1 /*
2 * Copyright (c) 2005, 2024, 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 "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/allocation.hpp"
32 #include "memory/metaspace.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/callnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/escape.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/macro.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/narrowptrnode.hpp"
46 #include "opto/castnode.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/macros.hpp"
49
50 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn, int invocation) :
51 // If ReduceAllocationMerges is enabled we might call split_through_phi during
52 // split_unique_types and that will create additional nodes that need to be
53 // pushed to the ConnectionGraph. The code below bumps the initial capacity of
54 // _nodes by 10% to account for these additional nodes. If capacity is exceeded
55 // the array will be reallocated.
56 _nodes(C->comp_arena(), C->do_reduce_allocation_merges() ? C->unique()*1.10 : C->unique(), C->unique(), nullptr),
57 _in_worklist(C->comp_arena()),
58 _next_pidx(0),
59 _collecting(true),
60 _verify(false),
61 _compile(C),
62 _igvn(igvn),
63 _invocation(invocation),
64 _build_iterations(0),
65 _build_time(0.),
66 _node_map(C->comp_arena()) {
67 // Add unknown java object.
68 add_java_object(C->top(), PointsToNode::GlobalEscape);
69 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
70 set_not_scalar_replaceable(phantom_obj NOT_PRODUCT(COMMA "Phantom object"));
71 // Add ConP and ConN null oop nodes
72 Node* oop_null = igvn->zerocon(T_OBJECT);
73 assert(oop_null->_idx < nodes_size(), "should be created already");
74 add_java_object(oop_null, PointsToNode::NoEscape);
75 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
76 set_not_scalar_replaceable(null_obj NOT_PRODUCT(COMMA "Null object"));
77 if (UseCompressedOops) {
78 Node* noop_null = igvn->zerocon(T_NARROWOOP);
79 assert(noop_null->_idx < nodes_size(), "should be created already");
80 map_ideal_node(noop_null, null_obj);
81 }
82 }
83
84 bool ConnectionGraph::has_candidates(Compile *C) {
85 // EA brings benefits only when the code has allocations and/or locks which
86 // are represented by ideal Macro nodes.
87 int cnt = C->macro_count();
88 for (int i = 0; i < cnt; i++) {
89 Node *n = C->macro_node(i);
90 if (n->is_Allocate()) {
91 return true;
92 }
93 if (n->is_Lock()) {
94 Node* obj = n->as_Lock()->obj_node()->uncast();
95 if (!(obj->is_Parm() || obj->is_Con())) {
96 return true;
97 }
98 }
99 if (n->is_CallStaticJava() &&
100 n->as_CallStaticJava()->is_boxing_method()) {
101 return true;
102 }
103 }
104 return false;
105 }
106
107 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
108 Compile::TracePhase tp("escapeAnalysis", &Phase::timers[Phase::_t_escapeAnalysis]);
109 ResourceMark rm;
110
111 // Add ConP and ConN null oop nodes before ConnectionGraph construction
112 // to create space for them in ConnectionGraph::_nodes[].
113 Node* oop_null = igvn->zerocon(T_OBJECT);
114 Node* noop_null = igvn->zerocon(T_NARROWOOP);
115 int invocation = 0;
116 if (C->congraph() != nullptr) {
117 invocation = C->congraph()->_invocation + 1;
118 }
119 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn, invocation);
120 // Perform escape analysis
121 if (congraph->compute_escape()) {
122 // There are non escaping objects.
123 C->set_congraph(congraph);
124 }
125 // Cleanup.
126 if (oop_null->outcnt() == 0) {
127 igvn->hash_delete(oop_null);
128 }
129 if (noop_null->outcnt() == 0) {
130 igvn->hash_delete(noop_null);
131 }
132 }
133
134 bool ConnectionGraph::compute_escape() {
135 Compile* C = _compile;
136 PhaseGVN* igvn = _igvn;
137
138 // Worklists used by EA.
139 Unique_Node_List delayed_worklist;
140 Unique_Node_List reducible_merges;
141 GrowableArray<Node*> alloc_worklist;
142 GrowableArray<Node*> ptr_cmp_worklist;
143 GrowableArray<MemBarStoreStoreNode*> storestore_worklist;
144 GrowableArray<ArrayCopyNode*> arraycopy_worklist;
145 GrowableArray<PointsToNode*> ptnodes_worklist;
146 GrowableArray<JavaObjectNode*> java_objects_worklist;
147 GrowableArray<JavaObjectNode*> non_escaped_allocs_worklist;
148 GrowableArray<FieldNode*> oop_fields_worklist;
149 GrowableArray<SafePointNode*> sfn_worklist;
150 GrowableArray<MergeMemNode*> mergemem_worklist;
151 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
152
153 { Compile::TracePhase tp("connectionGraph", &Phase::timers[Phase::_t_connectionGraph]);
154
155 // 1. Populate Connection Graph (CG) with PointsTo nodes.
156 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
157 // Initialize worklist
158 if (C->root() != nullptr) {
159 ideal_nodes.push(C->root());
160 }
161 // Processed ideal nodes are unique on ideal_nodes list
162 // but several ideal nodes are mapped to the phantom_obj.
163 // To avoid duplicated entries on the following worklists
164 // add the phantom_obj only once to them.
165 ptnodes_worklist.append(phantom_obj);
166 java_objects_worklist.append(phantom_obj);
167 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
168 Node* n = ideal_nodes.at(next);
169 if ((n->Opcode() == Op_LoadX || n->Opcode() == Op_StoreX) &&
170 !n->in(MemNode::Address)->is_AddP() &&
171 _igvn->type(n->in(MemNode::Address))->isa_oopptr()) {
172 // Load/Store at mark work address is at offset 0 so has no AddP which confuses EA
173 Node* addp = new AddPNode(n->in(MemNode::Address), n->in(MemNode::Address), _igvn->MakeConX(0));
174 _igvn->register_new_node_with_optimizer(addp);
175 _igvn->replace_input_of(n, MemNode::Address, addp);
176 ideal_nodes.push(addp);
177 _nodes.at_put_grow(addp->_idx, nullptr, nullptr);
178 }
179 // Create PointsTo nodes and add them to Connection Graph. Called
180 // only once per ideal node since ideal_nodes is Unique_Node list.
181 add_node_to_connection_graph(n, &delayed_worklist);
182 PointsToNode* ptn = ptnode_adr(n->_idx);
183 if (ptn != nullptr && ptn != phantom_obj) {
184 ptnodes_worklist.append(ptn);
185 if (ptn->is_JavaObject()) {
186 java_objects_worklist.append(ptn->as_JavaObject());
187 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
188 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
189 // Only allocations and java static calls results are interesting.
190 non_escaped_allocs_worklist.append(ptn->as_JavaObject());
191 }
192 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
193 oop_fields_worklist.append(ptn->as_Field());
194 }
195 }
196 // Collect some interesting nodes for further use.
197 switch (n->Opcode()) {
198 case Op_MergeMem:
199 // Collect all MergeMem nodes to add memory slices for
200 // scalar replaceable objects in split_unique_types().
201 mergemem_worklist.append(n->as_MergeMem());
202 break;
203 case Op_CmpP:
204 case Op_CmpN:
205 // Collect compare pointers nodes.
206 if (OptimizePtrCompare) {
207 ptr_cmp_worklist.append(n);
208 }
209 break;
210 case Op_MemBarStoreStore:
211 // Collect all MemBarStoreStore nodes so that depending on the
212 // escape status of the associated Allocate node some of them
213 // may be eliminated.
214 if (!UseStoreStoreForCtor || n->req() > MemBarNode::Precedent) {
215 storestore_worklist.append(n->as_MemBarStoreStore());
216 }
217 break;
218 case Op_MemBarRelease:
219 if (n->req() > MemBarNode::Precedent) {
220 record_for_optimizer(n);
221 }
222 break;
223 #ifdef ASSERT
224 case Op_AddP:
225 // Collect address nodes for graph verification.
226 addp_worklist.append(n);
227 break;
228 #endif
229 case Op_ArrayCopy:
230 // Keep a list of ArrayCopy nodes so if one of its input is non
231 // escaping, we can record a unique type
232 arraycopy_worklist.append(n->as_ArrayCopy());
233 break;
234 default:
235 // not interested now, ignore...
236 break;
237 }
238 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
239 Node* m = n->fast_out(i); // Get user
240 ideal_nodes.push(m);
241 }
242 if (n->is_SafePoint()) {
243 sfn_worklist.append(n->as_SafePoint());
244 }
245 }
246
247 #ifndef PRODUCT
248 if (_compile->directive()->TraceEscapeAnalysisOption) {
249 tty->print("+++++ Initial worklist for ");
250 _compile->method()->print_name();
251 tty->print_cr(" (ea_inv=%d)", _invocation);
252 for (int i = 0; i < ptnodes_worklist.length(); i++) {
253 PointsToNode* ptn = ptnodes_worklist.at(i);
254 ptn->dump();
255 }
256 tty->print_cr("+++++ Calculating escape states and scalar replaceability");
257 }
258 #endif
259
260 if (non_escaped_allocs_worklist.length() == 0) {
261 _collecting = false;
262 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
263 return false; // Nothing to do.
264 }
265 // Add final simple edges to graph.
266 while(delayed_worklist.size() > 0) {
267 Node* n = delayed_worklist.pop();
268 add_final_edges(n);
269 }
270
271 #ifdef ASSERT
272 if (VerifyConnectionGraph) {
273 // Verify that no new simple edges could be created and all
274 // local vars has edges.
275 _verify = true;
276 int ptnodes_length = ptnodes_worklist.length();
277 for (int next = 0; next < ptnodes_length; ++next) {
278 PointsToNode* ptn = ptnodes_worklist.at(next);
279 add_final_edges(ptn->ideal_node());
280 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
281 ptn->dump();
282 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
283 }
284 }
285 _verify = false;
286 }
287 #endif
288 // Bytecode analyzer BCEscapeAnalyzer, used for Call nodes
289 // processing, calls to CI to resolve symbols (types, fields, methods)
290 // referenced in bytecode. During symbol resolution VM may throw
291 // an exception which CI cleans and converts to compilation failure.
292 if (C->failing()) {
293 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
294 return false;
295 }
296
297 // 2. Finish Graph construction by propagating references to all
298 // java objects through graph.
299 if (!complete_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
300 java_objects_worklist, oop_fields_worklist)) {
301 // All objects escaped or hit time or iterations limits.
302 _collecting = false;
303 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
304 return false;
305 }
306
307 // 3. Adjust scalar_replaceable state of nonescaping objects and push
308 // scalar replaceable allocations on alloc_worklist for processing
309 // in split_unique_types().
310 GrowableArray<JavaObjectNode*> jobj_worklist;
311 int non_escaped_length = non_escaped_allocs_worklist.length();
312 bool found_nsr_alloc = false;
313 for (int next = 0; next < non_escaped_length; next++) {
314 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
315 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
316 Node* n = ptn->ideal_node();
317 if (n->is_Allocate()) {
318 n->as_Allocate()->_is_non_escaping = noescape;
319 }
320 if (noescape && ptn->scalar_replaceable()) {
321 adjust_scalar_replaceable_state(ptn, reducible_merges);
322 if (ptn->scalar_replaceable()) {
323 jobj_worklist.push(ptn);
324 } else {
325 found_nsr_alloc = true;
326 }
327 }
328 }
329
330 // Propagate NSR (Not Scalar Replaceable) state.
331 if (found_nsr_alloc) {
332 find_scalar_replaceable_allocs(jobj_worklist);
333 }
334
335 // alloc_worklist will be processed in reverse push order.
336 // Therefore the reducible Phis will be processed for last and that's what we
337 // want because by then the scalarizable inputs of the merge will already have
338 // an unique instance type.
339 for (uint i = 0; i < reducible_merges.size(); i++ ) {
340 Node* n = reducible_merges.at(i);
341 alloc_worklist.append(n);
342 }
343
344 for (int next = 0; next < jobj_worklist.length(); ++next) {
345 JavaObjectNode* jobj = jobj_worklist.at(next);
346 if (jobj->scalar_replaceable()) {
347 alloc_worklist.append(jobj->ideal_node());
348 }
349 }
350
351 #ifdef ASSERT
352 if (VerifyConnectionGraph) {
353 // Verify that graph is complete - no new edges could be added or needed.
354 verify_connection_graph(ptnodes_worklist, non_escaped_allocs_worklist,
355 java_objects_worklist, addp_worklist);
356 }
357 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
358 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
359 null_obj->edge_count() == 0 &&
360 !null_obj->arraycopy_src() &&
361 !null_obj->arraycopy_dst(), "sanity");
362 #endif
363
364 _collecting = false;
365
366 } // TracePhase t3("connectionGraph")
367
368 // 4. Optimize ideal graph based on EA information.
369 bool has_non_escaping_obj = (non_escaped_allocs_worklist.length() > 0);
370 if (has_non_escaping_obj) {
371 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
372 }
373
374 #ifndef PRODUCT
375 if (PrintEscapeAnalysis) {
376 dump(ptnodes_worklist); // Dump ConnectionGraph
377 }
378 #endif
379
380 #ifdef ASSERT
381 if (VerifyConnectionGraph) {
382 int alloc_length = alloc_worklist.length();
383 for (int next = 0; next < alloc_length; ++next) {
384 Node* n = alloc_worklist.at(next);
385 PointsToNode* ptn = ptnode_adr(n->_idx);
386 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
387 }
388 }
389
390 if (VerifyReduceAllocationMerges) {
391 for (uint i = 0; i < reducible_merges.size(); i++ ) {
392 Node* n = reducible_merges.at(i);
393 if (!can_reduce_phi(n->as_Phi())) {
394 TraceReduceAllocationMerges = true;
395 n->dump(2);
396 n->dump(-2);
397 assert(can_reduce_phi(n->as_Phi()), "Sanity: previous reducible Phi is no longer reducible before SUT.");
398 }
399 }
400 }
401 #endif
402
403 // 5. Separate memory graph for scalar replaceable allcations.
404 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
405 if (has_scalar_replaceable_candidates && EliminateAllocations) {
406 assert(C->do_aliasing(), "Aliasing should be enabled");
407 // Now use the escape information to create unique types for
408 // scalar replaceable objects.
409 split_unique_types(alloc_worklist, arraycopy_worklist, mergemem_worklist, reducible_merges);
410 if (C->failing()) {
411 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
412 return false;
413 }
414 C->print_method(PHASE_AFTER_EA, 2);
415
416 #ifdef ASSERT
417 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
418 tty->print("=== No allocations eliminated for ");
419 C->method()->print_short_name();
420 if (!EliminateAllocations) {
421 tty->print(" since EliminateAllocations is off ===");
422 } else if(!has_scalar_replaceable_candidates) {
423 tty->print(" since there are no scalar replaceable candidates ===");
424 }
425 tty->cr();
426 #endif
427 }
428
429 // 6. Reduce allocation merges used as debug information. This is done after
430 // split_unique_types because the methods used to create SafePointScalarObject
431 // need to traverse the memory graph to find values for object fields. We also
432 // set to null the scalarized inputs of reducible Phis so that the Allocate
433 // that they point can be later scalar replaced.
