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 OpaqueNotNull node between If and Bool nodes. But we could also have a sub class of IfNode,
592 // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
593 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
594 if (can_reduce) {
595 Node* iff_cmp = iff->in(1)->in(1);
596 int opc = iff_cmp->Opcode();
597 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
598 }
599 if (!can_reduce) {
600 #ifndef PRODUCT
601 if (TraceReduceAllocationMerges) {
602 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
603 n->dump(5);
604 }
605 #endif
606 return false;
607 }
608 }
609 }
610
611 if (!can_reduce_check_users(use, nesting+1)) {
612 return false;
613 }
614 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
615 if (!can_reduce_cmp(n, use)) {
616 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. CmpP/N %d isn't reducible.", n->_idx, _invocation, use->_idx);)
617 return false;
618 }
619 } else {
620 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. One of the uses is: %d %s", n->_idx, _invocation, use->_idx, use->Name());)
621 return false;
622 }
623 }
624
625 return true;
626 }
627
628 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
629 // only used in some certain code shapes. Check comments in
630 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
631 // details.
632 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
633 // If there was an error attempting to reduce allocation merges for this
634 // method we might have disabled the compilation and be retrying with RAM
635 // disabled.
636 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
637 return false;
638 }
639
640 const Type* phi_t = _igvn->type(ophi);
641 if (phi_t == nullptr ||
642 phi_t->make_ptr() == nullptr ||
643 phi_t->make_ptr()->isa_aryptr() != nullptr) {
644 return false;
645 }
646
647 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
648 return false;
649 }
650
651 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
652 return true;
653 }
654
655 // This method will return a CmpP/N that we need to use on the If controlling a
656 // CastPP after it was split. This method is only called on bases that are
657 // nullable therefore we always need a controlling if for the splitted CastPP.
658 //
659 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
660 // If the CastPP currently doesn't have a control then the CmpP/N will be
661 // against the null constant, otherwise it will be against the constant input of
662 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
663 // case because we have constraints on it and because the CastPP has a control
664 // input.
665 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
666 const Type* t = base->bottom_type();
667 Node* con = nullptr;
668
669 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
670 con = _igvn->zerocon(t->basic_type());
671 } else {
672 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
673 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
674 Node* bol = curr_ctrl->in(0)->in(1);
675 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
676 Node* curr_cmp = bol->in(1);
677 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
678 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
679 }
680
681 return CmpNode::make(base, con, t->basic_type());
682 }
683
684 // This method 'specializes' the CastPP passed as parameter to the base passed
685 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
686 // means that the CastPP now will be specific for a given base instead of a Phi.
687 // An If-Then-Else-Region block is inserted to control the CastPP. The control
688 // of the CastPP is a copy of the current one (if there is one) or a check
689 // against null.
690 //
691 // Before:
692 //
693 // C1 C2 ... Cn
694 // \ | /
695 // \ | /
696 // \ | /
697 // \ | /
698 // \ | /
699 // \ | /
700 // \|/
701 // Region B1 B2 ... Bn
702 // | \ | /
703 // | \ | /
704 // | \ | /
705 // | \ | /
706 // | \ | /
707 // | \ | /
708 // ---------------> Phi
709 // |
710 // X |
711 // | |
712 // | |
713 // ------> CastPP
714 //
715 // After (only partial illustration; base = B2, current_control = C2):
716 //
717 // C2
718 // |
719 // If
720 // / \
721 // / \
722 // T F
723 // /\ /
724 // / \ /
725 // / \ /
726 // C1 CastPP Reg Cn
727 // | | |
728 // | | |
729 // | | |
730 // -------------- | ----------
731 // | | |
732 // Region
733 //
734 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
735 Node* control_successor = current_control->unique_ctrl_out();
736 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
737 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
738 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
739 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
740 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
741 Node* end_region = _igvn->transform(new RegionNode(3));
742
743 // Insert the new if-else-region block into the graph
744 end_region->set_req(1, not_eq_control);
745 end_region->set_req(2, yes_eq_control);
746 control_successor->replace_edge(current_control, end_region, _igvn);
747
748 _igvn->_worklist.push(current_control);
749 _igvn->_worklist.push(control_successor);
750
751 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::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(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
784 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
785 Node* load = curr_load->clone();
786 load->set_req(0, nullptr);
787 load->set_req(1, mem);
788 load->set_req(2, addr);
789
790 if (cmp_region != nullptr) { // see comment on previous if
791 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
792 intermediate_phi->set_req(1, _igvn->transform(load));
793 load = intermediate_phi;
794 }
795
796 data_phi->set_req(i, _igvn->transform(load));
797 } else {
798 // Just use the default, which is already in phi
799 }
800 }
801
802 // Takes care of updating CG and split_unique_types worklists due
803 // to cloned AddP->Load.
804 updates_after_load_split(data_phi, curr_load, alloc_worklist);
805
806 return _igvn->transform(data_phi);
807 }
808
809 // This method only reduces CastPP fields loads; SafePoints are handled
810 // separately. The idea here is basically to clone the CastPP and place copies
811 // on each input of the Phi, including non-scalar replaceable inputs.
812 // Experimentation shows that the resulting IR graph is simpler that way than if
813 // we just split the cast through scalar-replaceable inputs.
814 //
815 // The reduction process requires that CastPP's control be one of:
816 // 1) no control,
817 // 2) the same region as Ophi, or
818 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
819 //
820 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
821 // flow. If the CastPP originally had an IfTrue/False control input then we'll
822 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
823 // juse use a CmpP/N against the null constant.
824 //
825 // The If-Then-Else-Region isn't always needed. For instance, if input to
826 // splitted cast was not nullable (or if it was the null constant) then we don't
827 // need (shouldn't) use a CastPP at all.
828 //
829 // After the casts are splitted we'll split the AddP->Loads through the Phi and
830 // connect them to the just split CastPPs.
831 //
832 // Before (CastPP control is same as Phi):
833 //
834 // Region Allocate Null Call
835 // | \ | /
836 // | \ | /
837 // | \ | /
838 // | \ | /
839 // | \ | /
840 // | \ | /
841 // ------------------> Phi # Oop Phi
842 // | |
843 // | |
844 // | |
845 // | |
846 // ----------------> CastPP
847 // |
848 // AddP
849 // |
850 // Load
851 //
852 // After (Very much simplified):
853 //
854 // Call Null
855 // \ /
856 // CmpP
857 // |
858 // Bool#NE
859 // |
860 // If
861 // / \
862 // T F
863 // / \ /
864 // / R
865 // CastPP |
866 // | |
867 // AddP |
868 // | |
869 // Load |
870 // \ | 0
871 // Allocate \ | /
872 // \ \ | /
873 // AddP Phi
874 // \ /
875 // Load /
876 // \ 0 /
877 // \ | /
878 // \|/
879 // Phi # "Field" Phi
880 //
881 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node*> &alloc_worklist) {
882 Node* ophi = curr_castpp->in(1);
883 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
884
885 // Identify which base should be used for AddP->Load later when spliting the
886 // CastPP->Loads through ophi. Three kind of values may be stored in this
887 // array, depending on the nullability status of the corresponding input in
888 // ophi.
889 //
890 // - nullptr: Meaning that the base is actually the null constant and therefore
891 // we won't try to load from it.
892 //
893 // - CFG Node: Meaning that the base is a CastPP that was specialized for
894 // this input of Ophi. I.e., we added an If->Then->Else-Region
895 // that will 'activate' the CastPp only when the input is not Null.
896 //
897 // - Other Node: Meaning that the base is not nullable and therefore we'll try
898 // to load directly from it.
899 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
900
901 for (uint i = 1; i < ophi->req(); i++) {
902 Node* base = ophi->in(i);
903 const Type* base_t = _igvn->type(base);
904
905 if (base_t->maybe_null()) {
906 if (base->is_Con()) {
907 // Nothing todo as bases_for_loads[i] is already null
908 } else {
909 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
910 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
911 }
912 } else {
913 bases_for_loads.at_put(i, base);
914 }
915 }
916
917 // Now let's split the CastPP->Loads through the Phi
918 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
919 Node* use = curr_castpp->raw_out(i);
920 if (use->is_AddP()) {
921 for (int j = use->outcnt()-1; j >= 0;) {
922 Node* use_use = use->raw_out(j);
923 assert(use_use->is_Load(), "Expected this to be a Load node.");
924
925 // We can't make an unconditional load from a nullable input. The
926 // 'split_castpp_load_through_phi` method will add an
927 // 'If-Then-Else-Region` around nullable bases and only load from them
928 // when the input is not null.
929 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
930 _igvn->replace_node(use_use, phi);
931
932 --j;
933 j = MIN2(j, (int)use->outcnt()-1);
934 }
935
936 _igvn->remove_dead_node(use);
937 }
938 --i;
939 i = MIN2(i, (int)curr_castpp->outcnt()-1);
940 }
941 }
942
943 // This method split a given CmpP/N through the Phi used in one of its inputs.
944 // As a result we convert a comparison with a pointer to a comparison with an
945 // integer.
946 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
947 // while the other must be a constant.
948 // The splitting process is basically just cloning the CmpP/N above the input
949 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
950 // can prove at compile time the result of the comparison.
951 //
952 // Before:
953 //
954 // in1 in2 ... inN
955 // \ | /
956 // \ | /
957 // \ | /
958 // \ | /
959 // \ | /
960 // \ | /
961 // Phi
962 // | Other
963 // | /
964 // | /
965 // | /
966 // CmpP/N
967 //
968 // After:
969 //
970 // in1 Other in2 Other inN Other
971 // | | | | | |
972 // \ | | | | |
973 // \ / | / | /
974 // CmpP/N CmpP/N CmpP/N
975 // Bool Bool Bool
976 // \ | /
977 // \ | /
978 // \ | /
979 // \ | /
980 // \ | /
981 // \ | /
982 // \ | /
983 // \ | /
984 // Phi
985 // |
986 // | Zero
987 // | /
988 // | /
989 // | /
990 // CmpI
991 //
992 //
993 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
994 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
995 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
996
997 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
998 Node* zero = _igvn->intcon(0);
999 Node* one = _igvn->intcon(1);
1000 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
1001
1002 // This Phi will merge the result of the Cmps split through the Phi
1003 Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
1004
1005 for (uint i=1; i<ophi->req(); i++) {
1006 Node* ophi_input = ophi->in(i);
1007 Node* res_phi_input = nullptr;
1008
1009 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1010 if (tcmp->singleton()) {
1011 if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
1012 (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
1013 res_phi_input = one;
1014 } else {
1015 res_phi_input = zero;
1016 }
1017 } else {
1018 Node* ncmp = _igvn->transform(cmp->clone());
1019 ncmp->set_req(1, ophi_input);
1020 ncmp->set_req(2, other);
1021 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1022 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1023 }
1024
1025 res_phi->set_req(i, res_phi_input);
1026 }
1027
1028 // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1029 Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1030 _igvn->replace_node(cmp, new_cmp);
1031 }
1032
1033 // Push the newly created AddP on alloc_worklist and patch
1034 // the connection graph. Note that the changes in the CG below
1035 // won't affect the ES of objects since the new nodes have the
1036 // same status as the old ones.
1037 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1038 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1039 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1040 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1041
1042 if (data_phi == nullptr || !data_phi->is_Phi()) {
1043 // Make this a retry?
1044 return ;
1045 }
1046
1047 Node* previous_addp = previous_load->in(MemNode::Address);
1048 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1049 for (uint i = 1; i < data_phi->req(); i++) {
1050 Node* new_load = data_phi->in(i);
1051
1052 if (new_load->is_Phi()) {
1053 // new_load is currently the "intermediate_phi" from an specialized
1054 // CastPP.
1055 new_load = new_load->in(1);
1056 }
1057
1058 // "new_load" might actually be a constant, parameter, etc.
1059 if (new_load->is_Load()) {
1060 Node* new_addp = new_load->in(MemNode::Address);
1061
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 JVMState *jvms = sfpt->jvms();
1277 uint merge_idx = (sfpt->req() - jvms->scloff());
1278 int debug_start = jvms->debug_start();
1279
1280 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1281 smerge->init_req(0, _compile->root());
1282 _igvn->register_new_node_with_optimizer(smerge);
1283
1284 // The next two inputs are:
1285 // (1) A copy of the original pointer to NSR objects.
1286 // (2) A selector, used to decide if we need to rematerialize an object
1287 // or use the pointer to a NSR object.
1288 // See more details of these fields in the declaration of SafePointScalarMergeNode
1289 sfpt->add_req(nsr_merge_pointer);
1290 sfpt->add_req(selector);
1291
1292 for (uint i = 1; i < ophi->req(); i++) {
1293 Node* base = ophi->in(i);
1294 JavaObjectNode* ptn = unique_java_object(base);
1295
1296 // If the base is not scalar replaceable we don't need to register information about
1297 // it at this time.
1298 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1299 continue;
1300 }
1301
1302 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1303 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt);
1304 if (sobj == nullptr) {
1305 return false;
1306 }
1307
1308 // Now make a pass over the debug information replacing any references
1309 // to the allocated object with "sobj"
1310 Node* ccpp = alloc->result_cast();
1311 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1312
1313 // Register the scalarized object as a candidate for reallocation
1314 smerge->add_req(sobj);
1315 }
1316
1317 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1318 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1319
1320 // The call to 'replace_edges_in_range' above might have removed the
1321 // reference to ophi that we need at _merge_pointer_idx. The line below make
1322 // sure the reference is maintained.
1323 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1324 _igvn->_worklist.push(sfpt);
1325 }
1326
1327 return true;
1328 }
1329
1330 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node*> &alloc_worklist) {
1331 bool delay = _igvn->delay_transform();
1332 _igvn->set_delay_transform(true);
1333 _igvn->hash_delete(ophi);
1334
1335 // Copying all users first because some will be removed and others won't.
1336 // Ophi also may acquire some new users as part of Cast reduction.
1337 // CastPPs also need to be processed before CmpPs.
1338 Unique_Node_List castpps;
1339 Unique_Node_List others;
1340 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1341 Node* use = ophi->fast_out(i);
1342
1343 if (use->is_CastPP()) {
1344 castpps.push(use);
1345 } else if (use->is_AddP() || use->is_Cmp()) {
1346 others.push(use);
1347 } else {
1348 // Safepoints to be processed later; other users aren't expected here
1349 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1350 }
1351 }
1352
1353 _compile->print_method(PHASE_EA_BEFORE_PHI_REDUCTION, 5, ophi);
1354
1355 // CastPPs need to be processed before Cmps because during the process of
1356 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1357 // by the If controlling the CastPP.
1358 for (uint i = 0; i < castpps.size(); i++) {
1359 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist);
1360 _compile->print_method(PHASE_EA_AFTER_PHI_CASTPP_REDUCTION, 6, castpps.at(i));
1361 }
1362
1363 for (uint i = 0; i < others.size(); i++) {
1364 Node* use = others.at(i);
1365
1366 if (use->is_AddP()) {
1367 reduce_phi_on_field_access(use, alloc_worklist);
1368 _compile->print_method(PHASE_EA_AFTER_PHI_ADDP_REDUCTION, 6, use);
1369 } else if(use->is_Cmp()) {
1370 reduce_phi_on_cmp(use);
1371 _compile->print_method(PHASE_EA_AFTER_PHI_CMP_REDUCTION, 6, use);
1372 }
1373 }
1374
1375 _igvn->set_delay_transform(delay);
1376 }
1377
1378 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1379 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1380 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1381 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1382 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1383
1384 for (uint i = 1; i < ophi->req(); i++) {
1385 Node* base = ophi->in(i);
1386 JavaObjectNode* ptn = unique_java_object(base);
1387
1388 if (ptn != nullptr && ptn->scalar_replaceable()) {
1389 new_phi->set_req(i, null_ptr);
1390 } else {
1391 new_phi->set_req(i, ophi->in(i));
1392 }
1393 }
1394
1395 for (int i = ophi->outcnt()-1; i >= 0;) {
1396 Node* out = ophi->raw_out(i);
1397
1398 if (out->is_ConstraintCast()) {
1399 const Type* out_t = _igvn->type(out)->make_ptr();
1400 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1401 bool change = out_new_t != out_t;
1402
1403 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1404 Node* out2 = out->raw_out(j);
1405 if (!out2->is_SafePoint()) {
1406 change = false;
1407 break;
1408 }
1409 }
1410
1411 if (change) {
1412 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::DependencyType::NonFloatingNarrowing, nullptr);
1413 _igvn->replace_node(out, new_cast);
1414 _igvn->register_new_node_with_optimizer(new_cast);
1415 }
1416 }
1417
1418 --i;
1419 i = MIN2(i, (int)ophi->outcnt()-1);
1420 }
1421
1422 _igvn->replace_node(ophi, new_phi);
1423 }
1424
1425 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1426 if (!C->do_reduce_allocation_merges()) return;
1427
1428 Unique_Node_List ideal_nodes;
1429 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1430 ideal_nodes.push(root);
1431
1432 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1433 Node* n = ideal_nodes.at(next);
1434
1435 if (n->is_SafePointScalarMerge()) {
1436 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1437
1438 // Validate inputs of merge
1439 for (uint i = 1; i < merge->req(); i++) {
1440 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1441 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1442 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1443 }
1444 }
1445
1446 // Validate users of merge
1447 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1448 Node* sfpt = merge->fast_out(i);
1449 if (sfpt->is_SafePoint()) {
1450 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1451
1452 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1453 assert(false, "SafePointScalarMerge nodes can't be nested.");
1454 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1455 }
1456 } else {
1457 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1458 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1459 }
1460 }
1461 }
1462
1463 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1464 Node* m = n->fast_out(i);
1465 ideal_nodes.push(m);
1466 }
1467 }
1468 }
1469
1470 // Returns true if there is an object in the scope of sfn that does not escape globally.
