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/metaspace.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/c2compiler.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/escape.hpp"
39 #include "opto/inlinetypenode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/castnode.hpp"
46 #include "opto/rootnode.hpp"
47 #include "utilities/macros.hpp"
48
49 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
50 // If ReduceAllocationMerges is enabled we might call split_through_phi during
51 // split_unique_types and that will create additional nodes that need to be
52 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
53 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
54 // the array will be reallocated.
55 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
56 _in_worklist(C->comp_arena()),
57 _next_pidx(0),
58 _collecting(true),
59 _verify(false),
60 _compile(C),
61 _igvn(igvn),
62 _invocation(invocation),
63 _build_iterations(0),
64 _build_time(0.),
65 _node_map(C->comp_arena()) {
66 // Add unknown java object.
67 add_java_object(C->top(), PointsToNode::GlobalEscape);
68 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
69 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
70 // Add ConP and ConN null oop nodes
71 Node* oop_null = igvn->zerocon(T_OBJECT);
72 assert(oop_null->_idx < nodes_size(), "should be created already");
73 add_java_object(oop_null, PointsToNode::NoEscape);
74 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
75 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
76 if (UseCompressedOops) {
77 Node* noop_null = igvn->zerocon(T_NARROWOOP);
78 assert(noop_null->_idx < nodes_size(), "should be created already");
79 map_ideal_node(noop_null, null_obj);
80 }
81 }
82
83 bool ConnectionGraph::has_candidates(Compile *C) {
84 // EA brings benefits only when the code has allocations and/or locks which
85 // are represented by ideal Macro nodes.
86 int cnt = C->macro_count();
87 for (int i = 0; i < cnt; i++) {
88 Node *n = C->macro_node(i);
89 if (n->is_Allocate()) {
90 return true;
91 }
92 if (n->is_Lock()) {
93 Node* obj = n->as_Lock()->obj_node()->uncast();
94 if (!(obj->is_Parm() || obj->is_Con())) {
95 return true;
96 }
97 }
98 if (n->is_CallStaticJava() &&
99 n->as_CallStaticJava()->is_boxing_method()) {
100 return true;
101 }
102 }
103 return false;
104 }
105
106 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
107 Compile::TracePhase tp(Phase::_t_escapeAnalysis);
108 ResourceMark rm;
109
110 // Add ConP and ConN null oop nodes before ConnectionGraph construction
111 // to create space for them in ConnectionGraph::_nodes[].
112 Node* oop_null = igvn->zerocon(T_OBJECT);
113 Node* noop_null = igvn->zerocon(T_NARROWOOP);
114 int invocation = 0;
115 if (C->congraph() != nullptr) {
116 invocation = C->congraph()->_invocation + 1;
117 }
118 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
119 // Perform escape analysis
120 if (congraph->compute_escape()) {
121 // There are non escaping objects.
122 C->set_congraph(congraph);
123 }
124 // Cleanup.
125 if (oop_null->outcnt() == 0) {
126 igvn->hash_delete(oop_null);
127 }
128 if (noop_null->outcnt() == 0) {
129 igvn->hash_delete(noop_null);
130 }
131 }
132
133 bool ConnectionGraph::compute_escape() {
134 Compile* C = _compile;
135 PhaseGVN* igvn = _igvn;
136
137 // Worklists used by EA.
138 Unique_Node_List delayed_worklist;
139 Unique_Node_List reducible_merges;
140 GrowableArray<Node*> alloc_worklist;
141 GrowableArray<Node*> ptr_cmp_worklist;
142 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
143 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
144 GrowableArray<PointsToNode*> ptnodes_worklist;
145 GrowableArray<JavaObjectNode*> java_objects_worklist;
146 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
147 GrowableArray<FieldNode*> oop_fields_worklist;
148 GrowableArray<SafePointNode*> sfn_worklist;
149 GrowableArray<MergeMemNode*> mergemem_worklist;
150 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
151
152 { Compile::TracePhase tp(Phase::_t_connectionGraph);
153
154 // 1. Populate Connection Graph (CG) with PointsTo nodes.
155 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
156 // Initialize worklist
157 if (C->root() != nullptr) {
158 ideal_nodes.push(C->root());
159 }
160 // Processed ideal nodes are unique on ideal_nodes list
161 // but several ideal nodes are mapped to the phantom_obj.
162 // To avoid duplicated entries on the following worklists
163 // add the phantom_obj only once to them.
164 ptnodes_worklist.append(phantom_obj);
165 java_objects_worklist.append(phantom_obj);
166 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
167 Node* n = ideal_nodes.at(next);
168 if ((n->Opcode() == Op_LoadX || n->Opcode() == Op_StoreX) &&
169 !n->in(MemNode::Address)->is_AddP() &&
170 _igvn->type(n->in(MemNode::Address))->isa_oopptr()) {
171 // Load/Store at mark work address is at offset 0 so has no AddP which confuses EA
172 Node* addp = new AddPNode(n->in(MemNode::Address), n->in(MemNode::Address), _igvn->MakeConX(0));
173 _igvn->register_new_node_with_optimizer(addp);
174 _igvn->replace_input_of(n, MemNode::Address, addp);
175 ideal_nodes.push(addp);
176 _nodes.at_put_grow(addp->_idx, nullptr, nullptr);
177 }
178 // Create PointsTo nodes and add them to Connection Graph. Called
179 // only once per ideal node since ideal_nodes is Unique_Node list.
180 add_node_to_connection_graph(n, &delayed_worklist);
181 PointsToNode* ptn = ptnode_adr(n->_idx);
182 if (ptn != nullptr && ptn != phantom_obj) {
183 ptnodes_worklist.append(ptn);
184 if (ptn->is_JavaObject()) {
185 java_objects_worklist.append(ptn->as_JavaObject());
186 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
187 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
188 // Only allocations and java static calls results are interesting.
189 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
190 }
191 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
192 oop_fields_worklist.append(ptn->as_Field());
193 }
194 }
195 // Collect some interesting nodes for further use.
196 switch (n->Opcode()) {
197 case Op_MergeMem:
198 // Collect all MergeMem nodes to add memory slices for
199 // scalar replaceable objects in split_unique_types().
200 mergemem_worklist.append(n->as_MergeMem());
201 break;
202 case Op_CmpP:
203 case Op_CmpN:
204 // Collect compare pointers nodes.
205 if (OptimizePtrCompare) {
206 ptr_cmp_worklist.append(n);
207 }
208 break;
209 case Op_MemBarStoreStore:
210 // Collect all MemBarStoreStore nodes so that depending on the
211 // escape status of the associated Allocate node some of them
212 // may be eliminated.
213 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
214 storestore_worklist.append(n->as_MemBarStoreStore());
215 }
216 break;
217 case Op_MemBarRelease:
218 if (n->req() > MemBarNode::Precedent) {
219 record_for_optimizer(n);
220 }
221 break;
222 #ifdef ASSERT
223 case Op_AddP:
224 // Collect address nodes for graph verification.
225 addp_worklist.append(n);
226 break;
227 #endif
228 case Op_ArrayCopy:
229 // Keep a list of ArrayCopy nodes so if one of its input is non
230 // escaping, we can record a unique type
231 arraycopy_worklist.append(n->as_ArrayCopy());
232 break;
233 default:
234 // not interested now, ignore...
235 break;
236 }
237 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
238 Node* m = n->fast_out(i); // Get user
239 ideal_nodes.push(m);
240 }
241 if (n->is_SafePoint()) {
242 sfn_worklist.append(n->as_SafePoint());
243 }
244 }
245
246 #ifndef PRODUCT
247 if (_compile->directive()->TraceEscapeAnalysisOption) {
248 tty->print("+++++ Initial worklist for ");
249 _compile->method()->print_name();
250 tty->print_cr(" (ea_inv=%d)", _invocation);
251 for (int i = 0; i < ptnodes_worklist.length(); i++) {
252 PointsToNode* ptn = ptnodes_worklist.at(i);
253 ptn->dump();
254 }
255 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
256 }
257 #endif
258
259 if (non_escaped_allocs_worklist.length() == 0) {
260 _collecting = false;
261 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
262 return false; // Nothing to do.
263 }
264 // Add final simple edges to graph.
265 while(delayed_worklist.size() > 0) {
266 Node* n = delayed_worklist.pop();
267 add_final_edges(n);
268 }
269
270 #ifdef ASSERT
271 if (VerifyConnectionGraph) {
272 // Verify that no new simple edges could be created and all
273 // local vars has edges.
274 _verify = true;
275 int ptnodes_length = ptnodes_worklist.length();
276 for (int next = 0; next < ptnodes_length; ++next) {
277 PointsToNode* ptn = ptnodes_worklist.at(next);
278 add_final_edges(ptn->ideal_node());
279 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
280 ptn->dump();
281 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
282 }
283 }
284 _verify = false;
285 }
286 #endif
287 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
288 // processing, calls to CI to resolve symbols (types, fields, methods)
289 // referenced in bytecode. During symbol resolution VM may throw
290 // an exception which CI cleans and converts to compilation failure.
291 if (C->failing()) {
292 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
293 return false;
294 }
295
296 // 2. Finish Graph construction by propagating references to all
297 // java objects through graph.
298 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
299 java_objects_worklist, oop_fields_worklist)) {
300 // All objects escaped or hit time or iterations limits.
301 _collecting = false;
302 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
303 return false;
304 }
305
306 // 3. Adjust scalar_replaceable state of nonescaping objects and push
307 // scalar replaceable allocations on alloc_worklist for processing
308 // in split_unique_types().
309 GrowableArray<JavaObjectNode*> jobj_worklist;
310 int non_escaped_length = non_escaped_allocs_worklist.length();
311 bool found_nsr_alloc = false;
312 for (int next = 0; next < non_escaped_length; next++) {
313 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
314 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
315 Node* n = ptn->ideal_node();
316 if (n->is_Allocate()) {
317 n->as_Allocate()->_is_non_escaping = noescape;
318 }
319 if (noescape && ptn->scalar_replaceable()) {
320 adjust_scalar_replaceable_state(ptn, reducible_merges);
321 if (ptn->scalar_replaceable()) {
322 jobj_worklist.push(ptn);
323 } else {
324 found_nsr_alloc = true;
325 }
326 }
327 }
328
329 // Propagate NSR (Not Scalar Replaceable) state.
330 if (found_nsr_alloc) {
331 find_scalar_replaceable_allocs(jobj_worklist, reducible_merges);
332 }
333
334 // alloc_worklist will be processed in reverse push order.
335 // Therefore the reducible Phis will be processed for last and that's what we
336 // want because by then the scalarizable inputs of the merge will already have
337 // an unique instance type.
338 for (uint i = 0; i < reducible_merges.size(); i++ ) {
339 Node* n = reducible_merges.at(i);
340 alloc_worklist.append(n);
341 }
342
343 for (int next = 0; next < jobj_worklist.length(); ++next) {
344 JavaObjectNode* jobj = jobj_worklist.at(next);
345 if (jobj->scalar_replaceable()) {
346 alloc_worklist.append(jobj->ideal_node());
347 }
348 }
349
350 #ifdef ASSERT
351 if (VerifyConnectionGraph) {
352 // Verify that graph is complete - no new edges could be added or needed.
353 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
354 java_objects_worklist, addp_worklist);
355 }
356 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
357 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
358 null_obj->edge_count() == 0 &&
359 !null_obj->arraycopy_src() &&
360 !null_obj->arraycopy_dst(), "sanity");
361 #endif
362
363 _collecting = false;
364
365 } // TracePhase t3("connectionGraph")
366
367 // 4. Optimize ideal graph based on EA information.
368 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
369 if (has_non_escaping_obj) {
370 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
371 }
372
373 #ifndef PRODUCT
374 if (PrintEscapeAnalysis) {
375 dump(ptnodes_worklist); // Dump ConnectionGraph
376 }
377 #endif
378
379 #ifdef ASSERT
380 if (VerifyConnectionGraph) {
381 int alloc_length = alloc_worklist.length();
382 for (int next = 0; next < alloc_length; ++next) {
383 Node* n = alloc_worklist.at(next);
384 PointsToNode* ptn = ptnode_adr(n->_idx);
385 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
386 }
387 }
388
389 if (VerifyReduceAllocationMerges) {
390 for (uint i = 0; i < reducible_merges.size(); i++ ) {
391 Node* n = reducible_merges.at(i);
392 if (!can_reduce_phi(n->as_Phi())) {
393 TraceReduceAllocationMerges = true;
394 n->dump(2);
395 n->dump(-2);
396 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
397 }
398 }
399 }
400 #endif
401
402 // 5. Separate memory graph for scalar replaceable allcations.
403 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
404 if (has_scalar_replaceable_candidates && EliminateAllocations) {
405 assert(C->do_aliasing(), "Aliasing should be enabled");
406 // Now use the escape information to create unique types for
407 // scalar replaceable objects.
408 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
409 if (C->failing()) {
410 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
411 return false;
412 }
413 C->print_method(PHASE_AFTER_EA, 2);
414
415 #ifdef ASSERT
416 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
417 tty->print("=== No allocations eliminated for ");
418 C->method()->print_short_name();
419 if (!EliminateAllocations) {
420 tty->print(" since EliminateAllocations is off ===");
421 } else if(!has_scalar_replaceable_candidates) {
422 tty->print(" since there are no scalar replaceable candidates ===");
423 }
424 tty->cr();
425 #endif
426 }
427
428 // 6. Reduce allocation merges used as debug information. This is done after
429 // split_unique_types because the methods used to create SafePointScalarObject
430 // need to traverse the memory graph to find values for object fields. We also
431 // set to null the scalarized inputs of reducible Phis so that the Allocate
432 // that they point can be later scalar replaced.
433 bool delay = _igvn->delay_transform();
434 _igvn->set_delay_transform(true);
435 for (uint i = 0; i < reducible_merges.size(); i++) {
436 Node* n = reducible_merges.at(i);
437 if (n->outcnt() > 0) {
438 if (!reduce_phi_on_safepoints(n->as_Phi())) {
439 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
440 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
441 return false;
442 }
443
444 // Now we set the scalar replaceable inputs of ophi to null, which is
445 // the last piece that would prevent it from being scalar replaceable.
446 reset_scalar_replaceable_entries(n->as_Phi());
447 }
448 }
449 _igvn->set_delay_transform(delay);
450
451 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
452 // java calls if they pass ArgEscape objects as parameters.
453 if (has_non_escaping_obj &&
454 (C->env()->should_retain_local_variables() ||
455 C->env()->jvmti_can_get_owned_monitor_info() ||
456 C->env()->jvmti_can_walk_any_space() ||
457 DeoptimizeObjectsALot)) {
458 int sfn_length = sfn_worklist.length();
459 for (int next = 0; next < sfn_length; next++) {
460 SafePointNode* sfn = sfn_worklist.at(next);
461 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
462 if (sfn->is_CallJava()) {
463 CallJavaNode* call = sfn->as_CallJava();
464 call->set_arg_escape(has_arg_escape(call));
465 }
466 }
467 }
468
469 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
470 return has_non_escaping_obj;
471 }
472
473 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
474 // if at least one scalar replaceable allocation participates in the merge.
475 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
476 bool found_sr_allocate = false;
477
478 for (uint i = 1; i < ophi->req(); i++) {
479 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
480 if (ptn != nullptr && ptn->scalar_replaceable()) {
481 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
482
483 // Don't handle arrays.
484 if (alloc->Opcode() != Op_Allocate) {
485 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
486 continue;
487 }
488
489 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
490 found_sr_allocate = true;
491 } else {
492 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
493 ptn->set_scalar_replaceable(false);
494 }
495 }
496 }
497
498 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);)
499 return found_sr_allocate;
500 }
501
502 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
503 // I require the 'other' input to be a constant so that I can move the Cmp
504 // around safely.
505 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
506 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
507 Node* left = cmp->in(1);
508 Node* right = cmp->in(2);
509
510 return (left == n || right == n) &&
511 (left->is_Con() || right->is_Con()) &&
512 cmp->outcnt() == 1;
513 }
514
515 // We are going to check if any of the SafePointScalarMerge entries
516 // in the SafePoint reference the Phi that we are checking.
517 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
518 JVMState *jvms = sfpt->jvms();
519
520 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
521 Node* sfpt_in = sfpt->in(i);
522 if (sfpt_in->is_SafePointScalarMerge()) {
523 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
524 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
525 if (nsr_ptr == n) {
526 return true;
527 }
528 }
529 }
530
531 return false;
532 }
533
534 // Check if we are able to untangle the merge. The following patterns are
535 // supported:
536 // - Phi -> SafePoints
537 // - Phi -> CmpP/N
538 // - Phi -> AddP -> Load
539 // - Phi -> CastPP -> SafePoints
540 // - Phi -> CastPP -> AddP -> Load
541 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
542 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
543 Node* use = n->fast_out(i);
544
545 if (use->is_SafePoint()) {
546 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
547 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
548 return false;
549 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
550 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
551 return false;
552 }
553 } else if (use->is_AddP()) {
554 Node* addp = use;
555 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
556 Node* use_use = addp->fast_out(j);
557 const Type* load_type = _igvn->type(use_use);
558
559 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
560 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());)
561 return false;
562 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
563 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());)
564 return false;
565 }
566 }
567 } else if (nesting > 0) {
568 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);)
569 return false;
570 } else if (use->is_CastPP()) {
571 const Type* cast_t = _igvn->type(use);
572 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
573 #ifndef PRODUCT
574 if (TraceReduceAllocationMerges) {
575 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
576 use->dump();
577 }
578 #endif
579 return false;
580 }
581
582 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
583 if (!is_trivial_control) {
584 // If it's not a trivial control then we check if we can reduce the
585 // CmpP/N used by the If controlling the cast.
586 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
587 Node* iff = use->in(0)->in(0);
588 // We may have an OpaqueNotNull node between If and Bool nodes. But we could also have a sub class of IfNode,
589 // for example, an OuterStripMinedLoopEnd or a Parse Predicate. Bail out in all these cases.
590 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
591 if (can_reduce) {
592 Node* iff_cmp = iff->in(1)->in(1);
593 int opc = iff_cmp->Opcode();
594 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
595 }
596 if (!can_reduce) {
597 #ifndef PRODUCT
598 if (TraceReduceAllocationMerges) {
599 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
600 n->dump(5);
601 }
602 #endif
603 return false;
604 }
605 }
606 }
607
608 if (!can_reduce_check_users(use, nesting+1)) {
609 return false;
610 }
611 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
612 if (!can_reduce_cmp(n, use)) {
613 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);)
614 return false;
615 }
616 } else {
617 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());)
618 return false;
619 }
620 }
621
622 return true;
623 }
624
625 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
626 // only used in some certain code shapes. Check comments in
627 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
628 // details.
629 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
630 // If there was an error attempting to reduce allocation merges for this
631 // method we might have disabled the compilation and be retrying with RAM
632 // disabled.
633 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
634 return false;
635 }
636
637 const Type* phi_t = _igvn->type(ophi);
638 if (phi_t == nullptr ||
639 phi_t->make_ptr() == nullptr ||
640 phi_t->make_ptr()->isa_aryptr() != nullptr) {
641 return false;
642 }
643
644 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
645 return false;
646 }
647
648 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
649 return true;
650 }
651
652 // This method will return a CmpP/N that we need to use on the If controlling a
653 // CastPP after it was split. This method is only called on bases that are
654 // nullable therefore we always need a controlling if for the splitted CastPP.
655 //
656 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
657 // If the CastPP currently doesn't have a control then the CmpP/N will be
658 // against the null constant, otherwise it will be against the constant input of
659 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
660 // case because we have constraints on it and because the CastPP has a control
661 // input.
662 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
663 const Type* t = base->bottom_type();
664 Node* con = nullptr;
665
666 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
667 con = _igvn->zerocon(t->basic_type());
668 } else {
669 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
670 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
671 Node* bol = curr_ctrl->in(0)->in(1);
672 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
673 Node* curr_cmp = bol->in(1);
674 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
675 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
676 }
677
678 return CmpNode::make(base, con, t->basic_type());
679 }
680
681 // This method 'specializes' the CastPP passed as parameter to the base passed
682 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
683 // means that the CastPP now will be specific for a given base instead of a Phi.
684 // An If-Then-Else-Region block is inserted to control the CastPP. The control
685 // of the CastPP is a copy of the current one (if there is one) or a check
686 // against null.