434 bool delay = _igvn->delay_transform();
435 _igvn->set_delay_transform(true);
436 for (uint i = 0; i < reducible_merges.size(); i++) {
437 Node* n = reducible_merges.at(i);
438 if (n->outcnt() > 0) {
439 if (!reduce_phi_on_safepoints(n->as_Phi())) {
440 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
441 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
442 return false;
443 }
444
445 // Now we set the scalar replaceable inputs of ophi to null, which is
446 // the last piece that would prevent it from being scalar replaceable.
447 reset_scalar_replaceable_entries(n->as_Phi());
448 }
449 }
450 _igvn->set_delay_transform(delay);
451
452 // Annotate at safepoints if they have <= ArgEscape objects in their scope and at
453 // java calls if they pass ArgEscape objects as parameters.
454 if (has_non_escaping_obj &&
455 (C->env()->should_retain_local_variables() ||
456 C->env()->jvmti_can_get_owned_monitor_info() ||
457 C->env()->jvmti_can_walk_any_space() ||
458 DeoptimizeObjectsALot)) {
459 int sfn_length = sfn_worklist.length();
460 for (int next = 0; next < sfn_length; next++) {
461 SafePointNode* sfn = sfn_worklist.at(next);
462 sfn->set_has_ea_local_in_scope(has_ea_local_in_scope(sfn));
463 if (sfn->is_CallJava()) {
464 CallJavaNode* call = sfn->as_CallJava();
465 call->set_arg_escape(has_arg_escape(call));
466 }
467 }
468 }
469
470 NOT_PRODUCT(escape_state_statistics(java_objects_worklist);)
471 return has_non_escaping_obj;
472 }
473
474 // Check if it's profitable to reduce the Phi passed as parameter. Returns true
475 // if at least one scalar replaceable allocation participates in the merge.
476 bool ConnectionGraph::can_reduce_phi_check_inputs(PhiNode* ophi) const {
477 bool found_sr_allocate = false;
478
479 for (uint i = 1; i < ophi->req(); i++) {
480 JavaObjectNode* ptn = unique_java_object(ophi->in(i));
481 if (ptn != nullptr && ptn->scalar_replaceable()) {
482 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
483
484 // Don't handle arrays.
485 if (alloc->Opcode() != Op_Allocate) {
486 assert(alloc->Opcode() == Op_AllocateArray, "Unexpected type of allocation.");
487 continue;
488 }
489
490 if (PhaseMacroExpand::can_eliminate_allocation(_igvn, alloc, nullptr)) {
491 found_sr_allocate = true;
492 } else {
493 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("%dth input of Phi %d is SR but can't be eliminated.", i, ophi->_idx);)
494 ptn->set_scalar_replaceable(false);
495 }
496 }
497 }
498
499 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);)
500 return found_sr_allocate;
501 }
502
503 // We can reduce the Cmp if it's a comparison between the Phi and a constant.
504 // I require the 'other' input to be a constant so that I can move the Cmp
505 // around safely.
506 bool ConnectionGraph::can_reduce_cmp(Node* n, Node* cmp) const {
507 assert(cmp->Opcode() == Op_CmpP || cmp->Opcode() == Op_CmpN, "not expected node: %s", cmp->Name());
508 Node* left = cmp->in(1);
509 Node* right = cmp->in(2);
510
511 return (left == n || right == n) &&
512 (left->is_Con() || right->is_Con()) &&
513 cmp->outcnt() == 1;
514 }
515
516 // We are going to check if any of the SafePointScalarMerge entries
517 // in the SafePoint reference the Phi that we are checking.
518 bool ConnectionGraph::has_been_reduced(PhiNode* n, SafePointNode* sfpt) const {
519 JVMState *jvms = sfpt->jvms();
520
521 for (uint i = jvms->debug_start(); i < jvms->debug_end(); i++) {
522 Node* sfpt_in = sfpt->in(i);
523 if (sfpt_in->is_SafePointScalarMerge()) {
524 SafePointScalarMergeNode* smerge = sfpt_in->as_SafePointScalarMerge();
525 Node* nsr_ptr = sfpt->in(smerge->merge_pointer_idx(jvms));
526 if (nsr_ptr == n) {
527 return true;
528 }
529 }
530 }
531
532 return false;
533 }
534
535 // Check if we are able to untangle the merge. The following patterns are
536 // supported:
537 // - Phi -> SafePoints
538 // - Phi -> CmpP/N
539 // - Phi -> AddP -> Load
540 // - Phi -> CastPP -> SafePoints
541 // - Phi -> CastPP -> AddP -> Load
542 bool ConnectionGraph::can_reduce_check_users(Node* n, uint nesting) const {
543 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
544 Node* use = n->fast_out(i);
545
546 if (use->is_SafePoint()) {
547 if (use->is_Call() && use->as_Call()->has_non_debug_use(n)) {
548 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. Call has non_debug_use().", n->_idx, _invocation);)
549 return false;
550 } else if (has_been_reduced(n->is_Phi() ? n->as_Phi() : n->as_CastPP()->in(1)->as_Phi(), use->as_SafePoint())) {
551 NOT_PRODUCT(if (TraceReduceAllocationMerges) tty->print_cr("Can NOT reduce Phi %d on invocation %d. It has already been reduced.", n->_idx, _invocation);)
552 return false;
553 }
554 } else if (use->is_AddP()) {
555 Node* addp = use;
556 for (DUIterator_Fast jmax, j = addp->fast_outs(jmax); j < jmax; j++) {
557 Node* use_use = addp->fast_out(j);
558 const Type* load_type = _igvn->type(use_use);
559
560 if (!use_use->is_Load() || !use_use->as_Load()->can_split_through_phi_base(_igvn)) {
561 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());)
562 return false;
563 } else if (load_type->isa_narrowklass() || load_type->isa_klassptr()) {
564 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());)
565 return false;
566 }
567 }
568 } else if (nesting > 0) {
569 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);)
570 return false;
571 } else if (use->is_CastPP()) {
572 const Type* cast_t = _igvn->type(use);
573 if (cast_t == nullptr || cast_t->make_ptr()->isa_instptr() == nullptr) {
574 #ifndef PRODUCT
575 if (TraceReduceAllocationMerges) {
576 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP is not to an instance.", n->_idx, _invocation);
577 use->dump();
578 }
579 #endif
580 return false;
581 }
582
583 bool is_trivial_control = use->in(0) == nullptr || use->in(0) == n->in(0);
584 if (!is_trivial_control) {
585 // If it's not a trivial control then we check if we can reduce the
586 // CmpP/N used by the If controlling the cast.
587 if (use->in(0)->is_IfTrue() || use->in(0)->is_IfFalse()) {
588 Node* iff = use->in(0)->in(0);
589 // We may have Opaque4 node between If and Bool nodes.
590 // Bail out in such case - we need to preserve Opaque4 for correct
591 // processing predicates after loop opts.
592 bool can_reduce = (iff->Opcode() == Op_If) && iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp();
593 if (can_reduce) {
594 Node* iff_cmp = iff->in(1)->in(1);
595 int opc = iff_cmp->Opcode();
596 can_reduce = (opc == Op_CmpP || opc == Op_CmpN) && can_reduce_cmp(n, iff_cmp);
597 }
598 if (!can_reduce) {
599 #ifndef PRODUCT
600 if (TraceReduceAllocationMerges) {
601 tty->print_cr("Can NOT reduce Phi %d on invocation %d. CastPP %d doesn't have simple control.", n->_idx, _invocation, use->_idx);
602 n->dump(5);
603 }
604 #endif
605 return false;
606 }
607 }
608 }
609
610 if (!can_reduce_check_users(use, nesting+1)) {
611 return false;
612 }
613 } else if (use->Opcode() == Op_CmpP || use->Opcode() == Op_CmpN) {
614 if (!can_reduce_cmp(n, use)) {
615 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);)
616 return false;
617 }
618 } else {
619 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());)
620 return false;
621 }
622 }
623
624 return true;
625 }
626
627 // Returns true if: 1) It's profitable to reduce the merge, and 2) The Phi is
628 // only used in some certain code shapes. Check comments in
629 // 'can_reduce_phi_inputs' and 'can_reduce_phi_users' for more
630 // details.
631 bool ConnectionGraph::can_reduce_phi(PhiNode* ophi) const {
632 // If there was an error attempting to reduce allocation merges for this
633 // method we might have disabled the compilation and be retrying with RAM
634 // disabled.
635 if (!_compile->do_reduce_allocation_merges() || ophi->region()->Opcode() != Op_Region) {
636 return false;
637 }
638
639 const Type* phi_t = _igvn->type(ophi);
640 if (phi_t == nullptr ||
641 phi_t->make_ptr() == nullptr ||
642 phi_t->make_ptr()->isa_aryptr() != nullptr) {
643 return false;
644 }
645
646 if (!can_reduce_phi_check_inputs(ophi) || !can_reduce_check_users(ophi, /* nesting: */ 0)) {
647 return false;
648 }
649
650 NOT_PRODUCT(if (TraceReduceAllocationMerges) { tty->print_cr("Can reduce Phi %d during invocation %d: ", ophi->_idx, _invocation); })
651 return true;
652 }
653
654 // This method will return a CmpP/N that we need to use on the If controlling a
655 // CastPP after it was split. This method is only called on bases that are
656 // nullable therefore we always need a controlling if for the splitted CastPP.
657 //
658 // 'curr_ctrl' is the control of the CastPP that we want to split through phi.
659 // If the CastPP currently doesn't have a control then the CmpP/N will be
660 // against the NULL constant, otherwise it will be against the constant input of
661 // the existing CmpP/N. It's guaranteed that there will be a CmpP/N in the later
662 // case because we have constraints on it and because the CastPP has a control
663 // input.
664 Node* ConnectionGraph::specialize_cmp(Node* base, Node* curr_ctrl) {
665 const Type* t = base->bottom_type();
666 Node* con = nullptr;
667
668 if (curr_ctrl == nullptr || curr_ctrl->is_Region()) {
669 con = _igvn->zerocon(t->basic_type());
670 } else {
671 // can_reduce_check_users() verified graph: true/false -> if -> bool -> cmp
672 assert(curr_ctrl->in(0)->Opcode() == Op_If, "unexpected node %s", curr_ctrl->in(0)->Name());
673 Node* bol = curr_ctrl->in(0)->in(1);
674 assert(bol->is_Bool(), "unexpected node %s", bol->Name());
675 Node* curr_cmp = bol->in(1);
676 assert(curr_cmp->Opcode() == Op_CmpP || curr_cmp->Opcode() == Op_CmpN, "unexpected node %s", curr_cmp->Name());
677 con = curr_cmp->in(1)->is_Con() ? curr_cmp->in(1) : curr_cmp->in(2);
678 }
679
680 return CmpNode::make(base, con, t->basic_type());
681 }
682
683 // This method 'specializes' the CastPP passed as parameter to the base passed
684 // as parameter. Note that the existing CastPP input is a Phi. "Specialize"
685 // means that the CastPP now will be specific for a given base instead of a Phi.
686 // An If-Then-Else-Region block is inserted to control the CastPP. The control
687 // of the CastPP is a copy of the current one (if there is one) or a check
688 // against NULL.
689 //
690 // Before:
691 //
692 // C1 C2 ... Cn
693 // \ | /
694 // \ | /
695 // \ | /
696 // \ | /
697 // \ | /
698 // \ | /
699 // \|/
700 // Region B1 B2 ... Bn
701 // | \ | /
702 // | \ | /
703 // | \ | /
704 // | \ | /
705 // | \ | /
706 // | \ | /
707 // ---------------> Phi
708 // |
709 // X |
710 // | |
711 // | |
712 // ------> CastPP
713 //
714 // After (only partial illustration; base = B2, current_control = C2):
715 //
716 // C2
717 // |
718 // If
719 // / \
720 // / \
721 // T F
722 // /\ /
723 // / \ /
724 // / \ /
725 // C1 CastPP Reg Cn
726 // | | |
727 // | | |
728 // | | |
729 // -------------- | ----------
730 // | | |
731 // Region
732 //
733 Node* ConnectionGraph::specialize_castpp(Node* castpp, Node* base, Node* current_control) {
734 Node* control_successor = current_control->unique_ctrl_out();
735 Node* cmp = _igvn->transform(specialize_cmp(base, castpp->in(0)));
736 Node* bol = _igvn->transform(new BoolNode(cmp, BoolTest::ne));
737 IfNode* if_ne = _igvn->transform(new IfNode(current_control, bol, PROB_MIN, COUNT_UNKNOWN))->as_If();
738 Node* not_eq_control = _igvn->transform(new IfTrueNode(if_ne));
739 Node* yes_eq_control = _igvn->transform(new IfFalseNode(if_ne));
740 Node* end_region = _igvn->transform(new RegionNode(3));
741
742 // Insert the new if-else-region block into the graph
743 end_region->set_req(1, not_eq_control);
744 end_region->set_req(2, yes_eq_control);
745 control_successor->replace_edge(current_control, end_region, _igvn);
746
747 _igvn->_worklist.push(current_control);
748 _igvn->_worklist.push(control_successor);
749
750 return _igvn->transform(ConstraintCastNode::make_cast_for_type(not_eq_control, base, _igvn->type(castpp), ConstraintCastNode::UnconditionalDependency, nullptr));
751 }
752
753 Node* ConnectionGraph::split_castpp_load_through_phi(Node* curr_addp, Node* curr_load, Node* region, GrowableArray<Node*>* bases_for_loads, GrowableArray<Node *> &alloc_worklist) {
754 const Type* load_type = _igvn->type(curr_load);
755 Node* nsr_value = _igvn->zerocon(load_type->basic_type());
756 Node* memory = curr_load->in(MemNode::Memory);
757
758 // The data_phi merging the loads needs to be nullable if
759 // we are loading pointers.
760 if (load_type->make_ptr() != nullptr) {
761 if (load_type->isa_narrowoop()) {
762 load_type = load_type->meet(TypeNarrowOop::NULL_PTR);
763 } else if (load_type->isa_ptr()) {
764 load_type = load_type->meet(TypePtr::NULL_PTR);
765 } else {
766 assert(false, "Unexpected load ptr type.");
767 }
768 }
769
770 Node* data_phi = PhiNode::make(region, nsr_value, load_type);
771
772 for (int i = 1; i < bases_for_loads->length(); i++) {
773 Node* base = bases_for_loads->at(i);
774 Node* cmp_region = nullptr;
775 if (base != nullptr) {
776 if (base->is_CFG()) { // means that we added a CastPP as child of this CFG node
777 cmp_region = base->unique_ctrl_out_or_null();
778 assert(cmp_region != nullptr, "There should be.");
779 base = base->find_out_with(Op_CastPP);
780 }
781
782 Node* addr = _igvn->transform(new AddPNode(base, base, curr_addp->in(AddPNode::Offset)));
783 Node* mem = (memory->is_Phi() && (memory->in(0) == region)) ? memory->in(i) : memory;
784 Node* load = curr_load->clone();
785 load->set_req(0, nullptr);
786 load->set_req(1, mem);
787 load->set_req(2, addr);
788
789 if (cmp_region != nullptr) { // see comment on previous if
790 Node* intermediate_phi = PhiNode::make(cmp_region, nsr_value, load_type);
791 intermediate_phi->set_req(1, _igvn->transform(load));
792 load = intermediate_phi;
793 }
794
795 data_phi->set_req(i, _igvn->transform(load));
796 } else {
797 // Just use the default, which is already in phi
798 }
799 }
800
801 // Takes care of updating CG and split_unique_types worklists due
802 // to cloned AddP->Load.