1471 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1472 Compile* C = _compile;
1473 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1474 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1475 DeoptimizeObjectsALot) {
1476 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1477 int num_locs = jvms->loc_size();
1478 for (int idx = 0; idx < num_locs; idx++) {
1479 Node* l = sfn->local(jvms, idx);
1480 if (not_global_escape(l)) {
1481 return true;
1482 }
1483 }
1484 }
1485 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1486 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1487 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1488 int num_mon = jvms->nof_monitors();
1489 for (int idx = 0; idx < num_mon; idx++) {
1490 Node* m = sfn->monitor_obj(jvms, idx);
1491 if (m != nullptr && not_global_escape(m)) {
1492 return true;
1493 }
1494 }
1495 }
1496 }
1497 return false;
1498 }
1499
1500 // Returns true if at least one of the arguments to the call is an object
1501 // that does not escape globally.
1502 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1503 if (call->method() != nullptr) {
1504 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1505 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1506 Node* p = call->in(idx);
1507 if (not_global_escape(p)) {
1508 return true;
1509 }
1510 }
1511 } else {
1512 const char* name = call->as_CallStaticJava()->_name;
1513 assert(name != nullptr, "no name");
1514 // no arg escapes through uncommon traps
1515 if (strcmp(name, "uncommon_trap") != 0) {
1516 // process_call_arguments() assumes that all arguments escape globally
1517 const TypeTuple* d = call->tf()->domain();
1518 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1519 const Type* at = d->field_at(i);
1520 if (at->isa_oopptr() != nullptr) {
1521 return true;
1522 }
1523 }
1524 }
1525 }
1526 return false;
1527 }
1528
1529
1530
1531 // Utility function for nodes that load an object
1532 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1533 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1534 // ThreadLocal has RawPtr type.
1535 const Type* t = _igvn->type(n);
1536 if (t->make_ptr() != nullptr) {
1537 Node* adr = n->in(MemNode::Address);
1538 #ifdef ASSERT
1539 if (!adr->is_AddP()) {
1540 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1541 } else {
1542 assert((ptnode_adr(adr->_idx) == nullptr ||
1543 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1544 }
1545 #endif
1546 add_local_var_and_edge(n, PointsToNode::NoEscape,
1547 adr, delayed_worklist);
1548 }
1549 }
1550
1551 // Populate Connection Graph with PointsTo nodes and create simple
1552 // connection graph edges.
1553 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1554 assert(!_verify, "this method should not be called for verification");
1555 PhaseGVN* igvn = _igvn;
1556 uint n_idx = n->_idx;
1557 PointsToNode* n_ptn = ptnode_adr(n_idx);
1558 if (n_ptn != nullptr) {
1559 return; // No need to redefine PointsTo node during first iteration.
1560 }
1561 int opcode = n->Opcode();
1562 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1563 if (gc_handled) {
1564 return; // Ignore node if already handled by GC.
1565 }
1566
1567 if (n->is_Call()) {
1568 // Arguments to allocation and locking don't escape.
1569 if (n->is_AbstractLock()) {
1570 // Put Lock and Unlock nodes on IGVN worklist to process them during
1571 // first IGVN optimization when escape information is still available.
1572 record_for_optimizer(n);
1573 } else if (n->is_Allocate()) {
1574 add_call_node(n->as_Call());
1575 record_for_optimizer(n);
1576 } else {
1577 if (n->is_CallStaticJava()) {
1578 const char* name = n->as_CallStaticJava()->_name;
1579 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1580 return; // Skip uncommon traps
1581 }
1582 }
1583 // Don't mark as processed since call's arguments have to be processed.
1584 delayed_worklist->push(n);
1585 // Check if a call returns an object.
1586 if ((n->as_Call()->returns_pointer() &&
1587 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1588 (n->is_CallStaticJava() &&
1589 n->as_CallStaticJava()->is_boxing_method())) {
1590 add_call_node(n->as_Call());
1591 }
1592 }
1593 return;
1594 }
1595 // Put this check here to process call arguments since some call nodes
1596 // point to phantom_obj.
1597 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1598 return; // Skip predefined nodes.
1599 }
1600 switch (opcode) {
1601 case Op_AddP: {
1602 Node* base = get_addp_base(n);
1603 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1604 // Field nodes are created for all field types. They are used in
1605 // adjust_scalar_replaceable_state() and split_unique_types().
1606 // Note, non-oop fields will have only base edges in Connection
1607 // Graph because such fields are not used for oop loads and stores.
1608 int offset = address_offset(n, igvn);
1609 add_field(n, PointsToNode::NoEscape, offset);
1610 if (ptn_base == nullptr) {
1611 delayed_worklist->push(n); // Process it later.
1612 } else {
1613 n_ptn = ptnode_adr(n_idx);
1614 add_base(n_ptn->as_Field(), ptn_base);
1615 }
1616 break;
1617 }
1618 case Op_CastX2P: {
1619 map_ideal_node(n, phantom_obj);
1620 break;
1621 }
1622 case Op_CastPP:
1623 case Op_CheckCastPP:
1624 case Op_EncodeP:
1625 case Op_DecodeN:
1626 case Op_EncodePKlass:
1627 case Op_DecodeNKlass: {
1628 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1629 break;
1630 }
1631 case Op_CMoveP: {
1632 add_local_var(n, PointsToNode::NoEscape);
1633 // Do not add edges during first iteration because some could be
1634 // not defined yet.
1635 delayed_worklist->push(n);
1636 break;
1637 }
1638 case Op_ConP:
1639 case Op_ConN:
1640 case Op_ConNKlass: {
1641 // assume all oop constants globally escape except for null
1642 PointsToNode::EscapeState es;
1643 const Type* t = igvn->type(n);
1644 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1645 es = PointsToNode::NoEscape;
1646 } else {
1647 es = PointsToNode::GlobalEscape;
1648 }
1649 PointsToNode* ptn_con = add_java_object(n, es);
1650 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1651 break;
1652 }
1653 case Op_CreateEx: {
1654 // assume that all exception objects globally escape
1655 map_ideal_node(n, phantom_obj);
1656 break;
1657 }
1658 case Op_LoadKlass:
1659 case Op_LoadNKlass: {
1660 // Unknown class is loaded
1661 map_ideal_node(n, phantom_obj);
1662 break;
1663 }
1664 case Op_LoadP:
1665 case Op_LoadN: {
1666 add_objload_to_connection_graph(n, delayed_worklist);
1667 break;
1668 }
1669 case Op_Parm: {
1670 map_ideal_node(n, phantom_obj);
1671 break;
1672 }
1673 case Op_PartialSubtypeCheck: {
1674 // Produces Null or notNull and is used in only in CmpP so
1675 // phantom_obj could be used.
1676 map_ideal_node(n, phantom_obj); // Result is unknown
1677 break;
1678 }
1679 case Op_Phi: {
1680 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1681 // ThreadLocal has RawPtr type.
1682 const Type* t = n->as_Phi()->type();
1683 if (t->make_ptr() != nullptr) {
1684 add_local_var(n, PointsToNode::NoEscape);
1685 // Do not add edges during first iteration because some could be
1686 // not defined yet.
1687 delayed_worklist->push(n);
1688 }
1689 break;
1690 }
1691 case Op_Proj: {
1692 // we are only interested in the oop result projection from a call
1693 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&1694 n->in(0)->as_Call()->returns_pointer()) {
1695 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1696 }
1697 break;
1698 }
1699 case Op_Rethrow: // Exception object escapes
1700 case Op_Return: {
1701 if (n->req() > TypeFunc::Parms &&
1702 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1703 // Treat Return value as LocalVar with GlobalEscape escape state.
1704 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1705 }
1706 break;
1707 }
1708 case Op_CompareAndExchangeP:
1709 case Op_CompareAndExchangeN:
1710 case Op_GetAndSetP:
1711 case Op_GetAndSetN: {
1712 add_objload_to_connection_graph(n, delayed_worklist);
1713 // fall-through
1714 }
1715 case Op_StoreP:
1716 case Op_StoreN:
1717 case Op_StoreNKlass:
1718 case Op_WeakCompareAndSwapP:
1719 case Op_WeakCompareAndSwapN:
1720 case Op_CompareAndSwapP:
1721 case Op_CompareAndSwapN: {
1722 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1723 break;
1724 }
1725 case Op_AryEq:
1726 case Op_CountPositives:
1727 case Op_StrComp:
1728 case Op_StrEquals:
1729 case Op_StrIndexOf:
1730 case Op_StrIndexOfChar:
1731 case Op_StrInflatedCopy:
1732 case Op_StrCompressedCopy:
1733 case Op_VectorizedHashCode:
1734 case Op_EncodeISOArray: {
1735 add_local_var(n, PointsToNode::ArgEscape);
1736 delayed_worklist->push(n); // Process it later.
1737 break;
1738 }
1739 case Op_ThreadLocal: {
1740 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1741 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1742 break;
1743 }
1744 case Op_Blackhole: {
1745 // All blackhole pointer arguments are globally escaping.
1746 // Only do this if there is at least one pointer argument.
1747 // Do not add edges during first iteration because some could be
1748 // not defined yet, defer to final step.
1749 for (uint i = 0; i < n->req(); i++) {
1750 Node* in = n->in(i);
1751 if (in != nullptr) {
1752 const Type* at = _igvn->type(in);
1753 if (!at->isa_ptr()) continue;
1754
1755 add_local_var(n, PointsToNode::GlobalEscape);
1756 delayed_worklist->push(n);
1757 break;
1758 }
1759 }
1760 break;
1761 }
1762 default:
1763 ; // Do nothing for nodes not related to EA.
1764 }
1765 return;
1766 }
1767
1768 // Add final simple edges to graph.
1769 void ConnectionGraph::add_final_edges(Node *n) {
1770 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1771 #ifdef ASSERT
1772 if (_verify && n_ptn->is_JavaObject())
1773 return; // This method does not change graph for JavaObject.
1774 #endif
1775
1776 if (n->is_Call()) {
1777 process_call_arguments(n->as_Call());
1778 return;
1779 }
1780 assert(n->is_Store() || n->is_LoadStore() ||
1781 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1782 "node should be registered already");
1783 int opcode = n->Opcode();
1784 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1785 if (gc_handled) {
1786 return; // Ignore node if already handled by GC.
1787 }
1788 switch (opcode) {
1789 case Op_AddP: {
1790 Node* base = get_addp_base(n);
1791 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1792 assert(ptn_base != nullptr, "field's base should be registered");
1793 add_base(n_ptn->as_Field(), ptn_base);
1794 break;
1795 }
1796 case Op_CastPP:
1797 case Op_CheckCastPP:
1798 case Op_EncodeP:
1799 case Op_DecodeN:
1800 case Op_EncodePKlass:
1801 case Op_DecodeNKlass: {
1802 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1803 break;
1804 }
1805 case Op_CMoveP: {
1806 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1807 Node* in = n->in(i);
1808 if (in == nullptr) {
1809 continue; // ignore null
1810 }
1811 Node* uncast_in = in->uncast();
1812 if (uncast_in->is_top() || uncast_in == n) {
1813 continue; // ignore top or inputs which go back this node
1814 }
1815 PointsToNode* ptn = ptnode_adr(in->_idx);
1816 assert(ptn != nullptr, "node should be registered");
1817 add_edge(n_ptn, ptn);
1818 }
1819 break;
1820 }
1821 case Op_LoadP:
1822 case Op_LoadN: {
1823 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1824 // ThreadLocal has RawPtr type.
1825 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1826 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1827 break;
1828 }
1829 case Op_Phi: {
1830 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1831 // ThreadLocal has RawPtr type.
1832 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1833 for (uint i = 1; i < n->req(); i++) {
1834 Node* in = n->in(i);
1835 if (in == nullptr) {
1836 continue; // ignore null
1837 }
1838 Node* uncast_in = in->uncast();
1839 if (uncast_in->is_top() || uncast_in == n) {
1840 continue; // ignore top or inputs which go back this node
1841 }
1842 PointsToNode* ptn = ptnode_adr(in->_idx);
1843 assert(ptn != nullptr, "node should be registered");
1844 add_edge(n_ptn, ptn);
1845 }
1846 break;
1847 }
1848 case Op_Proj: {
1849 // we are only interested in the oop result projection from a call1850 assert(n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&1851 n->in(0)->as_Call()->returns_pointer(), "Unexpected node type");1852 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1853 break;
1854 }
1855 case Op_Rethrow: // Exception object escapes
1856 case Op_Return: {
1857 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1858 "Unexpected node type");
1859 // Treat Return value as LocalVar with GlobalEscape escape state.
1860 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1861 break;
1862 }
1863 case Op_CompareAndExchangeP:
1864 case Op_CompareAndExchangeN:
1865 case Op_GetAndSetP:
1866 case Op_GetAndSetN:{
1867 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1868 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1869 // fall-through
1870 }
1871 case Op_CompareAndSwapP:
1872 case Op_CompareAndSwapN:
1873 case Op_WeakCompareAndSwapP:
1874 case Op_WeakCompareAndSwapN:
1875 case Op_StoreP:
1876 case Op_StoreN:
1877 case Op_StoreNKlass:{
1878 add_final_edges_unsafe_access(n, opcode);
1879 break;
1880 }
1881 case Op_VectorizedHashCode:
1882 case Op_AryEq:
1883 case Op_CountPositives:
1884 case Op_StrComp:
1885 case Op_StrEquals:
1886 case Op_StrIndexOf:
1887 case Op_StrIndexOfChar:
1888 case Op_StrInflatedCopy:
1889 case Op_StrCompressedCopy:
1890 case Op_EncodeISOArray: {
1891 // char[]/byte[] arrays passed to string intrinsic do not escape but
1892 // they are not scalar replaceable. Adjust escape state for them.
1893 // Start from in(2) edge since in(1) is memory edge.
1894 for (uint i = 2; i < n->req(); i++) {
1895 Node* adr = n->in(i);
1896 const Type* at = _igvn->type(adr);
1897 if (!adr->is_top() && at->isa_ptr()) {
1898 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1899 at->isa_ptr() != nullptr, "expecting a pointer");
1900 if (adr->is_AddP()) {
1901 adr = get_addp_base(adr);
1902 }
1903 PointsToNode* ptn = ptnode_adr(adr->_idx);
1904 assert(ptn != nullptr, "node should be registered");
1905 add_edge(n_ptn, ptn);
1906 }
1907 }
1908 break;
1909 }
1910 case Op_Blackhole: {
1911 // All blackhole pointer arguments are globally escaping.
1912 for (uint i = 0; i < n->req(); i++) {
1913 Node* in = n->in(i);
1914 if (in != nullptr) {
1915 const Type* at = _igvn->type(in);
1916 if (!at->isa_ptr()) continue;
1917
1918 if (in->is_AddP()) {
1919 in = get_addp_base(in);
1920 }
1921
1922 PointsToNode* ptn = ptnode_adr(in->_idx);
1923 assert(ptn != nullptr, "should be defined already");
1924 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1925 add_edge(n_ptn, ptn);
1926 }
1927 }
1928 break;
1929 }
1930 default: {
1931 // This method should be called only for EA specific nodes which may
1932 // miss some edges when they were created.
1933 #ifdef ASSERT
1934 n->dump(1);
1935 #endif
1936 guarantee(false, "unknown node");
1937 }
1938 }
1939 return;
1940 }
1941
1942 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1943 Node* adr = n->in(MemNode::Address);
1944 const Type* adr_type = _igvn->type(adr);
1945 adr_type = adr_type->make_ptr();
1946 if (adr_type == nullptr) {
1947 return; // skip dead nodes
1948 }
1949 if (adr_type->isa_oopptr()
1950 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1951 && adr_type == TypeRawPtr::NOTNULL
1952 && is_captured_store_address(adr))) {
1953 delayed_worklist->push(n); // Process it later.
1954 #ifdef ASSERT
1955 assert (adr->is_AddP(), "expecting an AddP");
1956 if (adr_type == TypeRawPtr::NOTNULL) {
1957 // Verify a raw address for a store captured by Initialize node.
1958 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1959 assert(offs != Type::OffsetBot, "offset must be a constant");
1960 }
1961 #endif
1962 } else {
1963 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1964 if (adr->is_BoxLock()) {
1965 return;
1966 }
1967 // Stored value escapes in unsafe access.
1968 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1969 delayed_worklist->push(n); // Process unsafe access later.
1970 return;
1971 }
1972 #ifdef ASSERT
1973 n->dump(1);
1974 assert(false, "not unsafe");
1975 #endif
1976 }
1977 }
1978
1979 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1980 Node* adr = n->in(MemNode::Address);
1981 const Type *adr_type = _igvn->type(adr);
1982 adr_type = adr_type->make_ptr();
1983 #ifdef ASSERT
1984 if (adr_type == nullptr) {
1985 n->dump(1);
1986 assert(adr_type != nullptr, "dead node should not be on list");
1987 return true;
1988 }
1989 #endif
1990
1991 if (adr_type->isa_oopptr()
1992 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1993 && adr_type == TypeRawPtr::NOTNULL
1994 && is_captured_store_address(adr))) {
1995 // Point Address to Value
1996 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1997 assert(adr_ptn != nullptr &&
1998 adr_ptn->as_Field()->is_oop(), "node should be registered");
1999 Node* val = n->in(MemNode::ValueIn);
2000 PointsToNode* ptn = ptnode_adr(val->_idx);
2001 assert(ptn != nullptr, "node should be registered");
2002 add_edge(adr_ptn, ptn);
2003 return true;
2004 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
2005 // Stored value escapes in unsafe access.
2006 Node* val = n->in(MemNode::ValueIn);
2007 PointsToNode* ptn = ptnode_adr(val->_idx);
2008 assert(ptn != nullptr, "node should be registered");
2009 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2010 // Add edge to object for unsafe access with offset.
