 continue;
2676 }
2677 // Added edge from Arraycopy node to arraycopy's source java object
2678 if (add_edge(use, jobj)) {
2679 jobj->set_arraycopy_src();
2680 new_edges++;
2681 }
2682 // and stop here.
2683 continue;
2684 }
2685 if (!add_edge(use, jobj)) {
2686 continue; // No new edge added, there was such edge already.
2687 }
2688 new_edges++;
2689 if (use->is_LocalVar()) {
2690 add_uses_to_worklist(use);
2691 if (use->arraycopy_dst()) {
2692 for (EdgeIterator i(use); i.has_next(); i.next()) {
2693 PointsToNode* e = i.get();
2694 if (e->is_Arraycopy()) {
2695 if (jobj == null_obj) { // null object does not have field edges
2696 continue;
2697 }
2698 // Add edge from arraycopy's destination java object to Arraycopy node.
2699 if (add_edge(jobj, e)) {
2700 new_edges++;
2701 jobj->set_arraycopy_dst();
2702 }
2703 }
2704 }
2705 }
2706 } else {
2707 // Added new edge to stored in field values.
2708 // Put on worklist all field's uses (loads) and
2709 // related field nodes (same base and offset).
2710 add_field_uses_to_worklist(use->as_Field());
2711 }
2712 }
2713 _worklist.clear();
2714 _in_worklist.reset();
2715 return new_edges;
2716 }
2717
2718 // Put on worklist all related field nodes.
2719 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2720 assert(field->is_oop(), "sanity");
2721 int offset = field->offset();
2722 add_uses_to_worklist(field);
2723 // Loop over all bases of this field and push on worklist Field nodes
2724 // with the same offset and base (since they may reference the same field).
2725 for (BaseIterator i(field); i.has_next(); i.next()) {
2726 PointsToNode* base = i.get();
2727 add_fields_to_worklist(field, base);
2728 // Check if the base was source object of arraycopy and go over arraycopy's
2729 // destination objects since values stored to a field of source object are
2730 // accessible by uses (loads) of fields of destination objects.
2731 if (base->arraycopy_src()) {
2732 for (UseIterator j(base); j.has_next(); j.next()) {
2733 PointsToNode* arycp = j.get();
2734 if (arycp->is_Arraycopy()) {
2735 for (UseIterator k(arycp); k.has_next(); k.next()) {
2736 PointsToNode* abase = k.get();
2737 if (abase->arraycopy_dst() && abase != base) {
2738 // Look for the same arraycopy reference.
2739 add_fields_to_worklist(field, abase);
2740 }
2741 }
2742 }
2743 }
2744 }
2745 }
2746 }
2747
2748 // Put on worklist all related field nodes.
2749 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2750 int offset = field->offset();
2751 if (base->is_LocalVar()) {
2752 for (UseIterator j(base); j.has_next(); j.next()) {
2753 PointsToNode* f = j.get();
2754 if (PointsToNode::is_base_use(f)) { // Field
2755 f = PointsToNode::get_use_node(f);
2756 if (f == field || !f->as_Field()->is_oop()) {
2757 continue;
2758 }
2759 int offs = f->as_Field()->offset();
2760 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2761 add_to_worklist(f);
2762 }
2763 }
2764 }
2765 } else {
2766 assert(base->is_JavaObject(), "sanity");
2767 if (// Skip phantom_object since it is only used to indicate that
2768 // this field's content globally escapes.
2769 (base != phantom_obj) &&
2770 // null object node does not have fields.
2771 (base != null_obj)) {
2772 for (EdgeIterator i(base); i.has_next(); i.next()) {
2773 PointsToNode* f = i.get();
2774 // Skip arraycopy edge since store to destination object field
2775 // does not update value in source object field.
2776 if (f->is_Arraycopy()) {
2777 assert(base->arraycopy_dst(), "sanity");
2778 continue;
2779 }
2780 if (f == field || !f->as_Field()->is_oop()) {
2781 continue;
2782 }
2783 int offs = f->as_Field()->offset();
2784 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2785 add_to_worklist(f);
2786 }
2787 }
2788 }
2789 }
2790 }
2791
2792 // Find fields which have unknown value.
2793 int ConnectionGraph::find_field_value(FieldNode* field) {
2794 // Escaped fields should have init value already.
2795 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2796 int new_edges = 0;
2797 for (BaseIterator i(field); i.has_next(); i.next()) {
2798 PointsToNode* base = i.get();
2799 if (base->is_JavaObject()) {
2800 // Skip Allocate's fields which will be processed later.
2801 if (base->ideal_node()->is_Allocate()) {
2802 return 0;
2803 }
2804 assert(base == null_obj, "only null ptr base expected here");
2805 }
2806 }
2807 if (add_edge(field, phantom_obj)) {
2808 // New edge was added
2809 new_edges++;
2810 add_field_uses_to_worklist(field);
2811 }
2812 return new_edges;
2813 }
2814
2815 // Find fields initializing values for allocations.
2816 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2817 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2818 Node* alloc = pta->ideal_node();
2819
2820 // Do nothing for Allocate nodes since its fields values are
2821 // "known" unless they are initialized by arraycopy/clone.
2822 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2823 return 0;
2824 }
2825 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
2826 #ifdef ASSERT
2827 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == nullptr) {2828 assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub(), "sanity");
2829 }
2830 #endif
2831 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2832 int new_edges = 0;
2833 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2834 PointsToNode* field = i.get();
2835 if (field->is_Field() && field->as_Field()->is_oop()) {
2836 if (add_edge(field, phantom_obj)) {
2837 // New edge was added
2838 new_edges++;
2839 add_field_uses_to_worklist(field->as_Field());
2840 }
2841 }
2842 }
2843 return new_edges;
2844 }
2845
2846 // Find fields initializing values for allocations.
2847 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2848 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2849 Node* alloc = pta->ideal_node();
2850 // Do nothing for Call nodes since its fields values are unknown.
2851 if (!alloc->is_Allocate()) {
2852 return 0;
2853 }
2854 InitializeNode* ini = alloc->as_Allocate()->initialization();
2855 bool visited_bottom_offset = false;
2856 GrowableArray<int> offsets_worklist;
2857 int new_edges = 0;
2858
2859 // Check if an oop field's initializing value is recorded and add
2860 // a corresponding null if field's value if it is not recorded.
2861 // Connection Graph does not record a default initialization by null
2862 // captured by Initialize node.
2863 //
2864 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2865 PointsToNode* field = i.get(); // Field (AddP)
2866 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2867 continue; // Not oop field
2868 }
2869 int offset = field->as_Field()->offset();
2870 if (offset == Type::OffsetBot) {
2871 if (!visited_bottom_offset) {
2872 // OffsetBot is used to reference array's element,
2873 // always add reference to null to all Field nodes since we don't
2874 // known which element is referenced.
2875 if (add_edge(field, null_obj)) {
2876 // New edge was added
2877 new_edges++;
2878 add_field_uses_to_worklist(field->as_Field());
2879 visited_bottom_offset = true;
2880 }
2881 }
2882 } else {
2883 // Check only oop fields.
2884 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2885 if (adr_type->isa_rawptr()) {
2886 #ifdef ASSERT
2887 // Raw pointers are used for initializing stores so skip it
2888 // since it should be recorded already
2889 Node* base = get_addp_base(field->ideal_node());
2890 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2891 #endif
2892 continue;
2893 }
2894 if (!offsets_worklist.contains(offset)) {
2895 offsets_worklist.append(offset);
2896 Node* value = nullptr;
2897 if (ini != nullptr) {
2898 // StoreP::value_basic_type() == T_ADDRESS
2899 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2900 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2901 // Make sure initializing store has the same type as this AddP.
2902 // This AddP may reference non existing field because it is on a
2903 // dead branch of bimorphic call which is not eliminated yet.
2904 if (store != nullptr && store->is_Store() &&
2905 store->as_Store()->value_basic_type() == ft) {
2906 value = store->in(MemNode::ValueIn);
2907 #ifdef ASSERT
2908 if (VerifyConnectionGraph) {
2909 // Verify that AddP already points to all objects the value points to.
2910 PointsToNode* val = ptnode_adr(value->_idx);
2911 assert((val != nullptr), "should be processed already");
2912 PointsToNode* missed_obj = nullptr;
2913 if (val->is_JavaObject()) {
2914 if (!field->points_to(val->as_JavaObject())) {
2915 missed_obj = val;
2916 }
2917 } else {
2918 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2919 tty->print_cr("----------init store has invalid value -----");
2920 store->dump();
2921 val->dump();
2922 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2923 }
2924 for (EdgeIterator j(val); j.has_next(); j.next()) {
2925 PointsToNode* obj = j.get();
2926 if (obj->is_JavaObject()) {
2927 if (!field->points_to(obj->as_JavaObject())) {
2928 missed_obj = obj;
2929 break;
2930 }
2931 }
2932 }
2933 }
2934 if (missed_obj != nullptr) {
2935 tty->print_cr("----------field---------------------------------");
2936 field->dump();
2937 tty->print_cr("----------missed referernce to object-----------");
2938 missed_obj->dump();
2939 tty->print_cr("----------object referernced by init store -----");
2940 store->dump();
2941 val->dump();
2942 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2943 }
2944 }
2945 #endif
2946 } else {
2947 // There could be initializing stores which follow allocation.
2948 // For example, a volatile field store is not collected
2949 // by Initialize node.
2950 //
2951 // Need to check for dependent loads to separate such stores from
2952 // stores which follow loads. For now, add initial value null so
2953 // that compare pointers optimization works correctly.
2954 }
2955 }
2956 if (value == nullptr) {
2957 // A field's initializing value was not recorded. Add null.
2958 if (add_edge(field, null_obj)) {
2959 // New edge was added
2960 new_edges++;
2961 add_field_uses_to_worklist(field->as_Field());
2962 }
2963 }
2964 }
2965 }
2966 }
2967 return new_edges;
2968 }
2969
2970 // Adjust scalar_replaceable state after Connection Graph is built.
2971 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2972 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2973 // returns true. If one of the constraints in this method set 'jobj' to NSR
2974 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2975 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2976 // that other object and the Phi will become a reducible Phi.
2977 // There could be multiple merges involving the same jobj.
2978 Unique_Node_List candidates;
2979
2980 // Search for non-escaping objects which are not scalar replaceable
2981 // and mark them to propagate the state to referenced objects.
2982
2983 for (UseIterator i(jobj); i.has_next(); i.next()) {
2984 PointsToNode* use = i.get();
2985 if (use->is_Arraycopy()) {
2986 continue;
2987 }
2988 if (use->is_Field()) {
2989 FieldNode* field = use->as_Field();
2990 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2991 // 1. An object is not scalar replaceable if the field into which it is
2992 // stored has unknown offset (stored into unknown element of an array).