803 updates_after_load_split(data_phi, curr_load, alloc_worklist);
804
805 return _igvn->transform(data_phi);
806 }
807
808 // This method only reduces CastPP fields loads; SafePoints are handled
809 // separately. The idea here is basically to clone the CastPP and place copies
810 // on each input of the Phi, including non-scalar replaceable inputs.
811 // Experimentation shows that the resulting IR graph is simpler that way than if
812 // we just split the cast through scalar-replaceable inputs.
813 //
814 // The reduction process requires that CastPP's control be one of:
815 // 1) no control,
816 // 2) the same region as Ophi, or
817 // 3) an IfTrue/IfFalse coming from an CmpP/N between Ophi and a constant.
818 //
819 // After splitting the CastPP we'll put it under an If-Then-Else-Region control
820 // flow. If the CastPP originally had an IfTrue/False control input then we'll
821 // use a similar CmpP/N to control the new If-Then-Else-Region. Otherwise, we'll
822 // juse use a CmpP/N against the NULL constant.
823 //
824 // The If-Then-Else-Region isn't always needed. For instance, if input to
825 // splitted cast was not nullable (or if it was the NULL constant) then we don't
826 // need (shouldn't) use a CastPP at all.
827 //
828 // After the casts are splitted we'll split the AddP->Loads through the Phi and
829 // connect them to the just split CastPPs.
830 //
831 // Before (CastPP control is same as Phi):
832 //
833 // Region Allocate Null Call
834 // | \ | /
835 // | \ | /
836 // | \ | /
837 // | \ | /
838 // | \ | /
839 // | \ | /
840 // ------------------> Phi # Oop Phi
841 // | |
842 // | |
843 // | |
844 // | |
845 // ----------------> CastPP
846 // |
847 // AddP
848 // |
849 // Load
850 //
851 // After (Very much simplified):
852 //
853 // Call NULL
854 // \ /
855 // CmpP
856 // |
857 // Bool#NE
858 // |
859 // If
860 // / \
861 // T F
862 // / \ /
863 // / R
864 // CastPP |
865 // | |
866 // AddP |
867 // | |
868 // Load |
869 // \ | 0
870 // Allocate \ | /
871 // \ \ | /
872 // AddP Phi
873 // \ /
874 // Load /
875 // \ 0 /
876 // \ | /
877 // \|/
878 // Phi # "Field" Phi
879 //
880 void ConnectionGraph::reduce_phi_on_castpp_field_load(Node* curr_castpp, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
881 Node* ophi = curr_castpp->in(1);
882 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
883
884 // Identify which base should be used for AddP->Load later when spliting the
885 // CastPP->Loads through ophi. Three kind of values may be stored in this
886 // array, depending on the nullability status of the corresponding input in
887 // ophi.
888 //
889 // - nullptr: Meaning that the base is actually the NULL constant and therefore
890 // we won't try to load from it.
891 //
892 // - CFG Node: Meaning that the base is a CastPP that was specialized for
893 // this input of Ophi. I.e., we added an If->Then->Else-Region
894 // that will 'activate' the CastPp only when the input is not Null.
895 //
896 // - Other Node: Meaning that the base is not nullable and therefore we'll try
897 // to load directly from it.
898 GrowableArray<Node*> bases_for_loads(ophi->req(), ophi->req(), nullptr);
899
900 for (uint i = 1; i < ophi->req(); i++) {
901 Node* base = ophi->in(i);
902 const Type* base_t = _igvn->type(base);
903
904 if (base_t->maybe_null()) {
905 if (base->is_Con()) {
906 // Nothing todo as bases_for_loads[i] is already nullptr
907 } else {
908 Node* new_castpp = specialize_castpp(curr_castpp, base, ophi->in(0)->in(i));
909 bases_for_loads.at_put(i, new_castpp->in(0)); // Use the ctrl of the new node just as a flag
910 }
911 } else {
912 bases_for_loads.at_put(i, base);
913 }
914 }
915
916 // Now let's split the CastPP->Loads through the Phi
917 for (int i = curr_castpp->outcnt()-1; i >= 0;) {
918 Node* use = curr_castpp->raw_out(i);
919 if (use->is_AddP()) {
920 for (int j = use->outcnt()-1; j >= 0;) {
921 Node* use_use = use->raw_out(j);
922 assert(use_use->is_Load(), "Expected this to be a Load node.");
923
924 // We can't make an unconditional load from a nullable input. The
925 // 'split_castpp_load_through_phi` method will add an
926 // 'If-Then-Else-Region` around nullable bases and only load from them
927 // when the input is not null.
928 Node* phi = split_castpp_load_through_phi(use, use_use, ophi->in(0), &bases_for_loads, alloc_worklist);
929 _igvn->replace_node(use_use, phi);
930
931 --j;
932 j = MIN2(j, (int)use->outcnt()-1);
933 }
934
935 _igvn->remove_dead_node(use);
936 }
937 --i;
938 i = MIN2(i, (int)curr_castpp->outcnt()-1);
939 }
940 }
941
942 // This method split a given CmpP/N through the Phi used in one of its inputs.
943 // As a result we convert a comparison with a pointer to a comparison with an
944 // integer.
945 // The only requirement is that one of the inputs of the CmpP/N must be a Phi
946 // while the other must be a constant.
947 // The splitting process is basically just cloning the CmpP/N above the input
948 // Phi. However, some (most) of the cloned CmpP/Ns won't be requred because we
949 // can prove at compile time the result of the comparison.
950 //
951 // Before:
952 //
953 // in1 in2 ... inN
954 // \ | /
955 // \ | /
956 // \ | /
957 // \ | /
958 // \ | /
959 // \ | /
960 // Phi
961 // | Other
962 // | /
963 // | /
964 // | /
965 // CmpP/N
966 //
967 // After:
968 //
969 // in1 Other in2 Other inN Other
970 // | | | | | |
971 // \ | | | | |
972 // \ / | / | /
973 // CmpP/N CmpP/N CmpP/N
974 // Bool Bool Bool
975 // \ | /
976 // \ | /
977 // \ | /
978 // \ | /
979 // \ | /
980 // \ | /
981 // \ | /
982 // \ | /
983 // Phi
984 // |
985 // | Zero
986 // | /
987 // | /
988 // | /
989 // CmpI
990 //
991 //
992 void ConnectionGraph::reduce_phi_on_cmp(Node* cmp) {
993 Node* ophi = cmp->in(1)->is_Con() ? cmp->in(2) : cmp->in(1);
994 assert(ophi->is_Phi(), "Expected this to be a Phi node.");
995
996 Node* other = cmp->in(1)->is_Con() ? cmp->in(1) : cmp->in(2);
997 Node* zero = _igvn->intcon(0);
998 BoolTest::mask mask = cmp->unique_out()->as_Bool()->_test._test;
999
1000 // This Phi will merge the result of the Cmps split through the Phi
1001 Node* res_phi = _igvn->transform(PhiNode::make(ophi->in(0), zero, TypeInt::INT));
1002
1003 for (uint i=1; i<ophi->req(); i++) {
1004 Node* ophi_input = ophi->in(i);
1005 Node* res_phi_input = nullptr;
1006
1007 const TypeInt* tcmp = optimize_ptr_compare(ophi_input, other);
1008 if (tcmp->singleton()) {
1009 res_phi_input = _igvn->makecon(tcmp);
1010 } else {
1011 Node* ncmp = _igvn->transform(cmp->clone());
1012 ncmp->set_req(1, ophi_input);
1013 ncmp->set_req(2, other);
1014 Node* bol = _igvn->transform(new BoolNode(ncmp, mask));
1015 res_phi_input = bol->as_Bool()->as_int_value(_igvn);
1016 }
1017
1018 res_phi->set_req(i, res_phi_input);
1019 }
1020
1021 Node* new_cmp = _igvn->transform(new CmpINode(res_phi, zero));
1022 _igvn->replace_node(cmp, new_cmp);
1023 }
1024
1025 // Push the newly created AddP on alloc_worklist and patch
1026 // the connection graph. Note that the changes in the CG below
1027 // won't affect the ES of objects since the new nodes have the
1028 // same status as the old ones.
1029 void ConnectionGraph::updates_after_load_split(Node* data_phi, Node* previous_load, GrowableArray<Node *> &alloc_worklist) {
1030 assert(data_phi != nullptr, "Output of split_through_phi is null.");
1031 assert(data_phi != previous_load, "Output of split_through_phi is same as input.");
1032 assert(data_phi->is_Phi(), "Output of split_through_phi isn't a Phi.");
1033
1034 if (data_phi == nullptr || !data_phi->is_Phi()) {
1035 // Make this a retry?
1036 return ;
1037 }
1038
1039 Node* previous_addp = previous_load->in(MemNode::Address);
1040 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1041 for (uint i = 1; i < data_phi->req(); i++) {
1042 Node* new_load = data_phi->in(i);
1043
1044 if (new_load->is_Phi()) {
1045 // new_load is currently the "intermediate_phi" from an specialized
1046 // CastPP.
1047 new_load = new_load->in(1);
1048 }
1049
1050 // "new_load" might actually be a constant, parameter, etc.
1051 if (new_load->is_Load()) {
1052 Node* new_addp = new_load->in(MemNode::Address);
1053 Node* base = get_addp_base(new_addp);
1054
1055 // The base might not be something that we can create an unique
1056 // type for. If that's the case we are done with that input.
1057 PointsToNode* jobj_ptn = unique_java_object(base);
1058 if (jobj_ptn == nullptr || !jobj_ptn->scalar_replaceable()) {
1059 continue;
1060 }
1061
1062 // Push to alloc_worklist since the base has an unique_type
1063 alloc_worklist.append_if_missing(new_addp);
1064
1065 // Now let's add the node to the connection graph
1066 _nodes.at_grow(new_addp->_idx, nullptr);
1067 add_field(new_addp, fn->escape_state(), fn->offset());
1068 add_base(ptnode_adr(new_addp->_idx)->as_Field(), ptnode_adr(base->_idx));
1069
1070 // If the load doesn't load an object then it won't be
1071 // part of the connection graph
1072 PointsToNode* curr_load_ptn = ptnode_adr(previous_load->_idx);
1073 if (curr_load_ptn != nullptr) {
1074 _nodes.at_grow(new_load->_idx, nullptr);
1075 add_local_var(new_load, curr_load_ptn->escape_state());
1076 add_edge(ptnode_adr(new_load->_idx), ptnode_adr(new_addp->_idx)->as_Field());
1077 }
1078 }
1079 }
1080 }
1081
1082 void ConnectionGraph::reduce_phi_on_field_access(Node* previous_addp, GrowableArray<Node *> &alloc_worklist) {
1083 // We'll pass this to 'split_through_phi' so that it'll do the split even
1084 // though the load doesn't have an unique instance type.
1085 bool ignore_missing_instance_id = true;
1086
1087 // All AddPs are present in the connection graph
1088 FieldNode* fn = ptnode_adr(previous_addp->_idx)->as_Field();
1089
1090 // Iterate over AddP looking for a Load
1091 for (int k = previous_addp->outcnt()-1; k >= 0;) {
1092 Node* previous_load = previous_addp->raw_out(k);
1093 if (previous_load->is_Load()) {
1094 Node* data_phi = previous_load->as_Load()->split_through_phi(_igvn, ignore_missing_instance_id);
1095
1096 // Takes care of updating CG and split_unique_types worklists due to cloned
1097 // AddP->Load.
1098 updates_after_load_split(data_phi, previous_load, alloc_worklist);
1099
1100 _igvn->replace_node(previous_load, data_phi);
1101 }
1102 --k;
1103 k = MIN2(k, (int)previous_addp->outcnt()-1);
1104 }
1105
1106 // Remove the old AddP from the processing list because it's dead now
1107 assert(previous_addp->outcnt() == 0, "AddP should be dead now.");
1108 alloc_worklist.remove_if_existing(previous_addp);
1109 }
1110
1111 // Create a 'selector' Phi based on the inputs of 'ophi'. If index 'i' of the
1112 // selector is:
1113 // -> a '-1' constant, the i'th input of the original Phi is NSR.
1114 // -> a 'x' constant >=0, the i'th input of of original Phi will be SR and
1115 // the info about the scalarized object will be at index x of ObjectMergeValue::possible_objects
1116 PhiNode* ConnectionGraph::create_selector(PhiNode* ophi) const {
1117 Node* minus_one = _igvn->register_new_node_with_optimizer(ConINode::make(-1));
1118 Node* selector = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), minus_one, TypeInt::INT));
1119 uint number_of_sr_objects = 0;
1120 for (uint i = 1; i < ophi->req(); i++) {
1121 Node* base = ophi->in(i);
1122 JavaObjectNode* ptn = unique_java_object(base);
1123
1124 if (ptn != nullptr && ptn->scalar_replaceable()) {
1125 Node* sr_obj_idx = _igvn->register_new_node_with_optimizer(ConINode::make(number_of_sr_objects));
1126 selector->set_req(i, sr_obj_idx);
1127 number_of_sr_objects++;
1128 }
1129 }
1130
1131 return selector->as_Phi();
1132 }
1133
1134 // Returns true if the AddP node 'n' has at least one base that is a reducible
1135 // merge. If the base is a CastPP/CheckCastPP then the input of the cast is
1136 // checked instead.
1137 bool ConnectionGraph::has_reducible_merge_base(AddPNode* n, Unique_Node_List &reducible_merges) {
1138 PointsToNode* ptn = ptnode_adr(n->_idx);
1139 if (ptn == nullptr || !ptn->is_Field() || ptn->as_Field()->base_count() < 2) {
1140 return false;
1141 }
1142
1143 for (BaseIterator i(ptn->as_Field()); i.has_next(); i.next()) {
1144 Node* base = i.get()->ideal_node();
1145
1146 if (reducible_merges.member(base)) {
1147 return true;
1148 }
1149
1150 if (base->is_CastPP() || base->is_CheckCastPP()) {
1151 base = base->in(1);
1152 if (reducible_merges.member(base)) {
1153 return true;
1154 }
1155 }
1156 }
1157
1158 return false;
1159 }
1160
1161 // This method will call its helper method to reduce SafePoint nodes that use
1162 // 'ophi' or a casted version of 'ophi'. All SafePoint nodes using the same
1163 // "version" of Phi use the same debug information (regarding the Phi).