2011 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2012 assert(adr_ptn != nullptr, "node should be registered");
2013 if (adr_ptn->is_Field()) {
2014 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2015 add_edge(adr_ptn, ptn);
2016 }
2017 return true;
2018 }
2019 #ifdef ASSERT
2020 n->dump(1);
2021 assert(false, "not unsafe");
2022 #endif
2023 return false;
2024 }
2025
2026 void ConnectionGraph::add_call_node(CallNode* call) {
2027 assert(call->returns_pointer(), "only for call which returns pointer");
2028 uint call_idx = call->_idx;
2029 if (call->is_Allocate()) {
2030 Node* k = call->in(AllocateNode::KlassNode);
2031 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2032 assert(kt != nullptr, "TypeKlassPtr required.");
2033 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2034 bool scalar_replaceable = true;
2035 NOT_PRODUCT(const char* nsr_reason = "");
2036 if (call->is_AllocateArray()) {
2037 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2038 es = PointsToNode::GlobalEscape;
2039 } else {
2040 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2041 if (length < 0) {
2042 // Not scalar replaceable if the length is not constant.
2043 scalar_replaceable = false;
2044 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2045 } else if (length > EliminateAllocationArraySizeLimit) {
2046 // Not scalar replaceable if the length is too big.
2047 scalar_replaceable = false;
2048 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2049 }
2050 }
2051 } else { // Allocate instance
2052 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2053 es = PointsToNode::GlobalEscape;
2054 } else {
2055 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2056 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2057 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2058 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2059 !ik->can_be_instantiated() ||
2060 ik->has_finalizer()) {
2061 es = PointsToNode::GlobalEscape;
2062 } else {
2063 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2064 if (nfields > EliminateAllocationFieldsLimit) {
2065 // Not scalar replaceable if there are too many fields.
2066 scalar_replaceable = false;
2067 NOT_PRODUCT(nsr_reason = "has too many fields");
2068 }
2069 }
2070 }
2071 }
2072 add_java_object(call, es);
2073 PointsToNode* ptn = ptnode_adr(call_idx);
2074 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2075 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2076 }
2077 } else if (call->is_CallStaticJava()) {
2078 // Call nodes could be different types:
2079 //
2080 // 1. CallDynamicJavaNode (what happened during call is unknown):
2081 //
2082 // - mapped to GlobalEscape JavaObject node if oop is returned;
2083 //
2084 // - all oop arguments are escaping globally;
2085 //
2086 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2087 //
2088 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2089 //
2090 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2091 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2092 // during call is returned;
2093 // - mapped to ArgEscape LocalVar node pointed to object arguments
2094 // which are returned and does not escape during call;
2095 //
2096 // - oop arguments escaping status is defined by bytecode analysis;
2097 //
2098 // For a static call, we know exactly what method is being called.
2099 // Use bytecode estimator to record whether the call's return value escapes.
2100 ciMethod* meth = call->as_CallJava()->method();
2101 if (meth == nullptr) {
2102 assert(call->as_CallStaticJava()->is_call_to_multianewarray_stub(), "TODO: add failed case check");
2103 // Returns a newly allocated non-escaped object.
2104 add_java_object(call, PointsToNode::NoEscape);
2105 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2106 } else if (meth->is_boxing_method()) {
2107 // Returns boxing object
2108 PointsToNode::EscapeState es;
2109 vmIntrinsics::ID intr = meth->intrinsic_id();
2110 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2111 // It does not escape if object is always allocated.
2112 es = PointsToNode::NoEscape;
2113 } else {
2114 // It escapes globally if object could be loaded from cache.
2115 es = PointsToNode::GlobalEscape;
2116 }
2117 add_java_object(call, es);
2118 if (es == PointsToNode::GlobalEscape) {
2119 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2120 }
2121 } else {
2122 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2123 call_analyzer->copy_dependencies(_compile->dependencies());
2124 if (call_analyzer->is_return_allocated()) {
2125 // Returns a newly allocated non-escaped object, simply
2126 // update dependency information.
2127 // Mark it as NoEscape so that objects referenced by
2128 // it's fields will be marked as NoEscape at least.
2129 add_java_object(call, PointsToNode::NoEscape);
2130 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2131 } else {
2132 // Determine whether any arguments are returned.
2133 const TypeTuple* d = call->tf()->domain();
2134 bool ret_arg = false;
2135 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2136 if (d->field_at(i)->isa_ptr() != nullptr &&
2137 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2138 ret_arg = true;
2139 break;
2140 }
2141 }
2142 if (ret_arg) {
2143 add_local_var(call, PointsToNode::ArgEscape);
2144 } else {
2145 // Returns unknown object.
2146 map_ideal_node(call, phantom_obj);
2147 }
2148 }
2149 }
2150 } else {
2151 // An other type of call, assume the worst case:
2152 // returned value is unknown and globally escapes.
2153 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2154 map_ideal_node(call, phantom_obj);
2155 }
2156 }
2157
2158 void ConnectionGraph::process_call_arguments(CallNode *call) {
2159 bool is_arraycopy = false;
2160 switch (call->Opcode()) {
2161 #ifdef ASSERT
2162 case Op_Allocate:
2163 case Op_AllocateArray:
2164 case Op_Lock:
2165 case Op_Unlock:
2166 assert(false, "should be done already");
2167 break;
2168 #endif
2169 case Op_ArrayCopy:
2170 case Op_CallLeafNoFP:
2171 // Most array copies are ArrayCopy nodes at this point but there
2172 // are still a few direct calls to the copy subroutines (See
2173 // PhaseStringOpts::copy_string())
2174 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2175 call->as_CallLeaf()->is_call_to_arraycopystub();
2176 // fall through
2177 case Op_CallLeafVector:
2178 case Op_CallLeaf: {
2179 // Stub calls, objects do not escape but they are not scale replaceable.
2180 // Adjust escape state for outgoing arguments.
2181 const TypeTuple * d = call->tf()->domain();
2182 bool src_has_oops = false;
2183 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2184 const Type* at = d->field_at(i);
2185 Node *arg = call->in(i);
2186 if (arg == nullptr) {
2187 continue;
2188 }
2189 const Type *aat = _igvn->type(arg);
2190 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2191 continue;
2192 }
2193 if (arg->is_AddP()) {
2194 //
2195 // The inline_native_clone() case when the arraycopy stub is called
2196 // after the allocation before Initialize and CheckCastPP nodes.
2197 // Or normal arraycopy for object arrays case.
2198 //
2199 // Set AddP's base (Allocate) as not scalar replaceable since
2200 // pointer to the base (with offset) is passed as argument.
2201 //
2202 arg = get_addp_base(arg);
2203 }
2204 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2205 assert(arg_ptn != nullptr, "should be registered");
2206 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2207 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2208 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2209 aat->isa_ptr() != nullptr, "expecting an Ptr");
2210 bool arg_has_oops = aat->isa_oopptr() &&
2211 (aat->isa_instptr() ||
2212 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)));
2213 if (i == TypeFunc::Parms) {
2214 src_has_oops = arg_has_oops;
2215 }
2216 //
2217 // src or dst could be j.l.Object when other is basic type array:
2218 //
2219 // arraycopy(char[],0,Object*,0,size);
2220 // arraycopy(Object*,0,char[],0,size);
2221 //
2222 // Don't add edges in such cases.
2223 //
2224 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2225 arg_has_oops && (i > TypeFunc::Parms);
2226 #ifdef ASSERT
2227 if (!(is_arraycopy ||
2228 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2229 (call->as_CallLeaf()->_name != nullptr &&
2230 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2231 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2232 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2233 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2234 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2235 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2236 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2237 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2238 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2239 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2240 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2241 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2242 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2243 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2244 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2245 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2246 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2247 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2248 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2249 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2250 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2251 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2252 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2253 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2254 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2255 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2256 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2257 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2258 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2259 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2260 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2261 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2262 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2263 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2264 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2265 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2266 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2267 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2268 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2269 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2270 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2271 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2272 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2273 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2274 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2275 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2276 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2277 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2278 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2279 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2280 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2281 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2282 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2283 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2284 ))) {
2285 call->dump();
2286 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2287 }
2288 #endif
2289 // Always process arraycopy's destination object since
2290 // we need to add all possible edges to references in
2291 // source object.
2292 if (arg_esc >= PointsToNode::ArgEscape &&
2293 !arg_is_arraycopy_dest) {
2294 continue;
2295 }
2296 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2297 if (call->is_ArrayCopy()) {
2298 ArrayCopyNode* ac = call->as_ArrayCopy();
2299 if (ac->is_clonebasic() ||
2300 ac->is_arraycopy_validated() ||
2301 ac->is_copyof_validated() ||
2302 ac->is_copyofrange_validated()) {
2303 es = PointsToNode::NoEscape;
2304 }
2305 }
2306 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2307 if (arg_is_arraycopy_dest) {
2308 Node* src = call->in(TypeFunc::Parms);
2309 if (src->is_AddP()) {
2310 src = get_addp_base(src);
2311 }
2312 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2313 assert(src_ptn != nullptr, "should be registered");
2314 // Special arraycopy edge:
2315 // Only escape state of destination object's fields affects
2316 // escape state of fields in source object.
2317 add_arraycopy(call, es, src_ptn, arg_ptn);
2318 }
2319 }
2320 }
2321 break;
2322 }
2323 case Op_CallStaticJava: {
2324 // For a static call, we know exactly what method is being called.
2325 // Use bytecode estimator to record the call's escape affects
2326 #ifdef ASSERT
2327 const char* name = call->as_CallStaticJava()->_name;
2328 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2329 #endif
2330 ciMethod* meth = call->as_CallJava()->method();
2331 if ((meth != nullptr) && meth->is_boxing_method()) {
2332 break; // Boxing methods do not modify any oops.
2333 }
2334 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2335 // fall-through if not a Java method or no analyzer information
2336 if (call_analyzer != nullptr) {
2337 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2338 const TypeTuple* d = call->tf()->domain();
2339 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2340 const Type* at = d->field_at(i);
2341 int k = i - TypeFunc::Parms;
2342 Node* arg = call->in(i);
2343 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2344 if (at->isa_ptr() != nullptr &&
2345 call_analyzer->is_arg_returned(k)) {
2346 // The call returns arguments.
2347 if (call_ptn != nullptr) { // Is call's result used?
2348 assert(call_ptn->is_LocalVar(), "node should be registered");
2349 assert(arg_ptn != nullptr, "node should be registered");
2350 add_edge(call_ptn, arg_ptn);
2351 }
2352 }
2353 if (at->isa_oopptr() != nullptr &&
2354 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2355 if (!call_analyzer->is_arg_stack(k)) {
2356 // The argument global escapes
2357 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2358 } else {
2359 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2360 if (!call_analyzer->is_arg_local(k)) {
2361 // The argument itself doesn't escape, but any fields might
2362 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2363 }
2364 }
2365 }
2366 }
2367 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2368 // The call returns arguments.
2369 assert(call_ptn->edge_count() > 0, "sanity");
2370 if (!call_analyzer->is_return_local()) {
2371 // Returns also unknown object.
2372 add_edge(call_ptn, phantom_obj);
2373 }
2374 }
2375 break;
2376 }
2377 }
2378 default: {
2379 // Fall-through here if not a Java method or no analyzer information
2380 // or some other type of call, assume the worst case: all arguments
2381 // globally escape.
2382 const TypeTuple* d = call->tf()->domain();
2383 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2384 const Type* at = d->field_at(i);
2385 if (at->isa_oopptr() != nullptr) {
2386 Node* arg = call->in(i);
2387 if (arg->is_AddP()) {
2388 arg = get_addp_base(arg);
2389 }
2390 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2391 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2392 }
2393 }
2394 }
2395 }
2396 }
2397
2398
2399 // Finish Graph construction.
2400 bool ConnectionGraph::complete_connection_graph(
2401 GrowableArray<PointsToNode*>& ptnodes_worklist,
2402 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2403 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2404 GrowableArray<FieldNode*>& oop_fields_worklist) {
2405 // Normally only 1-3 passes needed to build Connection Graph depending
2406 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2407 // Set limit to 20 to catch situation when something did go wrong and
2408 // bailout Escape Analysis.
2409 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2410 #define GRAPH_BUILD_ITER_LIMIT 20
2411
2412 // Propagate GlobalEscape and ArgEscape escape states and check that
2413 // we still have non-escaping objects. The method pushs on _worklist
2414 // Field nodes which reference phantom_object.
2415 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2416 return false; // Nothing to do.
2417 }
2418 // Now propagate references to all JavaObject nodes.
2419 int java_objects_length = java_objects_worklist.length();
2420 elapsedTimer build_time;
2421 build_time.start();
2422 elapsedTimer time;
2423 bool timeout = false;
2424 int new_edges = 1;
2425 int iterations = 0;
2426 do {
2427 while ((new_edges > 0) &&
2428 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2429 double start_time = time.seconds();
2430 time.start();
2431 new_edges = 0;
2432 // Propagate references to phantom_object for nodes pushed on _worklist
2433 // by find_non_escaped_objects() and find_field_value().
2434 new_edges += add_java_object_edges(phantom_obj, false);
2435 for (int next = 0; next < java_objects_length; ++next) {
2436 JavaObjectNode* ptn = java_objects_worklist.at(next);
2437 new_edges += add_java_object_edges(ptn, true);
2438
2439 #define SAMPLE_SIZE 4
2440 if ((next % SAMPLE_SIZE) == 0) {
2441 // Each 4 iterations calculate how much time it will take
2442 // to complete graph construction.
2443 time.stop();
2444 // Poll for requests from shutdown mechanism to quiesce compiler
2445 // because Connection graph construction may take long time.
2446 CompileBroker::maybe_block();
2447 double stop_time = time.seconds();
2448 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2449 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2450 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2451 timeout = true;
2452 break; // Timeout
2453 }
2454 start_time = stop_time;
2455 time.start();
2456 }
2457 #undef SAMPLE_SIZE
2458
2459 }
2460 if (timeout) break;
2461 if (new_edges > 0) {
2462 // Update escape states on each iteration if graph was updated.
2463 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2464 return false; // Nothing to do.
2465 }
2466 }
2467 time.stop();
2468 if (time.seconds() >= EscapeAnalysisTimeout) {
2469 timeout = true;
2470 break;
2471 }
2472 _compile->print_method(PHASE_EA_COMPLETE_CONNECTION_GRAPH_ITER, 5);
2473 }
2474 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2475 time.start();
2476 // Find fields which have unknown value.
2477 int fields_length = oop_fields_worklist.length();
2478 for (int next = 0; next < fields_length; next++) {
2479 FieldNode* field = oop_fields_worklist.at(next);
2480 if (field->edge_count() == 0) {
2481 new_edges += find_field_value(field);
2482 // This code may added new edges to phantom_object.
2483 // Need an other cycle to propagate references to phantom_object.
2484 }
2485 }
2486 time.stop();
2487 if (time.seconds() >= EscapeAnalysisTimeout) {
2488 timeout = true;
2489 break;
2490 }
2491 } else {
2492 new_edges = 0; // Bailout
2493 }
2494 } while (new_edges > 0);
2495
2496 build_time.stop();
2497 _build_time = build_time.seconds();
2498 _build_iterations = iterations;
2499
2500 // Bailout if passed limits.
2501 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2502 Compile* C = _compile;
2503 if (C->log() != nullptr) {
2504 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2505 C->log()->text("%s", timeout ? "time" : "iterations");
2506 C->log()->end_elem(" limit'");
2507 }
2508 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2509 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2510 // Possible infinite build_connection_graph loop,
2511 // bailout (no changes to ideal graph were made).
2512 return false;
2513 }
2514
2515 #undef GRAPH_BUILD_ITER_LIMIT
2516
2517 // Find fields initialized by null for non-escaping Allocations.
2518 int non_escaped_length = non_escaped_allocs_worklist.length();
2519 for (int next = 0; next < non_escaped_length; next++) {
2520 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2521 PointsToNode::EscapeState es = ptn->escape_state();
2522 assert(es <= PointsToNode::ArgEscape, "sanity");
2523 if (es == PointsToNode::NoEscape) {
2524 if (find_init_values_null(ptn, _igvn) > 0) {
2525 // Adding references to null object does not change escape states
2526 // since it does not escape. Also no fields are added to null object.
2527 add_java_object_edges(null_obj, false);
2528 }
2529 }
2530 Node* n = ptn->ideal_node();
2531 if (n->is_Allocate()) {
2532 // The object allocated by this Allocate node will never be
2533 // seen by an other thread. Mark it so that when it is
2534 // expanded no MemBarStoreStore is added.
2535 InitializeNode* ini = n->as_Allocate()->initialization();
2536 if (ini != nullptr)
2537 ini->set_does_not_escape();
2538 }
2539 }
2540 return true; // Finished graph construction.
2541 }
2542
2543 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2544 // and check that we still have non-escaping java objects.
2545 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2546 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2547 bool print_method) {
2548 GrowableArray<PointsToNode*> escape_worklist;
2549 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2550 int ptnodes_length = ptnodes_worklist.length();
2551 for (int next = 0; next < ptnodes_length; ++next) {
2552 PointsToNode* ptn = ptnodes_worklist.at(next);
2553 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2554 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2555 escape_worklist.push(ptn);
2556 }
2557 }
2558 // Set escape states to referenced nodes (edges list).
2559 while (escape_worklist.length() > 0) {
2560 PointsToNode* ptn = escape_worklist.pop();
2561 PointsToNode::EscapeState es = ptn->escape_state();
2562 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2563 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2564 es >= PointsToNode::ArgEscape) {
2565 // GlobalEscape or ArgEscape state of field means it has unknown value.
2566 if (add_edge(ptn, phantom_obj)) {
2567 // New edge was added
2568 add_field_uses_to_worklist(ptn->as_Field());
2569 }
2570 }
2571 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2572 PointsToNode* e = i.get();
2573 if (e->is_Arraycopy()) {
2574 assert(ptn->arraycopy_dst(), "sanity");
2575 // Propagate only fields escape state through arraycopy edge.
2576 if (e->fields_escape_state() < field_es) {
2577 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2578 escape_worklist.push(e);
2579 }
2580 } else if (es >= field_es) {
2581 // fields_escape_state is also set to 'es' if it is less than 'es'.