2993 if (field->offset() == Type::OffsetBot) {
2994 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2995 return;
2996 }
2997 for (BaseIterator i(field); i.has_next(); i.next()) {
2998 PointsToNode* base = i.get();
2999 // 2. An object is not scalar replaceable if the field into which it is
3000 // stored has multiple bases one of which is null.
3001 if ((base == null_obj) && (field->base_count() > 1)) {
3002 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
3003 return;
3004 }
3005 // 2.5. An object is not scalar replaceable if the field into which it is
3006 // stored has NSR base.
3007 if (!base->scalar_replaceable()) {
3008 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3009 return;
3010 }
3011 }
3012 }
3013 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3014 // 3. An object is not scalar replaceable if it is merged with other objects
3015 // and we can't remove the merge
3016 for (EdgeIterator j(use); j.has_next(); j.next()) {
3017 PointsToNode* ptn = j.get();
3018 if (ptn->is_JavaObject() && ptn != jobj) {
3019 Node* use_n = use->ideal_node();
3020
3021 // These other local vars may point to multiple objects through a Phi
3022 // In this case we skip them and see if we can reduce the Phi.
3023 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3024 use_n = use_n->in(1);
3025 }
3026
3027 // If it's already a candidate or confirmed reducible merge we can skip verification
3028 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3029 continue;
3030 }
3031
3032 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3033 candidates.push(use_n);
3034 } else {
3035 // Mark all objects as NSR if we can't remove the merge
3036 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3037 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3038 }
3039 }
3040 }
3041 if (!jobj->scalar_replaceable()) {
3042 return;
3043 }
3044 }
3045
3046 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3047 if (j.get()->is_Arraycopy()) {
3048 continue;
3049 }
3050
3051 // Non-escaping object node should point only to field nodes.
3052 FieldNode* field = j.get()->as_Field();
3053 int offset = field->as_Field()->offset();
3054
3055 // 4. An object is not scalar replaceable if it has a field with unknown
3056 // offset (array's element is accessed in loop).
3057 if (offset == Type::OffsetBot) {
3058 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3059 return;
3060 }
3061 // 5. Currently an object is not scalar replaceable if a LoadStore node
3062 // access its field since the field value is unknown after it.
3063 //
3064 Node* n = field->ideal_node();
3065
3066 // Test for an unsafe access that was parsed as maybe off heap
3067 // (with a CheckCastPP to raw memory).
3068 assert(n->is_AddP(), "expect an address computation");
3069 if (n->in(AddPNode::Base)->is_top() &&
3070 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3071 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3072 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3073 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3074 return;
3075 }
3076
3077 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3078 Node* u = n->fast_out(i);
3079 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3080 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3081 return;
3082 }
3083 }
3084
3085 // 6. Or the address may point to more then one object. This may produce
3086 // the false positive result (set not scalar replaceable)
3087 // since the flow-insensitive escape analysis can't separate
3088 // the case when stores overwrite the field's value from the case
3089 // when stores happened on different control branches.
3090 //
3091 // Note: it will disable scalar replacement in some cases:
3092 //
3093 // Point p[] = new Point[1];
3094 // p[0] = new Point(); // Will be not scalar replaced
3095 //
3096 // but it will save us from incorrect optimizations in next cases:
3097 //
3098 // Point p[] = new Point[1];
3099 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3100 //
3101 if (field->base_count() > 1 && candidates.size() == 0) {
3102 if (has_non_reducible_merge(field, reducible_merges)) {
3103 for (BaseIterator i(field); i.has_next(); i.next()) {
3104 PointsToNode* base = i.get();
3105 // Don't take into account LocalVar nodes which
3106 // may point to only one object which should be also
3107 // this field's base by now.
3108 if (base->is_JavaObject() && base != jobj) {
3109 // Mark all bases.
3110 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3111 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3112 }
3113 }
3114
3115 if (!jobj->scalar_replaceable()) {
3116 return;
3117 }
3118 }
3119 }
3120 }
3121
3122 // The candidate is truly a reducible merge only if none of the other
3123 // constraints ruled it as NSR. There could be multiple merges involving the
3124 // same jobj.
3125 assert(jobj->scalar_replaceable(), "sanity");
3126 for (uint i = 0; i < candidates.size(); i++ ) {
3127 Node* candidate = candidates.at(i);
3128 reducible_merges.push(candidate);
3129 }
3130 }
3131
3132 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3133 for (BaseIterator i(field); i.has_next(); i.next()) {
3134 Node* base = i.get()->ideal_node();
3135 if (base->is_Phi() && !reducible_merges.member(base)) {
3136 return true;
3137 }
3138 }
3139 return false;
3140 }
3141
3142 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3143 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3144
3145 // Look for 'phis' that refer to 'jobj' as the last
3146 // remaining scalar replaceable input.
3147 uint reducible_merges_cnt = reducible_merges.size();
3148 for (uint i = 0; i < reducible_merges_cnt; i++) {
3149 Node* phi = reducible_merges.at(i);
3150
3151 // This 'Phi' will be a 'good' if it still points to
3152 // at least one scalar replaceable object. Note that 'obj'
3153 // was/should be marked as NSR before calling this function.
3154 bool good_phi = false;
3155
3156 for (uint j = 1; j < phi->req(); j++) {
3157 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3158 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3159 good_phi = true;
3160 break;
3161 }
3162 }
3163
3164 if (!good_phi) {
3165 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3166 reducible_merges.remove(i);
3167
3168 // Decrement the index because the 'remove' call above actually
3169 // moves the last entry of the list to position 'i'.
3170 i--;
3171
3172 reducible_merges_cnt--;
3173 }
3174 }
3175 }
3176
3177 // Propagate NSR (Not scalar replaceable) state.
3178 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3179 int jobj_length = jobj_worklist.length();
3180 bool found_nsr_alloc = true;
3181 while (found_nsr_alloc) {
3182 found_nsr_alloc = false;
3183 for (int next = 0; next < jobj_length; ++next) {
3184 JavaObjectNode* jobj = jobj_worklist.at(next);
3185 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3186 PointsToNode* use = i.get();
3187 if (use->is_Field()) {
3188 FieldNode* field = use->as_Field();
3189 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3190 assert(field->offset() != Type::OffsetBot, "sanity");
3191 for (BaseIterator i(field); i.has_next(); i.next()) {
3192 PointsToNode* base = i.get();
3193 // An object is not scalar replaceable if the field into which
3194 // it is stored has NSR base.
3195 if ((base != null_obj) && !base->scalar_replaceable()) {
3196 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3197 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3198 // because there is no point in reducing a Phi that won't improve the number of SR
3199 // objects.
3200 revisit_reducible_phi_status(jobj, reducible_merges);
3201 found_nsr_alloc = true;
3202 break;
3203 }
3204 }
3205 } else if (use->is_LocalVar()) {
3206 Node* phi = use->ideal_node();
3207 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3208 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3209 reducible_merges.yank(phi);
3210 found_nsr_alloc = true;
3211 break;
3212 }
3213 }
3214 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3215 }
3216 }
3217 }
3218 }
3219
3220 #ifdef ASSERT
3221 void ConnectionGraph::verify_connection_graph(
3222 GrowableArray<PointsToNode*>& ptnodes_worklist,
3223 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3224 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3225 GrowableArray<Node*>& addp_worklist) {
3226 // Verify that graph is complete - no new edges could be added.
3227 int java_objects_length = java_objects_worklist.length();
3228 int non_escaped_length = non_escaped_allocs_worklist.length();
3229 int new_edges = 0;
3230 for (int next = 0; next < java_objects_length; ++next) {
3231 JavaObjectNode* ptn = java_objects_worklist.at(next);
3232 new_edges += add_java_object_edges(ptn, true);
3233 }
3234 assert(new_edges == 0, "graph was not complete");
3235 // Verify that escape state is final.
3236 int length = non_escaped_allocs_worklist.length();
3237 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3238 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3239 (non_escaped_length == length) &&
3240 (_worklist.length() == 0), "escape state was not final");
3241
3242 // Verify fields information.
3243 int addp_length = addp_worklist.length();
3244 for (int next = 0; next < addp_length; ++next ) {
3245 Node* n = addp_worklist.at(next);
3246 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3247 if (field->is_oop()) {
3248 // Verify that field has all bases
3249 Node* base = get_addp_base(n);
3250 PointsToNode* ptn = ptnode_adr(base->_idx);
3251 if (ptn->is_JavaObject()) {
3252 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3253 } else {
3254 assert(ptn->is_LocalVar(), "sanity");
3255 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3256 PointsToNode* e = i.get();
3257 if (e->is_JavaObject()) {
3258 assert(field->has_base(e->as_JavaObject()), "sanity");
3259 }
3260 }
3261 }
3262 // Verify that all fields have initializing values.
3263 if (field->edge_count() == 0) {
3264 tty->print_cr("----------field does not have references----------");
3265 field->dump();
3266 for (BaseIterator i(field); i.has_next(); i.next()) {
3267 PointsToNode* base = i.get();
3268 tty->print_cr("----------field has next base---------------------");
3269 base->dump();
3270 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3271 tty->print_cr("----------base has fields-------------------------");
3272 for (EdgeIterator j(base); j.has_next(); j.next()) {
3273 j.get()->dump();
3274 }
3275 tty->print_cr("----------base has references---------------------");
3276 for (UseIterator j(base); j.has_next(); j.next()) {
3277 j.get()->dump();
3278 }
3279 }
3280 }
3281 for (UseIterator i(field); i.has_next(); i.next()) {
3282 i.get()->dump();
3283 }
3284 assert(field->edge_count() > 0, "sanity");
3285 }
3286 }
3287 }
3288 }
3289 #endif
3290
3291 // Optimize ideal graph.
3292 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3293 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3294 Compile* C = _compile;
3295 PhaseIterGVN* igvn = _igvn;
3296 if (EliminateLocks) {
3297 // Mark locks before changing ideal graph.
3298 int cnt = C->macro_count();
3299 for (int i = 0; i < cnt; i++) {
3300 Node *n = C->macro_node(i);
3301 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3302 AbstractLockNode* alock = n->as_AbstractLock();
3303 if (!alock->is_non_esc_obj()) {
3304 if (can_eliminate_lock(alock)) {
3305 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3306 // The lock could be marked eliminated by lock coarsening
3307 // code during first IGVN before EA. Replace coarsened flag
3308 // to eliminate all associated locks/unlocks.