1164 // Therefore, I collect all safepoints and patch them all at once.
1165 //
1166 // The safepoints using the Phi node have to be processed before safepoints of
1167 // CastPP nodes. The reason is, when reducing a CastPP we add a reference (the
1168 // NSR merge pointer) to the input of the CastPP (i.e., the Phi) in the
1169 // safepoint. If we process CastPP's safepoints before Phi's safepoints the
1170 // algorithm that process Phi's safepoints will think that the added Phi
1171 // reference is a regular reference.
1172 bool ConnectionGraph::reduce_phi_on_safepoints(PhiNode* ophi) {
1173 PhiNode* selector = create_selector(ophi);
1174 Unique_Node_List safepoints;
1175 Unique_Node_List casts;
1176
1177 // Just collect the users of the Phis for later processing
1178 // in the needed order.
1179 for (uint i = 0; i < ophi->outcnt(); i++) {
1180 Node* use = ophi->raw_out(i);
1181 if (use->is_SafePoint()) {
1182 safepoints.push(use);
1183 } else if (use->is_CastPP()) {
1184 casts.push(use);
1185 } else {
1186 assert(use->outcnt() == 0, "Only CastPP & SafePoint users should be left.");
1187 }
1188 }
1189
1190 // Need to process safepoints using the Phi first
1191 if (!reduce_phi_on_safepoints_helper(ophi, nullptr, selector, safepoints)) {
1192 return false;
1193 }
1194
1195 // Now process CastPP->safepoints
1196 for (uint i = 0; i < casts.size(); i++) {
1197 Node* cast = casts.at(i);
1198 Unique_Node_List cast_sfpts;
1199
1200 for (DUIterator_Fast jmax, j = cast->fast_outs(jmax); j < jmax; j++) {
1201 Node* use_use = cast->fast_out(j);
1202 if (use_use->is_SafePoint()) {
1203 cast_sfpts.push(use_use);
1204 } else {
1205 assert(use_use->outcnt() == 0, "Only SafePoint users should be left.");
1206 }
1207 }
1208
1209 if (!reduce_phi_on_safepoints_helper(ophi, cast, selector, cast_sfpts)) {
1210 return false;
1211 }
1212 }
1213
1214 return true;
1215 }
1216
1217 // This method will create a SafePointScalarMERGEnode for each SafePoint in
1218 // 'safepoints'. It then will iterate on the inputs of 'ophi' and create a
1219 // SafePointScalarObjectNode for each scalar replaceable input. Each
1220 // SafePointScalarMergeNode may describe multiple scalar replaced objects -
1221 // check detailed description in SafePointScalarMergeNode class header.
1222 bool ConnectionGraph::reduce_phi_on_safepoints_helper(Node* ophi, Node* cast, Node* selector, Unique_Node_List& safepoints) {
1223 PhaseMacroExpand mexp(*_igvn);
1224 Node* original_sfpt_parent = cast != nullptr ? cast : ophi;
1225 const TypeOopPtr* merge_t = _igvn->type(original_sfpt_parent)->make_oopptr();
1226
1227 Node* nsr_merge_pointer = ophi;
1228 if (cast != nullptr) {
1229 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1230 nsr_merge_pointer = _igvn->transform(ConstraintCastNode::make_cast_for_type(cast->in(0), cast->in(1), new_t, ConstraintCastNode::RegularDependency, nullptr));
1231 }
1232
1233 for (uint spi = 0; spi < safepoints.size(); spi++) {
1234 SafePointNode* sfpt = safepoints.at(spi)->as_SafePoint();
1235 JVMState *jvms = sfpt->jvms();
1236 uint merge_idx = (sfpt->req() - jvms->scloff());
1237 int debug_start = jvms->debug_start();
1238
1239 SafePointScalarMergeNode* smerge = new SafePointScalarMergeNode(merge_t, merge_idx);
1240 smerge->init_req(0, _compile->root());
1241 _igvn->register_new_node_with_optimizer(smerge);
1242
1243 // The next two inputs are:
1244 // (1) A copy of the original pointer to NSR objects.
1245 // (2) A selector, used to decide if we need to rematerialize an object
1246 // or use the pointer to a NSR object.
1247 // See more details of these fields in the declaration of SafePointScalarMergeNode
1248 sfpt->add_req(nsr_merge_pointer);
1249 sfpt->add_req(selector);
1250
1251 for (uint i = 1; i < ophi->req(); i++) {
1252 Node* base = ophi->in(i);
1253 JavaObjectNode* ptn = unique_java_object(base);
1254
1255 // If the base is not scalar replaceable we don't need to register information about
1256 // it at this time.
1257 if (ptn == nullptr || !ptn->scalar_replaceable()) {
1258 continue;
1259 }
1260
1261 AllocateNode* alloc = ptn->ideal_node()->as_Allocate();
1262 Unique_Node_List value_worklist;
1263 #ifdef ASSERT
1264 const Type* res_type = alloc->result_cast()->bottom_type();
1265 if (res_type->is_inlinetypeptr() && !Compile::current()->has_circular_inline_type()) {
1266 PhiNode* phi = ophi->as_Phi();
1267 assert(!ophi->as_Phi()->can_push_inline_types_down(_igvn), "missed earlier scalarization opportunity");
1268 }
1269 #endif
1270 SafePointScalarObjectNode* sobj = mexp.create_scalarized_object_description(alloc, sfpt, &value_worklist);
1271 if (sobj == nullptr) {
1272 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1273 return false;
1274 }
1275
1276 // Now make a pass over the debug information replacing any references
1277 // to the allocated object with "sobj"
1278 Node* ccpp = alloc->result_cast();
1279 sfpt->replace_edges_in_range(ccpp, sobj, debug_start, jvms->debug_end(), _igvn);
1280
1281 // Register the scalarized object as a candidate for reallocation
1282 smerge->add_req(sobj);
1283
1284 // Scalarize inline types that were added to the safepoint.
1285 // Don't allow linking a constant oop (if available) for flat array elements
1286 // because Deoptimization::reassign_flat_array_elements needs field values.
1287 const bool allow_oop = !merge_t->is_flat();
1288 for (uint j = 0; j < value_worklist.size(); ++j) {
1289 InlineTypeNode* vt = value_worklist.at(j)->as_InlineType();
1290 vt->make_scalar_in_safepoints(_igvn, allow_oop);
1291 }
1292 }
1293
1294 // Replaces debug information references to "original_sfpt_parent" in "sfpt" with references to "smerge"
1295 sfpt->replace_edges_in_range(original_sfpt_parent, smerge, debug_start, jvms->debug_end(), _igvn);
1296
1297 // The call to 'replace_edges_in_range' above might have removed the
1298 // reference to ophi that we need at _merge_pointer_idx. The line below make
1299 // sure the reference is maintained.
1300 sfpt->set_req(smerge->merge_pointer_idx(jvms), nsr_merge_pointer);
1301 _igvn->_worklist.push(sfpt);
1302 }
1303
1304 return true;
1305 }
1306
1307 void ConnectionGraph::reduce_phi(PhiNode* ophi, GrowableArray<Node *> &alloc_worklist, GrowableArray<Node *> &memnode_worklist) {
1308 bool delay = _igvn->delay_transform();
1309 _igvn->set_delay_transform(true);
1310 _igvn->hash_delete(ophi);
1311
1312 // Copying all users first because some will be removed and others won't.
1313 // Ophi also may acquire some new users as part of Cast reduction.
1314 // CastPPs also need to be processed before CmpPs.
1315 Unique_Node_List castpps;
1316 Unique_Node_List others;
1317 for (DUIterator_Fast imax, i = ophi->fast_outs(imax); i < imax; i++) {
1318 Node* use = ophi->fast_out(i);
1319
1320 if (use->is_CastPP()) {
1321 castpps.push(use);
1322 } else if (use->is_AddP() || use->is_Cmp()) {
1323 others.push(use);
1324 } else if (use->is_SafePoint()) {
1325 // processed later
1326 } else {
1327 assert(use->is_SafePoint(), "Unexpected user of reducible Phi %d -> %d:%s:%d", ophi->_idx, use->_idx, use->Name(), use->outcnt());
1328 }
1329 }
1330
1331 // CastPPs need to be processed before Cmps because during the process of
1332 // splitting CastPPs we make reference to the inputs of the Cmp that is used
1333 // by the If controlling the CastPP.
1334 for (uint i = 0; i < castpps.size(); i++) {
1335 reduce_phi_on_castpp_field_load(castpps.at(i), alloc_worklist, memnode_worklist);
1336 }
1337
1338 for (uint i = 0; i < others.size(); i++) {
1339 Node* use = others.at(i);
1340
1341 if (use->is_AddP()) {
1342 reduce_phi_on_field_access(use, alloc_worklist);
1343 } else if(use->is_Cmp()) {
1344 reduce_phi_on_cmp(use);
1345 }
1346 }
1347
1348 _igvn->set_delay_transform(delay);
1349 }
1350
1351 void ConnectionGraph::reset_scalar_replaceable_entries(PhiNode* ophi) {
1352 Node* null_ptr = _igvn->makecon(TypePtr::NULL_PTR);
1353 const TypeOopPtr* merge_t = _igvn->type(ophi)->make_oopptr();
1354 const Type* new_t = merge_t->meet(TypePtr::NULL_PTR);
1355 Node* new_phi = _igvn->register_new_node_with_optimizer(PhiNode::make(ophi->region(), null_ptr, new_t));
1356
1357 for (uint i = 1; i < ophi->req(); i++) {
1358 Node* base = ophi->in(i);
1359 JavaObjectNode* ptn = unique_java_object(base);
1360
1361 if (ptn != nullptr && ptn->scalar_replaceable()) {
1362 new_phi->set_req(i, null_ptr);
1363 } else {
1364 new_phi->set_req(i, ophi->in(i));
1365 }
1366 }
1367
1368 for (int i = ophi->outcnt()-1; i >= 0;) {
1369 Node* out = ophi->raw_out(i);
1370
1371 if (out->is_ConstraintCast()) {
1372 const Type* out_t = _igvn->type(out)->make_ptr();
1373 const Type* out_new_t = out_t->meet(TypePtr::NULL_PTR);
1374 bool change = out_new_t != out_t;
1375
1376 for (int j = out->outcnt()-1; change && j >= 0; --j) {
1377 Node* out2 = out->raw_out(j);
1378 if (!out2->is_SafePoint()) {
1379 change = false;
1380 break;
1381 }
1382 }
1383
1384 if (change) {
1385 Node* new_cast = ConstraintCastNode::make_cast_for_type(out->in(0), out->in(1), out_new_t, ConstraintCastNode::StrongDependency, nullptr);
1386 _igvn->replace_node(out, new_cast);
1387 _igvn->register_new_node_with_optimizer(new_cast);
1388 }
1389 }
1390
1391 --i;
1392 i = MIN2(i, (int)ophi->outcnt()-1);
1393 }
1394
1395 _igvn->replace_node(ophi, new_phi);
1396 }
1397
1398 void ConnectionGraph::verify_ram_nodes(Compile* C, Node* root) {
1399 if (!C->do_reduce_allocation_merges()) return;
1400
1401 Unique_Node_List ideal_nodes;
1402 ideal_nodes.map(C->live_nodes(), nullptr); // preallocate space
1403 ideal_nodes.push(root);
1404
1405 for (uint next = 0; next < ideal_nodes.size(); ++next) {
1406 Node* n = ideal_nodes.at(next);
1407
1408 if (n->is_SafePointScalarMerge()) {
1409 SafePointScalarMergeNode* merge = n->as_SafePointScalarMerge();
1410
1411 // Validate inputs of merge
1412 for (uint i = 1; i < merge->req(); i++) {
1413 if (merge->in(i) != nullptr && !merge->in(i)->is_top() && !merge->in(i)->is_SafePointScalarObject()) {
1414 assert(false, "SafePointScalarMerge inputs should be null/top or SafePointScalarObject.");
1415 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1416 }
1417 }
1418
1419 // Validate users of merge
1420 for (DUIterator_Fast imax, i = merge->fast_outs(imax); i < imax; i++) {
1421 Node* sfpt = merge->fast_out(i);
1422 if (sfpt->is_SafePoint()) {
1423 int merge_idx = merge->merge_pointer_idx(sfpt->as_SafePoint()->jvms());
1424
1425 if (sfpt->in(merge_idx) != nullptr && sfpt->in(merge_idx)->is_SafePointScalarMerge()) {
1426 assert(false, "SafePointScalarMerge nodes can't be nested.");
1427 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1428 }
1429 } else {
1430 assert(false, "Only safepoints can use SafePointScalarMerge nodes.");
1431 C->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
1432 }
1433 }
1434 }
1435
1436 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1437 Node* m = n->fast_out(i);
1438 ideal_nodes.push(m);
1439 }
1440 }
1441 }
1442
1443 // Returns true if there is an object in the scope of sfn that does not escape globally.
1444 bool ConnectionGraph::has_ea_local_in_scope(SafePointNode* sfn) {
1445 Compile* C = _compile;
1446 for (JVMState* jvms = sfn->jvms(); jvms != nullptr; jvms = jvms->caller()) {
1447 if (C->env()->should_retain_local_variables() || C->env()->jvmti_can_walk_any_space() ||
1448 DeoptimizeObjectsALot) {
1449 // Jvmti agents can access locals. Must provide info about local objects at runtime.
1450 int num_locs = jvms->loc_size();
1451 for (int idx = 0; idx < num_locs; idx++) {
1452 Node* l = sfn->local(jvms, idx);
1453 if (not_global_escape(l)) {
1454 return true;
1455 }
1456 }
1457 }
1458 if (C->env()->jvmti_can_get_owned_monitor_info() ||
1459 C->env()->jvmti_can_walk_any_space() || DeoptimizeObjectsALot) {
1460 // Jvmti agents can read monitors. Must provide info about locked objects at runtime.
1461 int num_mon = jvms->nof_monitors();
1462 for (int idx = 0; idx < num_mon; idx++) {
1463 Node* m = sfn->monitor_obj(jvms, idx);
1464 if (m != nullptr && not_global_escape(m)) {
1465 return true;
1466 }
1467 }
1468 }
1469 }
1470 return false;
1471 }
1472
1473 // Returns true if at least one of the arguments to the call is an object
1474 // that does not escape globally.