2582 if (e->escape_state() < es) {
2583 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2584 escape_worklist.push(e);
2585 }
2586 } else {
2587 // Propagate field escape state.
2588 bool es_changed = false;
2589 if (e->fields_escape_state() < field_es) {
2590 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2591 es_changed = true;
2592 }
2593 if ((e->escape_state() < field_es) &&
2594 e->is_Field() && ptn->is_JavaObject() &&
2595 e->as_Field()->is_oop()) {
2596 // Change escape state of referenced fields.
2597 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2598 es_changed = true;
2599 } else if (e->escape_state() < es) {
2600 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2601 es_changed = true;
2602 }
2603 if (es_changed) {
2604 escape_worklist.push(e);
2605 }
2606 }
2607 if (print_method) {
2608 _compile->print_method(PHASE_EA_CONNECTION_GRAPH_PROPAGATE_ITER, 6, e->ideal_node());
2609 }
2610 }
2611 }
2612 // Remove escaped objects from non_escaped list.
2613 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2614 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2615 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2616 non_escaped_allocs_worklist.delete_at(next);
2617 }
2618 if (ptn->escape_state() == PointsToNode::NoEscape) {
2619 // Find fields in non-escaped allocations which have unknown value.
2620 find_init_values_phantom(ptn);
2621 }
2622 }
2623 return (non_escaped_allocs_worklist.length() > 0);
2624 }
2625
2626 // Add all references to JavaObject node by walking over all uses.
2627 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2628 int new_edges = 0;
2629 if (populate_worklist) {
2630 // Populate _worklist by uses of jobj's uses.
2631 for (UseIterator i(jobj); i.has_next(); i.next()) {
2632 PointsToNode* use = i.get();
2633 if (use->is_Arraycopy()) {
2634 continue;
2635 }
2636 add_uses_to_worklist(use);
2637 if (use->is_Field() && use->as_Field()->is_oop()) {
2638 // Put on worklist all field's uses (loads) and
2639 // related field nodes (same base and offset).
2640 add_field_uses_to_worklist(use->as_Field());
2641 }
2642 }
2643 }
2644 for (int l = 0; l < _worklist.length(); l++) {
2645 PointsToNode* use = _worklist.at(l);
2646 if (PointsToNode::is_base_use(use)) {
2647 // Add reference from jobj to field and from field to jobj (field's base).
2648 use = PointsToNode::get_use_node(use)->as_Field();
2649 if (add_base(use->as_Field(), jobj)) {
2650 new_edges++;
2651 }
2652 continue;
2653 }
2654 assert(!use->is_JavaObject(), "sanity");
2655 if (use->is_Arraycopy()) {
2656 if (jobj == null_obj) { // null object does not have field edges
2657 continue;
2658 }
2659 // Added edge from Arraycopy node to arraycopy's source java object
2660 if (add_edge(use, jobj)) {
2661 jobj->set_arraycopy_src();
2662 new_edges++;
2663 }
2664 // and stop here.
2665 continue;
2666 }
2667 if (!add_edge(use, jobj)) {
2668 continue; // No new edge added, there was such edge already.
2669 }
2670 new_edges++;
2671 if (use->is_LocalVar()) {
2672 add_uses_to_worklist(use);
2673 if (use->arraycopy_dst()) {
2674 for (EdgeIterator i(use); i.has_next(); i.next()) {
2675 PointsToNode* e = i.get();
2676 if (e->is_Arraycopy()) {
2677 if (jobj == null_obj) { // null object does not have field edges
2678 continue;
2679 }
2680 // Add edge from arraycopy's destination java object to Arraycopy node.
2681 if (add_edge(jobj, e)) {
2682 new_edges++;
2683 jobj->set_arraycopy_dst();
2684 }
2685 }
2686 }
2687 }
2688 } else {
2689 // Added new edge to stored in field values.
2690 // Put on worklist all field's uses (loads) and
2691 // related field nodes (same base and offset).
2692 add_field_uses_to_worklist(use->as_Field());
2693 }
2694 }
2695 _worklist.clear();
2696 _in_worklist.reset();
2697 return new_edges;
2698 }
2699
2700 // Put on worklist all related field nodes.
2701 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2702 assert(field->is_oop(), "sanity");
2703 int offset = field->offset();
2704 add_uses_to_worklist(field);
2705 // Loop over all bases of this field and push on worklist Field nodes
2706 // with the same offset and base (since they may reference the same field).
2707 for (BaseIterator i(field); i.has_next(); i.next()) {
2708 PointsToNode* base = i.get();
2709 add_fields_to_worklist(field, base);
2710 // Check if the base was source object of arraycopy and go over arraycopy's
2711 // destination objects since values stored to a field of source object are
2712 // accessible by uses (loads) of fields of destination objects.
2713 if (base->arraycopy_src()) {
2714 for (UseIterator j(base); j.has_next(); j.next()) {
2715 PointsToNode* arycp = j.get();
2716 if (arycp->is_Arraycopy()) {
2717 for (UseIterator k(arycp); k.has_next(); k.next()) {
2718 PointsToNode* abase = k.get();
2719 if (abase->arraycopy_dst() && abase != base) {
2720 // Look for the same arraycopy reference.
2721 add_fields_to_worklist(field, abase);
2722 }
2723 }
2724 }
2725 }
2726 }
2727 }
2728 }
2729
2730 // Put on worklist all related field nodes.
2731 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2732 int offset = field->offset();
2733 if (base->is_LocalVar()) {
2734 for (UseIterator j(base); j.has_next(); j.next()) {
2735 PointsToNode* f = j.get();
2736 if (PointsToNode::is_base_use(f)) { // Field
2737 f = PointsToNode::get_use_node(f);
2738 if (f == field || !f->as_Field()->is_oop()) {
2739 continue;
2740 }
2741 int offs = f->as_Field()->offset();
2742 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2743 add_to_worklist(f);
2744 }
2745 }
2746 }
2747 } else {
2748 assert(base->is_JavaObject(), "sanity");
2749 if (// Skip phantom_object since it is only used to indicate that
2750 // this field's content globally escapes.
2751 (base != phantom_obj) &&
2752 // null object node does not have fields.
2753 (base != null_obj)) {
2754 for (EdgeIterator i(base); i.has_next(); i.next()) {
2755 PointsToNode* f = i.get();
2756 // Skip arraycopy edge since store to destination object field
2757 // does not update value in source object field.
2758 if (f->is_Arraycopy()) {
2759 assert(base->arraycopy_dst(), "sanity");
2760 continue;
2761 }
2762 if (f == field || !f->as_Field()->is_oop()) {
2763 continue;
2764 }
2765 int offs = f->as_Field()->offset();
2766 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2767 add_to_worklist(f);
2768 }
2769 }
2770 }
2771 }
2772 }
2773
2774 // Find fields which have unknown value.
2775 int ConnectionGraph::find_field_value(FieldNode* field) {
2776 // Escaped fields should have init value already.
2777 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2778 int new_edges = 0;
2779 for (BaseIterator i(field); i.has_next(); i.next()) {
2780 PointsToNode* base = i.get();
2781 if (base->is_JavaObject()) {
2782 // Skip Allocate's fields which will be processed later.
2783 if (base->ideal_node()->is_Allocate()) {
2784 return 0;
2785 }
2786 assert(base == null_obj, "only null ptr base expected here");
2787 }
2788 }
2789 if (add_edge(field, phantom_obj)) {
2790 // New edge was added
2791 new_edges++;
2792 add_field_uses_to_worklist(field);
2793 }
2794 return new_edges;
2795 }
2796
2797 // Find fields initializing values for allocations.
2798 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2799 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2800 Node* alloc = pta->ideal_node();
2801
2802 // Do nothing for Allocate nodes since its fields values are
2803 // "known" unless they are initialized by arraycopy/clone.
2804 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2805 return 0;
2806 }
2807 assert(pta->arraycopy_dst() || alloc->as_CallStaticJava(), "sanity");
2808 #ifdef ASSERT
2809 if (!pta->arraycopy_dst() && alloc->as_CallStaticJava()->method() == nullptr) {2810 assert(alloc->as_CallStaticJava()->is_call_to_multianewarray_stub(), "sanity");
2811 }
2812 #endif
2813 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2814 int new_edges = 0;
2815 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2816 PointsToNode* field = i.get();
2817 if (field->is_Field() && field->as_Field()->is_oop()) {
2818 if (add_edge(field, phantom_obj)) {
2819 // New edge was added
2820 new_edges++;
2821 add_field_uses_to_worklist(field->as_Field());
2822 }
2823 }
2824 }
2825 return new_edges;
2826 }
2827
2828 // Find fields initializing values for allocations.
2829 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2830 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2831 Node* alloc = pta->ideal_node();
2832 // Do nothing for Call nodes since its fields values are unknown.
2833 if (!alloc->is_Allocate()) {
2834 return 0;
2835 }
2836 InitializeNode* ini = alloc->as_Allocate()->initialization();
2837 bool visited_bottom_offset = false;
2838 GrowableArray<int> offsets_worklist;
2839 int new_edges = 0;
2840
2841 // Check if an oop field's initializing value is recorded and add
2842 // a corresponding null if field's value if it is not recorded.
2843 // Connection Graph does not record a default initialization by null
2844 // captured by Initialize node.
2845 //
2846 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2847 PointsToNode* field = i.get(); // Field (AddP)
2848 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2849 continue; // Not oop field
2850 }
2851 int offset = field->as_Field()->offset();
2852 if (offset == Type::OffsetBot) {
2853 if (!visited_bottom_offset) {
2854 // OffsetBot is used to reference array's element,
2855 // always add reference to null to all Field nodes since we don't
2856 // known which element is referenced.
2857 if (add_edge(field, null_obj)) {
2858 // New edge was added
2859 new_edges++;
2860 add_field_uses_to_worklist(field->as_Field());
2861 visited_bottom_offset = true;
2862 }
2863 }
2864 } else {
2865 // Check only oop fields.
2866 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2867 if (adr_type->isa_rawptr()) {
2868 #ifdef ASSERT
2869 // Raw pointers are used for initializing stores so skip it
2870 // since it should be recorded already
2871 Node* base = get_addp_base(field->ideal_node());
2872 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2873 #endif
2874 continue;
2875 }
2876 if (!offsets_worklist.contains(offset)) {
2877 offsets_worklist.append(offset);
2878 Node* value = nullptr;
2879 if (ini != nullptr) {
2880 // StoreP::value_basic_type() == T_ADDRESS
2881 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2882 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2883 // Make sure initializing store has the same type as this AddP.
2884 // This AddP may reference non existing field because it is on a
2885 // dead branch of bimorphic call which is not eliminated yet.
2886 if (store != nullptr && store->is_Store() &&
2887 store->as_Store()->value_basic_type() == ft) {
2888 value = store->in(MemNode::ValueIn);
2889 #ifdef ASSERT
2890 if (VerifyConnectionGraph) {
2891 // Verify that AddP already points to all objects the value points to.
2892 PointsToNode* val = ptnode_adr(value->_idx);
2893 assert((val != nullptr), "should be processed already");
2894 PointsToNode* missed_obj = nullptr;
2895 if (val->is_JavaObject()) {
2896 if (!field->points_to(val->as_JavaObject())) {
2897 missed_obj = val;
2898 }
2899 } else {
2900 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2901 tty->print_cr("----------init store has invalid value -----");
2902 store->dump();
2903 val->dump();
2904 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2905 }
2906 for (EdgeIterator j(val); j.has_next(); j.next()) {
2907 PointsToNode* obj = j.get();
2908 if (obj->is_JavaObject()) {
2909 if (!field->points_to(obj->as_JavaObject())) {
2910 missed_obj = obj;
2911 break;
2912 }
2913 }
2914 }
2915 }
2916 if (missed_obj != nullptr) {
2917 tty->print_cr("----------field---------------------------------");
2918 field->dump();
2919 tty->print_cr("----------missed referernce to object-----------");
2920 missed_obj->dump();
2921 tty->print_cr("----------object referernced by init store -----");
2922 store->dump();
2923 val->dump();
2924 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2925 }
2926 }
2927 #endif
2928 } else {
2929 // There could be initializing stores which follow allocation.
2930 // For example, a volatile field store is not collected
2931 // by Initialize node.
2932 //
2933 // Need to check for dependent loads to separate such stores from
2934 // stores which follow loads. For now, add initial value null so
2935 // that compare pointers optimization works correctly.
2936 }
2937 }
2938 if (value == nullptr) {
2939 // A field's initializing value was not recorded. Add null.
2940 if (add_edge(field, null_obj)) {
2941 // New edge was added
2942 new_edges++;
2943 add_field_uses_to_worklist(field->as_Field());
2944 }
2945 }
2946 }
2947 }
2948 }
2949 return new_edges;
2950 }
2951
2952 // Adjust scalar_replaceable state after Connection Graph is built.
2953 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2954 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2955 // returns true. If one of the constraints in this method set 'jobj' to NSR
2956 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2957 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2958 // that other object and the Phi will become a reducible Phi.
2959 // There could be multiple merges involving the same jobj.
2960 Unique_Node_List candidates;
2961
2962 // Search for non-escaping objects which are not scalar replaceable
2963 // and mark them to propagate the state to referenced objects.
2964
2965 for (UseIterator i(jobj); i.has_next(); i.next()) {
2966 PointsToNode* use = i.get();
2967 if (use->is_Arraycopy()) {
2968 continue;
2969 }
2970 if (use->is_Field()) {
2971 FieldNode* field = use->as_Field();
2972 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2973 // 1. An object is not scalar replaceable if the field into which it is
2974 // stored has unknown offset (stored into unknown element of an array).
2975 if (field->offset() == Type::OffsetBot) {
2976 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2977 return;
2978 }
2979 for (BaseIterator i(field); i.has_next(); i.next()) {
2980 PointsToNode* base = i.get();
2981 // 2. An object is not scalar replaceable if the field into which it is
2982 // stored has multiple bases one of which is null.
2983 if ((base == null_obj) && (field->base_count() > 1)) {
2984 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2985 return;
2986 }
2987 // 2.5. An object is not scalar replaceable if the field into which it is
2988 // stored has NSR base.
2989 if (!base->scalar_replaceable()) {
2990 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2991 return;
2992 }
2993 }
2994 }
2995 assert(use->is_Field() || use->is_LocalVar(), "sanity");
2996 // 3. An object is not scalar replaceable if it is merged with other objects
2997 // and we can't remove the merge
2998 for (EdgeIterator j(use); j.has_next(); j.next()) {
2999 PointsToNode* ptn = j.get();
3000 if (ptn->is_JavaObject() && ptn != jobj) {
3001 Node* use_n = use->ideal_node();
3002
3003 // These other local vars may point to multiple objects through a Phi
3004 // In this case we skip them and see if we can reduce the Phi.
3005 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3006 use_n = use_n->in(1);
3007 }
3008
3009 // If it's already a candidate or confirmed reducible merge we can skip verification
3010 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3011 continue;
3012 }
3013
3014 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3015 candidates.push(use_n);
3016 } else {
3017 // Mark all objects as NSR if we can't remove the merge
3018 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3019 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3020 }
3021 }
3022 }
3023 if (!jobj->scalar_replaceable()) {
3024 return;
3025 }
3026 }
3027
3028 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3029 if (j.get()->is_Arraycopy()) {
3030 continue;
3031 }
3032
3033 // Non-escaping object node should point only to field nodes.
3034 FieldNode* field = j.get()->as_Field();
3035 int offset = field->as_Field()->offset();
3036
3037 // 4. An object is not scalar replaceable if it has a field with unknown
3038 // offset (array's element is accessed in loop).
3039 if (offset == Type::OffsetBot) {
3040 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3041 return;
3042 }
3043 // 5. Currently an object is not scalar replaceable if a LoadStore node
3044 // access its field since the field value is unknown after it.
3045 //
3046 Node* n = field->ideal_node();
3047
3048 // Test for an unsafe access that was parsed as maybe off heap
3049 // (with a CheckCastPP to raw memory).
3050 assert(n->is_AddP(), "expect an address computation");
3051 if (n->in(AddPNode::Base)->is_top() &&
3052 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3053 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3054 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3055 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3056 return;
3057 }
3058
3059 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3060 Node* u = n->fast_out(i);
3061 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3062 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3063 return;
3064 }
3065 }
3066
3067 // 6. Or the address may point to more then one object. This may produce
3068 // the false positive result (set not scalar replaceable)
3069 // since the flow-insensitive escape analysis can't separate
3070 // the case when stores overwrite the field's value from the case
3071 // when stores happened on different control branches.
3072 //
3073 // Note: it will disable scalar replacement in some cases:
3074 //
3075 // Point p[] = new Point[1];
3076 // p[0] = new Point(); // Will be not scalar replaced
3077 //
3078 // but it will save us from incorrect optimizations in next cases:
3079 //
3080 // Point p[] = new Point[1];
3081 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3082 //
3083 if (field->base_count() > 1 && candidates.size() == 0) {
3084 if (has_non_reducible_merge(field, reducible_merges)) {
3085 for (BaseIterator i(field); i.has_next(); i.next()) {
3086 PointsToNode* base = i.get();
3087 // Don't take into account LocalVar nodes which
3088 // may point to only one object which should be also
3089 // this field's base by now.
3090 if (base->is_JavaObject() && base != jobj) {
3091 // Mark all bases.
3092 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3093 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3094 }
3095 }
3096
3097 if (!jobj->scalar_replaceable()) {
3098 return;
3099 }
3100 }
3101 }
3102 }
3103
3104 // The candidate is truly a reducible merge only if none of the other
3105 // constraints ruled it as NSR. There could be multiple merges involving the
3106 // same jobj.