687 //
688 // Before:
689 //
690 // C1 C2 ... Cn
691 // \ | /
692 // \ | /
693 // \ | /
694 // \ | /
695 // \ | /
696 // \ | /
697 // \|/
698 // Region B1 B2 ... Bn
699 // | \ | /
700 // | \ | /
701 // | \ | /
702 // | \ | /
703 // | \ | /
704 // | \ | /
705 // ---------------> Phi
706 // |
707 // X |
708 // | |
709 // | |
710 // ------> CastPP
711 //
712 // After (only partial illustration; base = B2, current_control = C2):
713 //
714 // C2
715 // |
716 // If
717 // / \
718 // / \
719 // T F
720 // /\ /
721 // / \ /
722 // / \ /
723 // C1 CastPP Reg Cn
724 // | | |
725 // | | |
726 // | | |
727 // -------------- | ----------
728 // | | |
729 // Region
730 //
731 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
732 Node* control_successor = current_control->unique_ctrl_out();
733 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
734 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
735 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
736 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
737 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
738 Node* end_region = _igvn->transform(new RegionNode(3));
739
740 // Insert the new if-else-region block into the graph
741 end_region->set_req(1, not_eq_control);
742 end_region->set_req(2, yes_eq_control);
743 control_successor->replace_edge(current_control, end_region, _igvn);
744
745 _igvn->_worklist.push(current_control);
746 _igvn->_worklist.push(control_successor);
747
748 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::UnconditionalDependency, nullptr));
749 }
750
751 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
752 const Type* load_type = _igvn->type(curr_load);
753 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
754 Node* memory = curr_load->in(MemNode::Memory);
755
756 // The data_phi merging the loads needs to be nullable if
757 // we are loading pointers.
758 if (load_type->make_ptr() != nullptr) {
759 if (load_type->isa_narrowoop()) {
760 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
761 } else if (load_type->isa_ptr()) {
762 load_type = load_type->meet(TypePtr::NULL_PTR);
763 } else {
764 assert(false, "Unexpected load ptr type.");
765 }
766 }
767
768 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
769
770 for (int i = 1; i < bases_for_loads->length(); i++) {
771 Node* base = bases_for_loads->at(i);
772 Node* cmp_region = nullptr;
773 if (base != nullptr) {
774 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
775 cmp_region = base->unique_ctrl_out_or_null();
776 assert(cmp_region != nullptr, "There should be.");
777 base = base->find_out_with(Op_CastPP);
778 }
779
780 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
781 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
782 Node* load = curr_load->clone();
783 load->set_req(0, nullptr);
784 load->set_req(1, mem);
785 load->set_req(2, addr);
786
787 if (cmp_region != nullptr) { // see comment on previous if
788 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
789 intermediate_phi->set_req(1, _igvn->transform(load));
790 load = intermediate_phi;
791 }
792
793 data_phi->set_req(i, _igvn->transform(load));
794 } else {
795 // Just use the default, which is already in phi
796 }
797 }
798
799 // Takes care of updating CG and split_unique_types worklists due
800 // to cloned AddP->Load.
801 updates_after_load_split(data_phi, curr_load, alloc_worklist);
802
803 return _igvn->transform(data_phi);
804 }
805
806 // This method only reduces CastPP fields loads; SafePoints are handled
807 // separately. The idea here is basically to clone the CastPP and place copies
808 // on each input of the Phi, including non-scalar replaceable inputs.
809 // Experimentation shows that the resulting IR graph is simpler that way than if
810 // we just split the cast through scalar-replaceable inputs.
811 //
812 // The reduction process requires that CastPP's control be one of:
813 // 1) no control,
814 // 2) the same region as Ophi, or
815 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
816 //
817 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
818 // flow. If the CastPP originally had an IfTrue/False control input then we'll
819 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
820 // juse use a CmpP/N against the null constant.
821 //
822 // The If-Then-Else-Region isn't always needed. For instance, if input to
823 // splitted cast was not nullable (or if it was the null constant) then we don't
824 // need (shouldn't) use a CastPP at all.
825 //
826 // After the casts are splitted we'll split the AddP->Loads through the Phi and
827 // connect them to the just split CastPPs.
828 //
829 // Before (CastPP control is same as Phi):
830 //
831 // Region Allocate Null Call
832 // | \ | /
833 // | \ | /
834 // | \ | /
835 // | \ | /
836 // | \ | /
837 // | \ | /
838 // ------------------> Phi # Oop Phi
839 // | |
840 // | |
841 // | |
842 // | |
843 // ----------------> CastPP
844 // |
845 // AddP
846 // |
847 // Load
848 //
849 // After (Very much simplified):
850 //
851 // Call Null
852 // \ /
853 // CmpP
854 // |
855 // Bool#NE
856 // |
857 // If
858 // / \
859 // T F
860 // / \ /
861 // / R
862 // CastPP |
863 // | |
864 // AddP |
865 // | |
866 // Load |
867 // \ | 0
868 // Allocate \ | /
869 // \ \ | /
870 // AddP Phi
871 // \ /
872 // Load /
873 // \ 0 /
874 // \ | /
875 // \|/
876 // Phi # "Field" Phi
877 //
878 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
879 Node* ophi = curr_castpp->in(1);
880 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
881
882 // Identify which base should be used for AddP->Load later when spliting the
883 // CastPP->Loads through ophi. Three kind of values may be stored in this
884 // array, depending on the nullability status of the corresponding input in
885 // ophi.
886 //
887 // - nullptr: Meaning that the base is actually the null constant and therefore
888 // we won't try to load from it.
889 //
890 // - CFG Node: Meaning that the base is a CastPP that was specialized for
891 // this input of Ophi. I.e., we added an If->Then->Else-Region
892 // that will 'activate' the CastPp only when the input is not Null.
893 //
894 // - Other Node: Meaning that the base is not nullable and therefore we'll try
895 // to load directly from it.
896 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
897
898 for (uint i = 1; i < ophi->req(); i++) {
899 Node* base = ophi->in(i);
900 const Type* base_t = _igvn->type(base);
901
902 if (base_t->maybe_null()) {
903 if (base->is_Con()) {
904 // Nothing todo as bases_for_loads[i] is already null
905 } else {
906 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
907 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
908 }
909 } else {
910 bases_for_loads.at_put(i, base);
911 }
912 }
913
914 // Now let's split the CastPP->Loads through the Phi
915 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
916 Node* use = curr_castpp->raw_out(i);
917 if (use->is_AddP()) {
918 for (int j = use->outcnt()-1; j >= 0;) {
919 Node* use_use = use->raw_out(j);
920 assert(use_use->is_Load(), "Expected this to be a Load node.");
921
922 // We can't make an unconditional load from a nullable input. The
923 // 'split_castpp_load_through_phi` method will add an
924 // 'If-Then-Else-Region` around nullable bases and only load from them
925 // when the input is not null.
926 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
927 _igvn->replace_node(use_use, phi);
928
929 --j;
930 j = MIN2(j, (int)use->outcnt()-1);
931 }
932
933 _igvn->remove_dead_node(use);
934 }
935 --i;
936 i = MIN2(i, (int)curr_castpp->outcnt()-1);
937 }
938 }
939
940 // This method split a given CmpP/N through the Phi used in one of its inputs.
941 // As a result we convert a comparison with a pointer to a comparison with an
942 // integer.
943 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
944 // while the other must be a constant.
945 // The splitting process is basically just cloning the CmpP/N above the input
946 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
947 // can prove at compile time the result of the comparison.
948 //
949 // Before:
950 //
951 // in1 in2 ... inN
952 // \ | /
953 // \ | /
954 // \ | /
955 // \ | /
956 // \ | /
957 // \ | /
958 // Phi
959 // | Other
960 // | /
961 // | /
962 // | /
963 // CmpP/N
964 //
965 // After:
966 //
967 // in1 Other in2 Other inN Other
968 // | | | | | |
969 // \ | | | | |
970 // \ / | / | /
971 // CmpP/N CmpP/N CmpP/N
972 // Bool Bool Bool
973 // \ | /
974 // \ | /
975 // \ | /
976 // \ | /
977 // \ | /
978 // \ | /
979 // \ | /
980 // \ | /
981 // Phi
982 // |
983 // | Zero
984 // | /
985 // | /
986 // | /
987 // CmpI
988 //
989 //
990 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
991 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
992 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
993
994 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
995 Node* zero = _igvn->intcon(0);
996 Node* one = _igvn->intcon(1);
997 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
998
999 // This Phi will merge the result of the Cmps split through the Phi
1000 Node* res_phi = PhiNode::make(ophi->in(0), zero, TypeInt::INT);
1001
1002 for (uint i=1; i<ophi->req(); i++) {
1003 Node* ophi_input = ophi->in(i);
1004 Node* res_phi_input = nullptr;
1005
1006 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1007 if (tcmp->singleton()) {
1008 if ((mask == BoolTest::mask::eq && tcmp == TypeInt::CC_EQ) ||
1009 (mask == BoolTest::mask::ne && tcmp == TypeInt::CC_GT)) {
1010 res_phi_input = one;
1011 } else {
1012 res_phi_input = zero;
1013 }
1014 } else {
1015 Node* ncmp = _igvn->transform(cmp->clone());
1016 ncmp->set_req(1, ophi_input);
1017 ncmp->set_req(2, other);
1018 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1019 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1020 }
1021
1022 res_phi->set_req(i, res_phi_input);
1023 }
1024
1025 // This CMP always compares whether the output of "res_phi" is TRUE as far as the "mask".
1026 Node* new_cmp = _igvn->transform(new CmpINode(_igvn->transform(res_phi), (mask == BoolTest::mask::eq) ? one : zero));
1027 _igvn->replace_node(cmp, new_cmp);
1028 }
1029
1030 // Push the newly created AddP on alloc_worklist and patch
1031 // the connection graph. Note that the changes in the CG below
1032 // won't affect the ES of objects since the new nodes have the
1033 // same status as the old ones.
1034 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1035 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1036 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1037 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1038
1039 if (data_phi == nullptr || !data_phi->is_Phi()) {
1040 // Make this a retry?
1041 return ;
1042 }
1043
1044 Node* previous_addp = previous_load->in(MemNode::Address);
1045 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1046 for (uint i = 1; i < data_phi->req(); i++) {
1047 Node* new_load = data_phi->in(i);
1048
1049 if (new_load->is_Phi()) {
1050 // new_load is currently the "intermediate_phi" from an specialized
1051 // CastPP.
1052 new_load = new_load->in(1);
1053 }
1054
1055 // "new_load" might actually be a constant, parameter, etc.
1056 if (new_load->is_Load()) {
1057 Node* new_addp = new_load->in(MemNode::Address);
1058 Node* base = get_addp_base(new_addp);
1059
1060 // The base might not be something that we can create an unique
1061 // type for. If that's the case we are done with that input.
1062 PointsToNode* jobj_ptn = unique_java_object(base);
1063 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1064 continue;
1065 }
1066
1067 // Push to alloc_worklist since the base has an unique_type
1068 alloc_worklist.append_if_missing(new_addp);
1069
1070 // Now let's add the node to the connection graph
1071 _nodes.at_grow(new_addp->_idx, nullptr);
1072 add_field(new_addp, fn->escape_state(), fn->offset());
1073 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1074
1075 // If the load doesn't load an object then it won't be
1076 // part of the connection graph
1077 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1078 if (curr_load_ptn != nullptr) {
1079 _nodes.at_grow(new_load->_idx, nullptr);
1080 add_local_var(new_load, curr_load_ptn->escape_state());
1081 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1082 }
1083 }
1084 }
1085 }
1086
1087 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1088 // We'll pass this to 'split_through_phi' so that it'll do the split even
1089 // though the load doesn't have an unique instance type.
1090 bool ignore_missing_instance_id = true;
1091
1092 // All AddPs are present in the connection graph
1093 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1094
1095 // Iterate over AddP looking for a Load
1096 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1097 Node* previous_load = previous_addp->raw_out(k);
1098 if (previous_load->is_Load()) {
1099 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1100
1101 // Takes care of updating CG and split_unique_types worklists due to cloned
1102 // AddP->Load.
1103 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1104
1105 _igvn->replace_node(previous_load, data_phi);
1106 }
1107 --k;
1108 k = MIN2(k, (int)previous_addp->outcnt()-1);
1109 }
1110
1111 // Remove the old AddP from the processing list because it's dead now
1112 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1113 alloc_worklist.remove_if_existing(previous_addp);
1114 }
1115
1116 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1117 // selector is:
1118 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1119 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1120 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1121 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1122 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1123 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1124 uint number_of_sr_objects = 0;
1125 for (uint i = 1; i < ophi->req(); i++) {
1126 Node* base = ophi->in(i);
1127 JavaObjectNode* ptn = unique_java_object(base);
1128
1129 if (ptn != nullptr && ptn->scalar_replaceable()) {
1130 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1131 selector->set_req(i, sr_obj_idx);
1132 number_of_sr_objects++;
1133 }
1134 }
1135
1136 return selector->as_Phi();
1137 }
1138
1139 // Returns true if the AddP node 'n' has at least one base that is a reducible
1140 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1141 // checked instead.
1142 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1143 PointsToNode* ptn = ptnode_adr(n->_idx);
1144 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1145 return false;
1146 }
1147
1148 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1149 Node* base = i.get()->ideal_node();
1150
1151 if (reducible_merges.member(base)) {
1152 return true;
1153 }
1154
1155 if (base->is_CastPP() || base->is_CheckCastPP()) {
1156 base = base->in(1);
1157 if (reducible_merges.member(base)) {
1158 return true;
1159 }
1160 }
1161 }
1162
1163 return false;
1164 }
1165
1166 // This method will call its helper method to reduce SafePoint nodes that use
1167 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1168 // "version" of Phi use the same debug information (regarding the Phi).
1169 // Therefore, I collect all safepoints and patch them all at once.
1170 //
1171 // The safepoints using the Phi node have to be processed before safepoints of
1172 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1173 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1174 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1175 // algorithm that process Phi's safepoints will think that the added Phi
1176 // reference is a regular reference.
1177 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1178 PhiNode* selector = create_selector(ophi);
1179 Unique_Node_List safepoints;
1180 Unique_Node_List casts;
1181
1182 // Just collect the users of the Phis for later processing
1183 // in the needed order.
1184 for (uint i = 0; i < ophi->outcnt(); i++) {
1185 Node* use = ophi->raw_out(i);
1186 if (use->is_SafePoint()) {
1187 safepoints.push(use);
1188 } else if (use->is_CastPP()) {
1189 casts.push(use);
1190 } else {
1191 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1192 }
1193 }
1194
1195 // Need to process safepoints using the Phi first
1196 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1197 return false;
1198 }
1199
1200 // Now process CastPP->safepoints
1201 for (uint i = 0; i < casts.size(); i++) {
1202 Node* cast = casts.at(i);
1203 Unique_Node_List cast_sfpts;
1204
1205 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1206 Node* use_use = cast->fast_out(j);
1207 if (use_use->is_SafePoint()) {
1208 cast_sfpts.push(use_use);
1209 } else {
1210 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1211 }
1212 }
1213
1214 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1215 return false;
1216 }
1217 }
1218
1219 return true;
1220 }
1221
1222 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1223 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1224 // SafePointScalarObjectNode for each scalar replaceable input. Each
1225 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1226 // check detailed description in SafePointScalarMergeNode class header.
1227 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1228 PhaseMacroExpand mexp(*_igvn);
1229 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1230 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1231
1232 Node* nsr_merge_pointer = ophi;
1233 if (cast != nullptr) {
1234 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1235 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::RegularDependency, nullptr));
1236 }
1237
1238 for (uint spi = 0; spi < safepoints.size(); spi++) {
1239 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1240 JVMState *jvms = sfpt->jvms();
1241 uint merge_idx = (sfpt->req() - jvms->scloff());
1242 int debug_start = jvms->debug_start();
1243
1244 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1245 smerge->init_req(0, _compile->root());
1246 _igvn->register_new_node_with_optimizer(smerge);
1247
1248 // The next two inputs are:
1249 // (1) A copy of the original pointer to NSR objects.
1250 // (2) A selector, used to decide if we need to rematerialize an object
1251 // or use the pointer to a NSR object.
1252 // See more details of these fields in the declaration of SafePointScalarMergeNode
1253 sfpt->add_req(nsr_merge_pointer);
1254 sfpt->add_req(selector);
1255
1256 for (uint i = 1; i < ophi->req(); i++) {
1257 Node* base = ophi->in(i);
1258 JavaObjectNode* ptn = unique_java_object(base);
1259
1260 // If the base is not scalar replaceable we don't need to register information about
1261 // it at this time.
1262 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1263 continue;
1264 }
1265
1266 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1267 Unique_Node_List value_worklist;
1268 #ifdef ASSERT
1269 const Type* res_type = alloc->result_cast()->bottom_type();
1270 if (res_type->is_inlinetypeptr() && !Compile::current()->has_circular_inline_type()) {
1271 PhiNode* phi = ophi->as_Phi();
1272 assert(!ophi->as_Phi()->can_push_inline_types_down(_igvn), "missed earlier scalarization opportunity");
1273 }
1274 #endif
1275 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt, &value_worklist);
1276 if (sobj == nullptr) {
1277 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1278 return false;
1279 }
1280
1281 // Now make a pass over the debug information replacing any references
1282 // to the allocated object with "sobj"
1283 Node* ccpp = alloc->result_cast();
1284 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1285
1286 // Register the scalarized object as a candidate for reallocation
1287 smerge->add_req(sobj);
1288
1289 // Scalarize inline types that were added to the safepoint.
1290 // Don't allow linking a constant oop (if available) for flat array elements
1291 // because Deoptimization::reassign_flat_array_elements needs field values.
1292 const bool allow_oop = !merge_t->is_flat();
1293 for (uint j = 0; j < value_worklist.size(); ++j) {
1294 InlineTypeNode* vt = value_worklist.at(j)->as_InlineType();
1295 vt->make_scalar_in_safepoints(_igvn, allow_oop);
1296 }
1297 }
1298
1299 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1300 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1301
1302 // The call to 'replace_edges_in_range' above might have removed the
1303 // reference to ophi that we need at _merge_pointer_idx. The line below make
1304 // sure the reference is maintained.
1305 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1306 _igvn->_worklist.push(sfpt);
1307 }
1308
1309 return true;
1310 }
1311
1312 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
1313 bool delay = _igvn->delay_transform();
1314 _igvn->set_delay_transform(true);
1315 _igvn->hash_delete(ophi);
1316
1317 // Copying all users first because some will be removed and others won't.
1318 // Ophi also may acquire some new users as part of Cast reduction.
1319 // CastPPs also need to be processed before CmpPs.
1320 Unique_Node_List castpps;
1321 Unique_Node_List others;
1322 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1323 Node* use = ophi->fast_out(i);
1324
1325 if (use->is_CastPP()) {
1326 castpps.push(use);
1327 } else if (use->is_AddP() || use->is_Cmp()) {
1328 others.push(use);
1329 } else if (use->is_SafePoint()) {
1330 // processed later
1331 } else {
1332 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1333 }
1334 }
1335
1336 // CastPPs need to be processed before Cmps because during the process of
1337 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1338 // by the If controlling the CastPP.
1339 for (uint i = 0; i < castpps.size(); i++) {
1340 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist, memnode_worklist);
1341 }
1342
1343 for (uint i = 0; i < others.size(); i++) {
1344 Node* use = others.at(i);
1345
1346 if (use->is_AddP()) {
1347 reduce_phi_on_field_access(use, alloc_worklist);
1348 } else if(use->is_Cmp()) {
1349 reduce_phi_on_cmp(use);
1350 }
1351 }
1352
1353 _igvn->set_delay_transform(delay);
1354 }
1355
1356 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1357 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1358 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1359 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1360 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1361
1362 for (uint i = 1; i < ophi->req(); i++) {
1363 Node* base = ophi->in(i);
1364 JavaObjectNode* ptn = unique_java_object(base);
1365
1366 if (ptn != nullptr && ptn->scalar_replaceable()) {
1367 new_phi->set_req(i, null_ptr);
1368 } else {
1369 new_phi->set_req(i, ophi->in(i));
1370 }
1371 }
1372
1373 for (int i = ophi->outcnt()-1; i >= 0;) {
1374 Node* out = ophi->raw_out(i);
1375
1376 if (out->is_ConstraintCast()) {
1377 const Type* out_t = _igvn->type(out)->make_ptr();
1378 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1379 bool change = out_new_t != out_t;
1380
1381 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1382 Node* out2 = out->raw_out(j);
1383 if (!out2->is_SafePoint()) {
1384 change = false;
1385 break;
1386 }
1387 }
1388
1389 if (change) {
1390 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::StrongDependency, nullptr);
1391 _igvn->replace_node(out, new_cast);
1392 _igvn->register_new_node_with_optimizer(new_cast);
1393 }
1394 }
1395
1396 --i;
1397 i = MIN2(i, (int)ophi->outcnt()-1);
1398 }
1399
1400 _igvn->replace_node(ophi, new_phi);
1401 }
1402
1403 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1404 if (!C->do_reduce_allocation_merges()) return;
1405
1406 Unique_Node_List ideal_nodes;
1407 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1408 ideal_nodes.push(root);
1409
1410 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1411 Node* n = ideal_nodes.at(next);
1412
1413 if (n->is_SafePointScalarMerge()) {
1414 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1415
1416 // Validate inputs of merge
1417 for (uint i = 1; i < merge->req(); i++) {
1418 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1419 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1420 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1421 }
1422 }
1423
1424 // Validate users of merge
1425 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1426 Node* sfpt = merge->fast_out(i);
1427 if (sfpt->is_SafePoint()) {
1428 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1429
1430 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1431 assert(false, "SafePointScalarMerge nodes can't be nested.");
1432 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1433 }
1434 } else {
1435 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1436 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1437 }
1438 }
1439 }
1440
1441 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1442 Node* m = n->fast_out(i);
1443 ideal_nodes.push(m);
1444 }
1445 }
1446 }
1447
1448 // Returns true if there is an object in the scope of sfn that does not escape globally.