3309 #ifdef ASSERT
3310 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3311 #endif
3312 alock->set_non_esc_obj();
3313 }
3314 }
3315 }
3316 }
3317 }
3318
3319 if (OptimizePtrCompare) {
3320 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3321 Node *n = ptr_cmp_worklist.at(i);
3322 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3323 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3324 if (tcmp->singleton()) {
3325 Node* cmp = igvn->makecon(tcmp);
3326 #ifndef PRODUCT
3327 if (PrintOptimizePtrCompare) {
3328 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"));
3329 if (Verbose) {
3330 n->dump(1);
3331 }
3332 }
3333 #endif
3334 igvn->replace_node(n, cmp);
3335 }
3336 }
3337 }
3338
3339 // For MemBarStoreStore nodes added in library_call.cpp, check
3340 // escape status of associated AllocateNode and optimize out
3341 // MemBarStoreStore node if the allocated object never escapes.
3342 for (int i = 0; i < storestore_worklist.length(); i++) {
3343 Node* storestore = storestore_worklist.at(i);
3344 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3345 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3346 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);3347 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));3348 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));3349 igvn->register_new_node_with_optimizer(mb);3350 igvn->replace_node(storestore, mb);
3351 }
3352 }
3353 }
3354
3355 // Optimize objects compare.
3356 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3357 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3358 if (!OptimizePtrCompare) {
3359 return UNKNOWN;
3360 }
3361 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3362 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3363
3364 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3365 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3366 JavaObjectNode* jobj1 = unique_java_object(left);
3367 JavaObjectNode* jobj2 = unique_java_object(right);
3368
3369 // The use of this method during allocation merge reduction may cause 'left'
3370 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3371 // that doesn't reference an unique java object.
3372 if (ptn1 == nullptr || ptn2 == nullptr ||
3373 jobj1 == nullptr || jobj2 == nullptr) {
3374 return UNKNOWN;
3375 }
3376
3377 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3378 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3379
3380 // Check simple cases first.
3381 if (jobj1 != nullptr) {
3382 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3383 if (jobj1 == jobj2) {
3384 // Comparing the same not escaping object.
3385 return EQ;
3386 }
3387 Node* obj = jobj1->ideal_node();
3388 // Comparing not escaping allocation.
3389 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3390 !ptn2->points_to(jobj1)) {
3391 return NE; // This includes nullness check.
3392 }
3393 }
3394 }
3395 if (jobj2 != nullptr) {
3396 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3397 Node* obj = jobj2->ideal_node();
3398 // Comparing not escaping allocation.
3399 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3400 !ptn1->points_to(jobj2)) {
3401 return NE; // This includes nullness check.
3402 }
3403 }
3404 }
3405 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3406 jobj2 != nullptr && jobj2 != phantom_obj &&
3407 jobj1->ideal_node()->is_Con() &&
3408 jobj2->ideal_node()->is_Con()) {
3409 // Klass or String constants compare. Need to be careful with
3410 // compressed pointers - compare types of ConN and ConP instead of nodes.
3411 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3412 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3413 if (t1->make_ptr() == t2->make_ptr()) {
3414 return EQ;
3415 } else {
3416 return NE;
3417 }
3418 }
3419 if (ptn1->meet(ptn2)) {
3420 return UNKNOWN; // Sets are not disjoint
3421 }
3422
3423 // Sets are disjoint.
3424 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3425 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3426 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3427 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3428 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3429 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3430 // Check nullness of unknown object.
3431 return UNKNOWN;
3432 }
3433
3434 // Disjointness by itself is not sufficient since
3435 // alias analysis is not complete for escaped objects.
3436 // Disjoint sets are definitely unrelated only when
3437 // at least one set has only not escaping allocations.
3438 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3439 if (ptn1->non_escaping_allocation()) {
3440 return NE;
3441 }
3442 }
3443 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3444 if (ptn2->non_escaping_allocation()) {
3445 return NE;
3446 }
3447 }
3448 return UNKNOWN;
3449 }
3450
3451 // Connection Graph construction functions.
3452
3453 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3454 PointsToNode* ptadr = _nodes.at(n->_idx);
3455 if (ptadr != nullptr) {
3456 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3457 return;
3458 }
3459 Compile* C = _compile;
3460 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3461 map_ideal_node(n, ptadr);
3462 }
3463
3464 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3465 PointsToNode* ptadr = _nodes.at(n->_idx);
3466 if (ptadr != nullptr) {
3467 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3468 return ptadr;
3469 }
3470 Compile* C = _compile;
3471 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3472 map_ideal_node(n, ptadr);
3473 return ptadr;
3474 }
3475
3476 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3477 PointsToNode* ptadr = _nodes.at(n->_idx);
3478 if (ptadr != nullptr) {
3479 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3480 return;
3481 }
3482 bool unsafe = false;
3483 bool is_oop = is_oop_field(n, offset, &unsafe);
3484 if (unsafe) {
3485 es = PointsToNode::GlobalEscape;
3486 }
3487 Compile* C = _compile;
3488 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3489 map_ideal_node(n, field);
3490 }
3491
3492 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3493 PointsToNode* src, PointsToNode* dst) {
3494 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3495 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3496 PointsToNode* ptadr = _nodes.at(n->_idx);
3497 if (ptadr != nullptr) {
3498 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3499 return;
3500 }
3501 Compile* C = _compile;
3502 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3503 map_ideal_node(n, ptadr);
3504 // Add edge from arraycopy node to source object.
3505 (void)add_edge(ptadr, src);
3506 src->set_arraycopy_src();
3507 // Add edge from destination object to arraycopy node.
3508 (void)add_edge(dst, ptadr);
3509 dst->set_arraycopy_dst();
3510 }
3511
3512 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3513 const Type* adr_type = n->as_AddP()->bottom_type();
3514 BasicType bt = T_INT;
3515 if (offset == Type::OffsetBot) {
3516 // Check only oop fields.
3517 if (!adr_type->isa_aryptr() ||
3518 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3519 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3520 // OffsetBot is used to reference array's element. Ignore first AddP.
3521 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3522 bt = T_OBJECT;
3523 }
3524 }
3525 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3526 if (adr_type->isa_instptr()) {
3527 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
3528 if (field != nullptr) {
3529 bt = field->layout_type();
3530 } else {
3531 // Check for unsafe oop field access
3532 if (has_oop_node_outs(n)) {
3533 bt = T_OBJECT;
3534 (*unsafe) = true;
3535 }
3536 }
3537 } else if (adr_type->isa_aryptr()) {
3538 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3539 // Ignore array length load.
3540 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3541 // Ignore first AddP.
3542 } else {
3543 const Type* elemtype = adr_type->isa_aryptr()->elem();3544 bt = elemtype->array_element_basic_type();
3545 }
3546 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3547 // Allocation initialization, ThreadLocal field access, unsafe access
3548 if (has_oop_node_outs(n)) {
3549 bt = T_OBJECT;
3550 }
3551 }
3552 }
3553 // Note: T_NARROWOOP is not classed as a real reference type
3554 bool res = (is_reference_type(bt) || bt == T_NARROWOOP);
3555 assert(!has_oop_node_outs(n) || res, "sanity: AddP has oop outs, needs to be treated as oop field");
3556 return res;
3557 }
3558
3559 bool ConnectionGraph::has_oop_node_outs(Node* n) {
3560 return n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3561 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3562 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3563 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n);
3564 }
3565
3566 // Returns unique pointed java object or null.
3567 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3568 // If the node was created after the escape computation we can't answer.
3569 uint idx = n->_idx;
3570 if (idx >= nodes_size()) {
3571 return nullptr;
3572 }
3573 PointsToNode* ptn = ptnode_adr(idx);
3574 if (ptn == nullptr) {
3575 return nullptr;
3576 }
3577 if (ptn->is_JavaObject()) {
3578 return ptn->as_JavaObject();
3579 }
3580 assert(ptn->is_LocalVar(), "sanity");
3581 // Check all java objects it points to.
3582 JavaObjectNode* jobj = nullptr;
3583 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3584 PointsToNode* e = i.get();
3585 if (e->is_JavaObject()) {
3586 if (jobj == nullptr) {
3587 jobj = e->as_JavaObject();
3588 } else if (jobj != e) {
3589 return nullptr;
3590 }
3591 }
3592 }
3593 return jobj;
3594 }
3595
3596 // Return true if this node points only to non-escaping allocations.
3597 bool PointsToNode::non_escaping_allocation() {
3598 if (is_JavaObject()) {
3599 Node* n = ideal_node();
3600 if (n->is_Allocate() || n->is_CallStaticJava()) {
3601 return (escape_state() == PointsToNode::NoEscape);
3602 } else {
3603 return false;
3604 }
3605 }
3606 assert(is_LocalVar(), "sanity");
3607 // Check all java objects it points to.
3608 for (EdgeIterator i(this); i.has_next(); i.next()) {
3609 PointsToNode* e = i.get();
3610 if (e->is_JavaObject()) {
3611 Node* n = e->ideal_node();
3612 if ((e->escape_state() != PointsToNode::NoEscape) ||
3613 !(n->is_Allocate() || n->is_CallStaticJava())) {
3614 return false;
3615 }
3616 }
3617 }
3618 return true;
3619 }
3620
3621 // Return true if we know the node does not escape globally.
3622 bool ConnectionGraph::not_global_escape(Node *n) {
3623 assert(!_collecting, "should not call during graph construction");
3624 // If the node was created after the escape computation we can't answer.
3625 uint idx = n->_idx;
3626 if (idx >= nodes_size()) {
3627 return false;
3628 }
3629 PointsToNode* ptn = ptnode_adr(idx);
3630 if (ptn == nullptr) {
3631 return false; // not in congraph (e.g. ConI)
3632 }
3633 PointsToNode::EscapeState es = ptn->escape_state();
3634 // If we have already computed a value, return it.
3635 if (es >= PointsToNode::GlobalEscape) {
3636 return false;
3637 }
3638 if (ptn->is_JavaObject()) {
3639 return true; // (es < PointsToNode::GlobalEscape);
3640 }
3641 assert(ptn->is_LocalVar(), "sanity");
3642 // Check all java objects it points to.
3643 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3644 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3645 return false;
3646 }
3647 }
3648 return true;
3649 }
3650
3651 // Return true if locked object does not escape globally
3652 // and locked code region (identified by BoxLockNode) is balanced:
3653 // all compiled code paths have corresponding Lock/Unlock pairs.
3654 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3655 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3656 if (EliminateNestedLocks) {
3657 // We can mark whole locking region as Local only when only
3658 // one object is used for locking.
3659 alock->box_node()->as_BoxLock()->set_local();
3660 }
3661 return true;
3662 }
3663 return false;
3664 }
3665
3666 // Helper functions
3667
3668 // Return true if this node points to specified node or nodes it points to.
3669 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3670 if (is_JavaObject()) {
3671 return (this == ptn);
3672 }
3673 assert(is_LocalVar() || is_Field(), "sanity");
3674 for (EdgeIterator i(this); i.has_next(); i.next()) {
3675 if (i.get() == ptn) {
3676 return true;
3677 }
3678 }
3679 return false;
3680 }
3681
3682 // Return true if one node points to an other.