3107 assert(jobj->scalar_replaceable(), "sanity");
3108 for (uint i = 0; i < candidates.size(); i++ ) {
3109 Node* candidate = candidates.at(i);
3110 reducible_merges.push(candidate);
3111 }
3112 }
3113
3114 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3115 for (BaseIterator i(field); i.has_next(); i.next()) {
3116 Node* base = i.get()->ideal_node();
3117 if (base->is_Phi() && !reducible_merges.member(base)) {
3118 return true;
3119 }
3120 }
3121 return false;
3122 }
3123
3124 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3125 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3126
3127 // Look for 'phis' that refer to 'jobj' as the last
3128 // remaining scalar replaceable input.
3129 uint reducible_merges_cnt = reducible_merges.size();
3130 for (uint i = 0; i < reducible_merges_cnt; i++) {
3131 Node* phi = reducible_merges.at(i);
3132
3133 // This 'Phi' will be a 'good' if it still points to
3134 // at least one scalar replaceable object. Note that 'obj'
3135 // was/should be marked as NSR before calling this function.
3136 bool good_phi = false;
3137
3138 for (uint j = 1; j < phi->req(); j++) {
3139 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3140 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3141 good_phi = true;
3142 break;
3143 }
3144 }
3145
3146 if (!good_phi) {
3147 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3148 reducible_merges.remove(i);
3149
3150 // Decrement the index because the 'remove' call above actually
3151 // moves the last entry of the list to position 'i'.
3152 i--;
3153
3154 reducible_merges_cnt--;
3155 }
3156 }
3157 }
3158
3159 // Propagate NSR (Not scalar replaceable) state.
3160 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3161 int jobj_length = jobj_worklist.length();
3162 bool found_nsr_alloc = true;
3163 while (found_nsr_alloc) {
3164 found_nsr_alloc = false;
3165 for (int next = 0; next < jobj_length; ++next) {
3166 JavaObjectNode* jobj = jobj_worklist.at(next);
3167 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3168 PointsToNode* use = i.get();
3169 if (use->is_Field()) {
3170 FieldNode* field = use->as_Field();
3171 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3172 assert(field->offset() != Type::OffsetBot, "sanity");
3173 for (BaseIterator i(field); i.has_next(); i.next()) {
3174 PointsToNode* base = i.get();
3175 // An object is not scalar replaceable if the field into which
3176 // it is stored has NSR base.
3177 if ((base != null_obj) && !base->scalar_replaceable()) {
3178 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3179 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3180 // because there is no point in reducing a Phi that won't improve the number of SR
3181 // objects.
3182 revisit_reducible_phi_status(jobj, reducible_merges);
3183 found_nsr_alloc = true;
3184 break;
3185 }
3186 }
3187 } else if (use->is_LocalVar()) {
3188 Node* phi = use->ideal_node();
3189 if (phi->Opcode() == Op_Phi && reducible_merges.member(phi) && !can_reduce_phi(phi->as_Phi())) {
3190 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is merged in a non-reducible phi"));
3191 reducible_merges.yank(phi);
3192 found_nsr_alloc = true;
3193 break;
3194 }
3195 }
3196 _compile->print_method(PHASE_EA_PROPAGATE_NSR_ITER, 5, jobj->ideal_node());
3197 }
3198 }
3199 }
3200 }
3201
3202 #ifdef ASSERT
3203 void ConnectionGraph::verify_connection_graph(
3204 GrowableArray<PointsToNode*>& ptnodes_worklist,
3205 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3206 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3207 GrowableArray<Node*>& addp_worklist) {
3208 // Verify that graph is complete - no new edges could be added.
3209 int java_objects_length = java_objects_worklist.length();
3210 int non_escaped_length = non_escaped_allocs_worklist.length();
3211 int new_edges = 0;
3212 for (int next = 0; next < java_objects_length; ++next) {
3213 JavaObjectNode* ptn = java_objects_worklist.at(next);
3214 new_edges += add_java_object_edges(ptn, true);
3215 }
3216 assert(new_edges == 0, "graph was not complete");
3217 // Verify that escape state is final.
3218 int length = non_escaped_allocs_worklist.length();
3219 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist, /*print_method=*/ false);
3220 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3221 (non_escaped_length == length) &&
3222 (_worklist.length() == 0), "escape state was not final");
3223
3224 // Verify fields information.
3225 int addp_length = addp_worklist.length();
3226 for (int next = 0; next < addp_length; ++next ) {
3227 Node* n = addp_worklist.at(next);
3228 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3229 if (field->is_oop()) {
3230 // Verify that field has all bases
3231 Node* base = get_addp_base(n);
3232 PointsToNode* ptn = ptnode_adr(base->_idx);
3233 if (ptn->is_JavaObject()) {
3234 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3235 } else {
3236 assert(ptn->is_LocalVar(), "sanity");
3237 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3238 PointsToNode* e = i.get();
3239 if (e->is_JavaObject()) {
3240 assert(field->has_base(e->as_JavaObject()), "sanity");
3241 }
3242 }
3243 }
3244 // Verify that all fields have initializing values.
3245 if (field->edge_count() == 0) {
3246 tty->print_cr("----------field does not have references----------");
3247 field->dump();
3248 for (BaseIterator i(field); i.has_next(); i.next()) {
3249 PointsToNode* base = i.get();
3250 tty->print_cr("----------field has next base---------------------");
3251 base->dump();
3252 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3253 tty->print_cr("----------base has fields-------------------------");
3254 for (EdgeIterator j(base); j.has_next(); j.next()) {
3255 j.get()->dump();
3256 }
3257 tty->print_cr("----------base has references---------------------");
3258 for (UseIterator j(base); j.has_next(); j.next()) {
3259 j.get()->dump();
3260 }
3261 }
3262 }
3263 for (UseIterator i(field); i.has_next(); i.next()) {
3264 i.get()->dump();
3265 }
3266 assert(field->edge_count() > 0, "sanity");
3267 }
3268 }
3269 }
3270 }
3271 #endif
3272
3273 // Optimize ideal graph.
3274 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3275 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3276 Compile* C = _compile;
3277 PhaseIterGVN* igvn = _igvn;
3278 if (EliminateLocks) {
3279 // Mark locks before changing ideal graph.
3280 int cnt = C->macro_count();
3281 for (int i = 0; i < cnt; i++) {
3282 Node *n = C->macro_node(i);
3283 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3284 AbstractLockNode* alock = n->as_AbstractLock();
3285 if (!alock->is_non_esc_obj()) {
3286 if (can_eliminate_lock(alock)) {
3287 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3288 // The lock could be marked eliminated by lock coarsening
3289 // code during first IGVN before EA. Replace coarsened flag
3290 // to eliminate all associated locks/unlocks.
3291 #ifdef ASSERT
3292 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3293 #endif
3294 alock->set_non_esc_obj();
3295 }
3296 }
3297 }
3298 }
3299 }
3300
3301 if (OptimizePtrCompare) {
3302 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3303 Node *n = ptr_cmp_worklist.at(i);
3304 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3305 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3306 if (tcmp->singleton()) {
3307 Node* cmp = igvn->makecon(tcmp);
3308 #ifndef PRODUCT
3309 if (PrintOptimizePtrCompare) {
3310 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"));
3311 if (Verbose) {
3312 n->dump(1);
3313 }
3314 }
3315 #endif
3316 igvn->replace_node(n, cmp);
3317 }
3318 }
3319 }
3320
3321 // For MemBarStoreStore nodes added in library_call.cpp, check
3322 // escape status of associated AllocateNode and optimize out
3323 // MemBarStoreStore node if the allocated object never escapes.
3324 for (int i = 0; i < storestore_worklist.length(); i++) {
3325 Node* storestore = storestore_worklist.at(i);
3326 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3327 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3328 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);3329 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));3330 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));3331 igvn->register_new_node_with_optimizer(mb);3332 igvn->replace_node(storestore, mb);
3333 }
3334 }
3335 }
3336
3337 // Optimize objects compare.
3338 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3339 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3340 if (!OptimizePtrCompare) {
3341 return UNKNOWN;
3342 }
3343 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3344 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3345
3346 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3347 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3348 JavaObjectNode* jobj1 = unique_java_object(left);
3349 JavaObjectNode* jobj2 = unique_java_object(right);
3350
3351 // The use of this method during allocation merge reduction may cause 'left'
3352 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3353 // that doesn't reference an unique java object.
3354 if (ptn1 == nullptr || ptn2 == nullptr ||
3355 jobj1 == nullptr || jobj2 == nullptr) {
3356 return UNKNOWN;
3357 }
3358
3359 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3360 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3361
3362 // Check simple cases first.
3363 if (jobj1 != nullptr) {
3364 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3365 if (jobj1 == jobj2) {
3366 // Comparing the same not escaping object.
3367 return EQ;
3368 }
3369 Node* obj = jobj1->ideal_node();
3370 // Comparing not escaping allocation.
3371 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3372 !ptn2->points_to(jobj1)) {
3373 return NE; // This includes nullness check.
3374 }
3375 }
3376 }
3377 if (jobj2 != nullptr) {
3378 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3379 Node* obj = jobj2->ideal_node();
3380 // Comparing not escaping allocation.
3381 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3382 !ptn1->points_to(jobj2)) {
3383 return NE; // This includes nullness check.
3384 }
3385 }
3386 }
3387 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3388 jobj2 != nullptr && jobj2 != phantom_obj &&
3389 jobj1->ideal_node()->is_Con() &&
3390 jobj2->ideal_node()->is_Con()) {
3391 // Klass or String constants compare. Need to be careful with
3392 // compressed pointers - compare types of ConN and ConP instead of nodes.
3393 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3394 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3395 if (t1->make_ptr() == t2->make_ptr()) {
3396 return EQ;
3397 } else {
3398 return NE;
3399 }
3400 }
3401 if (ptn1->meet(ptn2)) {
3402 return UNKNOWN; // Sets are not disjoint
3403 }
3404
3405 // Sets are disjoint.
3406 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3407 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3408 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3409 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3410 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3411 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3412 // Check nullness of unknown object.
3413 return UNKNOWN;
3414 }
3415
3416 // Disjointness by itself is not sufficient since
3417 // alias analysis is not complete for escaped objects.
3418 // Disjoint sets are definitely unrelated only when
3419 // at least one set has only not escaping allocations.
3420 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3421 if (ptn1->non_escaping_allocation()) {
3422 return NE;
3423 }
3424 }
3425 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3426 if (ptn2->non_escaping_allocation()) {
3427 return NE;
3428 }
3429 }
3430 return UNKNOWN;
3431 }
3432
3433 // Connection Graph construction functions.
3434
3435 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3436 PointsToNode* ptadr = _nodes.at(n->_idx);
3437 if (ptadr != nullptr) {
3438 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3439 return;
3440 }
3441 Compile* C = _compile;
3442 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3443 map_ideal_node(n, ptadr);
3444 }
3445
3446 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3447 PointsToNode* ptadr = _nodes.at(n->_idx);
3448 if (ptadr != nullptr) {
3449 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3450 return ptadr;
3451 }
3452 Compile* C = _compile;
3453 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3454 map_ideal_node(n, ptadr);
3455 return ptadr;
3456 }
3457
3458 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3459 PointsToNode* ptadr = _nodes.at(n->_idx);
3460 if (ptadr != nullptr) {
3461 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3462 return;
3463 }
3464 bool unsafe = false;
3465 bool is_oop = is_oop_field(n, offset, &unsafe);
3466 if (unsafe) {
3467 es = PointsToNode::GlobalEscape;
3468 }
3469 Compile* C = _compile;
3470 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3471 map_ideal_node(n, field);
3472 }
3473
3474 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3475 PointsToNode* src, PointsToNode* dst) {
3476 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3477 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3478 PointsToNode* ptadr = _nodes.at(n->_idx);
3479 if (ptadr != nullptr) {
3480 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3481 return;
3482 }
3483 Compile* C = _compile;
3484 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3485 map_ideal_node(n, ptadr);
3486 // Add edge from arraycopy node to source object.
3487 (void)add_edge(ptadr, src);
3488 src->set_arraycopy_src();
3489 // Add edge from destination object to arraycopy node.
3490 (void)add_edge(dst, ptadr);
3491 dst->set_arraycopy_dst();
3492 }
3493
3494 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3495 const Type* adr_type = n->as_AddP()->bottom_type();
3496 BasicType bt = T_INT;
3497 if (offset == Type::OffsetBot) {
3498 // Check only oop fields.
3499 if (!adr_type->isa_aryptr() ||
3500 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3501 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3502 // OffsetBot is used to reference array's element. Ignore first AddP.
3503 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3504 bt = T_OBJECT;
3505 }
3506 }
3507 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3508 if (adr_type->isa_instptr()) {
3509 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
3510 if (field != nullptr) {
3511 bt = field->layout_type();
3512 } else {
3513 // Check for unsafe oop field access
3514 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3515 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3516 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3517 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3518 bt = T_OBJECT;
3519 (*unsafe) = true;
3520 }
3521 }
3522 } else if (adr_type->isa_aryptr()) {
3523 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3524 // Ignore array length load.
3525 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3526 // Ignore first AddP.
3527 } else {
3528 const Type* elemtype = adr_type->isa_aryptr()->elem();3529 bt = elemtype->array_element_basic_type();
3530 }
3531 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3532 // Allocation initialization, ThreadLocal field access, unsafe access
3533 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3534 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3535 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3536 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3537 bt = T_OBJECT;
3538 }
3539 }
3540 }
3541 // Note: T_NARROWOOP is not classed as a real reference type
3542 return (is_reference_type(bt) || bt == T_NARROWOOP);
3543 }
3544
3545 // Returns unique pointed java object or null.
3546 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3547 // If the node was created after the escape computation we can't answer.
3548 uint idx = n->_idx;
3549 if (idx >= nodes_size()) {
3550 return nullptr;
3551 }
3552 PointsToNode* ptn = ptnode_adr(idx);
3553 if (ptn == nullptr) {
3554 return nullptr;
3555 }
3556 if (ptn->is_JavaObject()) {
3557 return ptn->as_JavaObject();
3558 }
3559 assert(ptn->is_LocalVar(), "sanity");
3560 // Check all java objects it points to.
3561 JavaObjectNode* jobj = nullptr;
3562 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3563 PointsToNode* e = i.get();
3564 if (e->is_JavaObject()) {
3565 if (jobj == nullptr) {
3566 jobj = e->as_JavaObject();
3567 } else if (jobj != e) {
3568 return nullptr;
3569 }
3570 }
3571 }
3572 return jobj;
3573 }
3574
3575 // Return true if this node points only to non-escaping allocations.
3576 bool PointsToNode::non_escaping_allocation() {
3577 if (is_JavaObject()) {
3578 Node* n = ideal_node();
3579 if (n->is_Allocate() || n->is_CallStaticJava()) {
3580 return (escape_state() == PointsToNode::NoEscape);
3581 } else {
3582 return false;
3583 }
3584 }
3585 assert(is_LocalVar(), "sanity");
3586 // Check all java objects it points to.
3587 for (EdgeIterator i(this); i.has_next(); i.next()) {
3588 PointsToNode* e = i.get();
3589 if (e->is_JavaObject()) {
3590 Node* n = e->ideal_node();
3591 if ((e->escape_state() != PointsToNode::NoEscape) ||
3592 !(n->is_Allocate() || n->is_CallStaticJava())) {
3593 return false;
3594 }
3595 }
3596 }
3597 return true;
3598 }
3599
3600 // Return true if we know the node does not escape globally.
3601 bool ConnectionGraph::not_global_escape(Node *n) {
3602 assert(!_collecting, "should not call during graph construction");
3603 // If the node was created after the escape computation we can't answer.
3604 uint idx = n->_idx;
3605 if (idx >= nodes_size()) {
3606 return false;
3607 }
3608 PointsToNode* ptn = ptnode_adr(idx);
3609 if (ptn == nullptr) {
3610 return false; // not in congraph (e.g. ConI)
3611 }
3612 PointsToNode::EscapeState es = ptn->escape_state();
3613 // If we have already computed a value, return it.
3614 if (es >= PointsToNode::GlobalEscape) {
3615 return false;
3616 }
3617 if (ptn->is_JavaObject()) {
3618 return true; // (es < PointsToNode::GlobalEscape);
3619 }
3620 assert(ptn->is_LocalVar(), "sanity");
3621 // Check all java objects it points to.
3622 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3623 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3624 return false;
3625 }
3626 }
3627 return true;
3628 }
3629
3630 // Return true if locked object does not escape globally
3631 // and locked code region (identified by BoxLockNode) is balanced:
3632 // all compiled code paths have corresponding Lock/Unlock pairs.
3633 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3634 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3635 if (EliminateNestedLocks) {
3636 // We can mark whole locking region as Local only when only
3637 // one object is used for locking.
3638 alock->box_node()->as_BoxLock()->set_local();
3639 }
3640 return true;
3641 }
3642 return false;
3643 }
3644
3645 // Helper functions
3646
3647 // Return true if this node points to specified node or nodes it points to.
3648 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3649 if (is_JavaObject()) {
3650 return (this == ptn);
3651 }
3652 assert(is_LocalVar() || is_Field(), "sanity");
3653 for (EdgeIterator i(this); i.has_next(); i.next()) {
3654 if (i.get() == ptn) {
3655 return true;
3656 }
3657 }
3658 return false;
3659 }
3660
3661 // Return true if one node points to an other.
3662 bool PointsToNode::meet(PointsToNode* ptn) {
3663 if (this == ptn) {
3664 return true;
3665 } else if (ptn->is_JavaObject()) {
3666 return this->points_to(ptn->as_JavaObject());
3667 } else if (this->is_JavaObject()) {
3668 return ptn->points_to(this->as_JavaObject());
3669 }
3670 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3671 int ptn_count = ptn->edge_count();
3672 for (EdgeIterator i(this); i.has_next(); i.next()) {
3673 PointsToNode* this_e = i.get();
3674 for (int j = 0; j < ptn_count; j++) {
3675 if (this_e == ptn->edge(j)) {
3676 return true;
3677 }
3678 }
3679 }
3680 return false;
3681 }
3682
3683 #ifdef ASSERT
3684 // Return true if bases point to this java object.