1475 bool ConnectionGraph::has_arg_escape(CallJavaNode* call) {
1476 if (call->method() != nullptr) {
1477 uint max_idx = TypeFunc::Parms + call->method()->arg_size();
1478 for (uint idx = TypeFunc::Parms; idx < max_idx; idx++) {
1479 Node* p = call->in(idx);
1480 if (not_global_escape(p)) {
1481 return true;
1482 }
1483 }
1484 } else {
1485 const char* name = call->as_CallStaticJava()->_name;
1486 assert(name != nullptr, "no name");
1487 // no arg escapes through uncommon traps
1488 if (strcmp(name, "uncommon_trap") != 0) {
1489 // process_call_arguments() assumes that all arguments escape globally
1490 const TypeTuple* d = call->tf()->domain_sig();
1491 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1492 const Type* at = d->field_at(i);
1493 if (at->isa_oopptr() != nullptr) {
1494 return true;
1495 }
1496 }
1497 }
1498 }
1499 return false;
1500 }
1501
1502
1503
1504 // Utility function for nodes that load an object
1505 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1506 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1507 // ThreadLocal has RawPtr type.
1508 const Type* t = _igvn->type(n);
1509 if (t->make_ptr() != nullptr) {
1510 Node* adr = n->in(MemNode::Address);
1511 #ifdef ASSERT
1512 if (!adr->is_AddP()) {
1513 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
1514 } else {
1515 assert((ptnode_adr(adr->_idx) == nullptr ||
1516 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
1517 }
1518 #endif
1519 add_local_var_and_edge(n, PointsToNode::NoEscape,
1520 adr, delayed_worklist);
1521 }
1522 }
1523
1524 // Populate Connection Graph with PointsTo nodes and create simple
1525 // connection graph edges.
1526 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
1527 assert(!_verify, "this method should not be called for verification");
1528 PhaseGVN* igvn = _igvn;
1529 uint n_idx = n->_idx;
1530 PointsToNode* n_ptn = ptnode_adr(n_idx);
1531 if (n_ptn != nullptr) {
1532 return; // No need to redefine PointsTo node during first iteration.
1533 }
1534 int opcode = n->Opcode();
1535 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_to_con_graph(this, igvn, delayed_worklist, n, opcode);
1536 if (gc_handled) {
1537 return; // Ignore node if already handled by GC.
1538 }
1539
1540 if (n->is_Call()) {
1541 // Arguments to allocation and locking don't escape.
1542 if (n->is_AbstractLock()) {
1543 // Put Lock and Unlock nodes on IGVN worklist to process them during
1544 // first IGVN optimization when escape information is still available.
1545 record_for_optimizer(n);
1546 } else if (n->is_Allocate()) {
1547 add_call_node(n->as_Call());
1548 record_for_optimizer(n);
1549 } else {
1550 if (n->is_CallStaticJava()) {
1551 const char* name = n->as_CallStaticJava()->_name;
1552 if (name != nullptr && strcmp(name, "uncommon_trap") == 0) {
1553 return; // Skip uncommon traps
1554 }
1555 }
1556 // Don't mark as processed since call's arguments have to be processed.
1557 delayed_worklist->push(n);
1558 // Check if a call returns an object.
1559 if ((n->as_Call()->returns_pointer() &&
1560 n->as_Call()->proj_out_or_null(TypeFunc::Parms) != nullptr) ||
1561 (n->is_CallStaticJava() &&
1562 n->as_CallStaticJava()->is_boxing_method())) {
1563 add_call_node(n->as_Call());
1564 } else if (n->as_Call()->tf()->returns_inline_type_as_fields()) {
1565 bool returns_oop = false;
1566 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !returns_oop; i++) {
1567 ProjNode* pn = n->fast_out(i)->as_Proj();
1568 if (pn->_con >= TypeFunc::Parms && pn->bottom_type()->isa_ptr()) {
1569 returns_oop = true;
1570 }
1571 }
1572 if (returns_oop) {
1573 add_call_node(n->as_Call());
1574 }
1575 }
1576 }
1577 return;
1578 }
1579 // Put this check here to process call arguments since some call nodes
1580 // point to phantom_obj.
1581 if (n_ptn == phantom_obj || n_ptn == null_obj) {
1582 return; // Skip predefined nodes.
1583 }
1584 switch (opcode) {
1585 case Op_AddP: {
1586 Node* base = get_addp_base(n);
1587 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1588 // Field nodes are created for all field types. They are used in
1589 // adjust_scalar_replaceable_state() and split_unique_types().
1590 // Note, non-oop fields will have only base edges in Connection
1591 // Graph because such fields are not used for oop loads and stores.
1592 int offset = address_offset(n, igvn);
1593 add_field(n, PointsToNode::NoEscape, offset);
1594 if (ptn_base == nullptr) {
1595 delayed_worklist->push(n); // Process it later.
1596 } else {
1597 n_ptn = ptnode_adr(n_idx);
1598 add_base(n_ptn->as_Field(), ptn_base);
1599 }
1600 break;
1601 }
1602 case Op_CastX2P: {
1603 map_ideal_node(n, phantom_obj);
1604 break;
1605 }
1606 case Op_InlineType:
1607 case Op_CastPP:
1608 case Op_CheckCastPP:
1609 case Op_EncodeP:
1610 case Op_DecodeN:
1611 case Op_EncodePKlass:
1612 case Op_DecodeNKlass: {
1613 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1614 break;
1615 }
1616 case Op_CMoveP: {
1617 add_local_var(n, PointsToNode::NoEscape);
1618 // Do not add edges during first iteration because some could be
1619 // not defined yet.
1620 delayed_worklist->push(n);
1621 break;
1622 }
1623 case Op_ConP:
1624 case Op_ConN:
1625 case Op_ConNKlass: {
1626 // assume all oop constants globally escape except for null
1627 PointsToNode::EscapeState es;
1628 const Type* t = igvn->type(n);
1629 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
1630 es = PointsToNode::NoEscape;
1631 } else {
1632 es = PointsToNode::GlobalEscape;
1633 }
1634 PointsToNode* ptn_con = add_java_object(n, es);
1635 set_not_scalar_replaceable(ptn_con NOT_PRODUCT(COMMA "Constant pointer"));
1636 break;
1637 }
1638 case Op_CreateEx: {
1639 // assume that all exception objects globally escape
1640 map_ideal_node(n, phantom_obj);
1641 break;
1642 }
1643 case Op_LoadKlass:
1644 case Op_LoadNKlass: {
1645 // Unknown class is loaded
1646 map_ideal_node(n, phantom_obj);
1647 break;
1648 }
1649 case Op_LoadP:
1650 case Op_LoadN: {
1651 add_objload_to_connection_graph(n, delayed_worklist);
1652 break;
1653 }
1654 case Op_Parm: {
1655 map_ideal_node(n, phantom_obj);
1656 break;
1657 }
1658 case Op_PartialSubtypeCheck: {
1659 // Produces Null or notNull and is used in only in CmpP so
1660 // phantom_obj could be used.
1661 map_ideal_node(n, phantom_obj); // Result is unknown
1662 break;
1663 }
1664 case Op_Phi: {
1665 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1666 // ThreadLocal has RawPtr type.
1667 const Type* t = n->as_Phi()->type();
1668 if (t->make_ptr() != nullptr) {
1669 add_local_var(n, PointsToNode::NoEscape);
1670 // Do not add edges during first iteration because some could be
1671 // not defined yet.
1672 delayed_worklist->push(n);
1673 }
1674 break;
1675 }
1676 case Op_Proj: {
1677 // we are only interested in the oop result projection from a call
1678 if (n->as_Proj()->_con >= TypeFunc::Parms && n->in(0)->is_Call() &&
1679 (n->in(0)->as_Call()->returns_pointer() || n->bottom_type()->isa_ptr())) {
1680 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1681 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1682 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), delayed_worklist);
1683 }
1684 break;
1685 }
1686 case Op_Rethrow: // Exception object escapes
1687 case Op_Return: {
1688 if (n->req() > TypeFunc::Parms &&
1689 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
1690 // Treat Return value as LocalVar with GlobalEscape escape state.
1691 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), delayed_worklist);
1692 }
1693 break;
1694 }
1695 case Op_CompareAndExchangeP:
1696 case Op_CompareAndExchangeN:
1697 case Op_GetAndSetP:
1698 case Op_GetAndSetN: {
1699 add_objload_to_connection_graph(n, delayed_worklist);
1700 // fall-through
1701 }
1702 case Op_StoreP:
1703 case Op_StoreN:
1704 case Op_StoreNKlass:
1705 case Op_WeakCompareAndSwapP:
1706 case Op_WeakCompareAndSwapN:
1707 case Op_CompareAndSwapP:
1708 case Op_CompareAndSwapN: {
1709 add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1710 break;
1711 }
1712 case Op_AryEq:
1713 case Op_CountPositives:
1714 case Op_StrComp:
1715 case Op_StrEquals:
1716 case Op_StrIndexOf:
1717 case Op_StrIndexOfChar:
1718 case Op_StrInflatedCopy:
1719 case Op_StrCompressedCopy:
1720 case Op_VectorizedHashCode:
1721 case Op_EncodeISOArray: {
1722 add_local_var(n, PointsToNode::ArgEscape);
1723 delayed_worklist->push(n); // Process it later.
1724 break;
1725 }
1726 case Op_ThreadLocal: {
1727 PointsToNode* ptn_thr = add_java_object(n, PointsToNode::ArgEscape);
1728 set_not_scalar_replaceable(ptn_thr NOT_PRODUCT(COMMA "Constant pointer"));
1729 break;
1730 }
1731 case Op_Blackhole: {
1732 // All blackhole pointer arguments are globally escaping.
1733 // Only do this if there is at least one pointer argument.
1734 // Do not add edges during first iteration because some could be
1735 // not defined yet, defer to final step.
1736 for (uint i = 0; i < n->req(); i++) {
1737 Node* in = n->in(i);
1738 if (in != nullptr) {
1739 const Type* at = _igvn->type(in);
1740 if (!at->isa_ptr()) continue;
1741
1742 add_local_var(n, PointsToNode::GlobalEscape);
1743 delayed_worklist->push(n);
1744 break;
1745 }
1746 }
1747 break;
1748 }
1749 default:
1750 ; // Do nothing for nodes not related to EA.
1751 }
1752 return;
1753 }
1754
1755 // Add final simple edges to graph.
1756 void ConnectionGraph::add_final_edges(Node *n) {
1757 PointsToNode* n_ptn = ptnode_adr(n->_idx);
1758 #ifdef ASSERT
1759 if (_verify && n_ptn->is_JavaObject())
1760 return; // This method does not change graph for JavaObject.
1761 #endif
1762
1763 if (n->is_Call()) {
1764 process_call_arguments(n->as_Call());
1765 return;
1766 }
1767 assert(n->is_Store() || n->is_LoadStore() ||
1768 ((n_ptn != nullptr) && (n_ptn->ideal_node() != nullptr)),
1769 "node should be registered already");
1770 int opcode = n->Opcode();
1771 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->escape_add_final_edges(this, _igvn, n, opcode);
1772 if (gc_handled) {
1773 return; // Ignore node if already handled by GC.
1774 }
1775 switch (opcode) {
1776 case Op_AddP: {
1777 Node* base = get_addp_base(n);
1778 PointsToNode* ptn_base = ptnode_adr(base->_idx);
1779 assert(ptn_base != nullptr, "field's base should be registered");
1780 add_base(n_ptn->as_Field(), ptn_base);
1781 break;
1782 }
1783 case Op_InlineType:
1784 case Op_CastPP:
1785 case Op_CheckCastPP:
1786 case Op_EncodeP:
1787 case Op_DecodeN:
1788 case Op_EncodePKlass:
1789 case Op_DecodeNKlass: {
1790 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), nullptr);
1791 break;
1792 }
1793 case Op_CMoveP: {
1794 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
1795 Node* in = n->in(i);
1796 if (in == nullptr) {
1797 continue; // ignore null
1798 }
1799 Node* uncast_in = in->uncast();
1800 if (uncast_in->is_top() || uncast_in == n) {
1801 continue; // ignore top or inputs which go back this node
1802 }
1803 PointsToNode* ptn = ptnode_adr(in->_idx);
1804 assert(ptn != nullptr, "node should be registered");
1805 add_edge(n_ptn, ptn);
1806 }
1807 break;
1808 }
1809 case Op_LoadP:
1810 case Op_LoadN: {
1811 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1812 // ThreadLocal has RawPtr type.
1813 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1814 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1815 break;
1816 }
1817 case Op_Phi: {
1818 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1819 // ThreadLocal has RawPtr type.
1820 assert(n->as_Phi()->type()->make_ptr() != nullptr, "Unexpected node type");
1821 for (uint i = 1; i < n->req(); i++) {
1822 Node* in = n->in(i);
1823 if (in == nullptr) {
1824 continue; // ignore null
1825 }
1826 Node* uncast_in = in->uncast();
1827 if (uncast_in->is_top() || uncast_in == n) {
1828 continue; // ignore top or inputs which go back this node
1829 }
1830 PointsToNode* ptn = ptnode_adr(in->_idx);
1831 assert(ptn != nullptr, "node should be registered");
1832 add_edge(n_ptn, ptn);
1833 }
1834 break;
1835 }
1836 case Op_Proj: {
1837 // we are only interested in the oop result projection from a call
1838 assert((n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->as_Call()->returns_pointer()) ||
1839 n->in(0)->as_Call()->tf()->returns_inline_type_as_fields(), "what kind of oop return is it?");
1840 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), nullptr);
1841 break;
1842 }
1843 case Op_Rethrow: // Exception object escapes
1844 case Op_Return: {
1845 assert(n->req() > TypeFunc::Parms && _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr(),
1846 "Unexpected node type");
1847 // Treat Return value as LocalVar with GlobalEscape escape state.
1848 add_local_var_and_edge(n, PointsToNode::GlobalEscape, n->in(TypeFunc::Parms), nullptr);
1849 break;
1850 }
1851 case Op_CompareAndExchangeP:
1852 case Op_CompareAndExchangeN:
1853 case Op_GetAndSetP:
1854 case Op_GetAndSetN:{
1855 assert(_igvn->type(n)->make_ptr() != nullptr, "Unexpected node type");
1856 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(MemNode::Address), nullptr);
1857 // fall-through
1858 }
1859 case Op_CompareAndSwapP:
1860 case Op_CompareAndSwapN:
1861 case Op_WeakCompareAndSwapP:
1862 case Op_WeakCompareAndSwapN:
1863 case Op_StoreP:
1864 case Op_StoreN:
1865 case Op_StoreNKlass:{
1866 add_final_edges_unsafe_access(n, opcode);
1867 break;
1868 }
1869 case Op_VectorizedHashCode:
1870 case Op_AryEq:
1871 case Op_CountPositives:
1872 case Op_StrComp:
1873 case Op_StrEquals:
1874 case Op_StrIndexOf:
1875 case Op_StrIndexOfChar:
1876 case Op_StrInflatedCopy:
1877 case Op_StrCompressedCopy:
1878 case Op_EncodeISOArray: {
1879 // char[]/byte[] arrays passed to string intrinsic do not escape but
1880 // they are not scalar replaceable. Adjust escape state for them.
1881 // Start from in(2) edge since in(1) is memory edge.