3683 bool PointsToNode::meet(PointsToNode* ptn) {
3684 if (this == ptn) {
3685 return true;
3686 } else if (ptn->is_JavaObject()) {
3687 return this->points_to(ptn->as_JavaObject());
3688 } else if (this->is_JavaObject()) {
3689 return ptn->points_to(this->as_JavaObject());
3690 }
3691 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3692 int ptn_count = ptn->edge_count();
3693 for (EdgeIterator i(this); i.has_next(); i.next()) {
3694 PointsToNode* this_e = i.get();
3695 for (int j = 0; j < ptn_count; j++) {
3696 if (this_e == ptn->edge(j)) {
3697 return true;
3698 }
3699 }
3700 }
3701 return false;
3702 }
3703
3704 #ifdef ASSERT
3705 // Return true if bases point to this java object.
3706 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3707 for (BaseIterator i(this); i.has_next(); i.next()) {
3708 if (i.get() == jobj) {
3709 return true;
3710 }
3711 }
3712 return false;
3713 }
3714 #endif
3715
3716 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3717 // Handle simple case first.
3718 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3719 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3720 return true;
3721 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3722 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3723 Node* addp_use = addp->fast_out(i);
3724 if (addp_use->is_Store()) {
3725 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3726 if (addp_use->fast_out(j)->is_Initialize()) {
3727 return true;
3728 }
3729 }
3730 }
3731 }
3732 }
3733 return false;
3734 }
3735
3736 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3737 const Type *adr_type = phase->type(adr);
3738 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3739 // We are computing a raw address for a store captured by an Initialize
3740 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3741 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3742 assert(offs != Type::OffsetBot ||
3743 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3744 "offset must be a constant or it is initialization of array");
3745 return offs;
3746 }
3747 const TypePtr *t_ptr = adr_type->isa_ptr();3748 assert(t_ptr != nullptr, "must be a pointer type");3749 return t_ptr->offset();
3750 }
3751
3752 Node* ConnectionGraph::get_addp_base(Node *addp) {
3753 assert(addp->is_AddP(), "must be AddP");
3754 //
3755 // AddP cases for Base and Address inputs:
3756 // case #1. Direct object's field reference:
3757 // Allocate
3758 // |
3759 // Proj #5 ( oop result )
3760 // |
3761 // CheckCastPP (cast to instance type)
3762 // | |
3763 // AddP ( base == address )
3764 //
3765 // case #2. Indirect object's field reference:
3766 // Phi
3767 // |
3768 // CastPP (cast to instance type)
3769 // | |
3770 // AddP ( base == address )
3771 //
3772 // case #3. Raw object's field reference for Initialize node:
3773 // Allocate
3774 // |
3775 // Proj #5 ( oop result )
3776 // top |
3777 // \ |
3778 // AddP ( base == top )
3779 //
3780 // case #4. Array's element reference:
3781 // {CheckCastPP | CastPP}
3782 // | | |
3783 // | AddP ( array's element offset )
3784 // | |
3785 // AddP ( array's offset )
3786 //
3787 // case #5. Raw object's field reference for arraycopy stub call:
3788 // The inline_native_clone() case when the arraycopy stub is called
3789 // after the allocation before Initialize and CheckCastPP nodes.
3790 // Allocate
3791 // |
3792 // Proj #5 ( oop result )
3793 // | |
3794 // AddP ( base == address )
3795 //
3796 // case #6. Constant Pool, ThreadLocal, CastX2P, Klass, OSR buffer buf or
3797 // Raw object's field reference:
3798 // {ConP, ThreadLocal, CastX2P, raw Load, Parm0}
3799 // top |
3800 // \ |
3801 // AddP ( base == top )
3802 //
3803 // case #7. Klass's field reference.
3804 // LoadKlass
3805 // | |
3806 // AddP ( base == address )
3807 //
3808 // case #8. narrow Klass's field reference.
3809 // LoadNKlass
3810 // |
3811 // DecodeN
3812 // | |
3813 // AddP ( base == address )
3814 //
3815 // case #9. Mixed unsafe access
3816 // {instance}
3817 // |
3818 // CheckCastPP (raw)
3819 // top |
3820 // \ |
3821 // AddP ( base == top )
3822 //
3823 Node *base = addp->in(AddPNode::Base);
3824 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3825 base = addp->in(AddPNode::Address);
3826 while (base->is_AddP()) {
3827 // Case #6 (unsafe access) may have several chained AddP nodes.
3828 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3829 base = base->in(AddPNode::Address);
3830 }
3831 if (base->Opcode() == Op_CheckCastPP &&
3832 base->bottom_type()->isa_rawptr() &&
3833 _igvn->type(base->in(1))->isa_oopptr()) {
3834 base = base->in(1); // Case #9
3835 } else {
3836 Node* uncast_base = base->uncast();
3837 int opcode = uncast_base->Opcode();
3838 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3839 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3840 (_igvn->C->is_osr_compilation() && uncast_base->is_Parm() && uncast_base->as_Parm()->_con == TypeFunc::Parms)||
3841 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3842 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_klassptr() != nullptr)) ||
3843 is_captured_store_address(addp), "sanity");
3844 }
3845 }
3846 return base;
3847 }
3848
3849 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3850 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3851 Node* addp2 = addp->raw_out(0);
3852 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3853 addp2->in(AddPNode::Base) == n &&
3854 addp2->in(AddPNode::Address) == addp) {
3855 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3856 //
3857 // Find array's offset to push it on worklist first and
3858 // as result process an array's element offset first (pushed second)
3859 // to avoid CastPP for the array's offset.
3860 // Otherwise the inserted CastPP (LocalVar) will point to what
3861 // the AddP (Field) points to. Which would be wrong since
3862 // the algorithm expects the CastPP has the same point as
3863 // as AddP's base CheckCastPP (LocalVar).
3864 //
3865 // ArrayAllocation
3866 // |
3867 // CheckCastPP
3868 // |
3869 // memProj (from ArrayAllocation CheckCastPP)
3870 // | ||
3871 // | || Int (element index)
3872 // | || | ConI (log(element size))
3873 // | || | /
3874 // | || LShift
3875 // | || /
3876 // | AddP (array's element offset)
3877 // | |
3878 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3879 // | / /
3880 // AddP (array's offset)
3881 // |
3882 // Load/Store (memory operation on array's element)
3883 //
3884 return addp2;
3885 }
3886 return nullptr;
3887 }
3888
3889 //
3890 // Adjust the type and inputs of an AddP which computes the
3891 // address of a field of an instance
3892 //
3893 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3894 PhaseGVN* igvn = _igvn;
3895 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3896 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3897 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3898 if (t == nullptr) {
3899 // We are computing a raw address for a store captured by an Initialize
3900 // compute an appropriate address type (cases #3 and #5).
3901 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3902 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3903 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3904 assert(offs != Type::OffsetBot, "offset must be a constant");
3905 t = base_t->add_offset(offs)->is_oopptr();
3906 }
3907 int inst_id = base_t->instance_id();
3908 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3909 "old type must be non-instance or match new type");
3910
3911 // The type 't' could be subclass of 'base_t'.
3912 // As result t->offset() could be large then base_t's size and it will
3913 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3914 // constructor verifies correctness of the offset.
3915 //
3916 // It could happened on subclass's branch (from the type profiling
3917 // inlining) which was not eliminated during parsing since the exactness
3918 // of the allocation type was not propagated to the subclass type check.
3919 //
3920 // Or the type 't' could be not related to 'base_t' at all.
3921 // It could happened when CHA type is different from MDO type on a dead path
3922 // (for example, from instanceof check) which is not collapsed during parsing.
3923 //
3924 // Do nothing for such AddP node and don't process its users since
3925 // this code branch will go away.
3926 //
3927 if (!t->is_known_instance() &&
3928 !base_t->maybe_java_subtype_of(t)) {
3929 return false; // bail out
3930 }
3931 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
3932 // Do NOT remove the next line: ensure a new alias index is allocated
3933 // for the instance type. Note: C++ will not remove it since the call
3934 // has side effect.
3935 int alias_idx = _compile->get_alias_index(tinst);
3936 igvn->set_type(addp, tinst);
3937 // record the allocation in the node map
3938 set_map(addp, get_map(base->_idx));
3939 // Set addp's Base and Address to 'base'.
3940 Node *abase = addp->in(AddPNode::Base);
3941 Node *adr = addp->in(AddPNode::Address);
3942 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3943 adr->in(0)->_idx == (uint)inst_id) {
3944 // Skip AddP cases #3 and #5.
3945 } else {
3946 assert(!abase->is_top(), "sanity"); // AddP case #3
3947 if (abase != base) {
3948 igvn->hash_delete(addp);
3949 addp->set_req(AddPNode::Base, base);
3950 if (abase == adr) {
3951 addp->set_req(AddPNode::Address, base);
3952 } else {
3953 // AddP case #4 (adr is array's element offset AddP node)
3954 #ifdef ASSERT
3955 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3956 assert(adr->is_AddP() && atype != nullptr &&
3957 atype->instance_id() == inst_id, "array's element offset should be processed first");
3958 #endif
3959 }
3960 igvn->hash_insert(addp);
3961 }
3962 }
3963 // Put on IGVN worklist since at least addp's type was changed above.
3964 record_for_optimizer(addp);
3965 return true;
3966 }
3967
3968 //
3969 // Create a new version of orig_phi if necessary. Returns either the newly
3970 // created phi or an existing phi. Sets create_new to indicate whether a new
3971 // phi was created. Cache the last newly created phi in the node map.
3972 //
3973 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3974 Compile *C = _compile;
3975 PhaseGVN* igvn = _igvn;
3976 new_created = false;
3977 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3978 // nothing to do if orig_phi is bottom memory or matches alias_idx
3979 if (phi_alias_idx == alias_idx) {
3980 return orig_phi;
3981 }
3982 // Have we recently created a Phi for this alias index?
3983 PhiNode *result = get_map_phi(orig_phi->_idx);
3984 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3985 return result;
3986 }
3987 // Previous check may fail when the same wide memory Phi was split into Phis
3988 // for different memory slices. Search all Phis for this region.
3989 if (result != nullptr) {
3990 Node* region = orig_phi->in(0);
3991 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3992 Node* phi = region->fast_out(i);
3993 if (phi->is_Phi() &&
3994 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3995 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3996 return phi->as_Phi();
3997 }
3998 }
3999 }
4000 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4001 if (C->do_escape_analysis() == true && !C->failing()) {
4002 // Retry compilation without escape analysis.
4003 // If this is the first failure, the sentinel string will "stick"
4004 // to the Compile object, and the C2Compiler will see it and retry.