3685 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3686 for (BaseIterator i(this); i.has_next(); i.next()) {
3687 if (i.get() == jobj) {
3688 return true;
3689 }
3690 }
3691 return false;
3692 }
3693 #endif
3694
3695 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3696 // Handle simple case first.
3697 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3698 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3699 return true;
3700 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3701 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3702 Node* addp_use = addp->fast_out(i);
3703 if (addp_use->is_Store()) {
3704 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3705 if (addp_use->fast_out(j)->is_Initialize()) {
3706 return true;
3707 }
3708 }
3709 }
3710 }
3711 }
3712 return false;
3713 }
3714
3715 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3716 const Type *adr_type = phase->type(adr);
3717 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3718 // We are computing a raw address for a store captured by an Initialize
3719 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3720 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3721 assert(offs != Type::OffsetBot ||
3722 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3723 "offset must be a constant or it is initialization of array");
3724 return offs;
3725 }
3726 const TypePtr *t_ptr = adr_type->isa_ptr();3727 assert(t_ptr != nullptr, "must be a pointer type");3728 return t_ptr->offset();
3729 }
3730
3731 Node* ConnectionGraph::get_addp_base(Node *addp) {
3732 assert(addp->is_AddP(), "must be AddP");
3733 //
3734 // AddP cases for Base and Address inputs:
3735 // case #1. Direct object's field reference:
3736 // Allocate
3737 // |
3738 // Proj #5 ( oop result )
3739 // |
3740 // CheckCastPP (cast to instance type)
3741 // | |
3742 // AddP ( base == address )
3743 //
3744 // case #2. Indirect object's field reference:
3745 // Phi
3746 // |
3747 // CastPP (cast to instance type)
3748 // | |
3749 // AddP ( base == address )
3750 //
3751 // case #3. Raw object's field reference for Initialize node:
3752 // Allocate
3753 // |
3754 // Proj #5 ( oop result )
3755 // top |
3756 // \ |
3757 // AddP ( base == top )
3758 //
3759 // case #4. Array's element reference:
3760 // {CheckCastPP | CastPP}
3761 // | | |
3762 // | AddP ( array's element offset )
3763 // | |
3764 // AddP ( array's offset )
3765 //
3766 // case #5. Raw object's field reference for arraycopy stub call:
3767 // The inline_native_clone() case when the arraycopy stub is called
3768 // after the allocation before Initialize and CheckCastPP nodes.
3769 // Allocate
3770 // |
3771 // Proj #5 ( oop result )
3772 // | |
3773 // AddP ( base == address )
3774 //
3775 // case #6. Constant Pool, ThreadLocal, CastX2P or
3776 // Raw object's field reference:
3777 // {ConP, ThreadLocal, CastX2P, raw Load}
3778 // top |
3779 // \ |
3780 // AddP ( base == top )
3781 //
3782 // case #7. Klass's field reference.
3783 // LoadKlass
3784 // | |
3785 // AddP ( base == address )
3786 //
3787 // case #8. narrow Klass's field reference.
3788 // LoadNKlass
3789 // |
3790 // DecodeN
3791 // | |
3792 // AddP ( base == address )
3793 //
3794 // case #9. Mixed unsafe access
3795 // {instance}
3796 // |
3797 // CheckCastPP (raw)
3798 // top |
3799 // \ |
3800 // AddP ( base == top )
3801 //
3802 Node *base = addp->in(AddPNode::Base);
3803 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3804 base = addp->in(AddPNode::Address);
3805 while (base->is_AddP()) {
3806 // Case #6 (unsafe access) may have several chained AddP nodes.
3807 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3808 base = base->in(AddPNode::Address);
3809 }
3810 if (base->Opcode() == Op_CheckCastPP &&
3811 base->bottom_type()->isa_rawptr() &&
3812 _igvn->type(base->in(1))->isa_oopptr()) {
3813 base = base->in(1); // Case #9
3814 } else {
3815 Node* uncast_base = base->uncast();
3816 int opcode = uncast_base->Opcode();
3817 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3818 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3819 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3820 is_captured_store_address(addp), "sanity");
3821 }
3822 }
3823 return base;
3824 }
3825
3826 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3827 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3828 Node* addp2 = addp->raw_out(0);
3829 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3830 addp2->in(AddPNode::Base) == n &&
3831 addp2->in(AddPNode::Address) == addp) {
3832 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3833 //
3834 // Find array's offset to push it on worklist first and
3835 // as result process an array's element offset first (pushed second)
3836 // to avoid CastPP for the array's offset.
3837 // Otherwise the inserted CastPP (LocalVar) will point to what
3838 // the AddP (Field) points to. Which would be wrong since
3839 // the algorithm expects the CastPP has the same point as
3840 // as AddP's base CheckCastPP (LocalVar).
3841 //
3842 // ArrayAllocation
3843 // |
3844 // CheckCastPP
3845 // |
3846 // memProj (from ArrayAllocation CheckCastPP)
3847 // | ||
3848 // | || Int (element index)
3849 // | || | ConI (log(element size))
3850 // | || | /
3851 // | || LShift
3852 // | || /
3853 // | AddP (array's element offset)
3854 // | |
3855 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3856 // | / /
3857 // AddP (array's offset)
3858 // |
3859 // Load/Store (memory operation on array's element)
3860 //
3861 return addp2;
3862 }
3863 return nullptr;
3864 }
3865
3866 //
3867 // Adjust the type and inputs of an AddP which computes the
3868 // address of a field of an instance
3869 //
3870 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3871 PhaseGVN* igvn = _igvn;
3872 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3873 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3874 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3875 if (t == nullptr) {
3876 // We are computing a raw address for a store captured by an Initialize
3877 // compute an appropriate address type (cases #3 and #5).
3878 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3879 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3880 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3881 assert(offs != Type::OffsetBot, "offset must be a constant");
3882 t = base_t->add_offset(offs)->is_oopptr();
3883 }
3884 int inst_id = base_t->instance_id();
3885 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3886 "old type must be non-instance or match new type");
3887
3888 // The type 't' could be subclass of 'base_t'.
3889 // As result t->offset() could be large then base_t's size and it will
3890 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3891 // constructor verifies correctness of the offset.
3892 //
3893 // It could happened on subclass's branch (from the type profiling
3894 // inlining) which was not eliminated during parsing since the exactness
3895 // of the allocation type was not propagated to the subclass type check.
3896 //
3897 // Or the type 't' could be not related to 'base_t' at all.
3898 // It could happened when CHA type is different from MDO type on a dead path
3899 // (for example, from instanceof check) which is not collapsed during parsing.
3900 //
3901 // Do nothing for such AddP node and don't process its users since
3902 // this code branch will go away.
3903 //
3904 if (!t->is_known_instance() &&
3905 !base_t->maybe_java_subtype_of(t)) {
3906 return false; // bail out
3907 }
3908 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
3909 // Do NOT remove the next line: ensure a new alias index is allocated
3910 // for the instance type. Note: C++ will not remove it since the call
3911 // has side effect.
3912 int alias_idx = _compile->get_alias_index(tinst);
3913 igvn->set_type(addp, tinst);
3914 // record the allocation in the node map
3915 set_map(addp, get_map(base->_idx));
3916 // Set addp's Base and Address to 'base'.
3917 Node *abase = addp->in(AddPNode::Base);
3918 Node *adr = addp->in(AddPNode::Address);
3919 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3920 adr->in(0)->_idx == (uint)inst_id) {
3921 // Skip AddP cases #3 and #5.
3922 } else {
3923 assert(!abase->is_top(), "sanity"); // AddP case #3
3924 if (abase != base) {
3925 igvn->hash_delete(addp);
3926 addp->set_req(AddPNode::Base, base);
3927 if (abase == adr) {
3928 addp->set_req(AddPNode::Address, base);
3929 } else {
3930 // AddP case #4 (adr is array's element offset AddP node)
3931 #ifdef ASSERT
3932 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3933 assert(adr->is_AddP() && atype != nullptr &&
3934 atype->instance_id() == inst_id, "array's element offset should be processed first");
3935 #endif
3936 }
3937 igvn->hash_insert(addp);
3938 }
3939 }
3940 // Put on IGVN worklist since at least addp's type was changed above.
3941 record_for_optimizer(addp);
3942 return true;
3943 }
3944
3945 //
3946 // Create a new version of orig_phi if necessary. Returns either the newly
3947 // created phi or an existing phi. Sets create_new to indicate whether a new
3948 // phi was created. Cache the last newly created phi in the node map.
3949 //
3950 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3951 Compile *C = _compile;
3952 PhaseGVN* igvn = _igvn;
3953 new_created = false;
3954 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3955 // nothing to do if orig_phi is bottom memory or matches alias_idx
3956 if (phi_alias_idx == alias_idx) {
3957 return orig_phi;
3958 }
3959 // Have we recently created a Phi for this alias index?
3960 PhiNode *result = get_map_phi(orig_phi->_idx);
3961 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3962 return result;
3963 }
3964 // Previous check may fail when the same wide memory Phi was split into Phis
3965 // for different memory slices. Search all Phis for this region.
3966 if (result != nullptr) {
3967 Node* region = orig_phi->in(0);
3968 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3969 Node* phi = region->fast_out(i);
3970 if (phi->is_Phi() &&
3971 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3972 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3973 return phi->as_Phi();
3974 }
3975 }
3976 }
3977 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
3978 if (C->do_escape_analysis() == true && !C->failing()) {
3979 // Retry compilation without escape analysis.
3980 // If this is the first failure, the sentinel string will "stick"
3981 // to the Compile object, and the C2Compiler will see it and retry.
3982 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
3983 }
3984 return nullptr;
3985 }
3986 orig_phi_worklist.append_if_missing(orig_phi);
3987 const TypePtr *atype = C->get_adr_type(alias_idx);
3988 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
3989 C->copy_node_notes_to(result, orig_phi);
3990 igvn->set_type(result, result->bottom_type());
3991 record_for_optimizer(result);
3992 set_map(orig_phi, result);
3993 new_created = true;
3994 return result;
3995 }
3996
3997 //
3998 // Return a new version of Memory Phi "orig_phi" with the inputs having the
3999 // specified alias index.
4000 //
4001 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4002 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4003 Compile *C = _compile;
4004 PhaseGVN* igvn = _igvn;
4005 bool new_phi_created;
4006 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4007 if (!new_phi_created) {
4008 return result;
4009 }
4010 GrowableArray<PhiNode *> phi_list;
4011 GrowableArray<uint> cur_input;
4012 PhiNode *phi = orig_phi;
4013 uint idx = 1;
4014 bool finished = false;
4015 while(!finished) {
4016 while (idx < phi->req()) {
4017 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4018 if (mem != nullptr && mem->is_Phi()) {
4019 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4020 if (new_phi_created) {
4021 // found an phi for which we created a new split, push current one on worklist and begin
4022 // processing new one
4023 phi_list.push(phi);
4024 cur_input.push(idx);
4025 phi = mem->as_Phi();
4026 result = newphi;
4027 idx = 1;
4028 continue;
4029 } else {
4030 mem = newphi;
4031 }
4032 }
4033 if (C->failing()) {
4034 return nullptr;
4035 }
4036 result->set_req(idx++, mem);
4037 }
4038 #ifdef ASSERT
4039 // verify that the new Phi has an input for each input of the original
4040 assert( phi->req() == result->req(), "must have same number of inputs.");
4041 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4042 #endif
4043 // Check if all new phi's inputs have specified alias index.
4044 // Otherwise use old phi.
4045 for (uint i = 1; i < phi->req(); i++) {
4046 Node* in = result->in(i);
4047 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4048 }
4049 // we have finished processing a Phi, see if there are any more to do
4050 finished = (phi_list.length() == 0 );
4051 if (!finished) {
4052 phi = phi_list.pop();
4053 idx = cur_input.pop();
4054 PhiNode *prev_result = get_map_phi(phi->_idx);
4055 prev_result->set_req(idx++, result);
4056 result = prev_result;
4057 }
4058 }
4059 return result;
4060 }
4061
4062 //
4063 // The next methods are derived from methods in MemNode.
4064 //
4065 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4066 Node *mem = mmem;
4067 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4068 // means an array I have not precisely typed yet. Do not do any
4069 // alias stuff with it any time soon.
4070 if (toop->base() != Type::AnyPtr &&
4071 !(toop->isa_instptr() &&
4072 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4073 toop->offset() == Type::OffsetBot)) {
4074 mem = mmem->memory_at(alias_idx);
4075 // Update input if it is progress over what we have now
4076 }
4077 return mem;
4078 }
4079
4080 //
4081 // Move memory users to their memory slices.
4082 //
4083 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4084 Compile* C = _compile;
4085 PhaseGVN* igvn = _igvn;
4086 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4087 assert(tp != nullptr, "ptr type");
4088 int alias_idx = C->get_alias_index(tp);
4089 int general_idx = C->get_general_index(alias_idx);
4090
4091 // Move users first
4092 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4093 Node* use = n->fast_out(i);
4094 if (use->is_MergeMem()) {
4095 MergeMemNode* mmem = use->as_MergeMem();
4096 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4097 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4098 continue; // Nothing to do
4099 }
4100 // Replace previous general reference to mem node.
4101 uint orig_uniq = C->unique();
4102 Node* m = find_inst_mem(n, general_idx, orig_phis);
4103 assert(orig_uniq == C->unique(), "no new nodes");
4104 mmem->set_memory_at(general_idx, m);
4105 --imax;
4106 --i;
4107 } else if (use->is_MemBar()) {
4108 assert(!use->is_Initialize(), "initializing stores should not be moved");
4109 if (use->req() > MemBarNode::Precedent &&
4110 use->in(MemBarNode::Precedent) == n) {
4111 // Don't move related membars.
4112 record_for_optimizer(use);
4113 continue;
4114 }
4115 tp = use->as_MemBar()->adr_type()->isa_ptr();
4116 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4117 alias_idx == general_idx) {
4118 continue; // Nothing to do
4119 }
4120 // Move to general memory slice.
4121 uint orig_uniq = C->unique();
4122 Node* m = find_inst_mem(n, general_idx, orig_phis);
4123 assert(orig_uniq == C->unique(), "no new nodes");
4124 igvn->hash_delete(use);
4125 imax -= use->replace_edge(n, m, igvn);
4126 igvn->hash_insert(use);
4127 record_for_optimizer(use);
4128 --i;
4129 #ifdef ASSERT
4130 } else if (use->is_Mem()) {
4131 // Memory nodes should have new memory input.
4132 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4133 assert(tp != nullptr, "ptr type");
4134 int idx = C->get_alias_index(tp);
4135 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4136 "Following memory nodes should have new memory input or be on the same memory slice");
4137 } else if (use->is_Phi()) {
4138 // Phi nodes should be split and moved already.
4139 tp = use->as_Phi()->adr_type()->isa_ptr();
4140 assert(tp != nullptr, "ptr type");
4141 int idx = C->get_alias_index(tp);
4142 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4143 } else {
4144 use->dump();
4145 assert(false, "should not be here");
4146 #endif
4147 }
4148 }
4149 }
4150
4151 //
4152 // Search memory chain of "mem" to find a MemNode whose address
4153 // is the specified alias index.
4154 //
4155 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4156 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4157 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4158 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4159 return nullptr;
4160 }
4161 if (orig_mem == nullptr) {
4162 return orig_mem;
4163 }
4164 Compile* C = _compile;
4165 PhaseGVN* igvn = _igvn;
4166 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4167 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4168 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4169 Node *prev = nullptr;
4170 Node *result = orig_mem;
4171 while (prev != result) {
4172 prev = result;
4173 if (result == start_mem) {
4174 break; // hit one of our sentinels
4175 }
4176 if (result->is_Mem()) {
4177 const Type *at = igvn->type(result->in(MemNode::Address));
4178 if (at == Type::TOP) {
4179 break; // Dead
4180 }
4181 assert (at->isa_ptr() != nullptr, "pointer type required.");
4182 int idx = C->get_alias_index(at->is_ptr());
4183 if (idx == alias_idx) {
4184 break; // Found
4185 }
4186 if (!is_instance && (at->isa_oopptr() == nullptr ||
4187 !at->is_oopptr()->is_known_instance())) {
4188 break; // Do not skip store to general memory slice.
4189 }
4190 result = result->in(MemNode::Memory);
4191 }
4192 if (!is_instance) {
4193 continue; // don't search further for non-instance types
4194 }
4195 // skip over a call which does not affect this memory slice
4196 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4197 Node *proj_in = result->in(0);
4198 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4199 break; // hit one of our sentinels
4200 } else if (proj_in->is_Call()) {
4201 // ArrayCopy node processed here as well
4202 CallNode *call = proj_in->as_Call();
4203 if (!call->may_modify(toop, igvn)) {
4204 result = call->in(TypeFunc::Memory);
4205 }
4206 } else if (proj_in->is_Initialize()) {
4207 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4208 // Stop if this is the initialization for the object instance which
4209 // which contains this memory slice, otherwise skip over it.
4210 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4211 result = proj_in->in(TypeFunc::Memory);
4212 } else if (C->get_alias_index(result->adr_type()) != alias_idx) {
4213 assert(C->get_general_index(alias_idx) == C->get_alias_index(result->adr_type()), "should be projection for the same field/array element");
4214 result = get_map(result->_idx);
4215 assert(result != nullptr, "new projection should have been allocated");
4216 break;
4217 }
4218 } else if (proj_in->is_MemBar()) {
4219 // Check if there is an array copy for a clone
4220 // Step over GC barrier when ReduceInitialCardMarks is disabled
4221 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4222 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4223
4224 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4225 // Stop if it is a clone
4226 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4227 if (ac->may_modify(toop, igvn)) {
4228 break;
4229 }
4230 }
4231 result = proj_in->in(TypeFunc::Memory);
4232 }
4233 } else if (result->is_MergeMem()) {
4234 MergeMemNode *mmem = result->as_MergeMem();
4235 result = step_through_mergemem(mmem, alias_idx, toop);
4236 if (result == mmem->base_memory()) {
4237 // Didn't find instance memory, search through general slice recursively.