1882 for (uint i = 2; i < n->req(); i++) {
1883 Node* adr = n->in(i);
1884 const Type* at = _igvn->type(adr);
1885 if (!adr->is_top() && at->isa_ptr()) {
1886 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
1887 at->isa_ptr() != nullptr, "expecting a pointer");
1888 if (adr->is_AddP()) {
1889 adr = get_addp_base(adr);
1890 }
1891 PointsToNode* ptn = ptnode_adr(adr->_idx);
1892 assert(ptn != nullptr, "node should be registered");
1893 add_edge(n_ptn, ptn);
1894 }
1895 }
1896 break;
1897 }
1898 case Op_Blackhole: {
1899 // All blackhole pointer arguments are globally escaping.
1900 for (uint i = 0; i < n->req(); i++) {
1901 Node* in = n->in(i);
1902 if (in != nullptr) {
1903 const Type* at = _igvn->type(in);
1904 if (!at->isa_ptr()) continue;
1905
1906 if (in->is_AddP()) {
1907 in = get_addp_base(in);
1908 }
1909
1910 PointsToNode* ptn = ptnode_adr(in->_idx);
1911 assert(ptn != nullptr, "should be defined already");
1912 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "blackhole"));
1913 add_edge(n_ptn, ptn);
1914 }
1915 }
1916 break;
1917 }
1918 default: {
1919 // This method should be called only for EA specific nodes which may
1920 // miss some edges when they were created.
1921 #ifdef ASSERT
1922 n->dump(1);
1923 #endif
1924 guarantee(false, "unknown node");
1925 }
1926 }
1927 return;
1928 }
1929
1930 void ConnectionGraph::add_to_congraph_unsafe_access(Node* n, uint opcode, Unique_Node_List* delayed_worklist) {
1931 Node* adr = n->in(MemNode::Address);
1932 const Type* adr_type = _igvn->type(adr);
1933 adr_type = adr_type->make_ptr();
1934 if (adr_type == nullptr) {
1935 return; // skip dead nodes
1936 }
1937 if (adr_type->isa_oopptr()
1938 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1939 && adr_type == TypeRawPtr::NOTNULL
1940 && is_captured_store_address(adr))) {
1941 delayed_worklist->push(n); // Process it later.
1942 #ifdef ASSERT
1943 assert (adr->is_AddP(), "expecting an AddP");
1944 if (adr_type == TypeRawPtr::NOTNULL) {
1945 // Verify a raw address for a store captured by Initialize node.
1946 int offs = (int) _igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1947 assert(offs != Type::OffsetBot, "offset must be a constant");
1948 }
1949 #endif
1950 } else {
1951 // Ignore copy the displaced header to the BoxNode (OSR compilation).
1952 if (adr->is_BoxLock()) {
1953 return;
1954 }
1955 // Stored value escapes in unsafe access.
1956 if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1957 delayed_worklist->push(n); // Process unsafe access later.
1958 return;
1959 }
1960 #ifdef ASSERT
1961 n->dump(1);
1962 assert(false, "not unsafe");
1963 #endif
1964 }
1965 }
1966
1967 bool ConnectionGraph::add_final_edges_unsafe_access(Node* n, uint opcode) {
1968 Node* adr = n->in(MemNode::Address);
1969 const Type *adr_type = _igvn->type(adr);
1970 adr_type = adr_type->make_ptr();
1971 #ifdef ASSERT
1972 if (adr_type == nullptr) {
1973 n->dump(1);
1974 assert(adr_type != nullptr, "dead node should not be on list");
1975 return true;
1976 }
1977 #endif
1978
1979 if (adr_type->isa_oopptr()
1980 || ((opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass)
1981 && adr_type == TypeRawPtr::NOTNULL
1982 && is_captured_store_address(adr))) {
1983 // Point Address to Value
1984 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
1985 assert(adr_ptn != nullptr &&
1986 adr_ptn->as_Field()->is_oop(), "node should be registered");
1987 Node* val = n->in(MemNode::ValueIn);
1988 PointsToNode* ptn = ptnode_adr(val->_idx);
1989 assert(ptn != nullptr, "node should be registered");
1990 add_edge(adr_ptn, ptn);
1991 return true;
1992 } else if ((opcode == Op_StoreP) && adr_type->isa_rawptr()) {
1993 // Stored value escapes in unsafe access.
1994 Node* val = n->in(MemNode::ValueIn);
1995 PointsToNode* ptn = ptnode_adr(val->_idx);
1996 assert(ptn != nullptr, "node should be registered");
1997 set_escape_state(ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA "stored at raw address"));
1998 // Add edge to object for unsafe access with offset.
1999 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
2000 assert(adr_ptn != nullptr, "node should be registered");
2001 if (adr_ptn->is_Field()) {
2002 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
2003 add_edge(adr_ptn, ptn);
2004 }
2005 return true;
2006 }
2007 #ifdef ASSERT
2008 n->dump(1);
2009 assert(false, "not unsafe");
2010 #endif
2011 return false;
2012 }
2013
2014 void ConnectionGraph::add_call_node(CallNode* call) {
2015 assert(call->returns_pointer() || call->tf()->returns_inline_type_as_fields(), "only for call which returns pointer");
2016 uint call_idx = call->_idx;
2017 if (call->is_Allocate()) {
2018 Node* k = call->in(AllocateNode::KlassNode);
2019 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
2020 assert(kt != nullptr, "TypeKlassPtr required.");
2021 PointsToNode::EscapeState es = PointsToNode::NoEscape;
2022 bool scalar_replaceable = true;
2023 NOT_PRODUCT(const char* nsr_reason = "");
2024 if (call->is_AllocateArray()) {
2025 if (!kt->isa_aryklassptr()) { // StressReflectiveCode
2026 es = PointsToNode::GlobalEscape;
2027 } else {
2028 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2029 if (length < 0) {
2030 // Not scalar replaceable if the length is not constant.
2031 scalar_replaceable = false;
2032 NOT_PRODUCT(nsr_reason = "has a non-constant length");
2033 } else if (length > EliminateAllocationArraySizeLimit) {
2034 // Not scalar replaceable if the length is too big.
2035 scalar_replaceable = false;
2036 NOT_PRODUCT(nsr_reason = "has a length that is too big");
2037 }
2038 }
2039 } else { // Allocate instance
2040 if (!kt->isa_instklassptr()) { // StressReflectiveCode
2041 es = PointsToNode::GlobalEscape;
2042 } else {
2043 const TypeInstKlassPtr* ikt = kt->is_instklassptr();
2044 ciInstanceKlass* ik = ikt->klass_is_exact() ? ikt->exact_klass()->as_instance_klass() : ikt->instance_klass();
2045 if (ik->is_subclass_of(_compile->env()->Thread_klass()) ||
2046 ik->is_subclass_of(_compile->env()->Reference_klass()) ||
2047 !ik->can_be_instantiated() ||
2048 ik->has_finalizer()) {
2049 es = PointsToNode::GlobalEscape;
2050 } else {
2051 int nfields = ik->as_instance_klass()->nof_nonstatic_fields();
2052 if (nfields > EliminateAllocationFieldsLimit) {
2053 // Not scalar replaceable if there are too many fields.
2054 scalar_replaceable = false;
2055 NOT_PRODUCT(nsr_reason = "has too many fields");
2056 }
2057 }
2058 }
2059 }
2060 add_java_object(call, es);
2061 PointsToNode* ptn = ptnode_adr(call_idx);
2062 if (!scalar_replaceable && ptn->scalar_replaceable()) {
2063 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA nsr_reason));
2064 }
2065 } else if (call->is_CallStaticJava()) {
2066 // Call nodes could be different types:
2067 //
2068 // 1. CallDynamicJavaNode (what happened during call is unknown):
2069 //
2070 // - mapped to GlobalEscape JavaObject node if oop is returned;
2071 //
2072 // - all oop arguments are escaping globally;
2073 //
2074 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
2075 //
2076 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
2077 //
2078 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
2079 // - mapped to NoEscape JavaObject node if non-escaping object allocated
2080 // during call is returned;
2081 // - mapped to ArgEscape LocalVar node pointed to object arguments
2082 // which are returned and does not escape during call;
2083 //
2084 // - oop arguments escaping status is defined by bytecode analysis;
2085 //
2086 // For a static call, we know exactly what method is being called.
2087 // Use bytecode estimator to record whether the call's return value escapes.
2088 ciMethod* meth = call->as_CallJava()->method();
2089 if (meth == nullptr) {
2090 const char* name = call->as_CallStaticJava()->_name;
2091 assert(strncmp(name, "_multianewarray", 15) == 0 ||
2092 strncmp(name, "_load_unknown_inline", 20) == 0, "TODO: add failed case check");
2093 // Returns a newly allocated non-escaped object.
2094 add_java_object(call, PointsToNode::NoEscape);
2095 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of multinewarray"));
2096 } else if (meth->is_boxing_method()) {
2097 // Returns boxing object
2098 PointsToNode::EscapeState es;
2099 vmIntrinsics::ID intr = meth->intrinsic_id();
2100 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
2101 // It does not escape if object is always allocated.
2102 es = PointsToNode::NoEscape;
2103 } else {
2104 // It escapes globally if object could be loaded from cache.
2105 es = PointsToNode::GlobalEscape;
2106 }
2107 add_java_object(call, es);
2108 if (es == PointsToNode::GlobalEscape) {
2109 set_not_scalar_replaceable(ptnode_adr(call->_idx) NOT_PRODUCT(COMMA "object can be loaded from boxing cache"));
2110 }
2111 } else {
2112 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
2113 call_analyzer->copy_dependencies(_compile->dependencies());
2114 if (call_analyzer->is_return_allocated()) {
2115 // Returns a newly allocated non-escaped object, simply
2116 // update dependency information.
2117 // Mark it as NoEscape so that objects referenced by
2118 // it's fields will be marked as NoEscape at least.
2119 add_java_object(call, PointsToNode::NoEscape);
2120 set_not_scalar_replaceable(ptnode_adr(call_idx) NOT_PRODUCT(COMMA "is result of call"));
2121 } else {
2122 // Determine whether any arguments are returned.
2123 const TypeTuple* d = call->tf()->domain_cc();
2124 bool ret_arg = false;
2125 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2126 if (d->field_at(i)->isa_ptr() != nullptr &&
2127 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2128 ret_arg = true;
2129 break;
2130 }
2131 }
2132 if (ret_arg) {
2133 add_local_var(call, PointsToNode::ArgEscape);
2134 } else {
2135 // Returns unknown object.
2136 map_ideal_node(call, phantom_obj);
2137 }
2138 }
2139 }
2140 } else {
2141 // An other type of call, assume the worst case:
2142 // returned value is unknown and globally escapes.
2143 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
2144 map_ideal_node(call, phantom_obj);
2145 }
2146 }
2147
2148 void ConnectionGraph::process_call_arguments(CallNode *call) {
2149 bool is_arraycopy = false;
2150 switch (call->Opcode()) {
2151 #ifdef ASSERT
2152 case Op_Allocate:
2153 case Op_AllocateArray:
2154 case Op_Lock:
2155 case Op_Unlock:
2156 assert(false, "should be done already");
2157 break;
2158 #endif
2159 case Op_ArrayCopy:
2160 case Op_CallLeafNoFP:
2161 // Most array copies are ArrayCopy nodes at this point but there
2162 // are still a few direct calls to the copy subroutines (See
2163 // PhaseStringOpts::copy_string())
2164 is_arraycopy = (call->Opcode() == Op_ArrayCopy) ||
2165 call->as_CallLeaf()->is_call_to_arraycopystub();
2166 // fall through
2167 case Op_CallLeafVector:
2168 case Op_CallLeaf: {
2169 // Stub calls, objects do not escape but they are not scale replaceable.
2170 // Adjust escape state for outgoing arguments.
2171 const TypeTuple * d = call->tf()->domain_sig();
2172 bool src_has_oops = false;
2173 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2174 const Type* at = d->field_at(i);
2175 Node *arg = call->in(i);
2176 if (arg == nullptr) {
2177 continue;
2178 }
2179 const Type *aat = _igvn->type(arg);
2180 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr()) {
2181 continue;
2182 }
2183 if (arg->is_AddP()) {
2184 //
2185 // The inline_native_clone() case when the arraycopy stub is called
2186 // after the allocation before Initialize and CheckCastPP nodes.
2187 // Or normal arraycopy for object arrays case.
2188 //
2189 // Set AddP's base (Allocate) as not scalar replaceable since
2190 // pointer to the base (with offset) is passed as argument.
2191 //
2192 arg = get_addp_base(arg);
2193 }
2194 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2195 assert(arg_ptn != nullptr, "should be registered");
2196 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
2197 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
2198 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2199 aat->isa_ptr() != nullptr, "expecting an Ptr");
2200 bool arg_has_oops = aat->isa_oopptr() &&
2201 (aat->isa_instptr() ||
2202 (aat->isa_aryptr() && (aat->isa_aryptr()->elem() == Type::BOTTOM || aat->isa_aryptr()->elem()->make_oopptr() != nullptr)) ||
2203 (aat->isa_aryptr() && aat->isa_aryptr()->elem() != nullptr &&
2204 aat->isa_aryptr()->is_flat() &&
2205 aat->isa_aryptr()->elem()->inline_klass()->contains_oops()));
2206 if (i == TypeFunc::Parms) {
2207 src_has_oops = arg_has_oops;
2208 }
2209 //
2210 // src or dst could be j.l.Object when other is basic type array:
2211 //
2212 // arraycopy(char[],0,Object*,0,size);
2213 // arraycopy(Object*,0,char[],0,size);
2214 //
2215 // Don't add edges in such cases.