1449 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1450 Compile* C = _compile;
1451 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1452 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1453 DeoptimizeObjectsALot) {
1454 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1455 int num_locs = jvms->loc_size();
1456 for (int idx = 0; idx < num_locs; idx++) {
1457 Node* l = sfn->local(jvms, idx);
1458 if (not_global_escape(l)) {
1459 return true;
1460 }
1461 }
1462 }
1463 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1464 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1465 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1466 int num_mon = jvms->nof_monitors();
1467 for (int idx = 0; idx < num_mon; idx++) {
1468 Node* m = sfn->monitor_obj(jvms, idx);
1469 if (m != nullptr && not_global_escape(m)) {
1470 return true;
1471 }
1472 }
1473 }
1474 }
1475 return false;
1476 }
1477
1478 // Returns true if at least one of the arguments to the call is an object
1479 // that does not escape globally.
1480 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1481 if (call->method() != nullptr) {
1482 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1483 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1484 Node* p = call->in(idx);
1485 if (not_global_escape(p)) {
1486 return true;
1487 }
1488 }
1489 } else {
1490 const char* name = call->as_CallStaticJava()->_name;
1491 assert(name != nullptr, "no name");
1492 // no arg escapes through uncommon traps
1493 if (strcmp(name, "uncommon_trap") != 0) {
1494 // process_call_arguments() assumes that all arguments escape globally
1495 const TypeTuple* d = call->tf()->domain_sig();
1496 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1497 const Type* at = d->field_at(i);
1498 if (at->isa_oopptr() != nullptr) {
1499 return true;
1500 }
1501 }
1502 }
1503 }
1504 return false;
1505 }
1506
1507
1508
1509 // Utility function for nodes that load an object
1510 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1511 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1512 // ThreadLocal has RawPtr type.
1513 const Type* t = _igvn->type(n);
1514 if (t->make_ptr() != nullptr) {
1515 Node* adr = n->in(MemNode::Address);
1516 #ifdef ASSERT
1517 if (!adr->is_AddP()) {
1518 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1519 } else {
1520 assert((ptnode_adr(adr->_idx) == nullptr ||
1521 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1522 }
1523 #endif
1524 add_local_var_and_edge(n, PointsToNode::NoEscape,
1525 adr, delayed_worklist);
1526 }
1527 }
1528
1529 // Populate Connection Graph with PointsTo nodes and create simple
1530 // connection graph edges.
1531 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1532 assert(!_verify, "this method should not be called for verification");
1533 PhaseGVN* igvn = _igvn;
1534 uint n_idx = n->_idx;
1535 PointsToNode* n_ptn = ptnode_adr(n_idx);
1536 if (n_ptn != nullptr) {
1537 return; // No need to redefine PointsTo node during first iteration.
1538 }
1539 int opcode = n->Opcode();
1540 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1541 if (gc_handled) {
1542 return; // Ignore node if already handled by GC.
1543 }
1544
1545 if (n->is_Call()) {
1546 // Arguments to allocation and locking don't escape.
1547 if (n->is_AbstractLock()) {
1548 // Put Lock and Unlock nodes on IGVN worklist to process them during
1549 // first IGVN optimization when escape information is still available.
1550 record_for_optimizer(n);
1551 } else if (n->is_Allocate()) {
1552 add_call_node(n->as_Call());
1553 record_for_optimizer(n);
1554 } else {
1555 if (n->is_CallStaticJava()) {
1556 const char* name = n->as_CallStaticJava()->_name;
1557 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1558 return; // Skip uncommon traps
1559 }
1560 }
1561 // Don't mark as processed since call's arguments have to be processed.
1562 delayed_worklist->push(n);
1563 // Check if a call returns an object.
1564 if ((n->as_Call()->returns_pointer() &&
1565 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1566 (n->is_CallStaticJava() &&
1567 n->as_CallStaticJava()->is_boxing_method())) {
1568 add_call_node(n->as_Call());
1569 } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1570 bool returns_oop = false;
1571 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1572 ProjNode* pn = n->fast_out(i)->as_Proj();
1573 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1574 returns_oop = true;
1575 }
1576 }
1577 if (returns_oop) {
1578 add_call_node(n->as_Call());
1579 }
1580 }
1581 }
1582 return;
1583 }
1584 // Put this check here to process call arguments since some call nodes
1585 // point to phantom_obj.
1586 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1587 return; // Skip predefined nodes.
1588 }
1589 switch (opcode) {
1590 case Op_AddP: {
1591 Node* base = get_addp_base(n);
1592 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1593 // Field nodes are created for all field types. They are used in
1594 // adjust_scalar_replaceable_state() and split_unique_types().
1595 // Note, non-oop fields will have only base edges in Connection
1596 // Graph because such fields are not used for oop loads and stores.
1597 int offset = address_offset(n, igvn);
1598 add_field(n, PointsToNode::NoEscape, offset);
1599 if (ptn_base == nullptr) {
1600 delayed_worklist->push(n); // Process it later.
1601 } else {
1602 n_ptn = ptnode_adr(n_idx);
1603 add_base(n_ptn->as_Field(), ptn_base);
1604 }
1605 break;
1606 }
1607 case Op_CastX2P:
1608 case Op_CastI2N: {
1609 map_ideal_node(n, phantom_obj);
1610 break;
1611 }
1612 case Op_InlineType:
1613 case Op_CastPP:
1614 case Op_CheckCastPP:
1615 case Op_EncodeP:
1616 case Op_DecodeN:
1617 case Op_EncodePKlass:
1618 case Op_DecodeNKlass: {
1619 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1620 break;
1621 }
1622 case Op_CMoveP: {
1623 add_local_var(n, PointsToNode::NoEscape);
1624 // Do not add edges during first iteration because some could be
1625 // not defined yet.
1626 delayed_worklist->push(n);
1627 break;
1628 }
1629 case Op_ConP:
1630 case Op_ConN:
1631 case Op_ConNKlass: {
1632 // assume all oop constants globally escape except for null
1633 PointsToNode::EscapeState es;
1634 const Type* t = igvn->type(n);
1635 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1636 es = PointsToNode::NoEscape;
1637 } else {
1638 es = PointsToNode::GlobalEscape;
1639 }
1640 PointsToNode* ptn_con = add_java_object(n, es);
1641 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1642 break;
1643 }
1644 case Op_CreateEx: {
1645 // assume that all exception objects globally escape
1646 map_ideal_node(n, phantom_obj);
1647 break;
1648 }
1649 case Op_LoadKlass:
1650 case Op_LoadNKlass: {
1651 // Unknown class is loaded
1652 map_ideal_node(n, phantom_obj);
1653 break;
1654 }
1655 case Op_LoadP:
1656 case Op_LoadN: {
1657 add_objload_to_connection_graph(n, delayed_worklist);
1658 break;
1659 }
1660 case Op_Parm: {
1661 map_ideal_node(n, phantom_obj);
1662 break;
1663 }
1664 case Op_PartialSubtypeCheck: {
1665 // Produces Null or notNull and is used in only in CmpP so
1666 // phantom_obj could be used.
1667 map_ideal_node(n, phantom_obj); // Result is unknown
1668 break;
1669 }
1670 case Op_Phi: {
1671 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1672 // ThreadLocal has RawPtr type.
1673 const Type* t = n->as_Phi()->type();
1674 if (t->make_ptr() != nullptr) {
1675 add_local_var(n, PointsToNode::NoEscape);
1676 // Do not add edges during first iteration because some could be
1677 // not defined yet.
1678 delayed_worklist->push(n);
1679 }
1680 break;
1681 }
1682 case Op_Proj: {
1683 // we are only interested in the oop result projection from a call
1684 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
1685 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) {
1686 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1687 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1688 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1689 }
1690 break;
1691 }
1692 case Op_Rethrow: // Exception object escapes
1693 case Op_Return: {
1694 if (n->req() > TypeFunc::Parms &&
1695 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1696 // Treat Return value as LocalVar with GlobalEscape escape state.
1697 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1698 }
1699 break;
1700 }
1701 case Op_CompareAndExchangeP:
1702 case Op_CompareAndExchangeN:
1703 case Op_GetAndSetP:
1704 case Op_GetAndSetN: {
1705 add_objload_to_connection_graph(n, delayed_worklist);
1706 // fall-through
1707 }
1708 case Op_StoreP:
1709 case Op_StoreN:
1710 case Op_StoreNKlass:
1711 case Op_WeakCompareAndSwapP:
1712 case Op_WeakCompareAndSwapN:
1713 case Op_CompareAndSwapP:
1714 case Op_CompareAndSwapN: {
1715 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1716 break;
1717 }
1718 case Op_AryEq:
1719 case Op_CountPositives:
1720 case Op_StrComp:
1721 case Op_StrEquals:
1722 case Op_StrIndexOf:
1723 case Op_StrIndexOfChar:
1724 case Op_StrInflatedCopy:
1725 case Op_StrCompressedCopy:
1726 case Op_VectorizedHashCode:
1727 case Op_EncodeISOArray: {
1728 add_local_var(n, PointsToNode::ArgEscape);
1729 delayed_worklist->push(n); // Process it later.
1730 break;
1731 }
1732 case Op_ThreadLocal: {
1733 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1734 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1735 break;
1736 }
1737 case Op_Blackhole: {
1738 // All blackhole pointer arguments are globally escaping.
1739 // Only do this if there is at least one pointer argument.
1740 // Do not add edges during first iteration because some could be
1741 // not defined yet, defer to final step.
1742 for (uint i = 0; i < n->req(); i++) {
1743 Node* in = n->in(i);
1744 if (in != nullptr) {
1745 const Type* at = _igvn->type(in);
1746 if (!at->isa_ptr()) continue;
1747
1748 add_local_var(n, PointsToNode::GlobalEscape);
1749 delayed_worklist->push(n);
1750 break;
1751 }
1752 }
1753 break;
1754 }
1755 default:
1756 ; // Do nothing for nodes not related to EA.
1757 }
1758 return;
1759 }
1760
1761 // Add final simple edges to graph.
1762 void ConnectionGraph::add_final_edges(Node *n) {
1763 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1764 #ifdef ASSERT
1765 if (_verify && n_ptn->is_JavaObject())
1766 return; // This method does not change graph for JavaObject.
1767 #endif
1768
1769 if (n->is_Call()) {
1770 process_call_arguments(n->as_Call());
1771 return;
1772 }
1773 assert(n->is_Store() || n->is_LoadStore() ||
1774 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1775 "node should be registered already");
1776 int opcode = n->Opcode();
1777 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1778 if (gc_handled) {
1779 return; // Ignore node if already handled by GC.
1780 }
1781 switch (opcode) {
1782 case Op_AddP: {
1783 Node* base = get_addp_base(n);
1784 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1785 assert(ptn_base != nullptr, "field's base should be registered");
1786 add_base(n_ptn->as_Field(), ptn_base);
1787 break;
1788 }
1789 case Op_InlineType:
1790 case Op_CastPP:
1791 case Op_CheckCastPP:
1792 case Op_EncodeP:
1793 case Op_DecodeN:
1794 case Op_EncodePKlass:
1795 case Op_DecodeNKlass: {
1796 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1797 break;
1798 }
1799 case Op_CMoveP: {
1800 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1801 Node* in = n->in(i);
1802 if (in == nullptr) {
1803 continue; // ignore null
1804 }
1805 Node* uncast_in = in->uncast();
1806 if (uncast_in->is_top() || uncast_in == n) {
1807 continue; // ignore top or inputs which go back this node
1808 }
1809 PointsToNode* ptn = ptnode_adr(in->_idx);
1810 assert(ptn != nullptr, "node should be registered");
1811 add_edge(n_ptn, ptn);
1812 }
1813 break;
1814 }
1815 case Op_LoadP:
1816 case Op_LoadN: {
1817 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1818 // ThreadLocal has RawPtr type.
1819 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1820 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1821 break;
1822 }
1823 case Op_Phi: {
1824 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1825 // ThreadLocal has RawPtr type.
1826 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1827 for (uint i = 1; i < n->req(); i++) {
1828 Node* in = n->in(i);
1829 if (in == nullptr) {
1830 continue; // ignore null
1831 }
1832 Node* uncast_in = in->uncast();
1833 if (uncast_in->is_top() || uncast_in == n) {
1834 continue; // ignore top or inputs which go back this node
1835 }
1836 PointsToNode* ptn = ptnode_adr(in->_idx);
1837 assert(ptn != nullptr, "node should be registered");
1838 add_edge(n_ptn, ptn);
1839 }
1840 break;
1841 }
1842 case Op_Proj: {
1843 // we are only interested in the oop result projection from a call
1844 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1845 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1846 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1847 break;
1848 }
1849 case Op_Rethrow: // Exception object escapes
1850 case Op_Return: {
1851 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1852 "Unexpected node type");
1853 // Treat Return value as LocalVar with GlobalEscape escape state.
1854 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1855 break;
1856 }
1857 case Op_CompareAndExchangeP:
1858 case Op_CompareAndExchangeN:
1859 case Op_GetAndSetP:
1860 case Op_GetAndSetN:{
1861 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1862 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1863 // fall-through
1864 }
1865 case Op_CompareAndSwapP:
1866 case Op_CompareAndSwapN:
1867 case Op_WeakCompareAndSwapP:
1868 case Op_WeakCompareAndSwapN:
1869 case Op_StoreP:
1870 case Op_StoreN:
1871 case Op_StoreNKlass:{
1872 add_final_edges_unsafe_access(n, opcode);
1873 break;
1874 }
1875 case Op_VectorizedHashCode:
1876 case Op_AryEq:
1877 case Op_CountPositives:
1878 case Op_StrComp:
1879 case Op_StrEquals:
1880 case Op_StrIndexOf:
1881 case Op_StrIndexOfChar:
1882 case Op_StrInflatedCopy:
1883 case Op_StrCompressedCopy:
1884 case Op_EncodeISOArray: {
1885 // char[]/byte[] arrays passed to string intrinsic do not escape but
1886 // they are not scalar replaceable. Adjust escape state for them.
1887 // Start from in(2) edge since in(1) is memory edge.
1888 for (uint i = 2; i < n->req(); i++) {
1889 Node* adr = n->in(i);
1890 const Type* at = _igvn->type(adr);
1891 if (!adr->is_top() && at->isa_ptr()) {
1892 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1893 at->isa_ptr() != nullptr, "expecting a pointer");
1894 if (adr->is_AddP()) {
1895 adr = get_addp_base(adr);
1896 }
1897 PointsToNode* ptn = ptnode_adr(adr->_idx);
1898 assert(ptn != nullptr, "node should be registered");
1899 add_edge(n_ptn, ptn);
1900 }
1901 }
1902 break;
1903 }
1904 case Op_Blackhole: {
1905 // All blackhole pointer arguments are globally escaping.
1906 for (uint i = 0; i < n->req(); i++) {
1907 Node* in = n->in(i);
1908 if (in != nullptr) {
1909 const Type* at = _igvn->type(in);
1910 if (!at->isa_ptr()) continue;
1911
1912 if (in->is_AddP()) {
1913 in = get_addp_base(in);
1914 }
1915
1916 PointsToNode* ptn = ptnode_adr(in->_idx);
1917 assert(ptn != nullptr, "should be defined already");
1918 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1919 add_edge(n_ptn, ptn);
1920 }
1921 }
1922 break;
1923 }
1924 default: {
1925 // This method should be called only for EA specific nodes which may
1926 // miss some edges when they were created.
1927 #ifdef ASSERT
1928 n->dump(1);
1929 #endif
1930 guarantee(false, "unknown node");
1931 }
1932 }
1933 return;
1934 }
1935
1936 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1937 Node* adr = n->in(MemNode::Address);
1938 const Type* adr_type = _igvn->type(adr);
1939 adr_type = adr_type->make_ptr();
1940 if (adr_type == nullptr) {
1941 return; // skip dead nodes
1942 }
1943 if (adr_type->isa_oopptr()
1944 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1945 && adr_type == TypeRawPtr::NOTNULL
1946 && is_captured_store_address(adr))) {
1947 delayed_worklist->push(n); // Process it later.
1948 #ifdef ASSERT
1949 assert (adr->is_AddP(), "expecting an AddP");
1950 if (adr_type == TypeRawPtr::NOTNULL) {
1951 // Verify a raw address for a store captured by Initialize node.
1952 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1953 assert(offs != Type::OffsetBot, "offset must be a constant");
1954 }
1955 #endif
1956 } else {
1957 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1958 if (adr->is_BoxLock()) {
1959 return;
1960 }
1961 // Stored value escapes in unsafe access.
1962 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1963 delayed_worklist->push(n); // Process unsafe access later.
1964 return;
1965 }
1966 #ifdef ASSERT
1967 n->dump(1);
1968 assert(false, "not unsafe");
1969 #endif
1970 }
1971 }
1972
1973 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1974 Node* adr = n->in(MemNode::Address);
1975 const Type *adr_type = _igvn->type(adr);
1976 adr_type = adr_type->make_ptr();
1977 #ifdef ASSERT
1978 if (adr_type == nullptr) {
1979 n->dump(1);
1980 assert(adr_type != nullptr, "dead node should not be on list");
1981 return true;
1982 }
1983 #endif
1984
1985 if (adr_type->isa_oopptr()
1986 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1987 && adr_type == TypeRawPtr::NOTNULL
1988 && is_captured_store_address(adr))) {
1989 // Point Address to Value
1990 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1991 assert(adr_ptn != nullptr &&
1992 adr_ptn->as_Field()->is_oop(), "node should be registered");
1993 Node* val = n->in(MemNode::ValueIn);
1994 PointsToNode* ptn = ptnode_adr(val->_idx);
1995 assert(ptn != nullptr, "node should be registered");
1996 add_edge(adr_ptn, ptn);
1997 return true;
1998 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1999 // Stored value escapes in unsafe access.
2000 Node* val = n->in(MemNode::ValueIn);
2001 PointsToNode* ptn = ptnode_adr(val->_idx);
2002 assert(ptn != nullptr, "node should be registered");
2003 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
2004 // Add edge to object for unsafe access with offset.
2005 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2006 assert(adr_ptn != nullptr, "node should be registered");
2007 if (adr_ptn->is_Field()) {
2008 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2009 add_edge(adr_ptn, ptn);
2010 }
2011 return true;
2012 }
2013 #ifdef ASSERT
2014 n->dump(1);
2015 assert(false, "not unsafe");
2016 #endif
2017 return false;
2018 }
2019
2020 void ConnectionGraph::add_call_node(CallNode* call) {
2021 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2022 uint call_idx = call->_idx;
2023 if (call->is_Allocate()) {
2024 Node* k = call->in(AllocateNode::KlassNode);
2025 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2026 assert(kt != nullptr, "TypeKlassPtr required.");
2027 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2028 bool scalar_replaceable = true;
2029 NOT_PRODUCT(const char* nsr_reason = "");
2030 if (call->is_AllocateArray()) {
2031 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2032 es = PointsToNode::GlobalEscape;
2033 } else {
2034 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2035 if (length < 0) {
2036 // Not scalar replaceable if the length is not constant.
2037 scalar_replaceable = false;
2038 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2039 } else if (length > EliminateAllocationArraySizeLimit) {
2040 // Not scalar replaceable if the length is too big.
2041 scalar_replaceable = false;
2042 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2043 }
2044 }
2045 } else { // Allocate instance
2046 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2047 es = PointsToNode::GlobalEscape;
2048 } else {
2049 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2050 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2051 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2052 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2053 !ik->can_be_instantiated() ||
2054 ik->has_finalizer()) {
2055 es = PointsToNode::GlobalEscape;
2056 } else {
2057 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2058 if (nfields > EliminateAllocationFieldsLimit) {
2059 // Not scalar replaceable if there are too many fields.
2060 scalar_replaceable = false;
2061 NOT_PRODUCT(nsr_reason = "has too many fields");
2062 }
2063 }
2064 }
2065 }
2066 add_java_object(call, es);
2067 PointsToNode* ptn = ptnode_adr(call_idx);
2068 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2069 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2070 }
2071 } else if (call->is_CallStaticJava()) {
2072 // Call nodes could be different types:
2073 //
2074 // 1. CallDynamicJavaNode (what happened during call is unknown):
2075 //
2076 // - mapped to GlobalEscape JavaObject node if oop is returned;
2077 //
2078 // - all oop arguments are escaping globally;
2079 //
2080 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2081 //
2082 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2083 //
2084 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2085 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2086 // during call is returned;
2087 // - mapped to ArgEscape LocalVar node pointed to object arguments
2088 // which are returned and does not escape during call;
2089 //
2090 // - oop arguments escaping status is defined by bytecode analysis;
2091 //
2092 // For a static call, we know exactly what method is being called.
2093 // Use bytecode estimator to record whether the call's return value escapes.