4238 result = mmem->memory_at(C->get_general_index(alias_idx));
4239 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4240 if (C->failing()) {
4241 return nullptr;
4242 }
4243 mmem->set_memory_at(alias_idx, result);
4244 }
4245 } else if (result->is_Phi() &&
4246 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4247 Node *un = result->as_Phi()->unique_input(igvn);
4248 if (un != nullptr) {
4249 orig_phis.append_if_missing(result->as_Phi());
4250 result = un;
4251 } else {
4252 break;
4253 }
4254 } else if (result->is_ClearArray()) {
4255 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4256 // Can not bypass initialization of the instance
4257 // we are looking for.
4258 break;
4259 }
4260 // Otherwise skip it (the call updated 'result' value).
4261 } else if (result->Opcode() == Op_SCMemProj) {
4262 Node* mem = result->in(0);
4263 Node* adr = nullptr;
4264 if (mem->is_LoadStore()) {
4265 adr = mem->in(MemNode::Address);
4266 } else {
4267 assert(mem->Opcode() == Op_EncodeISOArray ||
4268 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4269 adr = mem->in(3); // Memory edge corresponds to destination array
4270 }
4271 const Type *at = igvn->type(adr);
4272 if (at != Type::TOP) {
4273 assert(at->isa_ptr() != nullptr, "pointer type required.");
4274 int idx = C->get_alias_index(at->is_ptr());
4275 if (idx == alias_idx) {
4276 // Assert in debug mode
4277 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4278 break; // In product mode return SCMemProj node
4279 }
4280 }
4281 result = mem->in(MemNode::Memory);
4282 } else if (result->Opcode() == Op_StrInflatedCopy) {
4283 Node* adr = result->in(3); // Memory edge corresponds to destination array
4284 const Type *at = igvn->type(adr);
4285 if (at != Type::TOP) {
4286 assert(at->isa_ptr() != nullptr, "pointer type required.");
4287 int idx = C->get_alias_index(at->is_ptr());
4288 if (idx == alias_idx) {
4289 // Assert in debug mode
4290 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4291 break; // In product mode return SCMemProj node
4292 }
4293 }
4294 result = result->in(MemNode::Memory);
4295 }
4296 }
4297 if (result->is_Phi()) {
4298 PhiNode *mphi = result->as_Phi();
4299 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4300 const TypePtr *t = mphi->adr_type();
4301 if (!is_instance) {
4302 // Push all non-instance Phis on the orig_phis worklist to update inputs
4303 // during Phase 4 if needed.
4304 orig_phis.append_if_missing(mphi);
4305 } else if (C->get_alias_index(t) != alias_idx) {
4306 // Create a new Phi with the specified alias index type.
4307 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4308 }
4309 }
4310 // the result is either MemNode, PhiNode, InitializeNode.
4311 return result;
4312 }
4313
4314 //
4315 // Convert the types of non-escaped object to instance types where possible,
4316 // propagate the new type information through the graph, and update memory
4317 // edges and MergeMem inputs to reflect the new type.
4318 //
4319 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4320 // The processing is done in 4 phases:
4321 //
4322 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4323 // types for the CheckCastPP for allocations where possible.
4324 // Propagate the new types through users as follows:
4325 // casts and Phi: push users on alloc_worklist
4326 // AddP: cast Base and Address inputs to the instance type
4327 // push any AddP users on alloc_worklist and push any memnode
4328 // users onto memnode_worklist.
4329 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4330 // search the Memory chain for a store with the appropriate type
4331 // address type. If a Phi is found, create a new version with
4332 // the appropriate memory slices from each of the Phi inputs.
4333 // For stores, process the users as follows:
4334 // MemNode: push on memnode_worklist
4335 // MergeMem: push on mergemem_worklist
4336 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4337 // moving the first node encountered of each instance type to the
4338 // the input corresponding to its alias index.
4339 // appropriate memory slice.
4340 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4341 //
4342 // In the following example, the CheckCastPP nodes are the cast of allocation
4343 // results and the allocation of node 29 is non-escaped and eligible to be an
4344 // instance type.
4345 //
4346 // We start with:
4347 //
4348 // 7 Parm #memory
4349 // 10 ConI "12"
4350 // 19 CheckCastPP "Foo"
4351 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4352 // 29 CheckCastPP "Foo"
4353 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4354 //
4355 // 40 StoreP 25 7 20 ... alias_index=4
4356 // 50 StoreP 35 40 30 ... alias_index=4
4357 // 60 StoreP 45 50 20 ... alias_index=4
4358 // 70 LoadP _ 60 30 ... alias_index=4
4359 // 80 Phi 75 50 60 Memory alias_index=4
4360 // 90 LoadP _ 80 30 ... alias_index=4
4361 // 100 LoadP _ 80 20 ... alias_index=4
4362 //
4363 //
4364 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4365 // and creating a new alias index for node 30. This gives:
4366 //
4367 // 7 Parm #memory
4368 // 10 ConI "12"
4369 // 19 CheckCastPP "Foo"
4370 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4371 // 29 CheckCastPP "Foo" iid=24
4372 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4373 //
4374 // 40 StoreP 25 7 20 ... alias_index=4
4375 // 50 StoreP 35 40 30 ... alias_index=6
4376 // 60 StoreP 45 50 20 ... alias_index=4
4377 // 70 LoadP _ 60 30 ... alias_index=6
4378 // 80 Phi 75 50 60 Memory alias_index=4
4379 // 90 LoadP _ 80 30 ... alias_index=6
4380 // 100 LoadP _ 80 20 ... alias_index=4
4381 //
4382 // In phase 2, new memory inputs are computed for the loads and stores,
4383 // And a new version of the phi is created. In phase 4, the inputs to
4384 // node 80 are updated and then the memory nodes are updated with the
4385 // values computed in phase 2. This results in:
4386 //
4387 // 7 Parm #memory
4388 // 10 ConI "12"
4389 // 19 CheckCastPP "Foo"
4390 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4391 // 29 CheckCastPP "Foo" iid=24
4392 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4393 //
4394 // 40 StoreP 25 7 20 ... alias_index=4
4395 // 50 StoreP 35 7 30 ... alias_index=6
4396 // 60 StoreP 45 40 20 ... alias_index=4
4397 // 70 LoadP _ 50 30 ... alias_index=6
4398 // 80 Phi 75 40 60 Memory alias_index=4
4399 // 120 Phi 75 50 50 Memory alias_index=6
4400 // 90 LoadP _ 120 30 ... alias_index=6
4401 // 100 LoadP _ 80 20 ... alias_index=4
4402 //
4403 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4404 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4405 GrowableArray<MergeMemNode*> &mergemem_worklist,
4406 Unique_Node_List &reducible_merges) {
4407 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4408 GrowableArray<Node *> memnode_worklist;
4409 GrowableArray<PhiNode *> orig_phis;
4410 PhaseIterGVN *igvn = _igvn;
4411 uint new_index_start = (uint) _compile->num_alias_types();
4412 VectorSet visited;
4413 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4414 uint unique_old = _compile->unique();
4415
4416 // Phase 1: Process possible allocations from alloc_worklist.
4417 // Create instance types for the CheckCastPP for allocations where possible.
4418 //
4419 // (Note: don't forget to change the order of the second AddP node on
4420 // the alloc_worklist if the order of the worklist processing is changed,
4421 // see the comment in find_second_addp().)
4422 //
4423 while (alloc_worklist.length() != 0) {
4424 Node *n = alloc_worklist.pop();
4425 uint ni = n->_idx;
4426 if (n->is_Call()) {
4427 CallNode *alloc = n->as_Call();
4428 // copy escape information to call node
4429 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4430 PointsToNode::EscapeState es = ptn->escape_state();
4431 // We have an allocation or call which returns a Java object,
4432 // see if it is non-escaped.
4433 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4434 continue;
4435 }
4436 // Find CheckCastPP for the allocate or for the return value of a call
4437 n = alloc->result_cast();
4438 if (n == nullptr) { // No uses except Initialize node
4439 if (alloc->is_Allocate()) {
4440 // Set the scalar_replaceable flag for allocation
4441 // so it could be eliminated if it has no uses.
4442 alloc->as_Allocate()->_is_scalar_replaceable = true;
4443 }
4444 continue;
4445 }
4446 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4447 // we could reach here for allocate case if one init is associated with many allocs.
4448 if (alloc->is_Allocate()) {
4449 alloc->as_Allocate()->_is_scalar_replaceable = false;
4450 }
4451 continue;
4452 }
4453
4454 // The inline code for Object.clone() casts the allocation result to
4455 // java.lang.Object and then to the actual type of the allocated
4456 // object. Detect this case and use the second cast.
4457 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4458 // the allocation result is cast to java.lang.Object and then
4459 // to the actual Array type.
4460 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4461 && (alloc->is_AllocateArray() ||
4462 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4463 Node *cast2 = nullptr;
4464 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4465 Node *use = n->fast_out(i);
4466 if (use->is_CheckCastPP()) {
4467 cast2 = use;
4468 break;
4469 }
4470 }
4471 if (cast2 != nullptr) {
4472 n = cast2;
4473 } else {
4474 // Non-scalar replaceable if the allocation type is unknown statically
4475 // (reflection allocation), the object can't be restored during
4476 // deoptimization without precise type.
4477 continue;
4478 }
4479 }
4480
4481 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4482 if (t == nullptr) {
4483 continue; // not a TypeOopPtr
4484 }
4485 if (!t->klass_is_exact()) {
4486 continue; // not an unique type
4487 }
4488 if (alloc->is_Allocate()) {
4489 // Set the scalar_replaceable flag for allocation
4490 // so it could be eliminated.
4491 alloc->as_Allocate()->_is_scalar_replaceable = true;
4492 }
4493 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4494 // in order for an object to be scalar-replaceable, it must be:
4495 // - a direct allocation (not a call returning an object)
4496 // - non-escaping
4497 // - eligible to be a unique type
4498 // - not determined to be ineligible by escape analysis
4499 set_map(alloc, n);
4500 set_map(n, alloc);
4501 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4502 igvn->hash_delete(n);
4503 igvn->set_type(n, tinst);
4504 n->raise_bottom_type(tinst);
4505 igvn->hash_insert(n);
4506 record_for_optimizer(n);
4507 // Allocate an alias index for the header fields. Accesses to
4508 // the header emitted during macro expansion wouldn't have
4509 // correct memory state otherwise.
4510 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4511 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4512 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4513 // Add a new NarrowMem projection for each existing NarrowMem projection with new adr type
4514 InitializeNode* init = alloc->as_Allocate()->initialization();
4515 assert(init != nullptr, "can't find Initialization node for this Allocate node");
4516 auto process_narrow_proj = [&](NarrowMemProjNode* proj) {
4517 const TypePtr* adr_type = proj->adr_type();
4518 const TypePtr* new_adr_type = tinst->add_offset(adr_type->offset());
4519 if (adr_type != new_adr_type && !init->already_has_narrow_mem_proj_with_adr_type(new_adr_type)) {
4520 DEBUG_ONLY( uint alias_idx = _compile->get_alias_index(new_adr_type); )
4521 assert(_compile->get_general_index(alias_idx) == _compile->get_alias_index(adr_type), "new adr type should be narrowed down from existing adr type");
4522 NarrowMemProjNode* new_proj = new NarrowMemProjNode(init, new_adr_type);
4523 igvn->set_type(new_proj, new_proj->bottom_type());
4524 record_for_optimizer(new_proj);
4525 set_map(proj, new_proj); // record it so ConnectionGraph::find_inst_mem() can find it
4526 }
4527 };
4528 init->for_each_narrow_mem_proj_with_new_uses(process_narrow_proj);
4529
4530 // First, put on the worklist all Field edges from Connection Graph
4531 // which is more accurate than putting immediate users from Ideal Graph.
4532 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4533 PointsToNode* tgt = e.get();
4534 if (tgt->is_Arraycopy()) {
4535 continue;
4536 }
4537 Node* use = tgt->ideal_node();
4538 assert(tgt->is_Field() && use->is_AddP(),
4539 "only AddP nodes are Field edges in CG");
4540 if (use->outcnt() > 0) { // Don't process dead nodes
4541 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4542 if (addp2 != nullptr) {
4543 assert(alloc->is_AllocateArray(),"array allocation was expected");
4544 alloc_worklist.append_if_missing(addp2);
4545 }
4546 alloc_worklist.append_if_missing(use);
4547 }
4548 }
4549
4550 // An allocation may have an Initialize which has raw stores. Scan
4551 // the users of the raw allocation result and push AddP users
4552 // on alloc_worklist.
4553 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4554 assert (raw_result != nullptr, "must have an allocation result");
4555 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4556 Node *use = raw_result->fast_out(i);
4557 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4558 Node* addp2 = find_second_addp(use, raw_result);
4559 if (addp2 != nullptr) {
4560 assert(alloc->is_AllocateArray(),"array allocation was expected");
4561 alloc_worklist.append_if_missing(addp2);
4562 }
4563 alloc_worklist.append_if_missing(use);
4564 } else if (use->is_MemBar()) {
4565 memnode_worklist.append_if_missing(use);
4566 }
4567 }
4568 }
4569 } else if (n->is_AddP()) {
4570 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4571 // This AddP will go away when we reduce the Phi
4572 continue;
4573 }
4574 Node* addp_base = get_addp_base(n);
4575 JavaObjectNode* jobj = unique_java_object(addp_base);
4576 if (jobj == nullptr || jobj == phantom_obj) {
4577 #ifdef ASSERT
4578 ptnode_adr(get_addp_base(n)->_idx)->dump();
4579 ptnode_adr(n->_idx)->dump();
4580 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4581 #endif
4582 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4583 return;
4584 }
4585 Node *base = get_map(jobj->idx()); // CheckCastPP node
4586 if (!split_AddP(n, base)) continue; // wrong type from dead path
4587 } else if (n->is_Phi() ||
4588 n->is_CheckCastPP() ||
4589 n->is_EncodeP() ||
4590 n->is_DecodeN() ||
4591 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4592 if (visited.test_set(n->_idx)) {
4593 assert(n->is_Phi(), "loops only through Phi's");
4594 continue; // already processed
4595 }
4596 // Reducible Phi's will be removed from the graph after split_unique_types
4597 // finishes. For now we just try to split out the SR inputs of the merge.
4598 Node* parent = n->in(1);
4599 if (reducible_merges.member(n)) {
4600 reduce_phi(n->as_Phi(), alloc_worklist);
4601 #ifdef ASSERT
4602 if (VerifyReduceAllocationMerges) {
4603 reduced_merges.push(n);
4604 }
4605 #endif
4606 continue;
4607 } else if (reducible_merges.member(parent)) {
4608 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4609 // part of reduce_merge.
4610 continue;
4611 }
4612 JavaObjectNode* jobj = unique_java_object(n);
4613 if (jobj == nullptr || jobj == phantom_obj) {
4614 #ifdef ASSERT
4615 ptnode_adr(n->_idx)->dump();
4616 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4617 #endif
4618 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4619 return;
4620 } else {
4621 Node *val = get_map(jobj->idx()); // CheckCastPP node
4622 TypeNode *tn = n->as_Type();
4623 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4624 assert(tinst != nullptr && tinst->is_known_instance() &&
4625 tinst->instance_id() == jobj->idx() , "instance type expected.");
4626
4627 const Type *tn_type = igvn->type(tn);
4628 const TypeOopPtr *tn_t;
4629 if (tn_type->isa_narrowoop()) {
4630 tn_t = tn_type->make_ptr()->isa_oopptr();
4631 } else {
4632 tn_t = tn_type->isa_oopptr();
4633 }
4634 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4635 if (tn_type->isa_narrowoop()) {
4636 tn_type = tinst->make_narrowoop();
4637 } else {
4638 tn_type = tinst;
4639 }
4640 igvn->hash_delete(tn);
4641 igvn->set_type(tn, tn_type);
4642 tn->set_type(tn_type);
4643 igvn->hash_insert(tn);
4644 record_for_optimizer(n);
4645 } else {
4646 assert(tn_type == TypePtr::NULL_PTR ||
4647 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4648 "unexpected type");
4649 continue; // Skip dead path with different type
4650 }
4651 }
4652 } else {
4653 DEBUG_ONLY(n->dump();)
4654 assert(false, "EA: unexpected node");
4655 continue;
4656 }
4657 // push allocation's users on appropriate worklist
4658 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4659 Node *use = n->fast_out(i);
4660 if(use->is_Mem() && use->in(MemNode::Address) == n) {
4661 // Load/store to instance's field
4662 memnode_worklist.append_if_missing(use);
4663 } else if (use->is_MemBar()) {
4664 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4665 memnode_worklist.append_if_missing(use);
4666 }
4667 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4668 Node* addp2 = find_second_addp(use, n);
4669 if (addp2 != nullptr) {
4670 alloc_worklist.append_if_missing(addp2);
4671 }
4672 alloc_worklist.append_if_missing(use);
4673 } else if (use->is_Phi() ||
4674 use->is_CheckCastPP() ||
4675 use->is_EncodeNarrowPtr() ||
4676 use->is_DecodeNarrowPtr() ||
4677 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4678 alloc_worklist.append_if_missing(use);
4679 #ifdef ASSERT
4680 } else if (use->is_Mem()) {
4681 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4682 } else if (use->is_MergeMem()) {
4683 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4684 } else if (use->is_SafePoint()) {
4685 // Look for MergeMem nodes for calls which reference unique allocation
4686 // (through CheckCastPP nodes) even for debug info.