2216 //
2217 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
2218 arg_has_oops && (i > TypeFunc::Parms);
2219 #ifdef ASSERT
2220 if (!(is_arraycopy ||
2221 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(call) ||
2222 (call->as_CallLeaf()->_name != nullptr &&
2223 (strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
2224 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32C") == 0 ||
2225 strcmp(call->as_CallLeaf()->_name, "updateBytesAdler32") == 0 ||
2226 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
2227 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
2228 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
2229 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0 ||
2230 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_encryptAESCrypt") == 0 ||
2231 strcmp(call->as_CallLeaf()->_name, "electronicCodeBook_decryptAESCrypt") == 0 ||
2232 strcmp(call->as_CallLeaf()->_name, "counterMode_AESCrypt") == 0 ||
2233 strcmp(call->as_CallLeaf()->_name, "galoisCounterMode_AESCrypt") == 0 ||
2234 strcmp(call->as_CallLeaf()->_name, "poly1305_processBlocks") == 0 ||
2235 strcmp(call->as_CallLeaf()->_name, "intpoly_montgomeryMult_P256") == 0 ||
2236 strcmp(call->as_CallLeaf()->_name, "intpoly_assign") == 0 ||
2237 strcmp(call->as_CallLeaf()->_name, "ghash_processBlocks") == 0 ||
2238 strcmp(call->as_CallLeaf()->_name, "chacha20Block") == 0 ||
2239 strcmp(call->as_CallLeaf()->_name, "encodeBlock") == 0 ||
2240 strcmp(call->as_CallLeaf()->_name, "decodeBlock") == 0 ||
2241 strcmp(call->as_CallLeaf()->_name, "md5_implCompress") == 0 ||
2242 strcmp(call->as_CallLeaf()->_name, "md5_implCompressMB") == 0 ||
2243 strcmp(call->as_CallLeaf()->_name, "sha1_implCompress") == 0 ||
2244 strcmp(call->as_CallLeaf()->_name, "sha1_implCompressMB") == 0 ||
2245 strcmp(call->as_CallLeaf()->_name, "sha256_implCompress") == 0 ||
2246 strcmp(call->as_CallLeaf()->_name, "sha256_implCompressMB") == 0 ||
2247 strcmp(call->as_CallLeaf()->_name, "sha512_implCompress") == 0 ||
2248 strcmp(call->as_CallLeaf()->_name, "sha512_implCompressMB") == 0 ||
2249 strcmp(call->as_CallLeaf()->_name, "sha3_implCompress") == 0 ||
2250 strcmp(call->as_CallLeaf()->_name, "sha3_implCompressMB") == 0 ||
2251 strcmp(call->as_CallLeaf()->_name, "multiplyToLen") == 0 ||
2252 strcmp(call->as_CallLeaf()->_name, "squareToLen") == 0 ||
2253 strcmp(call->as_CallLeaf()->_name, "mulAdd") == 0 ||
2254 strcmp(call->as_CallLeaf()->_name, "montgomery_multiply") == 0 ||
2255 strcmp(call->as_CallLeaf()->_name, "montgomery_square") == 0 ||
2256 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2257 strcmp(call->as_CallLeaf()->_name, "load_unknown_inline") == 0 ||
2258 strcmp(call->as_CallLeaf()->_name, "store_unknown_inline") == 0 ||
2259 strcmp(call->as_CallLeaf()->_name, "bigIntegerRightShiftWorker") == 0 ||
2260 strcmp(call->as_CallLeaf()->_name, "bigIntegerLeftShiftWorker") == 0 ||
2261 strcmp(call->as_CallLeaf()->_name, "vectorizedMismatch") == 0 ||
2262 strcmp(call->as_CallLeaf()->_name, "stringIndexOf") == 0 ||
2263 strcmp(call->as_CallLeaf()->_name, "arraysort_stub") == 0 ||
2264 strcmp(call->as_CallLeaf()->_name, "array_partition_stub") == 0 ||
2265 strcmp(call->as_CallLeaf()->_name, "get_class_id_intrinsic") == 0 ||
2266 strcmp(call->as_CallLeaf()->_name, "unsafe_setmemory") == 0)
2267 ))) {
2268 call->dump();
2269 fatal("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name);
2270 }
2271 #endif
2272 // Always process arraycopy's destination object since
2273 // we need to add all possible edges to references in
2274 // source object.
2275 if (arg_esc >= PointsToNode::ArgEscape &&
2276 !arg_is_arraycopy_dest) {
2277 continue;
2278 }
2279 PointsToNode::EscapeState es = PointsToNode::ArgEscape;
2280 if (call->is_ArrayCopy()) {
2281 ArrayCopyNode* ac = call->as_ArrayCopy();
2282 if (ac->is_clonebasic() ||
2283 ac->is_arraycopy_validated() ||
2284 ac->is_copyof_validated() ||
2285 ac->is_copyofrange_validated()) {
2286 es = PointsToNode::NoEscape;
2287 }
2288 }
2289 set_escape_state(arg_ptn, es NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2290 if (arg_is_arraycopy_dest) {
2291 Node* src = call->in(TypeFunc::Parms);
2292 if (src->is_AddP()) {
2293 src = get_addp_base(src);
2294 }
2295 PointsToNode* src_ptn = ptnode_adr(src->_idx);
2296 assert(src_ptn != nullptr, "should be registered");
2297 if (arg_ptn != src_ptn) {
2298 // Special arraycopy edge:
2299 // A destination object's field can't have the source object
2300 // as base since objects escape states are not related.
2301 // Only escape state of destination object's fields affects
2302 // escape state of fields in source object.
2303 add_arraycopy(call, es, src_ptn, arg_ptn);
2304 }
2305 }
2306 }
2307 }
2308 break;
2309 }
2310 case Op_CallStaticJava: {
2311 // For a static call, we know exactly what method is being called.
2312 // Use bytecode estimator to record the call's escape affects
2313 #ifdef ASSERT
2314 const char* name = call->as_CallStaticJava()->_name;
2315 assert((name == nullptr || strcmp(name, "uncommon_trap") != 0), "normal calls only");
2316 #endif
2317 ciMethod* meth = call->as_CallJava()->method();
2318 if ((meth != nullptr) && meth->is_boxing_method()) {
2319 break; // Boxing methods do not modify any oops.
2320 }
2321 BCEscapeAnalyzer* call_analyzer = (meth !=nullptr) ? meth->get_bcea() : nullptr;
2322 // fall-through if not a Java method or no analyzer information
2323 if (call_analyzer != nullptr) {
2324 PointsToNode* call_ptn = ptnode_adr(call->_idx);
2325 const TypeTuple* d = call->tf()->domain_cc();
2326 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2327 const Type* at = d->field_at(i);
2328 int k = i - TypeFunc::Parms;
2329 Node* arg = call->in(i);
2330 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
2331 if (at->isa_ptr() != nullptr &&
2332 call_analyzer->is_arg_returned(k)) {
2333 // The call returns arguments.
2334 if (call_ptn != nullptr) { // Is call's result used?
2335 assert(call_ptn->is_LocalVar(), "node should be registered");
2336 assert(arg_ptn != nullptr, "node should be registered");
2337 add_edge(call_ptn, arg_ptn);
2338 }
2339 }
2340 if (at->isa_oopptr() != nullptr &&
2341 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
2342 if (!call_analyzer->is_arg_stack(k)) {
2343 // The argument global escapes
2344 set_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2345 } else {
2346 set_escape_state(arg_ptn, PointsToNode::ArgEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2347 if (!call_analyzer->is_arg_local(k)) {
2348 // The argument itself doesn't escape, but any fields might
2349 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2350 }
2351 }
2352 }
2353 }
2354 if (call_ptn != nullptr && call_ptn->is_LocalVar()) {
2355 // The call returns arguments.
2356 assert(call_ptn->edge_count() > 0, "sanity");
2357 if (!call_analyzer->is_return_local()) {
2358 // Returns also unknown object.
2359 add_edge(call_ptn, phantom_obj);
2360 }
2361 }
2362 break;
2363 }
2364 }
2365 default: {
2366 // Fall-through here if not a Java method or no analyzer information
2367 // or some other type of call, assume the worst case: all arguments
2368 // globally escape.
2369 const TypeTuple* d = call->tf()->domain_cc();
2370 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2371 const Type* at = d->field_at(i);
2372 if (at->isa_oopptr() != nullptr) {
2373 Node* arg = call->in(i);
2374 if (arg->is_AddP()) {
2375 arg = get_addp_base(arg);
2376 }
2377 assert(ptnode_adr(arg->_idx) != nullptr, "should be defined already");
2378 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape NOT_PRODUCT(COMMA trace_arg_escape_message(call)));
2379 }
2380 }
2381 }
2382 }
2383 }
2384
2385
2386 // Finish Graph construction.
2387 bool ConnectionGraph::complete_connection_graph(
2388 GrowableArray<PointsToNode*>& ptnodes_worklist,
2389 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
2390 GrowableArray<JavaObjectNode*>& java_objects_worklist,
2391 GrowableArray<FieldNode*>& oop_fields_worklist) {
2392 // Normally only 1-3 passes needed to build Connection Graph depending
2393 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
2394 // Set limit to 20 to catch situation when something did go wrong and
2395 // bailout Escape Analysis.
2396 // Also limit build time to 20 sec (60 in debug VM), EscapeAnalysisTimeout flag.
2397 #define GRAPH_BUILD_ITER_LIMIT 20
2398
2399 // Propagate GlobalEscape and ArgEscape escape states and check that
2400 // we still have non-escaping objects. The method pushs on _worklist
2401 // Field nodes which reference phantom_object.
2402 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2403 return false; // Nothing to do.
2404 }
2405 // Now propagate references to all JavaObject nodes.
2406 int java_objects_length = java_objects_worklist.length();
2407 elapsedTimer build_time;
2408 build_time.start();
2409 elapsedTimer time;
2410 bool timeout = false;
2411 int new_edges = 1;
2412 int iterations = 0;
2413 do {
2414 while ((new_edges > 0) &&
2415 (iterations++ < GRAPH_BUILD_ITER_LIMIT)) {
2416 double start_time = time.seconds();
2417 time.start();
2418 new_edges = 0;
2419 // Propagate references to phantom_object for nodes pushed on _worklist
2420 // by find_non_escaped_objects() and find_field_value().
2421 new_edges += add_java_object_edges(phantom_obj, false);
2422 for (int next = 0; next < java_objects_length; ++next) {
2423 JavaObjectNode* ptn = java_objects_worklist.at(next);
2424 new_edges += add_java_object_edges(ptn, true);
2425
2426 #define SAMPLE_SIZE 4
2427 if ((next % SAMPLE_SIZE) == 0) {
2428 // Each 4 iterations calculate how much time it will take
2429 // to complete graph construction.
2430 time.stop();
2431 // Poll for requests from shutdown mechanism to quiesce compiler
2432 // because Connection graph construction may take long time.
2433 CompileBroker::maybe_block();
2434 double stop_time = time.seconds();
2435 double time_per_iter = (stop_time - start_time) / (double)SAMPLE_SIZE;
2436 double time_until_end = time_per_iter * (double)(java_objects_length - next);
2437 if ((start_time + time_until_end) >= EscapeAnalysisTimeout) {
2438 timeout = true;
2439 break; // Timeout
2440 }
2441 start_time = stop_time;
2442 time.start();
2443 }
2444 #undef SAMPLE_SIZE
2445
2446 }
2447 if (timeout) break;
2448 if (new_edges > 0) {
2449 // Update escape states on each iteration if graph was updated.
2450 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist)) {
2451 return false; // Nothing to do.
2452 }
2453 }
2454 time.stop();
2455 if (time.seconds() >= EscapeAnalysisTimeout) {
2456 timeout = true;
2457 break;
2458 }
2459 }
2460 if ((iterations < GRAPH_BUILD_ITER_LIMIT) && !timeout) {
2461 time.start();
2462 // Find fields which have unknown value.
2463 int fields_length = oop_fields_worklist.length();
2464 for (int next = 0; next < fields_length; next++) {
2465 FieldNode* field = oop_fields_worklist.at(next);
2466 if (field->edge_count() == 0) {
2467 new_edges += find_field_value(field);
2468 // This code may added new edges to phantom_object.
2469 // Need an other cycle to propagate references to phantom_object.
2470 }
2471 }
2472 time.stop();
2473 if (time.seconds() >= EscapeAnalysisTimeout) {
2474 timeout = true;
2475 break;
2476 }
2477 } else {
2478 new_edges = 0; // Bailout
2479 }
2480 } while (new_edges > 0);
2481
2482 build_time.stop();
2483 _build_time = build_time.seconds();
2484 _build_iterations = iterations;
2485
2486 // Bailout if passed limits.
2487 if ((iterations >= GRAPH_BUILD_ITER_LIMIT) || timeout) {
2488 Compile* C = _compile;
2489 if (C->log() != nullptr) {
2490 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
2491 C->log()->text("%s", timeout ? "time" : "iterations");
2492 C->log()->end_elem(" limit'");
2493 }
2494 assert(ExitEscapeAnalysisOnTimeout, "infinite EA connection graph build during invocation %d (%f sec, %d iterations) with %d nodes and worklist size %d",
2495 _invocation, _build_time, _build_iterations, nodes_size(), ptnodes_worklist.length());
2496 // Possible infinite build_connection_graph loop,
2497 // bailout (no changes to ideal graph were made).
2498 return false;
2499 }
2500
2501 #undef GRAPH_BUILD_ITER_LIMIT
2502
2503 // Find fields initialized by null for non-escaping Allocations.
2504 int non_escaped_length = non_escaped_allocs_worklist.length();
2505 for (int next = 0; next < non_escaped_length; next++) {
2506 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2507 PointsToNode::EscapeState es = ptn->escape_state();
2508 assert(es <= PointsToNode::ArgEscape, "sanity");
2509 if (es == PointsToNode::NoEscape) {
2510 if (find_init_values_null(ptn, _igvn) > 0) {
2511 // Adding references to null object does not change escape states
2512 // since it does not escape. Also no fields are added to null object.
2513 add_java_object_edges(null_obj, false);
2514 }
2515 }
2516 Node* n = ptn->ideal_node();
2517 if (n->is_Allocate()) {
2518 // The object allocated by this Allocate node will never be
2519 // seen by an other thread. Mark it so that when it is
2520 // expanded no MemBarStoreStore is added.
2521 InitializeNode* ini = n->as_Allocate()->initialization();
2522 if (ini != nullptr)
2523 ini->set_does_not_escape();
2524 }
2525 }
2526 return true; // Finished graph construction.
2527 }
2528
2529 // Propagate GlobalEscape and ArgEscape escape states to all nodes
2530 // and check that we still have non-escaping java objects.
2531 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
2532 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist) {
2533 GrowableArray<PointsToNode*> escape_worklist;
2534 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
2535 int ptnodes_length = ptnodes_worklist.length();
2536 for (int next = 0; next < ptnodes_length; ++next) {
2537 PointsToNode* ptn = ptnodes_worklist.at(next);
2538 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
2539 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
2540 escape_worklist.push(ptn);
2541 }
2542 }
2543 // Set escape states to referenced nodes (edges list).
2544 while (escape_worklist.length() > 0) {
2545 PointsToNode* ptn = escape_worklist.pop();
2546 PointsToNode::EscapeState es = ptn->escape_state();
2547 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
2548 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
2549 es >= PointsToNode::ArgEscape) {
2550 // GlobalEscape or ArgEscape state of field means it has unknown value.
2551 if (add_edge(ptn, phantom_obj)) {
2552 // New edge was added
2553 add_field_uses_to_worklist(ptn->as_Field());
2554 }
2555 }
2556 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2557 PointsToNode* e = i.get();
2558 if (e->is_Arraycopy()) {
2559 assert(ptn->arraycopy_dst(), "sanity");
2560 // Propagate only fields escape state through arraycopy edge.
2561 if (e->fields_escape_state() < field_es) {
2562 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2563 escape_worklist.push(e);
2564 }
2565 } else if (es >= field_es) {
2566 // fields_escape_state is also set to 'es' if it is less than 'es'.
2567 if (e->escape_state() < es) {
2568 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2569 escape_worklist.push(e);
2570 }
2571 } else {
2572 // Propagate field escape state.