2094 ciMethod* meth = call->as_CallJava()->method();
2095 if (meth == nullptr) {
2096 const char* name = call->as_CallStaticJava()->_name;
2097 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0 ||
2098 strncmp(name, "C2 Runtime load_unknown_inline", 30) == 0, "TODO: add failed case check");
2099 // Returns a newly allocated non-escaped object.
2100 add_java_object(call, PointsToNode::NoEscape);
2101 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2102 } else if (meth->is_boxing_method()) {
2103 // Returns boxing object
2104 PointsToNode::EscapeState es;
2105 vmIntrinsics::ID intr = meth->intrinsic_id();
2106 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2107 // It does not escape if object is always allocated.
2108 es = PointsToNode::NoEscape;
2109 } else {
2110 // It escapes globally if object could be loaded from cache.
2111 es = PointsToNode::GlobalEscape;
2112 }
2113 add_java_object(call, es);
2114 if (es == PointsToNode::GlobalEscape) {
2115 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2116 }
2117 } else {
2118 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2119 call_analyzer->copy_dependencies(_compile->dependencies());
2120 if (call_analyzer->is_return_allocated()) {
2121 // Returns a newly allocated non-escaped object, simply
2122 // update dependency information.
2123 // Mark it as NoEscape so that objects referenced by
2124 // it's fields will be marked as NoEscape at least.
2125 add_java_object(call, PointsToNode::NoEscape);
2126 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2127 } else {
2128 // Determine whether any arguments are returned.
2129 const TypeTuple* d = call->tf()->domain_cc();
2130 bool ret_arg = false;
2131 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2132 if (d->field_at(i)->isa_ptr() != nullptr &&
2133 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2134 ret_arg = true;
2135 break;
2136 }
2137 }
2138 if (ret_arg) {
2139 add_local_var(call, PointsToNode::ArgEscape);
2140 } else {
2141 // Returns unknown object.
2142 map_ideal_node(call, phantom_obj);
2143 }
2144 }
2145 }
2146 } else {
2147 // An other type of call, assume the worst case:
2148 // returned value is unknown and globally escapes.
2149 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2150 map_ideal_node(call, phantom_obj);
2151 }
2152 }
2153
2154 void ConnectionGraph::process_call_arguments(CallNode *call) {
2155 bool is_arraycopy = false;
2156 switch (call->Opcode()) {
2157 #ifdef ASSERT
2158 case Op_Allocate:
2159 case Op_AllocateArray:
2160 case Op_Lock:
2161 case Op_Unlock:
2162 assert(false, "should be done already");
2163 break;
2164 #endif
2165 case Op_ArrayCopy:
2166 case Op_CallLeafNoFP:
2167 // Most array copies are ArrayCopy nodes at this point but there
2168 // are still a few direct calls to the copy subroutines (See
2169 // PhaseStringOpts::copy_string())
2170 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2171 call->as_CallLeaf()->is_call_to_arraycopystub();
2172 // fall through
2173 case Op_CallLeafVector:
2174 case Op_CallLeaf: {
2175 // Stub calls, objects do not escape but they are not scale replaceable.
2176 // Adjust escape state for outgoing arguments.
2177 const TypeTuple * d = call->tf()->domain_sig();
2178 bool src_has_oops = false;
2179 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2180 const Type* at = d->field_at(i);
2181 Node *arg = call->in(i);
2182 if (arg == nullptr) {
2183 continue;
2184 }
2185 const Type *aat = _igvn->type(arg);
2186 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2187 continue;
2188 }
2189 if (arg->is_AddP()) {
2190 //
2191 // The inline_native_clone() case when the arraycopy stub is called
2192 // after the allocation before Initialize and CheckCastPP nodes.
2193 // Or normal arraycopy for object arrays case.
2194 //
2195 // Set AddP's base (Allocate) as not scalar replaceable since
2196 // pointer to the base (with offset) is passed as argument.
2197 //
2198 arg = get_addp_base(arg);
2199 }
2200 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2201 assert(arg_ptn != nullptr, "should be registered");
2202 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2203 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2204 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2205 aat->isa_ptr() != nullptr, "expecting an Ptr");
2206 bool arg_has_oops = aat->isa_oopptr() &&
2207 (aat->isa_instptr() ||
2208 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2209 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2210 aat->isa_aryptr()->is_flat() &&
2211 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2212 if (i == TypeFunc::Parms) {
2213 src_has_oops = arg_has_oops;
2214 }
2215 //
2216 // src or dst could be j.l.Object when other is basic type array:
2217 //
2218 // arraycopy(char[],0,Object*,0,size);
2219 // arraycopy(Object*,0,char[],0,size);
2220 //
2221 // Don't add edges in such cases.
2222 //
2223 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2224 arg_has_oops && (i > TypeFunc::Parms);
2225 #ifdef ASSERT
2226 if (!(is_arraycopy ||
2227 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2228 (call->as_CallLeaf()->_name != nullptr &&
2229 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2230 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2231 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2232 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2233 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2234 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2235 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2236 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2237 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2238 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2239 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2240 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2241 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2242 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2243 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2244 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2245 strcmp(call->as_CallLeaf()->_name, "kyberNtt") == 0 ||
2246 strcmp(call->as_CallLeaf()->_name, "kyberInverseNtt") == 0 ||
2247 strcmp(call->as_CallLeaf()->_name, "kyberNttMult") == 0 ||
2248 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_2") == 0 ||
2249 strcmp(call->as_CallLeaf()->_name, "kyberAddPoly_3") == 0 ||
2250 strcmp(call->as_CallLeaf()->_name, "kyber12To16") == 0 ||
2251 strcmp(call->as_CallLeaf()->_name, "kyberBarrettReduce") == 0 ||
2252 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostNtt") == 0 ||
2253 strcmp(call->as_CallLeaf()->_name, "dilithiumAlmostInverseNtt") == 0 ||
2254 strcmp(call->as_CallLeaf()->_name, "dilithiumNttMult") == 0 ||
2255 strcmp(call->as_CallLeaf()->_name, "dilithiumMontMulByConstant") == 0 ||
2256 strcmp(call->as_CallLeaf()->_name, "dilithiumDecomposePoly") == 0 ||
2257 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2258 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2259 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2260 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2261 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2262 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2263 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2264 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2265 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2266 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2267 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2268 strcmp(call->as_CallLeaf()->_name, "double_keccak") == 0 ||
2269 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2270 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2271 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2272 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2273 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2274 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2275 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2276 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2277 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2278 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2279 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2280 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2281 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2282 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2283 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2284 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2285 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2286 ))) {
2287 call->dump();
2288 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2289 }
2290 #endif
2291 // Always process arraycopy's destination object since
2292 // we need to add all possible edges to references in
2293 // source object.
2294 if (arg_esc >= PointsToNode::ArgEscape &&
2295 !arg_is_arraycopy_dest) {
2296 continue;
2297 }
2298 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2299 if (call->is_ArrayCopy()) {
2300 ArrayCopyNode* ac = call->as_ArrayCopy();
2301 if (ac->is_clonebasic() ||
2302 ac->is_arraycopy_validated() ||
2303 ac->is_copyof_validated() ||
2304 ac->is_copyofrange_validated()) {
2305 es = PointsToNode::NoEscape;
2306 }
2307 }
2308 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2309 if (arg_is_arraycopy_dest) {
2310 Node* src = call->in(TypeFunc::Parms);
2311 if (src->is_AddP()) {
2312 src = get_addp_base(src);
2313 }
2314 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2315 assert(src_ptn != nullptr, "should be registered");
2316 // Special arraycopy edge:
2317 // Only escape state of destination object's fields affects
2318 // escape state of fields in source object.
2319 add_arraycopy(call, es, src_ptn, arg_ptn);
2320 }
2321 }
2322 }
2323 break;
2324 }
2325 case Op_CallStaticJava: {
2326 // For a static call, we know exactly what method is being called.
2327 // Use bytecode estimator to record the call's escape affects
2328 #ifdef ASSERT
2329 const char* name = call->as_CallStaticJava()->_name;
2330 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2331 #endif
2332 ciMethod* meth = call->as_CallJava()->method();
2333 if ((meth != nullptr) && meth->is_boxing_method()) {
2334 break; // Boxing methods do not modify any oops.
2335 }
2336 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2337 // fall-through if not a Java method or no analyzer information
2338 if (call_analyzer != nullptr) {
2339 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2340 const TypeTuple* d = call->tf()->domain_cc();
2341 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2342 const Type* at = d->field_at(i);
2343 int k = i - TypeFunc::Parms;
2344 Node* arg = call->in(i);
2345 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2346 if (at->isa_ptr() != nullptr &&
2347 call_analyzer->is_arg_returned(k)) {
2348 // The call returns arguments.
2349 if (call_ptn != nullptr) { // Is call's result used?
2350 assert(call_ptn->is_LocalVar(), "node should be registered");
2351 assert(arg_ptn != nullptr, "node should be registered");
2352 add_edge(call_ptn, arg_ptn);
2353 }
2354 }
2355 if (at->isa_oopptr() != nullptr &&
2356 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2357 if (!call_analyzer->is_arg_stack(k)) {
2358 // The argument global escapes
2359 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2360 } else {
2361 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2362 if (!call_analyzer->is_arg_local(k)) {
2363 // The argument itself doesn't escape, but any fields might
2364 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2365 }
2366 }
2367 }
2368 }
2369 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2370 // The call returns arguments.
2371 assert(call_ptn->edge_count() > 0, "sanity");
2372 if (!call_analyzer->is_return_local()) {
2373 // Returns also unknown object.
2374 add_edge(call_ptn, phantom_obj);
2375 }
2376 }
2377 break;
2378 }
2379 }
2380 default: {
2381 // Fall-through here if not a Java method or no analyzer information
2382 // or some other type of call, assume the worst case: all arguments
2383 // globally escape.
2384 const TypeTuple* d = call->tf()->domain_cc();
2385 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2386 const Type* at = d->field_at(i);
2387 if (at->isa_oopptr() != nullptr) {
2388 Node* arg = call->in(i);
2389 if (arg->is_AddP()) {
2390 arg = get_addp_base(arg);
2391 }
2392 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2393 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2394 }
2395 }
2396 }
2397 }
2398 }
2399
2400
2401 // Finish Graph construction.
2402 bool ConnectionGraph::complete_connection_graph(
2403 GrowableArray<PointsToNode*>& ptnodes_worklist,
2404 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2405 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2406 GrowableArray<FieldNode*>& oop_fields_worklist) {
2407 // Normally only 1-3 passes needed to build Connection Graph depending
2408 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2409 // Set limit to 20 to catch situation when something did go wrong and
2410 // bailout Escape Analysis.
2411 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2412 #define GRAPH_BUILD_ITER_LIMIT 20
2413
2414 // Propagate GlobalEscape and ArgEscape escape states and check that
2415 // we still have non-escaping objects. The method pushs on _worklist
2416 // Field nodes which reference phantom_object.
2417 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2418 return false; // Nothing to do.
2419 }
2420 // Now propagate references to all JavaObject nodes.
2421 int java_objects_length = java_objects_worklist.length();
2422 elapsedTimer build_time;
2423 build_time.start();
2424 elapsedTimer time;
2425 bool timeout = false;
2426 int new_edges = 1;
2427 int iterations = 0;
2428 do {
2429 while ((new_edges > 0) &&
2430 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2431 double start_time = time.seconds();
2432 time.start();
2433 new_edges = 0;
2434 // Propagate references to phantom_object for nodes pushed on _worklist
2435 // by find_non_escaped_objects() and find_field_value().
2436 new_edges += add_java_object_edges(phantom_obj, false);
2437 for (int next = 0; next < java_objects_length; ++next) {
2438 JavaObjectNode* ptn = java_objects_worklist.at(next);
2439 new_edges += add_java_object_edges(ptn, true);
2440
2441 #define SAMPLE_SIZE 4
2442 if ((next % SAMPLE_SIZE) == 0) {
2443 // Each 4 iterations calculate how much time it will take
2444 // to complete graph construction.
2445 time.stop();
2446 // Poll for requests from shutdown mechanism to quiesce compiler
2447 // because Connection graph construction may take long time.
2448 CompileBroker::maybe_block();
2449 double stop_time = time.seconds();
2450 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2451 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2452 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2453 timeout = true;
2454 break; // Timeout
2455 }
2456 start_time = stop_time;
2457 time.start();
2458 }
2459 #undef SAMPLE_SIZE
2460
2461 }
2462 if (timeout) break;
2463 if (new_edges > 0) {
2464 // Update escape states on each iteration if graph was updated.
2465 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2466 return false; // Nothing to do.
2467 }
2468 }
2469 time.stop();
2470 if (time.seconds() >= EscapeAnalysisTimeout) {
2471 timeout = true;
2472 break;
2473 }
2474 }
2475 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2476 time.start();
2477 // Find fields which have unknown value.
2478 int fields_length = oop_fields_worklist.length();
2479 for (int next = 0; next < fields_length; next++) {
2480 FieldNode* field = oop_fields_worklist.at(next);
2481 if (field->edge_count() == 0) {
2482 new_edges += find_field_value(field);
2483 // This code may added new edges to phantom_object.
2484 // Need an other cycle to propagate references to phantom_object.
2485 }
2486 }
2487 time.stop();
2488 if (time.seconds() >= EscapeAnalysisTimeout) {
2489 timeout = true;
2490 break;
2491 }
2492 } else {
2493 new_edges = 0; // Bailout
2494 }
2495 } while (new_edges > 0);
2496
2497 build_time.stop();
2498 _build_time = build_time.seconds();
2499 _build_iterations = iterations;
2500
2501 // Bailout if passed limits.
2502 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2503 Compile* C = _compile;
2504 if (C->log() != nullptr) {
2505 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2506 C->log()->text("%s", timeout ? "time" : "iterations");
2507 C->log()->end_elem(" limit'");
2508 }
2509 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2510 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2511 // Possible infinite build_connection_graph loop,
2512 // bailout (no changes to ideal graph were made).
2513 return false;
2514 }
2515
2516 #undef GRAPH_BUILD_ITER_LIMIT
2517
2518 // Find fields initialized by null for non-escaping Allocations.
2519 int non_escaped_length = non_escaped_allocs_worklist.length();
2520 for (int next = 0; next < non_escaped_length; next++) {
2521 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2522 PointsToNode::EscapeState es = ptn->escape_state();
2523 assert(es <= PointsToNode::ArgEscape, "sanity");
2524 if (es == PointsToNode::NoEscape) {
2525 if (find_init_values_null(ptn, _igvn) > 0) {
2526 // Adding references to null object does not change escape states
2527 // since it does not escape. Also no fields are added to null object.
2528 add_java_object_edges(null_obj, false);
2529 }
2530 }
2531 Node* n = ptn->ideal_node();
2532 if (n->is_Allocate()) {
2533 // The object allocated by this Allocate node will never be
2534 // seen by an other thread. Mark it so that when it is
2535 // expanded no MemBarStoreStore is added.
2536 InitializeNode* ini = n->as_Allocate()->initialization();
2537 if (ini != nullptr)
2538 ini->set_does_not_escape();
2539 }
2540 }
2541 return true; // Finished graph construction.
2542 }
2543
2544 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2545 // and check that we still have non-escaping java objects.
2546 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2547 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist) {
2548 GrowableArray<PointsToNode*> escape_worklist;
2549 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2550 int ptnodes_length = ptnodes_worklist.length();
2551 for (int next = 0; next < ptnodes_length; ++next) {
2552 PointsToNode* ptn = ptnodes_worklist.at(next);
2553 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2554 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2555 escape_worklist.push(ptn);
2556 }
2557 }
2558 // Set escape states to referenced nodes (edges list).
2559 while (escape_worklist.length() > 0) {
2560 PointsToNode* ptn = escape_worklist.pop();
2561 PointsToNode::EscapeState es = ptn->escape_state();
2562 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2563 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2564 es >= PointsToNode::ArgEscape) {
2565 // GlobalEscape or ArgEscape state of field means it has unknown value.
2566 if (add_edge(ptn, phantom_obj)) {
2567 // New edge was added
2568 add_field_uses_to_worklist(ptn->as_Field());
2569 }
2570 }
2571 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2572 PointsToNode* e = i.get();
2573 if (e->is_Arraycopy()) {
2574 assert(ptn->arraycopy_dst(), "sanity");
2575 // Propagate only fields escape state through arraycopy edge.
2576 if (e->fields_escape_state() < field_es) {
2577 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2578 escape_worklist.push(e);
2579 }
2580 } else if (es >= field_es) {
2581 // fields_escape_state is also set to 'es' if it is less than 'es'.
2582 if (e->escape_state() < es) {
2583 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2584 escape_worklist.push(e);
2585 }
2586 } else {
2587 // Propagate field escape state.
2588 bool es_changed = false;
2589 if (e->fields_escape_state() < field_es) {
2590 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2591 es_changed = true;
2592 }
2593 if ((e->escape_state() < field_es) &&
2594 e->is_Field() && ptn->is_JavaObject() &&
2595 e->as_Field()->is_oop()) {
2596 // Change escape state of referenced fields.
2597 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2598 es_changed = true;
2599 } else if (e->escape_state() < es) {
2600 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2601 es_changed = true;
2602 }
2603 if (es_changed) {
2604 escape_worklist.push(e);
2605 }
2606 }
2607 }
2608 }
2609 // Remove escaped objects from non_escaped list.
2610 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2611 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2612 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2613 non_escaped_allocs_worklist.delete_at(next);
2614 }
2615 if (ptn->escape_state() == PointsToNode::NoEscape) {
2616 // Find fields in non-escaped allocations which have unknown value.
2617 find_init_values_phantom(ptn);
2618 }
2619 }
2620 return (non_escaped_allocs_worklist.length() > 0);
2621 }
2622
2623 // Add all references to JavaObject node by walking over all uses.
2624 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2625 int new_edges = 0;
2626 if (populate_worklist) {
2627 // Populate _worklist by uses of jobj's uses.
2628 for (UseIterator i(jobj); i.has_next(); i.next()) {
2629 PointsToNode* use = i.get();
2630 if (use->is_Arraycopy()) {
2631 continue;
2632 }
2633 add_uses_to_worklist(use);
2634 if (use->is_Field() && use->as_Field()->is_oop()) {
2635 // Put on worklist all field's uses (loads) and
2636 // related field nodes (same base and offset).
2637 add_field_uses_to_worklist(use->as_Field());
2638 }
2639 }
2640 }
2641 for (int l = 0; l < _worklist.length(); l++) {
2642 PointsToNode* use = _worklist.at(l);
2643 if (PointsToNode::is_base_use(use)) {
2644 // Add reference from jobj to field and from field to jobj (field's base).
2645 use = PointsToNode::get_use_node(use)->as_Field();
2646 if (add_base(use->as_Field(), jobj)) {
2647 new_edges++;
2648 }
2649 continue;
2650 }
2651 assert(!use->is_JavaObject(), "sanity");
2652 if (use->is_Arraycopy()) {
2653 if (jobj == null_obj) { // null object does not have field edges
2654 continue;
2655 }
2656 // Added edge from Arraycopy node to arraycopy's source java object
2657 if (add_edge(use, jobj)) {
2658 jobj->set_arraycopy_src();
2659 new_edges++;
2660 }
2661 // and stop here.
2662 continue;
2663 }
2664 if (!add_edge(use, jobj)) {
2665 continue; // No new edge added, there was such edge already.
2666 }
2667 new_edges++;
2668 if (use->is_LocalVar()) {
2669 add_uses_to_worklist(use);
2670 if (use->arraycopy_dst()) {
2671 for (EdgeIterator i(use); i.has_next(); i.next()) {
2672 PointsToNode* e = i.get();
2673 if (e->is_Arraycopy()) {
2674 if (jobj == null_obj) { // null object does not have field edges
2675 continue;
2676 }
2677 // Add edge from arraycopy's destination java object to Arraycopy node.
2678 if (add_edge(jobj, e)) {
2679 new_edges++;
2680 jobj->set_arraycopy_dst();
2681 }
2682 }
2683 }
2684 }
2685 } else {
2686 // Added new edge to stored in field values.
2687 // Put on worklist all field's uses (loads) and
2688 // related field nodes (same base and offset).
2689 add_field_uses_to_worklist(use->as_Field());
2690 }
2691 }
2692 _worklist.clear();
2693 _in_worklist.reset();
2694 return new_edges;
2695 }
2696
2697 // Put on worklist all related field nodes.
2698 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2699 assert(field->is_oop(), "sanity");
2700 int offset = field->offset();
2701 add_uses_to_worklist(field);
2702 // Loop over all bases of this field and push on worklist Field nodes
2703 // with the same offset and base (since they may reference the same field).
2704 for (BaseIterator i(field); i.has_next(); i.next()) {
2705 PointsToNode* base = i.get();
2706 add_fields_to_worklist(field, base);
2707 // Check if the base was source object of arraycopy and go over arraycopy's
2708 // destination objects since values stored to a field of source object are
2709 // accessible by uses (loads) of fields of destination objects.
