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