4005 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4006 }
4007 return nullptr;
4008 }
4009 orig_phi_worklist.append_if_missing(orig_phi);
4010 const TypePtr *atype = C->get_adr_type(alias_idx);
4011 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4012 C->copy_node_notes_to(result, orig_phi);
4013 igvn->set_type(result, result->bottom_type());
4014 record_for_optimizer(result);
4015 set_map(orig_phi, result);
4016 new_created = true;
4017 return result;
4018 }
4019
4020 //
4021 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4022 // specified alias index.
4023 //
4024 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4025 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4026 Compile *C = _compile;
4027 PhaseGVN* igvn = _igvn;
4028 bool new_phi_created;
4029 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4030 if (!new_phi_created) {
4031 return result;
4032 }
4033 GrowableArray<PhiNode *> phi_list;
4034 GrowableArray<uint> cur_input;
4035 PhiNode *phi = orig_phi;
4036 uint idx = 1;
4037 bool finished = false;
4038 while(!finished) {
4039 while (idx < phi->req()) {
4040 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4041 if (mem != nullptr && mem->is_Phi()) {
4042 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4043 if (new_phi_created) {
4044 // found an phi for which we created a new split, push current one on worklist and begin
4045 // processing new one
4046 phi_list.push(phi);
4047 cur_input.push(idx);
4048 phi = mem->as_Phi();
4049 result = newphi;
4050 idx = 1;
4051 continue;
4052 } else {
4053 mem = newphi;
4054 }
4055 }
4056 if (C->failing()) {
4057 return nullptr;
4058 }
4059 result->set_req(idx++, mem);
4060 }
4061 #ifdef ASSERT
4062 // verify that the new Phi has an input for each input of the original
4063 assert( phi->req() == result->req(), "must have same number of inputs.");
4064 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4065 #endif
4066 // Check if all new phi's inputs have specified alias index.
4067 // Otherwise use old phi.
4068 for (uint i = 1; i < phi->req(); i++) {
4069 Node* in = result->in(i);
4070 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4071 }
4072 // we have finished processing a Phi, see if there are any more to do
4073 finished = (phi_list.length() == 0 );
4074 if (!finished) {
4075 phi = phi_list.pop();
4076 idx = cur_input.pop();
4077 PhiNode *prev_result = get_map_phi(phi->_idx);
4078 prev_result->set_req(idx++, result);
4079 result = prev_result;
4080 }
4081 }
4082 return result;
4083 }
4084
4085 //
4086 // The next methods are derived from methods in MemNode.
4087 //
4088 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4089 Node *mem = mmem;
4090 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4091 // means an array I have not precisely typed yet. Do not do any
4092 // alias stuff with it any time soon.
4093 if (toop->base() != Type::AnyPtr &&
4094 !(toop->isa_instptr() &&
4095 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4096 toop->offset() == Type::OffsetBot)) {
4097 mem = mmem->memory_at(alias_idx);
4098 // Update input if it is progress over what we have now
4099 }
4100 return mem;
4101 }
4102
4103 //
4104 // Move memory users to their memory slices.
4105 //
4106 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4107 Compile* C = _compile;
4108 PhaseGVN* igvn = _igvn;
4109 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4110 assert(tp != nullptr, "ptr type");
4111 int alias_idx = C->get_alias_index(tp);
4112 int general_idx = C->get_general_index(alias_idx);
4113
4114 // Move users first
4115 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4116 Node* use = n->fast_out(i);
4117 if (use->is_MergeMem()) {
4118 MergeMemNode* mmem = use->as_MergeMem();
4119 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4120 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4121 continue; // Nothing to do
4122 }
4123 // Replace previous general reference to mem node.
4124 uint orig_uniq = C->unique();
4125 Node* m = find_inst_mem(n, general_idx, orig_phis);
4126 assert(orig_uniq == C->unique(), "no new nodes");
4127 mmem->set_memory_at(general_idx, m);
4128 --imax;
4129 --i;
4130 } else if (use->is_MemBar()) {
4131 assert(!use->is_Initialize(), "initializing stores should not be moved");
4132 if (use->req() > MemBarNode::Precedent &&
4133 use->in(MemBarNode::Precedent) == n) {
4134 // Don't move related membars.
4135 record_for_optimizer(use);
4136 continue;
4137 }
4138 tp = use->as_MemBar()->adr_type()->isa_ptr();
4139 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4140 alias_idx == general_idx) {
4141 continue; // Nothing to do
4142 }
4143 // Move to general memory slice.
4144 uint orig_uniq = C->unique();
4145 Node* m = find_inst_mem(n, general_idx, orig_phis);
4146 assert(orig_uniq == C->unique(), "no new nodes");
4147 igvn->hash_delete(use);
4148 imax -= use->replace_edge(n, m, igvn);
4149 igvn->hash_insert(use);
4150 record_for_optimizer(use);
4151 --i;
4152 #ifdef ASSERT
4153 } else if (use->is_Mem()) {
4154 // Memory nodes should have new memory input.
4155 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4156 assert(tp != nullptr, "ptr type");
4157 int idx = C->get_alias_index(tp);
4158 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4159 "Following memory nodes should have new memory input or be on the same memory slice");
4160 } else if (use->is_Phi()) {
4161 // Phi nodes should be split and moved already.
4162 tp = use->as_Phi()->adr_type()->isa_ptr();
4163 assert(tp != nullptr, "ptr type");
4164 int idx = C->get_alias_index(tp);
4165 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4166 } else {
4167 use->dump();
4168 assert(false, "should not be here");
4169 #endif
4170 }
4171 }
4172 }
4173
4174 //
4175 // Search memory chain of "mem" to find a MemNode whose address
4176 // is the specified alias index.
4177 //
4178 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4179 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4180 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4181 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4182 return nullptr;
4183 }
4184 if (orig_mem == nullptr) {
4185 return orig_mem;
4186 }
4187 Compile* C = _compile;
4188 PhaseGVN* igvn = _igvn;
4189 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4190 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4191 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4192 Node *prev = nullptr;
4193 Node *result = orig_mem;
4194 while (prev != result) {
4195 prev = result;
4196 if (result == start_mem) {
4197 break; // hit one of our sentinels
4198 }
4199 if (result->is_Mem()) {
4200 const Type *at = igvn->type(result->in(MemNode::Address));
4201 if (at == Type::TOP) {
4202 break; // Dead
4203 }
4204 assert (at->isa_ptr() != nullptr, "pointer type required.");
4205 int idx = C->get_alias_index(at->is_ptr());
4206 if (idx == alias_idx) {
4207 break; // Found
4208 }
4209 if (!is_instance && (at->isa_oopptr() == nullptr ||
4210 !at->is_oopptr()->is_known_instance())) {
4211 break; // Do not skip store to general memory slice.
4212 }
4213 result = result->in(MemNode::Memory);
4214 }
4215 if (!is_instance) {
4216 continue; // don't search further for non-instance types
4217 }
4218 // skip over a call which does not affect this memory slice
4219 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4220 Node *proj_in = result->in(0);
4221 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4222 break; // hit one of our sentinels
4223 } else if (proj_in->is_Call()) {
4224 // ArrayCopy node processed here as well
4225 CallNode *call = proj_in->as_Call();
4226 if (!call->may_modify(toop, igvn)) {
4227 result = call->in(TypeFunc::Memory);
4228 }
4229 } else if (proj_in->is_Initialize()) {
4230 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4231 // Stop if this is the initialization for the object instance which
4232 // which contains this memory slice, otherwise skip over it.
4233 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4234 result = proj_in->in(TypeFunc::Memory);
4235 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4236 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4237 result = get_map(result->_idx);
4238 assert(result != nullptr, "new projection should have been allocated");
4239 break;
4240 }
4241 } else if (proj_in->is_MemBar()) {
4242 // Check if there is an array copy for a clone
4243 // Step over GC barrier when ReduceInitialCardMarks is disabled
4244 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4245 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4246
4247 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4248 // Stop if it is a clone
4249 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4250 if (ac->may_modify(toop, igvn)) {
4251 break;
4252 }
4253 }
4254 result = proj_in->in(TypeFunc::Memory);
4255 }
4256 } else if (result->is_MergeMem()) {
4257 MergeMemNode *mmem = result->as_MergeMem();
4258 result = step_through_mergemem(mmem, alias_idx, toop);
4259 if (result == mmem->base_memory()) {
4260 // Didn't find instance memory, search through general slice recursively.
4261 result = mmem->memory_at(C->get_general_index(alias_idx));
4262 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4263 if (C->failing()) {
4264 return nullptr;
4265 }
4266 mmem->set_memory_at(alias_idx, result);
4267 }
4268 } else if (result->is_Phi() &&
4269 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4270 Node *un = result->as_Phi()->unique_input(igvn);
4271 if (un != nullptr) {
4272 orig_phis.append_if_missing(result->as_Phi());
4273 result = un;
4274 } else {
4275 break;
4276 }
4277 } else if (result->is_ClearArray()) {
4278 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4279 // Can not bypass initialization of the instance
4280 // we are looking for.
4281 break;
4282 }
4283 // Otherwise skip it (the call updated 'result' value).
4284 } else if (result->Opcode() == Op_SCMemProj) {
4285 Node* mem = result->in(0);
4286 Node* adr = nullptr;
4287 if (mem->is_LoadStore()) {
4288 adr = mem->in(MemNode::Address);
4289 } else {
4290 assert(mem->Opcode() == Op_EncodeISOArray ||
4291 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4292 adr = mem->in(3); // Memory edge corresponds to destination array
4293 }
4294 const Type *at = igvn->type(adr);
4295 if (at != Type::TOP) {
4296 assert(at->isa_ptr() != nullptr, "pointer type required.");
4297 int idx = C->get_alias_index(at->is_ptr());
4298 if (idx == alias_idx) {
4299 // Assert in debug mode
4300 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4301 break; // In product mode return SCMemProj node
4302 }
4303 }
4304 result = mem->in(MemNode::Memory);
4305 } else if (result->Opcode() == Op_StrInflatedCopy) {
4306 Node* adr = result->in(3); // Memory edge corresponds to destination array
4307 const Type *at = igvn->type(adr);
4308 if (at != Type::TOP) {
4309 assert(at->isa_ptr() != nullptr, "pointer type required.");
4310 int idx = C->get_alias_index(at->is_ptr());
4311 if (idx == alias_idx) {
4312 // Assert in debug mode
4313 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4314 break; // In product mode return SCMemProj node
4315 }
4316 }
4317 result = result->in(MemNode::Memory);
4318 }
4319 }
4320 if (result->is_Phi()) {
4321 PhiNode *mphi = result->as_Phi();
4322 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4323 const TypePtr *t = mphi->adr_type();
4324 if (!is_instance) {
4325 // Push all non-instance Phis on the orig_phis worklist to update inputs
4326 // during Phase 4 if needed.