2710 if (base->arraycopy_src()) {
2711 for (UseIterator j(base); j.has_next(); j.next()) {
2712 PointsToNode* arycp = j.get();
2713 if (arycp->is_Arraycopy()) {
2714 for (UseIterator k(arycp); k.has_next(); k.next()) {
2715 PointsToNode* abase = k.get();
2716 if (abase->arraycopy_dst() && abase != base) {
2717 // Look for the same arraycopy reference.
2718 add_fields_to_worklist(field, abase);
2719 }
2720 }
2721 }
2722 }
2723 }
2724 }
2725 }
2726
2727 // Put on worklist all related field nodes.
2728 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2729 int offset = field->offset();
2730 if (base->is_LocalVar()) {
2731 for (UseIterator j(base); j.has_next(); j.next()) {
2732 PointsToNode* f = j.get();
2733 if (PointsToNode::is_base_use(f)) { // Field
2734 f = PointsToNode::get_use_node(f);
2735 if (f == field || !f->as_Field()->is_oop()) {
2736 continue;
2737 }
2738 int offs = f->as_Field()->offset();
2739 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2740 add_to_worklist(f);
2741 }
2742 }
2743 }
2744 } else {
2745 assert(base->is_JavaObject(), "sanity");
2746 if (// Skip phantom_object since it is only used to indicate that
2747 // this field's content globally escapes.
2748 (base != phantom_obj) &&
2749 // null object node does not have fields.
2750 (base != null_obj)) {
2751 for (EdgeIterator i(base); i.has_next(); i.next()) {
2752 PointsToNode* f = i.get();
2753 // Skip arraycopy edge since store to destination object field
2754 // does not update value in source object field.
2755 if (f->is_Arraycopy()) {
2756 assert(base->arraycopy_dst(), "sanity");
2757 continue;
2758 }
2759 if (f == field || !f->as_Field()->is_oop()) {
2760 continue;
2761 }
2762 int offs = f->as_Field()->offset();
2763 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2764 add_to_worklist(f);
2765 }
2766 }
2767 }
2768 }
2769 }
2770
2771 // Find fields which have unknown value.
2772 int ConnectionGraph::find_field_value(FieldNode* field) {
2773 // Escaped fields should have init value already.
2774 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2775 int new_edges = 0;
2776 for (BaseIterator i(field); i.has_next(); i.next()) {
2777 PointsToNode* base = i.get();
2778 if (base->is_JavaObject()) {
2779 // Skip Allocate's fields which will be processed later.
2780 if (base->ideal_node()->is_Allocate()) {
2781 return 0;
2782 }
2783 assert(base == null_obj, "only null ptr base expected here");
2784 }
2785 }
2786 if (add_edge(field, phantom_obj)) {
2787 // New edge was added
2788 new_edges++;
2789 add_field_uses_to_worklist(field);
2790 }
2791 return new_edges;
2792 }
2793
2794 // Find fields initializing values for allocations.
2795 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2796 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2797 PointsToNode* init_val = phantom_obj;
2798 Node* alloc = pta->ideal_node();
2799
2800 // Do nothing for Allocate nodes since its fields values are
2801 // "known" unless they are initialized by arraycopy/clone.
2802 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2803 if (alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2804 // Null-free inline type arrays are initialized with an init value instead of null
2805 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::InitValue)->_idx);
2806 assert(init_val != nullptr, "init value should be registered");
2807 } else {
2808 return 0;
2809 }
2810 }
2811 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
2812 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
2813 #ifdef ASSERT
2814 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
2815 const char* name = alloc->as_CallStaticJava()->_name;
2816 assert(strncmp(name, "C2 Runtime multianewarray", 25) == 0 ||
2817 strncmp(name, "C2 Runtime load_unknown_inline", 30) == 0, "sanity");
2818 }
2819 #endif
2820 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2821 int new_edges = 0;
2822 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2823 PointsToNode* field = i.get();
2824 if (field->is_Field() && field->as_Field()->is_oop()) {
2825 if (add_edge(field, init_val)) {
2826 // New edge was added
2827 new_edges++;
2828 add_field_uses_to_worklist(field->as_Field());
2829 }
2830 }
2831 }
2832 return new_edges;
2833 }
2834
2835 // Find fields initializing values for allocations.
2836 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2837 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2838 Node* alloc = pta->ideal_node();
2839 // Do nothing for Call nodes since its fields values are unknown.
2840 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::InitValue) != nullptr) {
2841 return 0;
2842 }
2843 InitializeNode* ini = alloc->as_Allocate()->initialization();
2844 bool visited_bottom_offset = false;
2845 GrowableArray<int> offsets_worklist;
2846 int new_edges = 0;
2847
2848 // Check if an oop field's initializing value is recorded and add
2849 // a corresponding null if field's value if it is not recorded.
2850 // Connection Graph does not record a default initialization by null
2851 // captured by Initialize node.
2852 //
2853 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2854 PointsToNode* field = i.get(); // Field (AddP)
2855 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2856 continue; // Not oop field
2857 }
2858 int offset = field->as_Field()->offset();
2859 if (offset == Type::OffsetBot) {
2860 if (!visited_bottom_offset) {
2861 // OffsetBot is used to reference array's element,
2862 // always add reference to null to all Field nodes since we don't
2863 // known which element is referenced.
2864 if (add_edge(field, null_obj)) {
2865 // New edge was added
2866 new_edges++;
2867 add_field_uses_to_worklist(field->as_Field());
2868 visited_bottom_offset = true;
2869 }
2870 }
2871 } else {
2872 // Check only oop fields.
2873 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2874 if (adr_type->isa_rawptr()) {
2875 #ifdef ASSERT
2876 // Raw pointers are used for initializing stores so skip it
2877 // since it should be recorded already
2878 Node* base = get_addp_base(field->ideal_node());
2879 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2880 #endif
2881 continue;
2882 }
2883 if (!offsets_worklist.contains(offset)) {
2884 offsets_worklist.append(offset);
2885 Node* value = nullptr;
2886 if (ini != nullptr) {
2887 // StoreP::value_basic_type() == T_ADDRESS
2888 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2889 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2890 // Make sure initializing store has the same type as this AddP.
2891 // This AddP may reference non existing field because it is on a
2892 // dead branch of bimorphic call which is not eliminated yet.
2893 if (store != nullptr && store->is_Store() &&
2894 store->as_Store()->value_basic_type() == ft) {
2895 value = store->in(MemNode::ValueIn);
2896 #ifdef ASSERT
2897 if (VerifyConnectionGraph) {
2898 // Verify that AddP already points to all objects the value points to.
2899 PointsToNode* val = ptnode_adr(value->_idx);
2900 assert((val != nullptr), "should be processed already");
2901 PointsToNode* missed_obj = nullptr;
2902 if (val->is_JavaObject()) {
2903 if (!field->points_to(val->as_JavaObject())) {
2904 missed_obj = val;
2905 }
2906 } else {
2907 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2908 tty->print_cr("----------init store has invalid value -----");
2909 store->dump();
2910 val->dump();
2911 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2912 }
2913 for (EdgeIterator j(val); j.has_next(); j.next()) {
2914 PointsToNode* obj = j.get();
2915 if (obj->is_JavaObject()) {
2916 if (!field->points_to(obj->as_JavaObject())) {
2917 missed_obj = obj;
2918 break;
2919 }
2920 }
2921 }
2922 }
2923 if (missed_obj != nullptr) {
2924 tty->print_cr("----------field---------------------------------");
2925 field->dump();
2926 tty->print_cr("----------missed reference to object------------");
2927 missed_obj->dump();
2928 tty->print_cr("----------object referenced by init store-------");
2929 store->dump();
2930 val->dump();
2931 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2932 }
2933 }
2934 #endif
2935 } else {
2936 // There could be initializing stores which follow allocation.
2937 // For example, a volatile field store is not collected
2938 // by Initialize node.
2939 //
2940 // Need to check for dependent loads to separate such stores from
2941 // stores which follow loads. For now, add initial value null so
2942 // that compare pointers optimization works correctly.
2943 }
2944 }
2945 if (value == nullptr) {
2946 // A field's initializing value was not recorded. Add null.
2947 if (add_edge(field, null_obj)) {
2948 // New edge was added
2949 new_edges++;
2950 add_field_uses_to_worklist(field->as_Field());
2951 }
2952 }
2953 }
2954 }
2955 }
2956 return new_edges;
2957 }
2958
2959 // Adjust scalar_replaceable state after Connection Graph is built.
2960 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2961 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2962 // returns true. If one of the constraints in this method set 'jobj' to NSR
2963 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2964 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2965 // that other object and the Phi will become a reducible Phi.
2966 // There could be multiple merges involving the same jobj.
2967 Unique_Node_List candidates;
2968
2969 // Search for non-escaping objects which are not scalar replaceable
2970 // and mark them to propagate the state to referenced objects.
2971
2972 for (UseIterator i(jobj); i.has_next(); i.next()) {
2973 PointsToNode* use = i.get();
2974 if (use->is_Arraycopy()) {
2975 continue;
2976 }
2977 if (use->is_Field()) {
2978 FieldNode* field = use->as_Field();
2979 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2980 // 1. An object is not scalar replaceable if the field into which it is
2981 // stored has unknown offset (stored into unknown element of an array).
2982 if (field->offset() == Type::OffsetBot) {
2983 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2984 return;
2985 }
2986 for (BaseIterator i(field); i.has_next(); i.next()) {
2987 PointsToNode* base = i.get();
2988 // 2. An object is not scalar replaceable if the field into which it is
2989 // stored has multiple bases one of which is null.
2990 if ((base == null_obj) && (field->base_count() > 1)) {
2991 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2992 return;
2993 }
2994 // 2.5. An object is not scalar replaceable if the field into which it is
2995 // stored has NSR base.
2996 if (!base->scalar_replaceable()) {
2997 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2998 return;
2999 }
3000 }
3001 }
3002 assert(use->is_Field() || use->is_LocalVar(), "sanity");
3003 // 3. An object is not scalar replaceable if it is merged with other objects
3004 // and we can't remove the merge
3005 for (EdgeIterator j(use); j.has_next(); j.next()) {
3006 PointsToNode* ptn = j.get();
3007 if (ptn->is_JavaObject() && ptn != jobj) {
3008 Node* use_n = use->ideal_node();
3009
3010 // These other local vars may point to multiple objects through a Phi
3011 // In this case we skip them and see if we can reduce the Phi.
3012 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
3013 use_n = use_n->in(1);
3014 }
3015
3016 // If it's already a candidate or confirmed reducible merge we can skip verification
3017 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3018 continue;
3019 }
3020
3021 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3022 candidates.push(use_n);
3023 } else {
3024 // Mark all objects as NSR if we can't remove the merge
3025 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3026 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3027 }
3028 }
3029 }
3030 if (!jobj->scalar_replaceable()) {
3031 return;
3032 }
3033 }
3034
3035 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3036 if (j.get()->is_Arraycopy()) {
3037 continue;
3038 }
3039
3040 // Non-escaping object node should point only to field nodes.
3041 FieldNode* field = j.get()->as_Field();
3042 int offset = field->as_Field()->offset();
3043
3044 // 4. An object is not scalar replaceable if it has a field with unknown
3045 // offset (array's element is accessed in loop).
3046 if (offset == Type::OffsetBot) {
3047 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3048 return;
3049 }
3050 // 5. Currently an object is not scalar replaceable if a LoadStore node
3051 // access its field since the field value is unknown after it.
3052 //
3053 Node* n = field->ideal_node();
3054
3055 // Test for an unsafe access that was parsed as maybe off heap
3056 // (with a CheckCastPP to raw memory).
3057 assert(n->is_AddP(), "expect an address computation");
3058 if (n->in(AddPNode::Base)->is_top() &&
3059 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3060 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3061 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3062 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3063 return;
3064 }
3065
3066 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3067 Node* u = n->fast_out(i);
3068 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3069 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3070 return;
3071 }
3072 }
3073
3074 // 6. Or the address may point to more then one object. This may produce
3075 // the false positive result (set not scalar replaceable)
3076 // since the flow-insensitive escape analysis can't separate
3077 // the case when stores overwrite the field's value from the case
3078 // when stores happened on different control branches.
3079 //
3080 // Note: it will disable scalar replacement in some cases:
3081 //
3082 // Point p[] = new Point[1];
3083 // p[0] = new Point(); // Will be not scalar replaced
3084 //
3085 // but it will save us from incorrect optimizations in next cases:
3086 //
3087 // Point p[] = new Point[1];
3088 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3089 //
3090 if (field->base_count() > 1 && candidates.size() == 0) {
3091 if (has_non_reducible_merge(field, reducible_merges)) {
3092 for (BaseIterator i(field); i.has_next(); i.next()) {
3093 PointsToNode* base = i.get();
3094 // Don't take into account LocalVar nodes which
3095 // may point to only one object which should be also
3096 // this field's base by now.
3097 if (base->is_JavaObject() && base != jobj) {
3098 // Mark all bases.
3099 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3100 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3101 }
3102 }
3103
3104 if (!jobj->scalar_replaceable()) {
3105 return;
3106 }
3107 }
3108 }
3109 }
3110
3111 // The candidate is truly a reducible merge only if none of the other
3112 // constraints ruled it as NSR. There could be multiple merges involving the
3113 // same jobj.
3114 assert(jobj->scalar_replaceable(), "sanity");
3115 for (uint i = 0; i < candidates.size(); i++ ) {
3116 Node* candidate = candidates.at(i);
3117 reducible_merges.push(candidate);
3118 }
3119 }
3120
3121 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3122 for (BaseIterator i(field); i.has_next(); i.next()) {
3123 Node* base = i.get()->ideal_node();
3124 if (base->is_Phi() && !reducible_merges.member(base)) {
3125 return true;
3126 }
3127 }
3128 return false;
3129 }
3130
3131 void ConnectionGraph::revisit_reducible_phi_status(JavaObjectNode* jobj, Unique_Node_List& reducible_merges) {
3132 assert(jobj != nullptr && !jobj->scalar_replaceable(), "jobj should be set as NSR before calling this function.");
3133
3134 // Look for 'phis' that refer to 'jobj' as the last
3135 // remaining scalar replaceable input.
3136 uint reducible_merges_cnt = reducible_merges.size();
3137 for (uint i = 0; i < reducible_merges_cnt; i++) {
3138 Node* phi = reducible_merges.at(i);
3139
3140 // This 'Phi' will be a 'good' if it still points to
3141 // at least one scalar replaceable object. Note that 'obj'
3142 // was/should be marked as NSR before calling this function.
3143 bool good_phi = false;
3144
3145 for (uint j = 1; j < phi->req(); j++) {
3146 JavaObjectNode* phi_in_obj = unique_java_object(phi->in(j));
3147 if (phi_in_obj != nullptr && phi_in_obj->scalar_replaceable()) {
3148 good_phi = true;
3149 break;
3150 }
3151 }
3152
3153 if (!good_phi) {
3154 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Phi %d became non-reducible after node %d became NSR.", phi->_idx, jobj->ideal_node()->_idx);)
3155 reducible_merges.remove(i);
3156
3157 // Decrement the index because the 'remove' call above actually
3158 // moves the last entry of the list to position 'i'.
3159 i--;
3160
3161 reducible_merges_cnt--;
3162 }
3163 }
3164 }
3165
3166 // Propagate NSR (Not scalar replaceable) state.
3167 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist, Unique_Node_List &reducible_merges) {
3168 int jobj_length = jobj_worklist.length();
3169 bool found_nsr_alloc = true;
3170 while (found_nsr_alloc) {
3171 found_nsr_alloc = false;
3172 for (int next = 0; next < jobj_length; ++next) {
3173 JavaObjectNode* jobj = jobj_worklist.at(next);
3174 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3175 PointsToNode* use = i.get();
3176 if (use->is_Field()) {
3177 FieldNode* field = use->as_Field();
3178 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3179 assert(field->offset() != Type::OffsetBot, "sanity");
3180 for (BaseIterator i(field); i.has_next(); i.next()) {
3181 PointsToNode* base = i.get();
3182 // An object is not scalar replaceable if the field into which
3183 // it is stored has NSR base.
3184 if ((base != null_obj) && !base->scalar_replaceable()) {
3185 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3186 // Any merge that had only 'jobj' as scalar-replaceable will now be non-reducible,
3187 // because there is no point in reducing a Phi that won't improve the number of SR
3188 // objects.
3189 revisit_reducible_phi_status(jobj, reducible_merges);
3190 found_nsr_alloc = true;
3191 break;
3192 }
3193 }
3194 }
3195 }
3196 }
3197 }
3198 }
3199
3200 #ifdef ASSERT
3201 void ConnectionGraph::verify_connection_graph(
3202 GrowableArray<PointsToNode*>& ptnodes_worklist,
3203 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3204 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3205 GrowableArray<Node*>& addp_worklist) {
3206 // Verify that graph is complete - no new edges could be added.
3207 int java_objects_length = java_objects_worklist.length();
3208 int non_escaped_length = non_escaped_allocs_worklist.length();
3209 int new_edges = 0;
3210 for (int next = 0; next < java_objects_length; ++next) {
3211 JavaObjectNode* ptn = java_objects_worklist.at(next);
3212 new_edges += add_java_object_edges(ptn, true);
3213 }
3214 assert(new_edges == 0, "graph was not complete");
3215 // Verify that escape state is final.
3216 int length = non_escaped_allocs_worklist.length();
3217 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist);
3218 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3219 (non_escaped_length == length) &&
3220 (_worklist.length() == 0), "escape state was not final");
3221
3222 // Verify fields information.
3223 int addp_length = addp_worklist.length();
3224 for (int next = 0; next < addp_length; ++next ) {
3225 Node* n = addp_worklist.at(next);
3226 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3227 if (field->is_oop()) {
3228 // Verify that field has all bases
3229 Node* base = get_addp_base(n);
3230 PointsToNode* ptn = ptnode_adr(base->_idx);
3231 if (ptn->is_JavaObject()) {
3232 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3233 } else {
3234 assert(ptn->is_LocalVar(), "sanity");
3235 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3236 PointsToNode* e = i.get();
3237 if (e->is_JavaObject()) {
3238 assert(field->has_base(e->as_JavaObject()), "sanity");
3239 }
3240 }
3241 }
3242 // Verify that all fields have initializing values.
3243 if (field->edge_count() == 0) {
3244 tty->print_cr("----------field does not have references----------");
3245 field->dump();
3246 for (BaseIterator i(field); i.has_next(); i.next()) {
3247 PointsToNode* base = i.get();
3248 tty->print_cr("----------field has next base---------------------");
3249 base->dump();
3250 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3251 tty->print_cr("----------base has fields-------------------------");
3252 for (EdgeIterator j(base); j.has_next(); j.next()) {
3253 j.get()->dump();
3254 }
3255 tty->print_cr("----------base has references---------------------");
3256 for (UseIterator j(base); j.has_next(); j.next()) {
3257 j.get()->dump();
3258 }
3259 }
3260 }
3261 for (UseIterator i(field); i.has_next(); i.next()) {
3262 i.get()->dump();
3263 }
3264 assert(field->edge_count() > 0, "sanity");
3265 }
3266 }
3267 }
3268 }
3269 #endif
3270
3271 // Optimize ideal graph.
3272 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3273 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3274 Compile* C = _compile;
3275 PhaseIterGVN* igvn = _igvn;
3276 if (EliminateLocks) {
3277 // Mark locks before changing ideal graph.
3278 int cnt = C->macro_count();
3279 for (int i = 0; i < cnt; i++) {
3280 Node *n = C->macro_node(i);
3281 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3282 AbstractLockNode* alock = n->as_AbstractLock();
3283 if (!alock->is_non_esc_obj()) {
3284 const Type* obj_type = igvn->type(alock->obj_node());
3285 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3286 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3287 // The lock could be marked eliminated by lock coarsening
3288 // code during first IGVN before EA. Replace coarsened flag
3289 // to eliminate all associated locks/unlocks.
3290 #ifdef ASSERT
3291 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3292 #endif
3293 alock->set_non_esc_obj();
3294 }
3295 }
3296 }
3297 }
3298 }
3299
3300 if (OptimizePtrCompare) {
3301 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3302 Node *n = ptr_cmp_worklist.at(i);
3303 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3304 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3305 if (tcmp->singleton()) {
3306 Node* cmp = igvn->makecon(tcmp);
3307 #ifndef PRODUCT
3308 if (PrintOptimizePtrCompare) {
3309 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"));
3310 if (Verbose) {
3311 n->dump(1);
3312 }
3313 }
3314 #endif
3315 igvn->replace_node(n, cmp);
3316 }
3317 }
3318 }
3319
3320 // For MemBarStoreStore nodes added in library_call.cpp, check
3321 // escape status of associated AllocateNode and optimize out
3322 // MemBarStoreStore node if the allocated object never escapes.
3323 for (int i = 0; i < storestore_worklist.length(); i++) {
3324 Node* storestore = storestore_worklist.at(i);
3325 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3326 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3327 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3328 // Non-escaping inline type buffer allocations don't require a membar
3329 storestore->as_MemBar()->remove(_igvn);
3330 } else {
3331 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3332 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3333 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3334 igvn->register_new_node_with_optimizer(mb);
3335 igvn->replace_node(storestore, mb);
3336 }
3337 }
3338 }
3339 }
3340
3341 // Optimize objects compare.