4687 Node* m = use->in(TypeFunc::Memory);
4688 if (m->is_MergeMem()) {
4689 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4690 }
4691 } else if (use->Opcode() == Op_EncodeISOArray) {
4692 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4693 // EncodeISOArray overwrites destination array
4694 memnode_worklist.append_if_missing(use);
4695 }
4696 } else {
4697 uint op = use->Opcode();
4698 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4699 (use->in(MemNode::Memory) == n)) {
4700 // They overwrite memory edge corresponding to destination array,
4701 memnode_worklist.append_if_missing(use);
4702 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4703 op == Op_CastP2X ||
4704 op == Op_FastLock || op == Op_AryEq ||
4705 op == Op_StrComp || op == Op_CountPositives ||
4706 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4707 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4708 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4709 op == Op_SubTypeCheck ||
4710 op == Op_ReinterpretS2HF ||
4711 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4712 n->dump();
4713 use->dump();
4714 assert(false, "EA: missing allocation reference path");
4715 }
4716 #endif
4717 }
4718 }
4719
4720 }
4721
4722 #ifdef ASSERT
4723 if (VerifyReduceAllocationMerges) {
4724 for (uint i = 0; i < reducible_merges.size(); i++) {
4725 Node* phi = reducible_merges.at(i);
4726
4727 if (!reduced_merges.member(phi)) {
4728 phi->dump(2);
4729 phi->dump(-2);
4730 assert(false, "This reducible merge wasn't reduced.");
4731 }
4732
4733 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4734 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4735 Node* use = phi->fast_out(j);
4736 if (!use->is_SafePoint() && !use->is_CastPP()) {
4737 phi->dump(2);
4738 phi->dump(-2);
4739 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4740 }
4741 }
4742 }
4743 }
4744 #endif
4745
4746 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4747 // type, record it in the ArrayCopy node so we know what memory this
4748 // node uses/modified.
4749 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4750 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4751 Node* dest = ac->in(ArrayCopyNode::Dest);
4752 if (dest->is_AddP()) {
4753 dest = get_addp_base(dest);
4754 }
4755 JavaObjectNode* jobj = unique_java_object(dest);
4756 if (jobj != nullptr) {
4757 Node *base = get_map(jobj->idx());
4758 if (base != nullptr) {
4759 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4760 ac->_dest_type = base_t;
4761 }
4762 }
4763 Node* src = ac->in(ArrayCopyNode::Src);
4764 if (src->is_AddP()) {
4765 src = get_addp_base(src);
4766 }
4767 jobj = unique_java_object(src);
4768 if (jobj != nullptr) {
4769 Node* base = get_map(jobj->idx());
4770 if (base != nullptr) {
4771 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4772 ac->_src_type = base_t;
4773 }
4774 }
4775 }
4776
4777 // New alias types were created in split_AddP().
4778 uint new_index_end = (uint) _compile->num_alias_types();
4779
4780 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_1, 5);
4781
4782 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4783 // compute new values for Memory inputs (the Memory inputs are not
4784 // actually updated until phase 4.)
4785 if (memnode_worklist.length() == 0)
4786 return; // nothing to do
4787 while (memnode_worklist.length() != 0) {
4788 Node *n = memnode_worklist.pop();
4789 if (visited.test_set(n->_idx)) {
4790 continue;
4791 }
4792 if (n->is_Phi() || n->is_ClearArray()) {
4793 // we don't need to do anything, but the users must be pushed
4794 } else if (n->is_MemBar()) { // MemBar nodes
4795 if (!n->is_Initialize()) { // memory projections for Initialize pushed below (so we get to all their uses)
4796 // we don't need to do anything, but the users must be pushed
4797 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4798 if (n == nullptr) {
4799 continue;
4800 }
4801 }
4802 } else if (n->is_CallLeaf()) {
4803 // Runtime calls with narrow memory input (no MergeMem node)
4804 // get the memory projection
4805 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4806 if (n == nullptr) {
4807 continue;
4808 }
4809 } else if (n->Opcode() == Op_StrInflatedCopy) {
4810 // Check direct uses of StrInflatedCopy.
4811 // It is memory type Node - no special SCMemProj node.
4812 } else if (n->Opcode() == Op_StrCompressedCopy ||
4813 n->Opcode() == Op_EncodeISOArray) {
4814 // get the memory projection
4815 n = n->find_out_with(Op_SCMemProj);
4816 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4817 } else if (n->is_Proj()) {
4818 assert(n->in(0)->is_Initialize(), "we only push memory projections for Initialize");
4819 } else {
4820 #ifdef ASSERT
4821 if (!n->is_Mem()) {
4822 n->dump();
4823 }
4824 assert(n->is_Mem(), "memory node required.");
4825 #endif
4826 Node *addr = n->in(MemNode::Address);
4827 const Type *addr_t = igvn->type(addr);
4828 if (addr_t == Type::TOP) {
4829 continue;
4830 }
4831 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4832 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4833 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4834 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4835 if (_compile->failing()) {
4836 return;
4837 }
4838 if (mem != n->in(MemNode::Memory)) {
4839 // We delay the memory edge update since we need old one in
4840 // MergeMem code below when instances memory slices are separated.
4841 set_map(n, mem);
4842 }
4843 if (n->is_Load()) {
4844 continue; // don't push users
4845 } else if (n->is_LoadStore()) {
4846 // get the memory projection
4847 n = n->find_out_with(Op_SCMemProj);
4848 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4849 }
4850 }
4851 // push user on appropriate worklist
4852 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4853 Node *use = n->fast_out(i);
4854 if (use->is_Phi() || use->is_ClearArray()) {
4855 memnode_worklist.append_if_missing(use);
4856 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4857 memnode_worklist.append_if_missing(use);
4858 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4859 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4860 memnode_worklist.append_if_missing(use);
4861 }
4862 } else if (use->is_Proj()) {
4863 assert(n->is_Initialize(), "We only push projections of Initialize");
4864 if (use->as_Proj()->_con == TypeFunc::Memory) { // Ignore precedent edge
4865 memnode_worklist.append_if_missing(use);
4866 }
4867 #ifdef ASSERT
4868 } else if(use->is_Mem()) {
4869 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4870 } else if (use->is_MergeMem()) {
4871 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4872 } else if (use->Opcode() == Op_EncodeISOArray) {
4873 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4874 // EncodeISOArray overwrites destination array
4875 memnode_worklist.append_if_missing(use);
4876 }
4877 } else {
4878 uint op = use->Opcode();
4879 if ((use->in(MemNode::Memory) == n) &&
4880 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4881 // They overwrite memory edge corresponding to destination array,
4882 memnode_worklist.append_if_missing(use);
4883 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4884 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4885 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4886 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar)) {
4887 n->dump();
4888 use->dump();
4889 assert(false, "EA: missing memory path");
4890 }
4891 #endif
4892 }
4893 }
4894 }
4895
4896 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4897 // Walk each memory slice moving the first node encountered of each
4898 // instance type to the input corresponding to its alias index.
4899 uint length = mergemem_worklist.length();
4900 for( uint next = 0; next < length; ++next ) {
4901 MergeMemNode* nmm = mergemem_worklist.at(next);
4902 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4903 // Note: we don't want to use MergeMemStream here because we only want to
4904 // scan inputs which exist at the start, not ones we add during processing.
4905 // Note 2: MergeMem may already contains instance memory slices added
4906 // during find_inst_mem() call when memory nodes were processed above.
4907 igvn->hash_delete(nmm);
4908 uint nslices = MIN2(nmm->req(), new_index_start);
4909 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4910 Node* mem = nmm->in(i);
4911 Node* cur = nullptr;
4912 if (mem == nullptr || mem->is_top()) {
4913 continue;
4914 }
4915 // First, update mergemem by moving memory nodes to corresponding slices
4916 // if their type became more precise since this mergemem was created.
4917 while (mem->is_Mem()) {
4918 const Type* at = igvn->type(mem->in(MemNode::Address));
4919 if (at != Type::TOP) {
4920 assert (at->isa_ptr() != nullptr, "pointer type required.");
4921 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4922 if (idx == i) {
4923 if (cur == nullptr) {
4924 cur = mem;
4925 }
4926 } else {
4927 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4928 nmm->set_memory_at(idx, mem);
4929 }
4930 }
4931 }
4932 mem = mem->in(MemNode::Memory);
4933 }
4934 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4935 // Find any instance of the current type if we haven't encountered
4936 // already a memory slice of the instance along the memory chain.
4937 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4938 if((uint)_compile->get_general_index(ni) == i) {
4939 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4940 if (nmm->is_empty_memory(m)) {
4941 Node* result = find_inst_mem(mem, ni, orig_phis);
4942 if (_compile->failing()) {
4943 return;
4944 }
4945 nmm->set_memory_at(ni, result);
4946 }
4947 }
4948 }
4949 }
4950 // Find the rest of instances values
4951 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4952 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4953 Node* result = step_through_mergemem(nmm, ni, tinst);
4954 if (result == nmm->base_memory()) {
4955 // Didn't find instance memory, search through general slice recursively.
4956 result = nmm->memory_at(_compile->get_general_index(ni));
4957 result = find_inst_mem(result, ni, orig_phis);
4958 if (_compile->failing()) {
4959 return;
4960 }
4961 nmm->set_memory_at(ni, result);
4962 }
4963 }
4964
4965 // If we have crossed the 3/4 point of max node limit it's too risky
4966 // to continue with EA/SR because we might hit the max node limit.
4967 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4968 if (_compile->do_reduce_allocation_merges()) {
4969 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4970 } else if (_invocation > 0) {
4971 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4972 } else {
4973 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4974 }
4975 return;
4976 }
4977
4978 igvn->hash_insert(nmm);
4979 record_for_optimizer(nmm);
4980 }
4981
4982 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_3, 5);
4983
4984 // Phase 4: Update the inputs of non-instance memory Phis and
4985 // the Memory input of memnodes
4986 // First update the inputs of any non-instance Phi's from
4987 // which we split out an instance Phi. Note we don't have
4988 // to recursively process Phi's encountered on the input memory
4989 // chains as is done in split_memory_phi() since they will
4990 // also be processed here.
4991 for (int j = 0; j < orig_phis.length(); j++) {
4992 PhiNode *phi = orig_phis.at(j);
4993 int alias_idx = _compile->get_alias_index(phi->adr_type());
4994 igvn->hash_delete(phi);
4995 for (uint i = 1; i < phi->req(); i++) {
4996 Node *mem = phi->in(i);
4997 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
4998 if (_compile->failing()) {
4999 return;
5000 }
5001 if (mem != new_mem) {
5002 phi->set_req(i, new_mem);
5003 }
5004 }
5005 igvn->hash_insert(phi);
5006 record_for_optimizer(phi);
5007 }
5008
5009 // Update the memory inputs of MemNodes with the value we computed
5010 // in Phase 2 and move stores memory users to corresponding memory slices.
5011 // Disable memory split verification code until the fix for 6984348.
5012 // Currently it produces false negative results since it does not cover all cases.
5013 #if 0 // ifdef ASSERT
5014 visited.Reset();
5015 Node_Stack old_mems(arena, _compile->unique() >> 2);
5016 #endif
5017 for (uint i = 0; i < ideal_nodes.size(); i++) {
5018 Node* n = ideal_nodes.at(i);
5019 Node* nmem = get_map(n->_idx);
5020 assert(nmem != nullptr, "sanity");
5021 if (n->is_Mem()) {
5022 #if 0 // ifdef ASSERT
5023 Node* old_mem = n->in(MemNode::Memory);
5024 if (!visited.test_set(old_mem->_idx)) {
5025 old_mems.push(old_mem, old_mem->outcnt());
5026 }
5027 #endif
5028 assert(n->in(MemNode::Memory) != nmem, "sanity");
5029 if (!n->is_Load()) {
5030 // Move memory users of a store first.
5031 move_inst_mem(n, orig_phis);
5032 }
5033 // Now update memory input
5034 igvn->hash_delete(n);
5035 n->set_req(MemNode::Memory, nmem);
5036 igvn->hash_insert(n);
5037 record_for_optimizer(n);
5038 } else {
5039 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5040 n->is_AddP() || n->is_Phi() || n->is_NarrowMemProj(), "unknown node used for set_map()");
5041 }
5042 }
5043 #if 0 // ifdef ASSERT
5044 // Verify that memory was split correctly
5045 while (old_mems.is_nonempty()) {
5046 Node* old_mem = old_mems.node();
5047 uint old_cnt = old_mems.index();
5048 old_mems.pop();
5049 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5050 }
5051 #endif
5052 _compile->print_method(PHASE_EA_AFTER_SPLIT_UNIQUE_TYPES_4, 5);
5053 }
5054
5055 #ifndef PRODUCT
5056 int ConnectionGraph::_no_escape_counter = 0;
5057 int ConnectionGraph::_arg_escape_counter = 0;
5058 int ConnectionGraph::_global_escape_counter = 0;
5059
5060 static const char *node_type_names[] = {
5061 "UnknownType",
5062 "JavaObject",
5063 "LocalVar",
5064 "Field",
5065 "Arraycopy"
5066 };
5067
5068 static const char *esc_names[] = {
5069 "UnknownEscape",
5070 "NoEscape",
5071 "ArgEscape",
5072 "GlobalEscape"
5073 };
5074
5075 const char* PointsToNode::esc_name() const {
5076 return esc_names[(int)escape_state()];
5077 }
5078
5079 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5080 NodeType nt = node_type();
5081 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5082 if (print_state) {
5083 EscapeState es = escape_state();
5084 EscapeState fields_es = fields_escape_state();
5085 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5086 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5087 out->print("NSR ");
5088 }
5089 }
5090 }
5091
5092 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5093 dump_header(print_state, out);
5094 if (is_Field()) {
5095 FieldNode* f = (FieldNode*)this;
5096 if (f->is_oop()) {
5097 out->print("oop ");
5098 }
5099 if (f->offset() > 0) {
5100 out->print("+%d ", f->offset());
5101 }
5102 out->print("(");
5103 for (BaseIterator i(f); i.has_next(); i.next()) {
5104 PointsToNode* b = i.get();
5105 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5106 }
5107 out->print(" )");
5108 }
5109 out->print("[");
5110 for (EdgeIterator i(this); i.has_next(); i.next()) {
5111 PointsToNode* e = i.get();
5112 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5113 }
5114 out->print(" [");
5115 for (UseIterator i(this); i.has_next(); i.next()) {
5116 PointsToNode* u = i.get();
5117 bool is_base = false;
5118 if (PointsToNode::is_base_use(u)) {
5119 is_base = true;
5120 u = PointsToNode::get_use_node(u)->as_Field();
5121 }
5122 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5123 }
5124 out->print(" ]] ");
5125 if (_node == nullptr) {
5126 out->print("<null>%s", newline ? "\n" : "");
5127 } else {
5128 _node->dump(newline ? "\n" : "", false, out);
5129 }
5130 }
5131
5132 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5133 bool first = true;
5134 int ptnodes_length = ptnodes_worklist.length();
5135 for (int i = 0; i < ptnodes_length; i++) {
5136 PointsToNode *ptn = ptnodes_worklist.at(i);
5137 if (ptn == nullptr || !ptn->is_JavaObject()) {
5138 continue;
5139 }
5140 PointsToNode::EscapeState es = ptn->escape_state();
5141 if ((es != PointsToNode::NoEscape) && !Verbose) {
5142 continue;
5143 }
5144 Node* n = ptn->ideal_node();
5145 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5146 n->as_CallStaticJava()->is_boxing_method())) {
5147 if (first) {
5148 tty->cr();
5149 tty->print("======== Connection graph for ");
5150 _compile->method()->print_short_name();
5151 tty->cr();
5152 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5153 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5154 tty->cr();
5155 first = false;
5156 }
5157 ptn->dump();
5158 // Print all locals and fields which reference this allocation
5159 for (UseIterator j(ptn); j.has_next(); j.next()) {
5160 PointsToNode* use = j.get();
5161 if (use->is_LocalVar()) {
5162 use->dump(Verbose);
5163 } else if (Verbose) {
5164 use->dump();
5165 }
5166 }
5167 tty->cr();
5168 }
5169 }
5170 }
5171
5172 void ConnectionGraph::print_statistics() {
5173 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));
5174 }
5175
5176 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5177 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5178 return;
5179 }
5180 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5181 JavaObjectNode* ptn = java_objects_worklist.at(next);
5182 if (ptn->ideal_node()->is_Allocate()) {
5183 if (ptn->escape_state() == PointsToNode::NoEscape) {
5184 AtomicAccess::inc(&ConnectionGraph::_no_escape_counter);
5185 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5186 AtomicAccess::inc(&ConnectionGraph::_arg_escape_counter);
5187 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5188 AtomicAccess::inc(&ConnectionGraph::_global_escape_counter);
5189 } else {
5190 assert(false, "Unexpected Escape State");
5191 }
5192 }
5193 }
5194 }
5195
5196 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5197 if (_compile->directive()->TraceEscapeAnalysisOption) {
5198 assert(ptn != nullptr, "should not be null");
5199 assert(reason != nullptr, "should not be null");
5200 ptn->dump_header(true);
5201 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5202 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5203 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5204 }
5205 }
5206
5207 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5208 if (_compile->directive()->TraceEscapeAnalysisOption) {
5209 stringStream ss;
5210 ss.print("propagated from: ");
5211 from->dump(true, &ss, false);
5212 return ss.as_string();
5213 } else {
5214 return nullptr;
5215 }
5216 }
5217
5218 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5219 if (_compile->directive()->TraceEscapeAnalysisOption) {
5220 stringStream ss;
5221 ss.print("escapes as arg to:");
5222 call->dump("", false, &ss);
5223 return ss.as_string();
5224 } else {
5225 return nullptr;
5226 }
5227 }
5228
5229 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5230 if (_compile->directive()->TraceEscapeAnalysisOption) {
5231 stringStream ss;
5232 ss.print("is merged with other object: ");
5233 other->dump_header(true, &ss);
5234 return ss.as_string();
5235 } else {
5236 return nullptr;
5237 }
5238 }
5239
5240 #endif
5241
5242 void ConnectionGraph::record_for_optimizer(Node *n) {
5243 _igvn->_worklist.push(n);
5244 _igvn->add_users_to_worklist(n);
5245 }
--- EOF ---