4327 orig_phis.append_if_missing(mphi);
4328 } else if (C->get_alias_index(t) != alias_idx) {
4329 // Create a new Phi with the specified alias index type.
4330 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4331 }
4332 }
4333 // the result is either MemNode, PhiNode, InitializeNode.
4334 return result;
4335 }
4336
4337 //
4338 // Convert the types of non-escaped object to instance types where possible,
4339 // propagate the new type information through the graph, and update memory
4340 // edges and MergeMem inputs to reflect the new type.
4341 //
4342 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4343 // The processing is done in 4 phases:
4344 //
4345 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4346 // types for the CheckCastPP for allocations where possible.
4347 // Propagate the new types through users as follows:
4348 // casts and Phi: push users on alloc_worklist
4349 // AddP: cast Base and Address inputs to the instance type
4350 // push any AddP users on alloc_worklist and push any memnode
4351 // users onto memnode_worklist.
4352 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4353 // search the Memory chain for a store with the appropriate type
4354 // address type. If a Phi is found, create a new version with
4355 // the appropriate memory slices from each of the Phi inputs.
4356 // For stores, process the users as follows:
4357 // MemNode: push on memnode_worklist
4358 // MergeMem: push on mergemem_worklist
4359 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4360 // moving the first node encountered of each instance type to the
4361 // the input corresponding to its alias index.
4362 // appropriate memory slice.
4363 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4364 //
4365 // In the following example, the CheckCastPP nodes are the cast of allocation
4366 // results and the allocation of node 29 is non-escaped and eligible to be an
4367 // instance type.
4368 //
4369 // We start with:
4370 //
4371 // 7 Parm #memory
4372 // 10 ConI "12"
4373 // 19 CheckCastPP "Foo"
4374 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4375 // 29 CheckCastPP "Foo"
4376 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4377 //
4378 // 40 StoreP 25 7 20 ... alias_index=4
4379 // 50 StoreP 35 40 30 ... alias_index=4
4380 // 60 StoreP 45 50 20 ... alias_index=4
4381 // 70 LoadP _ 60 30 ... alias_index=4
4382 // 80 Phi 75 50 60 Memory alias_index=4
4383 // 90 LoadP _ 80 30 ... alias_index=4
4384 // 100 LoadP _ 80 20 ... alias_index=4
4385 //
4386 //
4387 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4388 // and creating a new alias index for node 30. This gives:
4389 //
4390 // 7 Parm #memory
4391 // 10 ConI "12"
4392 // 19 CheckCastPP "Foo"
4393 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4394 // 29 CheckCastPP "Foo" iid=24
4395 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4396 //
4397 // 40 StoreP 25 7 20 ... alias_index=4
4398 // 50 StoreP 35 40 30 ... alias_index=6
4399 // 60 StoreP 45 50 20 ... alias_index=4
4400 // 70 LoadP _ 60 30 ... alias_index=6
4401 // 80 Phi 75 50 60 Memory alias_index=4
4402 // 90 LoadP _ 80 30 ... alias_index=6
4403 // 100 LoadP _ 80 20 ... alias_index=4
4404 //
4405 // In phase 2, new memory inputs are computed for the loads and stores,
4406 // And a new version of the phi is created. In phase 4, the inputs to
4407 // node 80 are updated and then the memory nodes are updated with the
4408 // values computed in phase 2. This results in:
4409 //
4410 // 7 Parm #memory
4411 // 10 ConI "12"
4412 // 19 CheckCastPP "Foo"
4413 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4414 // 29 CheckCastPP "Foo" iid=24
4415 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4416 //
4417 // 40 StoreP 25 7 20 ... alias_index=4
4418 // 50 StoreP 35 7 30 ... alias_index=6
4419 // 60 StoreP 45 40 20 ... alias_index=4
4420 // 70 LoadP _ 50 30 ... alias_index=6
4421 // 80 Phi 75 40 60 Memory alias_index=4
4422 // 120 Phi 75 50 50 Memory alias_index=6
4423 // 90 LoadP _ 120 30 ... alias_index=6
4424 // 100 LoadP _ 80 20 ... alias_index=4
4425 //
4426 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4427 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4428 GrowableArray<MergeMemNode*> &mergemem_worklist,
4429 Unique_Node_List &reducible_merges) {
4430 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4431 GrowableArray<Node *> memnode_worklist;
4432 GrowableArray<PhiNode *> orig_phis;
4433 PhaseIterGVN *igvn = _igvn;
4434 uint new_index_start = (uint) _compile->num_alias_types();
4435 VectorSet visited;
4436 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4437
4438 // Phase 1: Process possible allocations from alloc_worklist.
4439 // Create instance types for the CheckCastPP for allocations where possible.
4440 //
4441 // (Note: don't forget to change the order of the second AddP node on
4442 // the alloc_worklist if the order of the worklist processing is changed,
4443 // see the comment in find_second_addp().)
4444 //
4445 while (alloc_worklist.length() != 0) {
4446 Node *n = alloc_worklist.pop();
4447 uint ni = n->_idx;
4448 if (n->is_Call()) {
4449 CallNode *alloc = n->as_Call();
4450 // copy escape information to call node
4451 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4452 PointsToNode::EscapeState es = ptn->escape_state();
4453 // We have an allocation or call which returns a Java object,
4454 // see if it is non-escaped.
4455 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4456 continue;
4457 }
4458 // Find CheckCastPP for the allocate or for the return value of a call
4459 n = alloc->result_cast();
4460 if (n == nullptr) { // No uses except Initialize node
4461 if (alloc->is_Allocate()) {
4462 // Set the scalar_replaceable flag for allocation
4463 // so it could be eliminated if it has no uses.
4464 alloc->as_Allocate()->_is_scalar_replaceable = true;
4465 }
4466 continue;
4467 }
4468 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4469 // we could reach here for allocate case if one init is associated with many allocs.
4470 if (alloc->is_Allocate()) {
4471 alloc->as_Allocate()->_is_scalar_replaceable = false;
4472 }
4473 continue;
4474 }
4475
4476 // The inline code for Object.clone() casts the allocation result to
4477 // java.lang.Object and then to the actual type of the allocated
4478 // object. Detect this case and use the second cast.
4479 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4480 // the allocation result is cast to java.lang.Object and then
4481 // to the actual Array type.
4482 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4483 && (alloc->is_AllocateArray() ||
4484 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4485 Node *cast2 = nullptr;
4486 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4487 Node *use = n->fast_out(i);
4488 if (use->is_CheckCastPP()) {
4489 cast2 = use;
4490 break;
4491 }
4492 }
4493 if (cast2 != nullptr) {
4494 n = cast2;
4495 } else {
4496 // Non-scalar replaceable if the allocation type is unknown statically
4497 // (reflection allocation), the object can't be restored during
4498 // deoptimization without precise type.
4499 continue;
4500 }
4501 }
4502
4503 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4504 if (t == nullptr) {
4505 continue; // not a TypeOopPtr
4506 }
4507 if (!t->klass_is_exact()) {
4508 continue; // not an unique type
4509 }
4510 if (alloc->is_Allocate()) {
4511 // Set the scalar_replaceable flag for allocation
4512 // so it could be eliminated.
4513 alloc->as_Allocate()->_is_scalar_replaceable = true;
4514 }
4515 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4516 // in order for an object to be scalar-replaceable, it must be:
4517 // - a direct allocation (not a call returning an object)
4518 // - non-escaping
4519 // - eligible to be a unique type
4520 // - not determined to be ineligible by escape analysis
4521 set_map(alloc, n);
4522 set_map(n, alloc);
4523 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4524 igvn->hash_delete(n);
4525 igvn->set_type(n, tinst);
4526 n->raise_bottom_type(tinst);
4527 igvn->hash_insert(n);
4528 record_for_optimizer(n);
4529 // Allocate an alias index for the header fields. Accesses to
4530 // the header emitted during macro expansion wouldn't have
4531 // correct memory state otherwise.
4532 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4533 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4534 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4535 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4536 InitializeNode* init = alloc->as_Allocate()->initialization();
4537 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4538 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4539 const TypePtr* adr_type = proj->adr_type();
4540 const TypePtr* new_adr_type = tinst->add_offset(adr_type->offset());
4541 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4542 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4543 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4544 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4545 igvn->set_type(new_proj, new_proj->bottom_type());
4546 record_for_optimizer(new_proj);
4547 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4548 }
4549 };
4550 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4551
4552 // First, put on the worklist all Field edges from Connection Graph
4553 // which is more accurate than putting immediate users from Ideal Graph.
4554 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4555 PointsToNode* tgt = e.get();
4556 if (tgt->is_Arraycopy()) {
4557 continue;
4558 }
4559 Node* use = tgt->ideal_node();
4560 assert(tgt->is_Field() && use->is_AddP(),
4561 "only AddP nodes are Field edges in CG");
4562 if (use->outcnt() > 0) { // Don't process dead nodes
4563 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4564 if (addp2 != nullptr) {
4565 assert(alloc->is_AllocateArray(),"array allocation was expected");
4566 alloc_worklist.append_if_missing(addp2);
4567 }
4568 alloc_worklist.append_if_missing(use);
4569 }
4570 }
4571
4572 // An allocation may have an Initialize which has raw stores. Scan
4573 // the users of the raw allocation result and push AddP users
4574 // on alloc_worklist.
4575 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4576 assert (raw_result != nullptr, "must have an allocation result");
4577 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4578 Node *use = raw_result->fast_out(i);
4579 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4580 Node* addp2 = find_second_addp(use, raw_result);
4581 if (addp2 != nullptr) {
4582 assert(alloc->is_AllocateArray(),"array allocation was expected");
4583 alloc_worklist.append_if_missing(addp2);
4584 }
4585 alloc_worklist.append_if_missing(use);
4586 } else if (use->is_MemBar()) {
4587 memnode_worklist.append_if_missing(use);
4588 }
4589 }
4590 }
4591 } else if (n->is_AddP()) {
4592 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4593 // This AddP will go away when we reduce the Phi
4594 continue;
4595 }
4596 Node* addp_base = get_addp_base(n);
4597 JavaObjectNode* jobj = unique_java_object(addp_base);
4598 if (jobj == nullptr || jobj == phantom_obj) {
4599 #ifdef ASSERT
4600 ptnode_adr(get_addp_base(n)->_idx)->dump();
4601 ptnode_adr(n->_idx)->dump();
4602 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4603 #endif
4604 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4605 return;
4606 }
4607 Node *base = get_map(jobj->idx()); // CheckCastPP node
4608 if (!split_AddP(n, base)) continue; // wrong type from dead path
4609 } else if (n->is_Phi() ||
4610 n->is_CheckCastPP() ||
4611 n->is_EncodeP() ||
4612 n->is_DecodeN() ||
4613 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4614 if (visited.test_set(n->_idx)) {
4615 assert(n->is_Phi(), "loops only through Phi's");
4616 continue; // already processed
4617 }
4618 // Reducible Phi's will be removed from the graph after split_unique_types
4619 // finishes. For now we just try to split out the SR inputs of the merge.