3342 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3343 assert(OptimizePtrCompare, "sanity");
3344 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3345 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3346 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3347
3348 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3349 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3350 JavaObjectNode* jobj1 = unique_java_object(left);
3351 JavaObjectNode* jobj2 = unique_java_object(right);
3352
3353 // The use of this method during allocation merge reduction may cause 'left'
3354 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3355 // that doesn't reference an unique java object.
3356 if (ptn1 == nullptr || ptn2 == nullptr ||
3357 jobj1 == nullptr || jobj2 == nullptr) {
3358 return UNKNOWN;
3359 }
3360
3361 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3362 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3363
3364 // Check simple cases first.
3365 if (jobj1 != nullptr) {
3366 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3367 if (jobj1 == jobj2) {
3368 // Comparing the same not escaping object.
3369 return EQ;
3370 }
3371 Node* obj = jobj1->ideal_node();
3372 // Comparing not escaping allocation.
3373 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3374 !ptn2->points_to(jobj1)) {
3375 return NE; // This includes nullness check.
3376 }
3377 }
3378 }
3379 if (jobj2 != nullptr) {
3380 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3381 Node* obj = jobj2->ideal_node();
3382 // Comparing not escaping allocation.
3383 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3384 !ptn1->points_to(jobj2)) {
3385 return NE; // This includes nullness check.
3386 }
3387 }
3388 }
3389 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3390 jobj2 != nullptr && jobj2 != phantom_obj &&
3391 jobj1->ideal_node()->is_Con() &&
3392 jobj2->ideal_node()->is_Con()) {
3393 // Klass or String constants compare. Need to be careful with
3394 // compressed pointers - compare types of ConN and ConP instead of nodes.
3395 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3396 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3397 if (t1->make_ptr() == t2->make_ptr()) {
3398 return EQ;
3399 } else {
3400 return NE;
3401 }
3402 }
3403 if (ptn1->meet(ptn2)) {
3404 return UNKNOWN; // Sets are not disjoint
3405 }
3406
3407 // Sets are disjoint.
3408 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3409 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3410 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3411 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3412 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3413 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3414 // Check nullness of unknown object.
3415 return UNKNOWN;
3416 }
3417
3418 // Disjointness by itself is not sufficient since
3419 // alias analysis is not complete for escaped objects.
3420 // Disjoint sets are definitely unrelated only when
3421 // at least one set has only not escaping allocations.
3422 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3423 if (ptn1->non_escaping_allocation()) {
3424 return NE;
3425 }
3426 }
3427 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3428 if (ptn2->non_escaping_allocation()) {
3429 return NE;
3430 }
3431 }
3432 return UNKNOWN;
3433 }
3434
3435 // Connection Graph construction functions.
3436
3437 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3438 PointsToNode* ptadr = _nodes.at(n->_idx);
3439 if (ptadr != nullptr) {
3440 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3441 return;
3442 }
3443 Compile* C = _compile;
3444 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3445 map_ideal_node(n, ptadr);
3446 }
3447
3448 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3449 PointsToNode* ptadr = _nodes.at(n->_idx);
3450 if (ptadr != nullptr) {
3451 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3452 return ptadr;
3453 }
3454 Compile* C = _compile;
3455 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3456 map_ideal_node(n, ptadr);
3457 return ptadr;
3458 }
3459
3460 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3461 PointsToNode* ptadr = _nodes.at(n->_idx);
3462 if (ptadr != nullptr) {
3463 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3464 return;
3465 }
3466 bool unsafe = false;
3467 bool is_oop = is_oop_field(n, offset, &unsafe);
3468 if (unsafe) {
3469 es = PointsToNode::GlobalEscape;
3470 }
3471 Compile* C = _compile;
3472 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3473 map_ideal_node(n, field);
3474 }
3475
3476 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3477 PointsToNode* src, PointsToNode* dst) {
3478 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3479 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3480 PointsToNode* ptadr = _nodes.at(n->_idx);
3481 if (ptadr != nullptr) {
3482 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3483 return;
3484 }
3485 Compile* C = _compile;
3486 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3487 map_ideal_node(n, ptadr);
3488 // Add edge from arraycopy node to source object.
3489 (void)add_edge(ptadr, src);
3490 src->set_arraycopy_src();
3491 // Add edge from destination object to arraycopy node.
3492 (void)add_edge(dst, ptadr);
3493 dst->set_arraycopy_dst();
3494 }
3495
3496 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3497 const Type* adr_type = n->as_AddP()->bottom_type();
3498 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3499 BasicType bt = T_INT;
3500 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3501 // Check only oop fields.
3502 if (!adr_type->isa_aryptr() ||
3503 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3504 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3505 // OffsetBot is used to reference array's element. Ignore first AddP.
3506 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3507 bt = T_OBJECT;
3508 }
3509 }
3510 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3511 if (adr_type->isa_instptr()) {
3512 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3513 if (field != nullptr) {
3514 bt = field->layout_type();
3515 } else {
3516 // Check for unsafe oop field access
3517 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3518 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3519 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3520 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3521 bt = T_OBJECT;
3522 (*unsafe) = true;
3523 }
3524 }
3525 } else if (adr_type->isa_aryptr()) {
3526 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3527 // Ignore array length load.
3528 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3529 // Ignore first AddP.
3530 } else {
3531 const Type* elemtype = adr_type->is_aryptr()->elem();
3532 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3533 ciInlineKlass* vk = elemtype->inline_klass();
3534 field_offset += vk->payload_offset();
3535 ciField* field = vk->get_field_by_offset(field_offset, false);
3536 if (field != nullptr) {
3537 bt = field->layout_type();
3538 } else {
3539 assert(field_offset == vk->payload_offset() + vk->null_marker_offset_in_payload(), "no field or null marker of %s at offset %d", vk->name()->as_utf8(), field_offset);
3540 bt = T_BOOLEAN;
3541 }
3542 } else {
3543 bt = elemtype->array_element_basic_type();
3544 }
3545 }
3546 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3547 // Allocation initialization, ThreadLocal field access, unsafe access
3548 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3549 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3550 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3551 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3552 bt = T_OBJECT;
3553 }
3554 }
3555 }
3556 // Note: T_NARROWOOP is not classed as a real reference type
3557 return (is_reference_type(bt) || bt == T_NARROWOOP);
3558 }
3559
3560 // Returns unique pointed java object or null.
3561 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3562 // If the node was created after the escape computation we can't answer.
3563 uint idx = n->_idx;
3564 if (idx >= nodes_size()) {
3565 return nullptr;
3566 }
3567 PointsToNode* ptn = ptnode_adr(idx);
3568 if (ptn == nullptr) {
3569 return nullptr;
3570 }
3571 if (ptn->is_JavaObject()) {
3572 return ptn->as_JavaObject();
3573 }
3574 assert(ptn->is_LocalVar(), "sanity");
3575 // Check all java objects it points to.
3576 JavaObjectNode* jobj = nullptr;
3577 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3578 PointsToNode* e = i.get();
3579 if (e->is_JavaObject()) {
3580 if (jobj == nullptr) {
3581 jobj = e->as_JavaObject();
3582 } else if (jobj != e) {
3583 return nullptr;
3584 }
3585 }
3586 }
3587 return jobj;
3588 }
3589
3590 // Return true if this node points only to non-escaping allocations.
3591 bool PointsToNode::non_escaping_allocation() {
3592 if (is_JavaObject()) {
3593 Node* n = ideal_node();
3594 if (n->is_Allocate() || n->is_CallStaticJava()) {
3595 return (escape_state() == PointsToNode::NoEscape);
3596 } else {
3597 return false;
3598 }
3599 }
3600 assert(is_LocalVar(), "sanity");
3601 // Check all java objects it points to.
3602 for (EdgeIterator i(this); i.has_next(); i.next()) {
3603 PointsToNode* e = i.get();
3604 if (e->is_JavaObject()) {
3605 Node* n = e->ideal_node();
3606 if ((e->escape_state() != PointsToNode::NoEscape) ||
3607 !(n->is_Allocate() || n->is_CallStaticJava())) {
3608 return false;
3609 }
3610 }
3611 }
3612 return true;
3613 }
3614
3615 // Return true if we know the node does not escape globally.
3616 bool ConnectionGraph::not_global_escape(Node *n) {
3617 assert(!_collecting, "should not call during graph construction");
3618 // If the node was created after the escape computation we can't answer.
3619 uint idx = n->_idx;
3620 if (idx >= nodes_size()) {
3621 return false;
3622 }
3623 PointsToNode* ptn = ptnode_adr(idx);
3624 if (ptn == nullptr) {
3625 return false; // not in congraph (e.g. ConI)
3626 }
3627 PointsToNode::EscapeState es = ptn->escape_state();
3628 // If we have already computed a value, return it.
3629 if (es >= PointsToNode::GlobalEscape) {
3630 return false;
3631 }
3632 if (ptn->is_JavaObject()) {
3633 return true; // (es < PointsToNode::GlobalEscape);
3634 }
3635 assert(ptn->is_LocalVar(), "sanity");
3636 // Check all java objects it points to.
3637 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3638 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3639 return false;
3640 }
3641 }
3642 return true;
3643 }
3644
3645 // Return true if locked object does not escape globally
3646 // and locked code region (identified by BoxLockNode) is balanced:
3647 // all compiled code paths have corresponding Lock/Unlock pairs.
3648 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3649 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3650 if (EliminateNestedLocks) {
3651 // We can mark whole locking region as Local only when only
3652 // one object is used for locking.
3653 alock->box_node()->as_BoxLock()->set_local();
3654 }
3655 return true;
3656 }
3657 return false;
3658 }
3659
3660 // Helper functions
3661
3662 // Return true if this node points to specified node or nodes it points to.
3663 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3664 if (is_JavaObject()) {
3665 return (this == ptn);
3666 }
3667 assert(is_LocalVar() || is_Field(), "sanity");
3668 for (EdgeIterator i(this); i.has_next(); i.next()) {
3669 if (i.get() == ptn) {
3670 return true;
3671 }
3672 }
3673 return false;
3674 }
3675
3676 // Return true if one node points to an other.
3677 bool PointsToNode::meet(PointsToNode* ptn) {
3678 if (this == ptn) {
3679 return true;
3680 } else if (ptn->is_JavaObject()) {
3681 return this->points_to(ptn->as_JavaObject());
3682 } else if (this->is_JavaObject()) {
3683 return ptn->points_to(this->as_JavaObject());
3684 }
3685 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3686 int ptn_count = ptn->edge_count();
3687 for (EdgeIterator i(this); i.has_next(); i.next()) {
3688 PointsToNode* this_e = i.get();
3689 for (int j = 0; j < ptn_count; j++) {
3690 if (this_e == ptn->edge(j)) {
3691 return true;
3692 }
3693 }
3694 }
3695 return false;
3696 }
3697
3698 #ifdef ASSERT
3699 // Return true if bases point to this java object.
3700 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3701 for (BaseIterator i(this); i.has_next(); i.next()) {
3702 if (i.get() == jobj) {
3703 return true;
3704 }
3705 }
3706 return false;
3707 }
3708 #endif
3709
3710 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3711 // Handle simple case first.
3712 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3713 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3714 return true;
3715 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3716 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3717 Node* addp_use = addp->fast_out(i);
3718 if (addp_use->is_Store()) {
3719 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3720 if (addp_use->fast_out(j)->is_Initialize()) {
3721 return true;
3722 }
3723 }
3724 }
3725 }
3726 }
3727 return false;
3728 }
3729
3730 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3731 const Type *adr_type = phase->type(adr);
3732 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3733 // We are computing a raw address for a store captured by an Initialize
3734 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3735 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3736 assert(offs != Type::OffsetBot ||
3737 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3738 "offset must be a constant or it is initialization of array");
3739 return offs;
3740 }
3741 return adr_type->is_ptr()->flat_offset();
3742 }
3743
3744 Node* ConnectionGraph::get_addp_base(Node *addp) {
3745 assert(addp->is_AddP(), "must be AddP");
3746 //
3747 // AddP cases for Base and Address inputs:
3748 // case #1. Direct object's field reference:
3749 // Allocate
3750 // |
3751 // Proj #5 ( oop result )
3752 // |
3753 // CheckCastPP (cast to instance type)
3754 // | |
3755 // AddP ( base == address )
3756 //
3757 // case #2. Indirect object's field reference:
3758 // Phi
3759 // |
3760 // CastPP (cast to instance type)
3761 // | |
3762 // AddP ( base == address )
3763 //
3764 // case #3. Raw object's field reference for Initialize node:
3765 // Allocate
3766 // |
3767 // Proj #5 ( oop result )
3768 // top |
3769 // \ |
3770 // AddP ( base == top )
3771 //
3772 // case #4. Array's element reference:
3773 // {CheckCastPP | CastPP}
3774 // | | |
3775 // | AddP ( array's element offset )
3776 // | |
3777 // AddP ( array's offset )
3778 //
3779 // case #5. Raw object's field reference for arraycopy stub call:
3780 // The inline_native_clone() case when the arraycopy stub is called
3781 // after the allocation before Initialize and CheckCastPP nodes.
3782 // Allocate
3783 // |
3784 // Proj #5 ( oop result )
3785 // | |
3786 // AddP ( base == address )
3787 //
3788 // case #6. Constant Pool, ThreadLocal, CastX2P or
3789 // Raw object's field reference:
3790 // {ConP, ThreadLocal, CastX2P, raw Load}
3791 // top |
3792 // \ |
3793 // AddP ( base == top )
3794 //
3795 // case #7. Klass's field reference.
3796 // LoadKlass
3797 // | |
3798 // AddP ( base == address )
3799 //
3800 // case #8. narrow Klass's field reference.
3801 // LoadNKlass
3802 // |
3803 // DecodeN
3804 // | |
3805 // AddP ( base == address )
3806 //
3807 // case #9. Mixed unsafe access
3808 // {instance}
3809 // |
3810 // CheckCastPP (raw)
3811 // top |
3812 // \ |
3813 // AddP ( base == top )
3814 //
3815 Node *base = addp->in(AddPNode::Base);
3816 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3817 base = addp->in(AddPNode::Address);
3818 while (base->is_AddP()) {
3819 // Case #6 (unsafe access) may have several chained AddP nodes.
3820 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3821 base = base->in(AddPNode::Address);
3822 }
3823 if (base->Opcode() == Op_CheckCastPP &&
3824 base->bottom_type()->isa_rawptr() &&
3825 _igvn->type(base->in(1))->isa_oopptr()) {
3826 base = base->in(1); // Case #9
3827 } else {
3828 Node* uncast_base = base->uncast();
3829 int opcode = uncast_base->Opcode();
3830 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3831 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3832 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3833 is_captured_store_address(addp), "sanity");
3834 }
3835 }
3836 return base;
3837 }
3838
3839 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3840 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3841 Node* addp2 = addp->raw_out(0);
3842 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3843 addp2->in(AddPNode::Base) == n &&
3844 addp2->in(AddPNode::Address) == addp) {
3845 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3846 //
3847 // Find array's offset to push it on worklist first and
3848 // as result process an array's element offset first (pushed second)
3849 // to avoid CastPP for the array's offset.
3850 // Otherwise the inserted CastPP (LocalVar) will point to what
3851 // the AddP (Field) points to. Which would be wrong since
3852 // the algorithm expects the CastPP has the same point as
3853 // as AddP's base CheckCastPP (LocalVar).
3854 //
3855 // ArrayAllocation
3856 // |
3857 // CheckCastPP
3858 // |
3859 // memProj (from ArrayAllocation CheckCastPP)
3860 // | ||
3861 // | || Int (element index)
3862 // | || | ConI (log(element size))
3863 // | || | /
3864 // | || LShift
3865 // | || /
3866 // | AddP (array's element offset)
3867 // | |
3868 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3869 // | / /
3870 // AddP (array's offset)
3871 // |
3872 // Load/Store (memory operation on array's element)
3873 //
3874 return addp2;
3875 }
3876 return nullptr;
3877 }
3878
3879 //
3880 // Adjust the type and inputs of an AddP which computes the
3881 // address of a field of an instance
3882 //
3883 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3884 PhaseGVN* igvn = _igvn;
3885 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3886 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3887 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3888 if (t == nullptr) {
3889 // We are computing a raw address for a store captured by an Initialize
3890 // compute an appropriate address type (cases #3 and #5).
3891 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3892 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3893 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3894 assert(offs != Type::OffsetBot, "offset must be a constant");
3895 if (base_t->isa_aryptr() != nullptr) {
3896 // In the case of a flat inline type array, each field has its
3897 // own slice so we need to extract the field being accessed from
3898 // the address computation
3899 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
3900 } else {
3901 t = base_t->add_offset(offs)->is_oopptr();
3902 }
3903 }
3904 int inst_id = base_t->instance_id();
3905 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3906 "old type must be non-instance or match new type");
3907
3908 // The type 't' could be subclass of 'base_t'.
3909 // As result t->offset() could be large then base_t's size and it will
3910 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3911 // constructor verifies correctness of the offset.
3912 //
3913 // It could happened on subclass's branch (from the type profiling
3914 // inlining) which was not eliminated during parsing since the exactness
3915 // of the allocation type was not propagated to the subclass type check.
3916 //
3917 // Or the type 't' could be not related to 'base_t' at all.
3918 // It could happen when CHA type is different from MDO type on a dead path
3919 // (for example, from instanceof check) which is not collapsed during parsing.
3920 //
3921 // Do nothing for such AddP node and don't process its users since
3922 // this code branch will go away.
3923 //
3924 if (!t->is_known_instance() &&
3925 !base_t->maybe_java_subtype_of(t)) {
3926 return false; // bail out
3927 }
3928 const TypePtr* tinst = base_t->add_offset(t->offset());
3929 if (tinst->isa_aryptr() && t->isa_aryptr()) {
3930 // In the case of a flat inline type array, each field has its
3931 // own slice so we need to keep track of the field being accessed.
3932 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
3933 // Keep array properties (not flat/null-free)
3934 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
3935 if (tinst == nullptr) {
3936 return false; // Skip dead path with inconsistent properties
3937 }
3938 }
3939
3940 // Do NOT remove the next line: ensure a new alias index is allocated
3941 // for the instance type. Note: C++ will not remove it since the call
3942 // has side effect.
3943 int alias_idx = _compile->get_alias_index(tinst);
3944 igvn->set_type(addp, tinst);
3945 // record the allocation in the node map
3946 set_map(addp, get_map(base->_idx));
3947 // Set addp's Base and Address to 'base'.
3948 Node *abase = addp->in(AddPNode::Base);
3949 Node *adr = addp->in(AddPNode::Address);
3950 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3951 adr->in(0)->_idx == (uint)inst_id) {
3952 // Skip AddP cases #3 and #5.
3953 } else {
3954 assert(!abase->is_top(), "sanity"); // AddP case #3
3955 if (abase != base) {
3956 igvn->hash_delete(addp);
3957 addp->set_req(AddPNode::Base, base);
3958 if (abase == adr) {
3959 addp->set_req(AddPNode::Address, base);
3960 } else {
3961 // AddP case #4 (adr is array's element offset AddP node)
3962 #ifdef ASSERT
3963 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3964 assert(adr->is_AddP() && atype != nullptr &&
3965 atype->instance_id() == inst_id, "array's element offset should be processed first");
3966 #endif
3967 }
3968 igvn->hash_insert(addp);
3969 }
3970 }
3971 // Put on IGVN worklist since at least addp's type was changed above.
3972 record_for_optimizer(addp);
3973 return true;
3974 }
3975
3976 //
3977 // Create a new version of orig_phi if necessary. Returns either the newly
3978 // created phi or an existing phi. Sets create_new to indicate whether a new
3979 // phi was created. Cache the last newly created phi in the node map.
3980 //
3981 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3982 Compile *C = _compile;
3983 PhaseGVN* igvn = _igvn;
3984 new_created = false;
3985 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3986 // nothing to do if orig_phi is bottom memory or matches alias_idx
3987 if (phi_alias_idx == alias_idx) {
3988 return orig_phi;
3989 }
3990 // Have we recently created a Phi for this alias index?
3991 PhiNode *result = get_map_phi(orig_phi->_idx);
3992 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3993 return result;
3994 }
3995 // Previous check may fail when the same wide memory Phi was split into Phis
3996 // for different memory slices. Search all Phis for this region.
3997 if (result != nullptr) {
3998 Node* region = orig_phi->in(0);
3999 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
4000 Node* phi = region->fast_out(i);
4001 if (phi->is_Phi() &&
4002 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
4003 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
4004 return phi->as_Phi();
4005 }
4006 }
4007 }
4008 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
4009 if (C->do_escape_analysis() == true && !C->failing()) {
4010 // Retry compilation without escape analysis.
4011 // If this is the first failure, the sentinel string will "stick"
4012 // to the Compile object, and the C2Compiler will see it and retry.
4013 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4014 }
4015 return nullptr;
4016 }
4017 orig_phi_worklist.append_if_missing(orig_phi);
4018 const TypePtr *atype = C->get_adr_type(alias_idx);
4019 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
4020 C->copy_node_notes_to(result, orig_phi);
4021 igvn->set_type(result, result->bottom_type());
4022 record_for_optimizer(result);
4023 set_map(orig_phi, result);
4024 new_created = true;
4025 return result;
4026 }
4027
4028 //
4029 // Return a new version of Memory Phi "orig_phi" with the inputs having the
4030 // specified alias index.
4031 //
4032 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
4033 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
4034 Compile *C = _compile;
4035 PhaseGVN* igvn = _igvn;
4036 bool new_phi_created;
4037 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
4038 if (!new_phi_created) {
4039 return result;
4040 }
4041 GrowableArray<PhiNode *> phi_list;
4042 GrowableArray<uint> cur_input;
4043 PhiNode *phi = orig_phi;
4044 uint idx = 1;
4045 bool finished = false;
4046 while(!finished) {
4047 while (idx < phi->req()) {
4048 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
4049 if (mem != nullptr && mem->is_Phi()) {
4050 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
4051 if (new_phi_created) {
4052 // found an phi for which we created a new split, push current one on worklist and begin
4053 // processing new one
4054 phi_list.push(phi);
4055 cur_input.push(idx);
4056 phi = mem->as_Phi();
4057 result = newphi;
4058 idx = 1;
4059 continue;
4060 } else {
4061 mem = newphi;
4062 }
4063 }
4064 if (C->failing()) {
4065 return nullptr;
4066 }
4067 result->set_req(idx++, mem);
4068 }
4069 #ifdef ASSERT
4070 // verify that the new Phi has an input for each input of the original
4071 assert( phi->req() == result->req(), "must have same number of inputs.");
4072 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4073 #endif
4074 // Check if all new phi's inputs have specified alias index.
4075 // Otherwise use old phi.
4076 for (uint i = 1; i < phi->req(); i++) {
4077 Node* in = result->in(i);
4078 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4079 }
4080 // we have finished processing a Phi, see if there are any more to do
4081 finished = (phi_list.length() == 0 );
4082 if (!finished) {
4083 phi = phi_list.pop();
4084 idx = cur_input.pop();
4085 PhiNode *prev_result = get_map_phi(phi->_idx);
4086 prev_result->set_req(idx++, result);
4087 result = prev_result;
4088 }
4089 }
4090 return result;
4091 }
4092
4093 //
4094 // The next methods are derived from methods in MemNode.
4095 //
4096 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4097 Node *mem = mmem;
4098 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4099 // means an array I have not precisely typed yet. Do not do any
4100 // alias stuff with it any time soon.
4101 if (toop->base() != Type::AnyPtr &&
4102 !(toop->isa_instptr() &&
4103 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4104 toop->offset() == Type::OffsetBot)) {
4105 mem = mmem->memory_at(alias_idx);
4106 // Update input if it is progress over what we have now
4107 }
4108 return mem;
4109 }
4110
4111 //
4112 // Move memory users to their memory slices.
4113 //
4114 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4115 Compile* C = _compile;
4116 PhaseGVN* igvn = _igvn;
4117 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4118 assert(tp != nullptr, "ptr type");
4119 int alias_idx = C->get_alias_index(tp);
4120 int general_idx = C->get_general_index(alias_idx);
4121
4122 // Move users first
4123 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4124 Node* use = n->fast_out(i);
4125 if (use->is_MergeMem()) {
4126 MergeMemNode* mmem = use->as_MergeMem();
4127 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4128 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4129 continue; // Nothing to do
4130 }
4131 // Replace previous general reference to mem node.
4132 uint orig_uniq = C->unique();
4133 Node* m = find_inst_mem(n, general_idx, orig_phis);
4134 assert(orig_uniq == C->unique(), "no new nodes");
4135 mmem->set_memory_at(general_idx, m);
4136 --imax;
4137 --i;
4138 } else if (use->is_MemBar()) {
4139 assert(!use->is_Initialize(), "initializing stores should not be moved");
4140 if (use->req() > MemBarNode::Precedent &&
4141 use->in(MemBarNode::Precedent) == n) {
4142 // Don't move related membars.
4143 record_for_optimizer(use);
4144 continue;
4145 }
4146 tp = use->as_MemBar()->adr_type()->isa_ptr();
4147 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4148 alias_idx == general_idx) {
4149 continue; // Nothing to do
4150 }
4151 // Move to general memory slice.
4152 uint orig_uniq = C->unique();
4153 Node* m = find_inst_mem(n, general_idx, orig_phis);
4154 assert(orig_uniq == C->unique(), "no new nodes");
4155 igvn->hash_delete(use);
4156 imax -= use->replace_edge(n, m, igvn);
4157 igvn->hash_insert(use);
4158 record_for_optimizer(use);
4159 --i;
4160 #ifdef ASSERT
4161 } else if (use->is_Mem()) {
4162 // Memory nodes should have new memory input.
4163 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4164 assert(tp != nullptr, "ptr type");
4165 int idx = C->get_alias_index(tp);
4166 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4167 "Following memory nodes should have new memory input or be on the same memory slice");
4168 } else if (use->is_Phi()) {
4169 // Phi nodes should be split and moved already.
4170 tp = use->as_Phi()->adr_type()->isa_ptr();
4171 assert(tp != nullptr, "ptr type");
4172 int idx = C->get_alias_index(tp);
4173 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4174 } else {
4175 use->dump();
4176 assert(false, "should not be here");
4177 #endif
4178 }
4179 }
4180 }
4181
4182 //
4183 // Search memory chain of "mem" to find a MemNode whose address
4184 // is the specified alias index.
4185 //
4186 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4187 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4188 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4189 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4190 return nullptr;
4191 }
4192 if (orig_mem == nullptr) {
4193 return orig_mem;
4194 }
4195 Compile* C = _compile;
4196 PhaseGVN* igvn = _igvn;
4197 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4198 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4199 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4200 Node *prev = nullptr;
4201 Node *result = orig_mem;
4202 while (prev != result) {
4203 prev = result;
4204 if (result == start_mem) {
4205 break; // hit one of our sentinels
4206 }
4207 if (result->is_Mem()) {
4208 const Type *at = igvn->type(result->in(MemNode::Address));
4209 if (at == Type::TOP) {
4210 break; // Dead
4211 }
4212 assert (at->isa_ptr() != nullptr, "pointer type required.");
4213 int idx = C->get_alias_index(at->is_ptr());
4214 if (idx == alias_idx) {
4215 break; // Found
4216 }
4217 if (!is_instance && (at->isa_oopptr() == nullptr ||
4218 !at->is_oopptr()->is_known_instance())) {
4219 break; // Do not skip store to general memory slice.
4220 }
4221 result = result->in(MemNode::Memory);
4222 }
4223 if (!is_instance) {
4224 continue; // don't search further for non-instance types
4225 }
4226 // skip over a call which does not affect this memory slice
4227 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4228 Node *proj_in = result->in(0);
4229 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4230 break; // hit one of our sentinels
4231 } else if (proj_in->is_Call()) {
4232 // ArrayCopy node processed here as well
4233 CallNode *call = proj_in->as_Call();
4234 if (!call->may_modify(toop, igvn)) {
4235 result = call->in(TypeFunc::Memory);
4236 }
4237 } else if (proj_in->is_Initialize()) {
4238 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4239 // Stop if this is the initialization for the object instance which
4240 // which contains this memory slice, otherwise skip over it.
4241 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4242 result = proj_in->in(TypeFunc::Memory);
4243 }
4244 } else if (proj_in->is_MemBar()) {
4245 // Check if there is an array copy for a clone
4246 // Step over GC barrier when ReduceInitialCardMarks is disabled
4247 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4248 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4249
4250 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4251 // Stop if it is a clone
4252 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4253 if (ac->may_modify(toop, igvn)) {
4254 break;
4255 }
4256 }
4257 result = proj_in->in(TypeFunc::Memory);
4258 }
4259 } else if (result->is_MergeMem()) {
4260 MergeMemNode *mmem = result->as_MergeMem();
4261 result = step_through_mergemem(mmem, alias_idx, toop);
4262 if (result == mmem->base_memory()) {
4263 // Didn't find instance memory, search through general slice recursively.
4264 result = mmem->memory_at(C->get_general_index(alias_idx));
4265 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4266 if (C->failing()) {
4267 return nullptr;
4268 }
4269 mmem->set_memory_at(alias_idx, result);
4270 }
4271 } else if (result->is_Phi() &&
4272 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4273 Node *un = result->as_Phi()->unique_input(igvn);
4274 if (un != nullptr) {
4275 orig_phis.append_if_missing(result->as_Phi());
4276 result = un;
4277 } else {
4278 break;
4279 }
4280 } else if (result->is_ClearArray()) {
4281 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4282 // Can not bypass initialization of the instance
4283 // we are looking for.
4284 break;
4285 }
4286 // Otherwise skip it (the call updated 'result' value).
4287 } else if (result->Opcode() == Op_SCMemProj) {
4288 Node* mem = result->in(0);
4289 Node* adr = nullptr;
4290 if (mem->is_LoadStore()) {
4291 adr = mem->in(MemNode::Address);
4292 } else {
4293 assert(mem->Opcode() == Op_EncodeISOArray ||
4294 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4295 adr = mem->in(3); // Memory edge corresponds to destination array
4296 }
4297 const Type *at = igvn->type(adr);
4298 if (at != Type::TOP) {
4299 assert(at->isa_ptr() != nullptr, "pointer type required.");
4300 int idx = C->get_alias_index(at->is_ptr());
4301 if (idx == alias_idx) {
4302 // Assert in debug mode
4303 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4304 break; // In product mode return SCMemProj node
4305 }
4306 }
4307 result = mem->in(MemNode::Memory);
4308 } else if (result->Opcode() == Op_StrInflatedCopy) {
4309 Node* adr = result->in(3); // Memory edge corresponds to destination array
4310 const Type *at = igvn->type(adr);
4311 if (at != Type::TOP) {
4312 assert(at->isa_ptr() != nullptr, "pointer type required.");
4313 int idx = C->get_alias_index(at->is_ptr());
4314 if (idx == alias_idx) {
4315 // Assert in debug mode
4316 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4317 break; // In product mode return SCMemProj node
4318 }
4319 }
4320 result = result->in(MemNode::Memory);
4321 }
4322 }
4323 if (result->is_Phi()) {
4324 PhiNode *mphi = result->as_Phi();
4325 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4326 const TypePtr *t = mphi->adr_type();
4327 if (!is_instance) {
4328 // Push all non-instance Phis on the orig_phis worklist to update inputs
4329 // during Phase 4 if needed.
4330 orig_phis.append_if_missing(mphi);
4331 } else if (C->get_alias_index(t) != alias_idx) {
4332 // Create a new Phi with the specified alias index type.
4333 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4334 }
4335 }
4336 // the result is either MemNode, PhiNode, InitializeNode.
4337 return result;
4338 }
4339
4340 //
4341 // Convert the types of non-escaped object to instance types where possible,
4342 // propagate the new type information through the graph, and update memory
4343 // edges and MergeMem inputs to reflect the new type.
4344 //
4345 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4346 // The processing is done in 4 phases:
4347 //
4348 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4349 // types for the CheckCastPP for allocations where possible.
4350 // Propagate the new types through users as follows:
4351 // casts and Phi: push users on alloc_worklist
4352 // AddP: cast Base and Address inputs to the instance type
4353 // push any AddP users on alloc_worklist and push any memnode
4354 // users onto memnode_worklist.
4355 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4356 // search the Memory chain for a store with the appropriate type
4357 // address type. If a Phi is found, create a new version with
4358 // the appropriate memory slices from each of the Phi inputs.
4359 // For stores, process the users as follows:
4360 // MemNode: push on memnode_worklist
4361 // MergeMem: push on mergemem_worklist
4362 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4363 // moving the first node encountered of each instance type to the
4364 // the input corresponding to its alias index.
4365 // appropriate memory slice.
4366 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4367 //
4368 // In the following example, the CheckCastPP nodes are the cast of allocation
4369 // results and the allocation of node 29 is non-escaped and eligible to be an
4370 // instance type.
4371 //
4372 // We start with:
4373 //
4374 // 7 Parm #memory
4375 // 10 ConI "12"
4376 // 19 CheckCastPP "Foo"
4377 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4378 // 29 CheckCastPP "Foo"
4379 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4380 //
4381 // 40 StoreP 25 7 20 ... alias_index=4
4382 // 50 StoreP 35 40 30 ... alias_index=4
4383 // 60 StoreP 45 50 20 ... alias_index=4
4384 // 70 LoadP _ 60 30 ... alias_index=4
4385 // 80 Phi 75 50 60 Memory alias_index=4
4386 // 90 LoadP _ 80 30 ... alias_index=4
4387 // 100 LoadP _ 80 20 ... alias_index=4
4388 //
4389 //
4390 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4391 // and creating a new alias index for node 30. This gives:
4392 //
4393 // 7 Parm #memory
4394 // 10 ConI "12"
4395 // 19 CheckCastPP "Foo"
4396 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4397 // 29 CheckCastPP "Foo" iid=24
4398 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4399 //
4400 // 40 StoreP 25 7 20 ... alias_index=4
4401 // 50 StoreP 35 40 30 ... alias_index=6
4402 // 60 StoreP 45 50 20 ... alias_index=4
4403 // 70 LoadP _ 60 30 ... alias_index=6
4404 // 80 Phi 75 50 60 Memory alias_index=4
4405 // 90 LoadP _ 80 30 ... alias_index=6
4406 // 100 LoadP _ 80 20 ... alias_index=4
4407 //
4408 // In phase 2, new memory inputs are computed for the loads and stores,
4409 // And a new version of the phi is created. In phase 4, the inputs to
4410 // node 80 are updated and then the memory nodes are updated with the
4411 // values computed in phase 2. This results in:
4412 //
4413 // 7 Parm #memory
4414 // 10 ConI "12"
4415 // 19 CheckCastPP "Foo"
4416 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4417 // 29 CheckCastPP "Foo" iid=24
4418 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4419 //
4420 // 40 StoreP 25 7 20 ... alias_index=4
4421 // 50 StoreP 35 7 30 ... alias_index=6
4422 // 60 StoreP 45 40 20 ... alias_index=4
4423 // 70 LoadP _ 50 30 ... alias_index=6
4424 // 80 Phi 75 40 60 Memory alias_index=4
4425 // 120 Phi 75 50 50 Memory alias_index=6
4426 // 90 LoadP _ 120 30 ... alias_index=6
4427 // 100 LoadP _ 80 20 ... alias_index=4
4428 //
4429 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4430 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4431 GrowableArray<MergeMemNode*> &mergemem_worklist,
4432 Unique_Node_List &reducible_merges) {
4433 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4434 GrowableArray<Node *> memnode_worklist;
4435 GrowableArray<PhiNode *> orig_phis;
4436 PhaseIterGVN *igvn = _igvn;
4437 uint new_index_start = (uint) _compile->num_alias_types();
4438 VectorSet visited;
4439 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4440 uint unique_old = _compile->unique();
4441
4442 // Phase 1: Process possible allocations from alloc_worklist.
4443 // Create instance types for the CheckCastPP for allocations where possible.
4444 //
4445 // (Note: don't forget to change the order of the second AddP node on
4446 // the alloc_worklist if the order of the worklist processing is changed,
4447 // see the comment in find_second_addp().)
4448 //
4449 while (alloc_worklist.length() != 0) {
4450 Node *n = alloc_worklist.pop();
4451 uint ni = n->_idx;
4452 if (n->is_Call()) {
4453 CallNode *alloc = n->as_Call();
4454 // copy escape information to call node
4455 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4456 PointsToNode::EscapeState es = ptn->escape_state();
4457 // We have an allocation or call which returns a Java object,
4458 // see if it is non-escaped.
4459 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4460 continue;
4461 }
4462 // Find CheckCastPP for the allocate or for the return value of a call
4463 n = alloc->result_cast();
4464 if (n == nullptr) { // No uses except Initialize node
4465 if (alloc->is_Allocate()) {
4466 // Set the scalar_replaceable flag for allocation
4467 // so it could be eliminated if it has no uses.
4468 alloc->as_Allocate()->_is_scalar_replaceable = true;
4469 }
4470 continue;
4471 }
4472 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4473 // we could reach here for allocate case if one init is associated with many allocs.
4474 if (alloc->is_Allocate()) {
4475 alloc->as_Allocate()->_is_scalar_replaceable = false;
4476 }
4477 continue;
4478 }
4479
4480 // The inline code for Object.clone() casts the allocation result to
4481 // java.lang.Object and then to the actual type of the allocated
4482 // object. Detect this case and use the second cast.
4483 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4484 // the allocation result is cast to java.lang.Object and then
4485 // to the actual Array type.
4486 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4487 && (alloc->is_AllocateArray() ||
4488 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4489 Node *cast2 = nullptr;
4490 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4491 Node *use = n->fast_out(i);
4492 if (use->is_CheckCastPP()) {
4493 cast2 = use;
4494 break;
4495 }
4496 }
4497 if (cast2 != nullptr) {
4498 n = cast2;
4499 } else {
4500 // Non-scalar replaceable if the allocation type is unknown statically
4501 // (reflection allocation), the object can't be restored during
4502 // deoptimization without precise type.
4503 continue;
4504 }
4505 }
4506
4507 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4508 if (t == nullptr) {
4509 continue; // not a TypeOopPtr
4510 }
4511 if (!t->klass_is_exact()) {
4512 continue; // not an unique type
4513 }
4514 if (alloc->is_Allocate()) {
4515 // Set the scalar_replaceable flag for allocation
4516 // so it could be eliminated.
4517 alloc->as_Allocate()->_is_scalar_replaceable = true;
4518 }
4519 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4520 // in order for an object to be scalar-replaceable, it must be:
4521 // - a direct allocation (not a call returning an object)
4522 // - non-escaping
4523 // - eligible to be a unique type
4524 // - not determined to be ineligible by escape analysis
4525 set_map(alloc, n);
4526 set_map(n, alloc);
4527 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4528 igvn->hash_delete(n);
4529 igvn->set_type(n, tinst);
4530 n->raise_bottom_type(tinst);
4531 igvn->hash_insert(n);
4532 record_for_optimizer(n);
4533 // Allocate an alias index for the header fields. Accesses to
4534 // the header emitted during macro expansion wouldn't have
4535 // correct memory state otherwise.
4536 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4537 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4538 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4539
4540 // First, put on the worklist all Field edges from Connection Graph
4541 // which is more accurate than putting immediate users from Ideal Graph.
4542 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4543 PointsToNode* tgt = e.get();
4544 if (tgt->is_Arraycopy()) {
4545 continue;
4546 }
4547 Node* use = tgt->ideal_node();
4548 assert(tgt->is_Field() && use->is_AddP(),
4549 "only AddP nodes are Field edges in CG");
4550 if (use->outcnt() > 0) { // Don't process dead nodes
4551 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4552 if (addp2 != nullptr) {
4553 assert(alloc->is_AllocateArray(),"array allocation was expected");
4554 alloc_worklist.append_if_missing(addp2);
4555 }
4556 alloc_worklist.append_if_missing(use);
4557 }
4558 }
4559
4560 // An allocation may have an Initialize which has raw stores. Scan
4561 // the users of the raw allocation result and push AddP users
4562 // on alloc_worklist.
4563 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4564 assert (raw_result != nullptr, "must have an allocation result");
4565 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4566 Node *use = raw_result->fast_out(i);
4567 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4568 Node* addp2 = find_second_addp(use, raw_result);
4569 if (addp2 != nullptr) {
4570 assert(alloc->is_AllocateArray(),"array allocation was expected");
4571 alloc_worklist.append_if_missing(addp2);
4572 }
4573 alloc_worklist.append_if_missing(use);
4574 } else if (use->is_MemBar()) {
4575 memnode_worklist.append_if_missing(use);
4576 }
4577 }
4578 }
4579 } else if (n->is_AddP()) {
4580 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4581 // This AddP will go away when we reduce the Phi
4582 continue;
4583 }
4584 Node* addp_base = get_addp_base(n);
4585 JavaObjectNode* jobj = unique_java_object(addp_base);
4586 if (jobj == nullptr || jobj == phantom_obj) {
4587 #ifdef ASSERT
4588 ptnode_adr(get_addp_base(n)->_idx)->dump();
4589 ptnode_adr(n->_idx)->dump();
4590 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4591 #endif
4592 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4593 return;
4594 }
4595 Node *base = get_map(jobj->idx()); // CheckCastPP node
4596 if (!split_AddP(n, base)) continue; // wrong type from dead path
4597 } else if (n->is_Phi() ||
4598 n->is_CheckCastPP() ||
4599 n->is_EncodeP() ||
4600 n->is_DecodeN() ||
4601 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4602 if (visited.test_set(n->_idx)) {
4603 assert(n->is_Phi(), "loops only through Phi's");
4604 continue; // already processed
4605 }
4606 // Reducible Phi's will be removed from the graph after split_unique_types
4607 // finishes. For now we just try to split out the SR inputs of the merge.