4620 Node* parent = n->in(1);
4621 if (reducible_merges.member(n)) {
4622 reduce_phi(n->as_Phi(), alloc_worklist);
4623 #ifdef ASSERT
4624 if (VerifyReduceAllocationMerges) {
4625 reduced_merges.push(n);
4626 }
4627 #endif
4628 continue;
4629 } else if (reducible_merges.member(parent)) {
4630 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4631 // part of reduce_merge.
4632 continue;
4633 }
4634 JavaObjectNode* jobj = unique_java_object(n);
4635 if (jobj == nullptr || jobj == phantom_obj) {
4636 #ifdef ASSERT
4637 ptnode_adr(n->_idx)->dump();
4638 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4639 #endif
4640 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4641 return;
4642 } else {
4643 Node *val = get_map(jobj->idx()); // CheckCastPP node
4644 TypeNode *tn = n->as_Type();
4645 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4646 assert(tinst != nullptr && tinst->is_known_instance() &&
4647 tinst->instance_id() == jobj->idx() , "instance type expected.");
4648
4649 const Type *tn_type = igvn->type(tn);
4650 const TypeOopPtr *tn_t;
4651 if (tn_type->isa_narrowoop()) {
4652 tn_t = tn_type->make_ptr()->isa_oopptr();
4653 } else {
4654 tn_t = tn_type->isa_oopptr();
4655 }
4656 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4657 if (tn_type->isa_narrowoop()) {
4658 tn_type = tinst->make_narrowoop();
4659 } else {
4660 tn_type = tinst;
4661 }
4662 igvn->hash_delete(tn);
4663 igvn->set_type(tn, tn_type);
4664 tn->set_type(tn_type);
4665 igvn->hash_insert(tn);
4666 record_for_optimizer(n);
4667 } else {
4668 assert(tn_type == TypePtr::NULL_PTR ||
4669 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4670 "unexpected type");
4671 continue; // Skip dead path with different type
4672 }
4673 }
4674 } else {
4675 DEBUG_ONLY(n->dump();)
4676 assert(false, "EA: unexpected node");
4677 continue;
4678 }
4679 // push allocation's users on appropriate worklist
4680 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4681 Node *use = n->fast_out(i);
4682 if(use->is_Mem() && use->in(MemNode::Address) == n) {
4683 // Load/store to instance's field
4684 memnode_worklist.append_if_missing(use);
4685 } else if (use->is_MemBar()) {
4686 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4687 memnode_worklist.append_if_missing(use);
4688 }
4689 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4690 Node* addp2 = find_second_addp(use, n);
4691 if (addp2 != nullptr) {
4692 alloc_worklist.append_if_missing(addp2);
4693 }
4694 alloc_worklist.append_if_missing(use);
4695 } else if (use->is_Phi() ||
4696 use->is_CheckCastPP() ||
4697 use->is_EncodeNarrowPtr() ||
4698 use->is_DecodeNarrowPtr() ||
4699 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4700 alloc_worklist.append_if_missing(use);
4701 #ifdef ASSERT
4702 } else if (use->is_Mem()) {
4703 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4704 } else if (use->is_MergeMem()) {
4705 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4706 } else if (use->is_SafePoint()) {
4707 // Look for MergeMem nodes for calls which reference unique allocation
4708 // (through CheckCastPP nodes) even for debug info.
4709 Node* m = use->in(TypeFunc::Memory);
4710 if (m->is_MergeMem()) {
4711 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4712 }
4713 } else if (use->Opcode() == Op_EncodeISOArray) {
4714 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4715 // EncodeISOArray overwrites destination array
4716 memnode_worklist.append_if_missing(use);
4717 }
4718 } else {
4719 uint op = use->Opcode();
4720 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4721 (use->in(MemNode::Memory) == n)) {
4722 // They overwrite memory edge corresponding to destination array,
4723 memnode_worklist.append_if_missing(use);
4724 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4725 op == Op_CastP2X ||
4726 op == Op_FastLock || op == Op_AryEq ||
4727 op == Op_StrComp || op == Op_CountPositives ||
4728 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4729 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4730 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4731 op == Op_SubTypeCheck ||
4732 op == Op_ReinterpretS2HF ||
4733 op == Op_ReachabilityFence ||
4734 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4735 n->dump();
4736 use->dump();
4737 assert(false, "EA: missing allocation reference path");
4738 }
4739 #endif
4740 }
4741 }
4742
4743 }
4744
4745 #ifdef ASSERT
4746 if (VerifyReduceAllocationMerges) {
4747 for (uint i = 0; i < reducible_merges.size(); i++) {
4748 Node* phi = reducible_merges.at(i);
4749
4750 if (!reduced_merges.member(phi)) {
4751 phi->dump(2);
4752 phi->dump(-2);
4753 assert(false, "This reducible merge wasn't reduced.");
4754 }
4755
4756 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4757 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4758 Node* use = phi->fast_out(j);
4759 if (!use->is_SafePoint() && !use->is_CastPP()) {
4760 phi->dump(2);
4761 phi->dump(-2);
4762 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4763 }
4764 }
4765 }
4766 }
4767 #endif
4768
4769 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4770 // type, record it in the ArrayCopy node so we know what memory this
4771 // node uses/modified.
4772 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4773 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4774 Node* dest = ac->in(ArrayCopyNode::Dest);
4775 if (dest->is_AddP()) {
4776 dest = get_addp_base(dest);
4777 }
4778 JavaObjectNode* jobj = unique_java_object(dest);
4779 if (jobj != nullptr) {
4780 Node *base = get_map(jobj->idx());
4781 if (base != nullptr) {
4782 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4783 ac->_dest_type = base_t;
4784 }
4785 }
4786 Node* src = ac->in(ArrayCopyNode::Src);
4787 if (src->is_AddP()) {
4788 src = get_addp_base(src);
4789 }
4790 jobj = unique_java_object(src);
4791 if (jobj != nullptr) {
4792 Node* base = get_map(jobj->idx());
4793 if (base != nullptr) {
4794 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4795 ac->_src_type = base_t;
4796 }
4797 }
4798 }
4799
4800 // New alias types were created in split_AddP().
4801 uint new_index_end = (uint) _compile->num_alias_types();
4802
4803 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
4804
4805 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4806 // compute new values for Memory inputs (the Memory inputs are not
4807 // actually updated until phase 4.)
4808 if (memnode_worklist.length() == 0)
4809 return; // nothing to do
4810 while (memnode_worklist.length() != 0) {
4811 Node *n = memnode_worklist.pop();
4812 if (visited.test_set(n->_idx)) {
4813 continue;
4814 }
4815 if (n->is_Phi() || n->is_ClearArray()) {
4816 // we don't need to do anything, but the users must be pushed
4817 } else if (n->is_MemBar()) { // MemBar nodes
4818 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
4819 // we don't need to do anything, but the users must be pushed
4820 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4821 if (n == nullptr) {
4822 continue;
4823 }
4824 }
4825 } else if (n->is_CallLeaf()) {
4826 // Runtime calls with narrow memory input (no MergeMem node)
4827 // get the memory projection
4828 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4829 if (n == nullptr) {
4830 continue;
4831 }
4832 } else if (n->Opcode() == Op_StrInflatedCopy) {
4833 // Check direct uses of StrInflatedCopy.
4834 // It is memory type Node - no special SCMemProj node.
4835 } else if (n->Opcode() == Op_StrCompressedCopy ||
4836 n->Opcode() == Op_EncodeISOArray) {
4837 // get the memory projection
4838 n = n->find_out_with(Op_SCMemProj);
4839 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4840 } else if (n->is_Proj()) {
4841 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
4842 } else {
4843 #ifdef ASSERT
4844 if (!n->is_Mem()) {
4845 n->dump();
4846 }
4847 assert(n->is_Mem(), "memory node required.");
4848 #endif
4849 Node *addr = n->in(MemNode::Address);
4850 const Type *addr_t = igvn->type(addr);
4851 if (addr_t == Type::TOP) {
4852 continue;
4853 }
4854 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4855 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4856 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4857 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4858 if (_compile->failing()) {
4859 return;
4860 }
4861 if (mem != n->in(MemNode::Memory)) {
4862 // We delay the memory edge update since we need old one in
4863 // MergeMem code below when instances memory slices are separated.
4864 set_map(n, mem);
4865 }
4866 if (n->is_Load()) {
4867 continue; // don't push users
4868 } else if (n->is_LoadStore()) {
4869 // get the memory projection
4870 n = n->find_out_with(Op_SCMemProj);
4871 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4872 }
4873 }
4874 // push user on appropriate worklist
4875 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4876 Node *use = n->fast_out(i);
4877 if (use->is_Phi() || use->is_ClearArray()) {
4878 memnode_worklist.append_if_missing(use);
4879 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4880 memnode_worklist.append_if_missing(use);
4881 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4882 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4883 memnode_worklist.append_if_missing(use);
4884 }
4885 } else if (use->is_Proj()) {
4886 assert(n->is_Initialize(), "We only push projections of Initialize");
4887 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
4888 memnode_worklist.append_if_missing(use);
4889 }
4890 #ifdef ASSERT
4891 } else if(use->is_Mem()) {
4892 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4893 } else if (use->is_MergeMem()) {
4894 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4895 } else if (use->Opcode() == Op_EncodeISOArray) {
4896 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4897 // EncodeISOArray overwrites destination array
4898 memnode_worklist.append_if_missing(use);
4899 }
4900 } else {
4901 uint op = use->Opcode();
4902 if ((use->in(MemNode::Memory) == n) &&
4903 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4904 // They overwrite memory edge corresponding to destination array,
4905 memnode_worklist.append_if_missing(use);
4906 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4907 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4908 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4909 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
4910 n->dump();
4911 use->dump();
4912 assert(false, "EA: missing memory path");
4913 }
4914 #endif
4915 }
4916 }
4917 }
4918
4919 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4920 // Walk each memory slice moving the first node encountered of each
4921 // instance type to the input corresponding to its alias index.
4922 uint length = mergemem_worklist.length();
4923 for( uint next = 0; next < length; ++next ) {
4924 MergeMemNode* nmm = mergemem_worklist.at(next);
4925 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4926 // Note: we don't want to use MergeMemStream here because we only want to
4927 // scan inputs which exist at the start, not ones we add during processing.