4608 Node* parent = n->in(1);
4609 if (reducible_merges.member(n)) {
4610 reduce_phi(n->as_Phi(), alloc_worklist, memnode_worklist);
4611 #ifdef ASSERT
4612 if (VerifyReduceAllocationMerges) {
4613 reduced_merges.push(n);
4614 }
4615 #endif
4616 continue;
4617 } else if (reducible_merges.member(parent)) {
4618 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4619 // part of reduce_merge.
4620 continue;
4621 }
4622 JavaObjectNode* jobj = unique_java_object(n);
4623 if (jobj == nullptr || jobj == phantom_obj) {
4624 #ifdef ASSERT
4625 ptnode_adr(n->_idx)->dump();
4626 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4627 #endif
4628 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4629 return;
4630 } else {
4631 Node *val = get_map(jobj->idx()); // CheckCastPP node
4632 TypeNode *tn = n->as_Type();
4633 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4634 assert(tinst != nullptr && tinst->is_known_instance() &&
4635 tinst->instance_id() == jobj->idx() , "instance type expected.");
4636
4637 const Type *tn_type = igvn->type(tn);
4638 const TypeOopPtr *tn_t;
4639 if (tn_type->isa_narrowoop()) {
4640 tn_t = tn_type->make_ptr()->isa_oopptr();
4641 } else {
4642 tn_t = tn_type->isa_oopptr();
4643 }
4644 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4645 if (tn_t->isa_aryptr()) {
4646 // Keep array properties (not flat/null-free)
4647 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
4648 if (tinst == nullptr) {
4649 continue; // Skip dead path with inconsistent properties
4650 }
4651 }
4652 if (tn_type->isa_narrowoop()) {
4653 tn_type = tinst->make_narrowoop();
4654 } else {
4655 tn_type = tinst;
4656 }
4657 igvn->hash_delete(tn);
4658 igvn->set_type(tn, tn_type);
4659 tn->set_type(tn_type);
4660 igvn->hash_insert(tn);
4661 record_for_optimizer(n);
4662 } else {
4663 assert(tn_type == TypePtr::NULL_PTR ||
4664 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4665 "unexpected type");
4666 continue; // Skip dead path with different type
4667 }
4668 }
4669 } else {
4670 DEBUG_ONLY(n->dump();)
4671 assert(false, "EA: unexpected node");
4672 continue;
4673 }
4674 // push allocation's users on appropriate worklist
4675 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4676 Node *use = n->fast_out(i);
4677 if (use->is_Mem() && use->in(MemNode::Address) == n) {
4678 // Load/store to instance's field
4679 memnode_worklist.append_if_missing(use);
4680 } else if (use->is_MemBar()) {
4681 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4682 memnode_worklist.append_if_missing(use);
4683 }
4684 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4685 Node* addp2 = find_second_addp(use, n);
4686 if (addp2 != nullptr) {
4687 alloc_worklist.append_if_missing(addp2);
4688 }
4689 alloc_worklist.append_if_missing(use);
4690 } else if (use->is_Phi() ||
4691 use->is_CheckCastPP() ||
4692 use->is_EncodeNarrowPtr() ||
4693 use->is_DecodeNarrowPtr() ||
4694 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4695 alloc_worklist.append_if_missing(use);
4696 #ifdef ASSERT
4697 } else if (use->is_Mem()) {
4698 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4699 } else if (use->is_MergeMem()) {
4700 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4701 } else if (use->is_SafePoint()) {
4702 // Look for MergeMem nodes for calls which reference unique allocation
4703 // (through CheckCastPP nodes) even for debug info.
4704 Node* m = use->in(TypeFunc::Memory);
4705 if (m->is_MergeMem()) {
4706 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4707 }
4708 } else if (use->Opcode() == Op_EncodeISOArray) {
4709 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4710 // EncodeISOArray overwrites destination array
4711 memnode_worklist.append_if_missing(use);
4712 }
4713 } else if (use->Opcode() == Op_Return) {
4714 // Allocation is referenced by field of returned inline type
4715 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
4716 } else {
4717 uint op = use->Opcode();
4718 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4719 (use->in(MemNode::Memory) == n)) {
4720 // They overwrite memory edge corresponding to destination array,
4721 memnode_worklist.append_if_missing(use);
4722 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4723 op == Op_CastP2X ||
4724 op == Op_FastLock || op == Op_AryEq ||
4725 op == Op_StrComp || op == Op_CountPositives ||
4726 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4727 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4728 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4729 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
4730 op == Op_ReinterpretS2HF ||
4731 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4732 n->dump();
4733 use->dump();
4734 assert(false, "EA: missing allocation reference path");
4735 }
4736 #endif
4737 }
4738 }
4739
4740 }
4741
4742 #ifdef ASSERT
4743 if (VerifyReduceAllocationMerges) {
4744 for (uint i = 0; i < reducible_merges.size(); i++) {
4745 Node* phi = reducible_merges.at(i);
4746
4747 if (!reduced_merges.member(phi)) {
4748 phi->dump(2);
4749 phi->dump(-2);
4750 assert(false, "This reducible merge wasn't reduced.");
4751 }
4752
4753 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4754 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4755 Node* use = phi->fast_out(j);
4756 if (!use->is_SafePoint() && !use->is_CastPP()) {
4757 phi->dump(2);
4758 phi->dump(-2);
4759 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4760 }
4761 }
4762 }
4763 }
4764 #endif
4765
4766 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4767 // type, record it in the ArrayCopy node so we know what memory this
4768 // node uses/modified.
4769 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4770 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4771 Node* dest = ac->in(ArrayCopyNode::Dest);
4772 if (dest->is_AddP()) {
4773 dest = get_addp_base(dest);
4774 }
4775 JavaObjectNode* jobj = unique_java_object(dest);
4776 if (jobj != nullptr) {
4777 Node *base = get_map(jobj->idx());
4778 if (base != nullptr) {
4779 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4780 ac->_dest_type = base_t;
4781 }
4782 }
4783 Node* src = ac->in(ArrayCopyNode::Src);
4784 if (src->is_AddP()) {
4785 src = get_addp_base(src);
4786 }
4787 jobj = unique_java_object(src);
4788 if (jobj != nullptr) {
4789 Node* base = get_map(jobj->idx());
4790 if (base != nullptr) {
4791 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4792 ac->_src_type = base_t;
4793 }
4794 }
4795 }
4796
4797 // New alias types were created in split_AddP().
4798 uint new_index_end = (uint) _compile->num_alias_types();
4799
4800 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4801 // compute new values for Memory inputs (the Memory inputs are not
4802 // actually updated until phase 4.)
4803 if (memnode_worklist.length() == 0)
4804 return; // nothing to do
4805 while (memnode_worklist.length() != 0) {
4806 Node *n = memnode_worklist.pop();
4807 if (visited.test_set(n->_idx)) {
4808 continue;
4809 }
4810 if (n->is_Phi() || n->is_ClearArray()) {
4811 // we don't need to do anything, but the users must be pushed
4812 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
4813 // we don't need to do anything, but the users must be pushed
4814 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4815 if (n == nullptr) {
4816 continue;
4817 }
4818 } else if (n->is_CallLeaf()) {
4819 // Runtime calls with narrow memory input (no MergeMem node)
4820 // get the memory projection
4821 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4822 if (n == nullptr) {
4823 continue;
4824 }
4825 } else if (n->Opcode() == Op_StrInflatedCopy) {
4826 // Check direct uses of StrInflatedCopy.
4827 // It is memory type Node - no special SCMemProj node.
4828 } else if (n->Opcode() == Op_StrCompressedCopy ||
4829 n->Opcode() == Op_EncodeISOArray) {
4830 // get the memory projection
4831 n = n->find_out_with(Op_SCMemProj);
4832 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4833 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
4834 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
4835 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
4836 } else {
4837 #ifdef ASSERT
4838 if (!n->is_Mem()) {
4839 n->dump();
4840 }
4841 assert(n->is_Mem(), "memory node required.");
4842 #endif
4843 Node *addr = n->in(MemNode::Address);
4844 const Type *addr_t = igvn->type(addr);
4845 if (addr_t == Type::TOP) {
4846 continue;
4847 }
4848 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4849 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4850 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4851 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4852 if (_compile->failing()) {
4853 return;
4854 }
4855 if (mem != n->in(MemNode::Memory)) {
4856 // We delay the memory edge update since we need old one in
4857 // MergeMem code below when instances memory slices are separated.
4858 set_map(n, mem);
4859 }
4860 if (n->is_Load()) {
4861 continue; // don't push users
4862 } else if (n->is_LoadStore()) {
4863 // get the memory projection
4864 n = n->find_out_with(Op_SCMemProj);
4865 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4866 }
4867 }
4868 // push user on appropriate worklist
4869 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4870 Node *use = n->fast_out(i);
4871 if (use->is_Phi() || use->is_ClearArray()) {
4872 memnode_worklist.append_if_missing(use);
4873 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4874 memnode_worklist.append_if_missing(use);
4875 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4876 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4877 memnode_worklist.append_if_missing(use);
4878 }
4879 #ifdef ASSERT
4880 } else if (use->is_Mem()) {
4881 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4882 } else if (use->is_MergeMem()) {
4883 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4884 } else if (use->Opcode() == Op_EncodeISOArray) {
4885 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4886 // EncodeISOArray overwrites destination array
4887 memnode_worklist.append_if_missing(use);
4888 }
4889 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
4890 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
4891 // store_unknown_inline overwrites destination array
4892 memnode_worklist.append_if_missing(use);
4893 } else {
4894 uint op = use->Opcode();
4895 if ((use->in(MemNode::Memory) == n) &&
4896 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4897 // They overwrite memory edge corresponding to destination array,
4898 memnode_worklist.append_if_missing(use);
4899 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4900 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4901 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4902 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
4903 n->dump();
4904 use->dump();
4905 assert(false, "EA: missing memory path");
4906 }
4907 #endif
4908 }
4909 }
4910 }
4911
4912 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4913 // Walk each memory slice moving the first node encountered of each
4914 // instance type to the input corresponding to its alias index.
4915 uint length = mergemem_worklist.length();
4916 for( uint next = 0; next < length; ++next ) {
4917 MergeMemNode* nmm = mergemem_worklist.at(next);
4918 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4919 // Note: we don't want to use MergeMemStream here because we only want to
4920 // scan inputs which exist at the start, not ones we add during processing.
4921 // Note 2: MergeMem may already contains instance memory slices added
4922 // during find_inst_mem() call when memory nodes were processed above.
4923 igvn->hash_delete(nmm);
4924 uint nslices = MIN2(nmm->req(), new_index_start);
4925 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4926 Node* mem = nmm->in(i);
4927 Node* cur = nullptr;
4928 if (mem == nullptr || mem->is_top()) {
4929 continue;
4930 }
4931 // First, update mergemem by moving memory nodes to corresponding slices
4932 // if their type became more precise since this mergemem was created.
4933 while (mem->is_Mem()) {
4934 const Type *at = igvn->type(mem->in(MemNode::Address));
4935 if (at != Type::TOP) {
4936 assert (at->isa_ptr() != nullptr, "pointer type required.");
4937 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4938 if (idx == i) {
4939 if (cur == nullptr) {
4940 cur = mem;
4941 }
4942 } else {
4943 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4944 nmm->set_memory_at(idx, mem);
4945 }
4946 }
4947 }
4948 mem = mem->in(MemNode::Memory);
4949 }
4950 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4951 // Find any instance of the current type if we haven't encountered
4952 // already a memory slice of the instance along the memory chain.
4953 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4954 if((uint)_compile->get_general_index(ni) == i) {
4955 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4956 if (nmm->is_empty_memory(m)) {
4957 Node* result = find_inst_mem(mem, ni, orig_phis);
4958 if (_compile->failing()) {
4959 return;
4960 }
4961 nmm->set_memory_at(ni, result);
4962 }
4963 }
4964 }
4965 }
4966 // Find the rest of instances values
4967 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4968 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4969 Node* result = step_through_mergemem(nmm, ni, tinst);
4970 if (result == nmm->base_memory()) {
4971 // Didn't find instance memory, search through general slice recursively.
4972 result = nmm->memory_at(_compile->get_general_index(ni));
4973 result = find_inst_mem(result, ni, orig_phis);
4974 if (_compile->failing()) {
4975 return;
4976 }
4977 nmm->set_memory_at(ni, result);
4978 }
4979 }
4980
4981 // If we have crossed the 3/4 point of max node limit it's too risky
4982 // to continue with EA/SR because we might hit the max node limit.
4983 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4984 if (_compile->do_reduce_allocation_merges()) {
4985 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4986 } else if (_invocation > 0) {
4987 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4988 } else {
4989 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4990 }
4991 return;
4992 }
4993
4994 igvn->hash_insert(nmm);
4995 record_for_optimizer(nmm);
4996 }
4997
4998 // Phase 4: Update the inputs of non-instance memory Phis and
4999 // the Memory input of memnodes
5000 // First update the inputs of any non-instance Phi's from
5001 // which we split out an instance Phi. Note we don't have
5002 // to recursively process Phi's encountered on the input memory
5003 // chains as is done in split_memory_phi() since they will
5004 // also be processed here.
5005 for (int j = 0; j < orig_phis.length(); j++) {
5006 PhiNode *phi = orig_phis.at(j);
5007 int alias_idx = _compile->get_alias_index(phi->adr_type());
5008 igvn->hash_delete(phi);
5009 for (uint i = 1; i < phi->req(); i++) {
5010 Node *mem = phi->in(i);
5011 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
5012 if (_compile->failing()) {
5013 return;
5014 }
5015 if (mem != new_mem) {
5016 phi->set_req(i, new_mem);
5017 }
5018 }
5019 igvn->hash_insert(phi);
5020 record_for_optimizer(phi);
5021 }
5022
5023 // Update the memory inputs of MemNodes with the value we computed
5024 // in Phase 2 and move stores memory users to corresponding memory slices.
5025 // Disable memory split verification code until the fix for 6984348.
5026 // Currently it produces false negative results since it does not cover all cases.
5027 #if 0 // ifdef ASSERT
5028 visited.Reset();
5029 Node_Stack old_mems(arena, _compile->unique() >> 2);
5030 #endif
5031 for (uint i = 0; i < ideal_nodes.size(); i++) {
5032 Node* n = ideal_nodes.at(i);
5033 Node* nmem = get_map(n->_idx);
5034 assert(nmem != nullptr, "sanity");
5035 if (n->is_Mem()) {
5036 #if 0 // ifdef ASSERT
5037 Node* old_mem = n->in(MemNode::Memory);
5038 if (!visited.test_set(old_mem->_idx)) {
5039 old_mems.push(old_mem, old_mem->outcnt());
5040 }
5041 #endif
5042 assert(n->in(MemNode::Memory) != nmem, "sanity");
5043 if (!n->is_Load()) {
5044 // Move memory users of a store first.
5045 move_inst_mem(n, orig_phis);
5046 }
5047 // Now update memory input
5048 igvn->hash_delete(n);
5049 n->set_req(MemNode::Memory, nmem);
5050 igvn->hash_insert(n);
5051 record_for_optimizer(n);
5052 } else {
5053 assert(n->is_Allocate() || n->is_CheckCastPP() ||
5054 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
5055 }
5056 }
5057 #if 0 // ifdef ASSERT
5058 // Verify that memory was split correctly
5059 while (old_mems.is_nonempty()) {
5060 Node* old_mem = old_mems.node();
5061 uint old_cnt = old_mems.index();
5062 old_mems.pop();
5063 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5064 }
5065 #endif
5066 }
5067
5068 #ifndef PRODUCT
5069 int ConnectionGraph::_no_escape_counter = 0;
5070 int ConnectionGraph::_arg_escape_counter = 0;
5071 int ConnectionGraph::_global_escape_counter = 0;
5072
5073 static const char *node_type_names[] = {
5074 "UnknownType",
5075 "JavaObject",
5076 "LocalVar",
5077 "Field",
5078 "Arraycopy"
5079 };
5080
5081 static const char *esc_names[] = {
5082 "UnknownEscape",
5083 "NoEscape",
5084 "ArgEscape",
5085 "GlobalEscape"
5086 };
5087
5088 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5089 NodeType nt = node_type();
5090 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5091 if (print_state) {
5092 EscapeState es = escape_state();
5093 EscapeState fields_es = fields_escape_state();
5094 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5095 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5096 out->print("NSR ");
5097 }
5098 }
5099 }
5100
5101 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5102 dump_header(print_state, out);
5103 if (is_Field()) {
5104 FieldNode* f = (FieldNode*)this;
5105 if (f->is_oop()) {
5106 out->print("oop ");
5107 }
5108 if (f->offset() > 0) {
5109 out->print("+%d ", f->offset());
5110 }
5111 out->print("(");
5112 for (BaseIterator i(f); i.has_next(); i.next()) {
5113 PointsToNode* b = i.get();
5114 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5115 }
5116 out->print(" )");
5117 }
5118 out->print("[");
5119 for (EdgeIterator i(this); i.has_next(); i.next()) {
5120 PointsToNode* e = i.get();
5121 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5122 }
5123 out->print(" [");
5124 for (UseIterator i(this); i.has_next(); i.next()) {
5125 PointsToNode* u = i.get();
5126 bool is_base = false;
5127 if (PointsToNode::is_base_use(u)) {
5128 is_base = true;
5129 u = PointsToNode::get_use_node(u)->as_Field();
5130 }
5131 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5132 }
5133 out->print(" ]] ");
5134 if (_node == nullptr) {
5135 out->print("<null>%s", newline ? "\n" : "");
5136 } else {
5137 _node->dump(newline ? "\n" : "", false, out);
5138 }
5139 }
5140
5141 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5142 bool first = true;
5143 int ptnodes_length = ptnodes_worklist.length();
5144 for (int i = 0; i < ptnodes_length; i++) {
5145 PointsToNode *ptn = ptnodes_worklist.at(i);
5146 if (ptn == nullptr || !ptn->is_JavaObject()) {
5147 continue;
5148 }
5149 PointsToNode::EscapeState es = ptn->escape_state();
5150 if ((es != PointsToNode::NoEscape) && !Verbose) {
5151 continue;
5152 }
5153 Node* n = ptn->ideal_node();
5154 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5155 n->as_CallStaticJava()->is_boxing_method())) {
5156 if (first) {
5157 tty->cr();
5158 tty->print("======== Connection graph for ");
5159 _compile->method()->print_short_name();
5160 tty->cr();
5161 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5162 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5163 tty->cr();
5164 first = false;
5165 }
5166 ptn->dump();
5167 // Print all locals and fields which reference this allocation
5168 for (UseIterator j(ptn); j.has_next(); j.next()) {
5169 PointsToNode* use = j.get();
5170 if (use->is_LocalVar()) {
5171 use->dump(Verbose);
5172 } else if (Verbose) {
5173 use->dump();
5174 }
5175 }
5176 tty->cr();
5177 }
5178 }
5179 }
5180
5181 void ConnectionGraph::print_statistics() {
5182 tty->print_cr("No escape = %d, Arg escape = %d, Global escape = %d", Atomic::load(&_no_escape_counter), Atomic::load(&_arg_escape_counter), Atomic::load(&_global_escape_counter));
5183 }
5184
5185 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5186 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5187 return;
5188 }
5189 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5190 JavaObjectNode* ptn = java_objects_worklist.at(next);
5191 if (ptn->ideal_node()->is_Allocate()) {
5192 if (ptn->escape_state() == PointsToNode::NoEscape) {
5193 Atomic::inc(&ConnectionGraph::_no_escape_counter);
5194 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5195 Atomic::inc(&ConnectionGraph::_arg_escape_counter);
5196 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5197 Atomic::inc(&ConnectionGraph::_global_escape_counter);
5198 } else {
5199 assert(false, "Unexpected Escape State");
5200 }
5201 }
5202 }
5203 }
5204
5205 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5206 if (_compile->directive()->TraceEscapeAnalysisOption) {
5207 assert(ptn != nullptr, "should not be null");
5208 assert(reason != nullptr, "should not be null");
5209 ptn->dump_header(true);
5210 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5211 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5212 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5213 }
5214 }
5215
5216 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5217 if (_compile->directive()->TraceEscapeAnalysisOption) {
5218 stringStream ss;
5219 ss.print("propagated from: ");
5220 from->dump(true, &ss, false);
5221 return ss.as_string();
5222 } else {
5223 return nullptr;
5224 }
5225 }
5226
5227 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5228 if (_compile->directive()->TraceEscapeAnalysisOption) {
5229 stringStream ss;
5230 ss.print("escapes as arg to:");
5231 call->dump("", false, &ss);
5232 return ss.as_string();
5233 } else {
5234 return nullptr;
5235 }
5236 }
5237
5238 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5239 if (_compile->directive()->TraceEscapeAnalysisOption) {
5240 stringStream ss;
5241 ss.print("is merged with other object: ");
5242 other->dump_header(true, &ss);
5243 return ss.as_string();
5244 } else {
5245 return nullptr;
5246 }
5247 }
5248
5249 #endif
5250
5251 void ConnectionGraph::record_for_optimizer(Node *n) {
5252 _igvn->_worklist.push(n);
5253 _igvn->add_users_to_worklist(n);
5254 }