4928 // Note 2: MergeMem may already contains instance memory slices added
4929 // during find_inst_mem() call when memory nodes were processed above.
4930 igvn->hash_delete(nmm);
4931 uint nslices = MIN2(nmm->req(), new_index_start);
4932 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4933 Node* mem = nmm->in(i);
4934 Node* cur = nullptr;
4935 if (mem == nullptr || mem->is_top()) {
4936 continue;
4937 }
4938 // First, update mergemem by moving memory nodes to corresponding slices
4939 // if their type became more precise since this mergemem was created.
4940 while (mem->is_Mem()) {
4941 const Type* at = igvn->type(mem->in(MemNode::Address));
4942 if (at != Type::TOP) {
4943 assert (at->isa_ptr() != nullptr, "pointer type required.");
4944 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4945 if (idx == i) {
4946 if (cur == nullptr) {
4947 cur = mem;
4948 }
4949 } else {
4950 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4951 nmm->set_memory_at(idx, mem);
4952 }
4953 }
4954 }
4955 mem = mem->in(MemNode::Memory);
4956 }
4957 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4958 // Find any instance of the current type if we haven't encountered
4959 // already a memory slice of the instance along the memory chain.
4960 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4961 if((uint)_compile->get_general_index(ni) == i) {
4962 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4963 if (nmm->is_empty_memory(m)) {
4964 Node* result = find_inst_mem(mem, ni, orig_phis);
4965 if (_compile->failing()) {
4966 return;
4967 }
4968 nmm->set_memory_at(ni, result);
4969 }
4970 }
4971 }
4972 }
4973 // Find the rest of instances values
4974 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4975 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4976 Node* result = step_through_mergemem(nmm, ni, tinst);
4977 if (result == nmm->base_memory()) {
4978 // Didn't find instance memory, search through general slice recursively.
4979 result = nmm->memory_at(_compile->get_general_index(ni));
4980 result = find_inst_mem(result, ni, orig_phis);
4981 if (_compile->failing()) {
4982 return;
4983 }
4984 nmm->set_memory_at(ni, result);
4985 }
4986 }
4987
4988 // If we have crossed the 3/4 point of max node limit it's too risky
4989 // to continue with EA/SR because we might hit the max node limit.
4990 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4991 if (_compile->do_reduce_allocation_merges()) {
4992 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4993 } else if (_invocation > 0) {
4994 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4995 } else {
4996 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4997 }
4998 return;
4999 }
5000
5001 igvn->hash_insert(nmm);
5002 record_for_optimizer(nmm);
5003 }
5004
5005 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
5006
5007 // Phase 4: Update the inputs of non-instance memory Phis and
5008 // the Memory input of memnodes
5009 // First update the inputs of any non-instance Phi's from
5010 // which we split out an instance Phi. Note we don't have
5011 // to recursively process Phi's encountered on the input memory
5012 // chains as is done in split_memory_phi() since they will
5013 // also be processed here.
5014 for (int j = 0; j < orig_phis.length(); j++) {
5015 PhiNode *phi = orig_phis.at(j);
5016 int alias_idx = _compile->get_alias_index(phi->adr_type());
5017 igvn->hash_delete(phi);
5018 for (uint i = 1; i < phi->req(); i++) {
5019 Node *mem = phi->in(i);
5020 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5021 if (_compile->failing()) {
5022 return;
5023 }
5024 if (mem != new_mem) {
5025 phi->set_req(i, new_mem);
5026 }
5027 }
5028 igvn->hash_insert(phi);
5029 record_for_optimizer(phi);
5030 }
5031
5032 // Update the memory inputs of MemNodes with the value we computed
5033 // in Phase 2 and move stores memory users to corresponding memory slices.
5034 // Disable memory split verification code until the fix for 6984348.
5035 // Currently it produces false negative results since it does not cover all cases.
5036 #if 0 // ifdef ASSERT
5037 visited.Reset();
5038 Node_Stack old_mems(arena, _compile->unique() >> 2);
5039 #endif
5040 for (uint i = 0; i < ideal_nodes.size(); i++) {
5041 Node* n = ideal_nodes.at(i);
5042 Node* nmem = get_map(n->_idx);
5043 assert(nmem != nullptr, "sanity");
5044 if (n->is_Mem()) {
5045 #if 0 // ifdef ASSERT
5046 Node* old_mem = n->in(MemNode::Memory);
5047 if (!visited.test_set(old_mem->_idx)) {
5048 old_mems.push(old_mem, old_mem->outcnt());
5049 }
5050 #endif
5051 assert(n->in(MemNode::Memory) != nmem, "sanity");
5052 if (!n->is_Load()) {
5053 // Move memory users of a store first.
5054 move_inst_mem(n, orig_phis);
5055 }
5056 // Now update memory input
5057 igvn->hash_delete(n);
5058 n->set_req(MemNode::Memory, nmem);
5059 igvn->hash_insert(n);
5060 record_for_optimizer(n);
5061 } else {
5062 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5063 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5064 }
5065 }
5066 #if 0 // ifdef ASSERT
5067 // Verify that memory was split correctly
5068 while (old_mems.is_nonempty()) {
5069 Node* old_mem = old_mems.node();
5070 uint old_cnt = old_mems.index();
5071 old_mems.pop();
5072 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5073 }
5074 #endif
5075 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5076 }
5077
5078 #ifndef PRODUCT
5079 int ConnectionGraph::_no_escape_counter = 0;
5080 int ConnectionGraph::_arg_escape_counter = 0;
5081 int ConnectionGraph::_global_escape_counter = 0;
5082
5083 static const char *node_type_names[] = {
5084 "UnknownType",
5085 "JavaObject",
5086 "LocalVar",
5087 "Field",
5088 "Arraycopy"
5089 };
5090
5091 static const char *esc_names[] = {
5092 "UnknownEscape",
5093 "NoEscape",
5094 "ArgEscape",
5095 "GlobalEscape"
5096 };
5097
5098 const char* PointsToNode::esc_name() const {
5099 return esc_names[(int)escape_state()];
5100 }
5101
5102 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5103 NodeType nt = node_type();
5104 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5105 if (print_state) {
5106 EscapeState es = escape_state();
5107 EscapeState fields_es = fields_escape_state();
5108 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5109 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5110 out->print("NSR ");
5111 }
5112 }
5113 }
5114
5115 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5116 dump_header(print_state, out);
5117 if (is_Field()) {
5118 FieldNode* f = (FieldNode*)this;
5119 if (f->is_oop()) {
5120 out->print("oop ");
5121 }
5122 if (f->offset() > 0) {
5123 out->print("+%d ", f->offset());
5124 }
5125 out->print("(");
5126 for (BaseIterator i(f); i.has_next(); i.next()) {
5127 PointsToNode* b = i.get();
5128 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5129 }
5130 out->print(" )");
5131 }
5132 out->print("[");
5133 for (EdgeIterator i(this); i.has_next(); i.next()) {
5134 PointsToNode* e = i.get();
5135 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5136 }
5137 out->print(" [");
5138 for (UseIterator i(this); i.has_next(); i.next()) {
5139 PointsToNode* u = i.get();
5140 bool is_base = false;
5141 if (PointsToNode::is_base_use(u)) {
5142 is_base = true;
5143 u = PointsToNode::get_use_node(u)->as_Field();
5144 }
5145 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5146 }
5147 out->print(" ]] ");
5148 if (_node == nullptr) {
5149 out->print("<null>%s", newline ? "\n" : "");
5150 } else {
5151 _node->dump(newline ? "\n" : "", false, out);
5152 }
5153 }
5154
5155 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5156 bool first = true;
5157 int ptnodes_length = ptnodes_worklist.length();
5158 for (int i = 0; i < ptnodes_length; i++) {
5159 PointsToNode *ptn = ptnodes_worklist.at(i);
5160 if (ptn == nullptr || !ptn->is_JavaObject()) {
5161 continue;
5162 }
5163 PointsToNode::EscapeState es = ptn->escape_state();
5164 if ((es != PointsToNode::NoEscape) && !Verbose) {
5165 continue;
5166 }
5167 Node* n = ptn->ideal_node();
5168 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5169 n->as_CallStaticJava()->is_boxing_method())) {
5170 if (first) {
5171 tty->cr();
5172 tty->print("======== Connection graph for ");
5173 _compile->method()->print_short_name();
5174 tty->cr();
5175 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5176 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5177 tty->cr();
5178 first = false;
5179 }
5180 ptn->dump();
5181 // Print all locals and fields which reference this allocation
5182 for (UseIterator j(ptn); j.has_next(); j.next()) {
5183 PointsToNode* use = j.get();
5184 if (use->is_LocalVar()) {
5185 use->dump(Verbose);
5186 } else if (Verbose) {
5187 use->dump();
5188 }
5189 }
5190 tty->cr();
5191 }
5192 }
5193 }
5194
5195 void ConnectionGraph::print_statistics() {
5196 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));
5197 }
5198
5199 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5200 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5201 return;
5202 }
5203 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5204 JavaObjectNode* ptn = java_objects_worklist.at(next);
5205 if (ptn->ideal_node()->is_Allocate()) {
5206 if (ptn->escape_state() == PointsToNode::NoEscape) {
5207 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5208 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5209 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5210 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5211 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5212 } else {
5213 assert(false, "Unexpected Escape State");
5214 }
5215 }
5216 }
5217 }
5218
5219 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5220 if (_compile->directive()->TraceEscapeAnalysisOption) {
5221 assert(ptn != nullptr, "should not be null");
5222 assert(reason != nullptr, "should not be null");
5223 ptn->dump_header(true);
5224 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5225 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5226 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5227 }
5228 }
5229
5230 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5231 if (_compile->directive()->TraceEscapeAnalysisOption) {
5232 stringStream ss;
5233 ss.print("propagated from: ");
5234 from->dump(true, &ss, false);
5235 return ss.as_string();
5236 } else {
5237 return nullptr;
5238 }
5239 }
5240
5241 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5242 if (_compile->directive()->TraceEscapeAnalysisOption) {
5243 stringStream ss;
5244 ss.print("escapes as arg to:");
5245 call->dump("", false, &ss);
5246 return ss.as_string();
5247 } else {
5248 return nullptr;
5249 }
5250 }
5251
5252 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5253 if (_compile->directive()->TraceEscapeAnalysisOption) {
5254 stringStream ss;
5255 ss.print("is merged with other object: ");
5256 other->dump_header(true, &ss);
5257 return ss.as_string();
5258 } else {
5259 return nullptr;
5260 }
5261 }
5262
5263 #endif
5264
5265 void ConnectionGraph::record_for_optimizer(Node *n) {
5266 _igvn->_worklist.push(n);
5267 _igvn->add_users_to_worklist(n);
5268 }
--- EOF ---