2573 bool es_changed = false;
2574 if (e->fields_escape_state() < field_es) {
2575 set_fields_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2576 es_changed = true;
2577 }
2578 if ((e->escape_state() < field_es) &&
2579 e->is_Field() && ptn->is_JavaObject() &&
2580 e->as_Field()->is_oop()) {
2581 // Change escape state of referenced fields.
2582 set_escape_state(e, field_es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2583 es_changed = true;
2584 } else if (e->escape_state() < es) {
2585 set_escape_state(e, es NOT_PRODUCT(COMMA trace_propagate_message(ptn)));
2586 es_changed = true;
2587 }
2588 if (es_changed) {
2589 escape_worklist.push(e);
2590 }
2591 }
2592 }
2593 }
2594 // Remove escaped objects from non_escaped list.
2595 for (int next = non_escaped_allocs_worklist.length()-1; next >= 0 ; --next) {
2596 JavaObjectNode* ptn = non_escaped_allocs_worklist.at(next);
2597 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
2598 non_escaped_allocs_worklist.delete_at(next);
2599 }
2600 if (ptn->escape_state() == PointsToNode::NoEscape) {
2601 // Find fields in non-escaped allocations which have unknown value.
2602 find_init_values_phantom(ptn);
2603 }
2604 }
2605 return (non_escaped_allocs_worklist.length() > 0);
2606 }
2607
2608 // Add all references to JavaObject node by walking over all uses.
2609 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
2610 int new_edges = 0;
2611 if (populate_worklist) {
2612 // Populate _worklist by uses of jobj's uses.
2613 for (UseIterator i(jobj); i.has_next(); i.next()) {
2614 PointsToNode* use = i.get();
2615 if (use->is_Arraycopy()) {
2616 continue;
2617 }
2618 add_uses_to_worklist(use);
2619 if (use->is_Field() && use->as_Field()->is_oop()) {
2620 // Put on worklist all field's uses (loads) and
2621 // related field nodes (same base and offset).
2622 add_field_uses_to_worklist(use->as_Field());
2623 }
2624 }
2625 }
2626 for (int l = 0; l < _worklist.length(); l++) {
2627 PointsToNode* use = _worklist.at(l);
2628 if (PointsToNode::is_base_use(use)) {
2629 // Add reference from jobj to field and from field to jobj (field's base).
2630 use = PointsToNode::get_use_node(use)->as_Field();
2631 if (add_base(use->as_Field(), jobj)) {
2632 new_edges++;
2633 }
2634 continue;
2635 }
2636 assert(!use->is_JavaObject(), "sanity");
2637 if (use->is_Arraycopy()) {
2638 if (jobj == null_obj) { // null object does not have field edges
2639 continue;
2640 }
2641 // Added edge from Arraycopy node to arraycopy's source java object
2642 if (add_edge(use, jobj)) {
2643 jobj->set_arraycopy_src();
2644 new_edges++;
2645 }
2646 // and stop here.
2647 continue;
2648 }
2649 if (!add_edge(use, jobj)) {
2650 continue; // No new edge added, there was such edge already.
2651 }
2652 new_edges++;
2653 if (use->is_LocalVar()) {
2654 add_uses_to_worklist(use);
2655 if (use->arraycopy_dst()) {
2656 for (EdgeIterator i(use); i.has_next(); i.next()) {
2657 PointsToNode* e = i.get();
2658 if (e->is_Arraycopy()) {
2659 if (jobj == null_obj) { // null object does not have field edges
2660 continue;
2661 }
2662 // Add edge from arraycopy's destination java object to Arraycopy node.
2663 if (add_edge(jobj, e)) {
2664 new_edges++;
2665 jobj->set_arraycopy_dst();
2666 }
2667 }
2668 }
2669 }
2670 } else {
2671 // Added new edge to stored in field values.
2672 // Put on worklist all field's uses (loads) and
2673 // related field nodes (same base and offset).
2674 add_field_uses_to_worklist(use->as_Field());
2675 }
2676 }
2677 _worklist.clear();
2678 _in_worklist.reset();
2679 return new_edges;
2680 }
2681
2682 // Put on worklist all related field nodes.
2683 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
2684 assert(field->is_oop(), "sanity");
2685 int offset = field->offset();
2686 add_uses_to_worklist(field);
2687 // Loop over all bases of this field and push on worklist Field nodes
2688 // with the same offset and base (since they may reference the same field).
2689 for (BaseIterator i(field); i.has_next(); i.next()) {
2690 PointsToNode* base = i.get();
2691 add_fields_to_worklist(field, base);
2692 // Check if the base was source object of arraycopy and go over arraycopy's
2693 // destination objects since values stored to a field of source object are
2694 // accessible by uses (loads) of fields of destination objects.
2695 if (base->arraycopy_src()) {
2696 for (UseIterator j(base); j.has_next(); j.next()) {
2697 PointsToNode* arycp = j.get();
2698 if (arycp->is_Arraycopy()) {
2699 for (UseIterator k(arycp); k.has_next(); k.next()) {
2700 PointsToNode* abase = k.get();
2701 if (abase->arraycopy_dst() && abase != base) {
2702 // Look for the same arraycopy reference.
2703 add_fields_to_worklist(field, abase);
2704 }
2705 }
2706 }
2707 }
2708 }
2709 }
2710 }
2711
2712 // Put on worklist all related field nodes.
2713 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
2714 int offset = field->offset();
2715 if (base->is_LocalVar()) {
2716 for (UseIterator j(base); j.has_next(); j.next()) {
2717 PointsToNode* f = j.get();
2718 if (PointsToNode::is_base_use(f)) { // Field
2719 f = PointsToNode::get_use_node(f);
2720 if (f == field || !f->as_Field()->is_oop()) {
2721 continue;
2722 }
2723 int offs = f->as_Field()->offset();
2724 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2725 add_to_worklist(f);
2726 }
2727 }
2728 }
2729 } else {
2730 assert(base->is_JavaObject(), "sanity");
2731 if (// Skip phantom_object since it is only used to indicate that
2732 // this field's content globally escapes.
2733 (base != phantom_obj) &&
2734 // null object node does not have fields.
2735 (base != null_obj)) {
2736 for (EdgeIterator i(base); i.has_next(); i.next()) {
2737 PointsToNode* f = i.get();
2738 // Skip arraycopy edge since store to destination object field
2739 // does not update value in source object field.
2740 if (f->is_Arraycopy()) {
2741 assert(base->arraycopy_dst(), "sanity");
2742 continue;
2743 }
2744 if (f == field || !f->as_Field()->is_oop()) {
2745 continue;
2746 }
2747 int offs = f->as_Field()->offset();
2748 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
2749 add_to_worklist(f);
2750 }
2751 }
2752 }
2753 }
2754 }
2755
2756 // Find fields which have unknown value.
2757 int ConnectionGraph::find_field_value(FieldNode* field) {
2758 // Escaped fields should have init value already.
2759 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
2760 int new_edges = 0;
2761 for (BaseIterator i(field); i.has_next(); i.next()) {
2762 PointsToNode* base = i.get();
2763 if (base->is_JavaObject()) {
2764 // Skip Allocate's fields which will be processed later.
2765 if (base->ideal_node()->is_Allocate()) {
2766 return 0;
2767 }
2768 assert(base == null_obj, "only null ptr base expected here");
2769 }
2770 }
2771 if (add_edge(field, phantom_obj)) {
2772 // New edge was added
2773 new_edges++;
2774 add_field_uses_to_worklist(field);
2775 }
2776 return new_edges;
2777 }
2778
2779 // Find fields initializing values for allocations.
2780 int ConnectionGraph::find_init_values_phantom(JavaObjectNode* pta) {
2781 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2782 PointsToNode* init_val = phantom_obj;
2783 Node* alloc = pta->ideal_node();
2784
2785 // Do nothing for Allocate nodes since its fields values are
2786 // "known" unless they are initialized by arraycopy/clone.
2787 if (alloc->is_Allocate() && !pta->arraycopy_dst()) {
2788 if (alloc->as_Allocate()->in(AllocateNode::DefaultValue) != nullptr) {
2789 // Non-flat inline type arrays are initialized with
2790 // the default value instead of null. Handle them here.
2791 init_val = ptnode_adr(alloc->as_Allocate()->in(AllocateNode::DefaultValue)->_idx);
2792 assert(init_val != nullptr, "default value should be registered");
2793 } else {
2794 return 0;
2795 }
2796 }
2797 // Non-escaped allocation returned from Java or runtime call has unknown values in fields.
2798 assert(pta->arraycopy_dst() || alloc->is_CallStaticJava() || init_val != phantom_obj, "sanity");
2799 #ifdef ASSERT
2800 if (alloc->is_CallStaticJava() && alloc->as_CallStaticJava()->method() == nullptr) {
2801 const char* name = alloc->as_CallStaticJava()->_name;
2802 assert(strncmp(name, "_multianewarray", 15) == 0 ||
2803 strncmp(name, "_load_unknown_inline", 20) == 0, "sanity");
2804 }
2805 #endif
2806 // Non-escaped allocation returned from Java or runtime call have unknown values in fields.
2807 int new_edges = 0;
2808 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2809 PointsToNode* field = i.get();
2810 if (field->is_Field() && field->as_Field()->is_oop()) {
2811 if (add_edge(field, init_val)) {
2812 // New edge was added
2813 new_edges++;
2814 add_field_uses_to_worklist(field->as_Field());
2815 }
2816 }
2817 }
2818 return new_edges;
2819 }
2820
2821 // Find fields initializing values for allocations.
2822 int ConnectionGraph::find_init_values_null(JavaObjectNode* pta, PhaseValues* phase) {
2823 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
2824 Node* alloc = pta->ideal_node();
2825 // Do nothing for Call nodes since its fields values are unknown.
2826 if (!alloc->is_Allocate() || alloc->as_Allocate()->in(AllocateNode::DefaultValue) != nullptr) {
2827 return 0;
2828 }
2829 InitializeNode* ini = alloc->as_Allocate()->initialization();
2830 bool visited_bottom_offset = false;
2831 GrowableArray<int> offsets_worklist;
2832 int new_edges = 0;
2833
2834 // Check if an oop field's initializing value is recorded and add
2835 // a corresponding null if field's value if it is not recorded.
2836 // Connection Graph does not record a default initialization by null
2837 // captured by Initialize node.
2838 //
2839 for (EdgeIterator i(pta); i.has_next(); i.next()) {
2840 PointsToNode* field = i.get(); // Field (AddP)
2841 if (!field->is_Field() || !field->as_Field()->is_oop()) {
2842 continue; // Not oop field
2843 }
2844 int offset = field->as_Field()->offset();
2845 if (offset == Type::OffsetBot) {
2846 if (!visited_bottom_offset) {
2847 // OffsetBot is used to reference array's element,
2848 // always add reference to null to all Field nodes since we don't
2849 // known which element is referenced.
2850 if (add_edge(field, null_obj)) {
2851 // New edge was added
2852 new_edges++;
2853 add_field_uses_to_worklist(field->as_Field());
2854 visited_bottom_offset = true;
2855 }
2856 }
2857 } else {
2858 // Check only oop fields.
2859 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
2860 if (adr_type->isa_rawptr()) {
2861 #ifdef ASSERT
2862 // Raw pointers are used for initializing stores so skip it
2863 // since it should be recorded already
2864 Node* base = get_addp_base(field->ideal_node());
2865 assert(adr_type->isa_rawptr() && is_captured_store_address(field->ideal_node()), "unexpected pointer type");
2866 #endif
2867 continue;
2868 }
2869 if (!offsets_worklist.contains(offset)) {
2870 offsets_worklist.append(offset);
2871 Node* value = nullptr;
2872 if (ini != nullptr) {
2873 // StoreP::memory_type() == T_ADDRESS
2874 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
2875 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
2876 // Make sure initializing store has the same type as this AddP.
2877 // This AddP may reference non existing field because it is on a
2878 // dead branch of bimorphic call which is not eliminated yet.
2879 if (store != nullptr && store->is_Store() &&
2880 store->as_Store()->memory_type() == ft) {
2881 value = store->in(MemNode::ValueIn);
2882 #ifdef ASSERT
2883 if (VerifyConnectionGraph) {
2884 // Verify that AddP already points to all objects the value points to.
2885 PointsToNode* val = ptnode_adr(value->_idx);
2886 assert((val != nullptr), "should be processed already");
2887 PointsToNode* missed_obj = nullptr;
2888 if (val->is_JavaObject()) {
2889 if (!field->points_to(val->as_JavaObject())) {
2890 missed_obj = val;
2891 }
2892 } else {
2893 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
2894 tty->print_cr("----------init store has invalid value -----");
2895 store->dump();
2896 val->dump();
2897 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
2898 }
2899 for (EdgeIterator j(val); j.has_next(); j.next()) {
2900 PointsToNode* obj = j.get();
2901 if (obj->is_JavaObject()) {
2902 if (!field->points_to(obj->as_JavaObject())) {
2903 missed_obj = obj;
2904 break;
2905 }
2906 }
2907 }
2908 }
2909 if (missed_obj != nullptr) {
2910 tty->print_cr("----------field---------------------------------");
2911 field->dump();
2912 tty->print_cr("----------missed reference to object------------");
2913 missed_obj->dump();
2914 tty->print_cr("----------object referenced by init store-------");
2915 store->dump();
2916 val->dump();
2917 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
2918 }
2919 }
2920 #endif
2921 } else {
2922 // There could be initializing stores which follow allocation.
2923 // For example, a volatile field store is not collected
2924 // by Initialize node.
2925 //
2926 // Need to check for dependent loads to separate such stores from
2927 // stores which follow loads. For now, add initial value null so
2928 // that compare pointers optimization works correctly.
2929 }
2930 }
2931 if (value == nullptr) {
2932 // A field's initializing value was not recorded. Add null.
2933 if (add_edge(field, null_obj)) {
2934 // New edge was added
2935 new_edges++;
2936 add_field_uses_to_worklist(field->as_Field());
2937 }
2938 }
2939 }
2940 }
2941 }
2942 return new_edges;
2943 }
2944
2945 // Adjust scalar_replaceable state after Connection Graph is built.
2946 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj, Unique_Node_List &reducible_merges) {
2947 // A Phi 'x' is a _candidate_ to be reducible if 'can_reduce_phi(x)'
2948 // returns true. If one of the constraints in this method set 'jobj' to NSR
2949 // then the candidate Phi is discarded. If the Phi has another SR 'jobj' as
2950 // input, 'adjust_scalar_replaceable_state' will eventually be called with
2951 // that other object and the Phi will become a reducible Phi.
2952 // There could be multiple merges involving the same jobj.
2953 Unique_Node_List candidates;
2954
2955 // Search for non-escaping objects which are not scalar replaceable
2956 // and mark them to propagate the state to referenced objects.
2957
2958 for (UseIterator i(jobj); i.has_next(); i.next()) {
2959 PointsToNode* use = i.get();
2960 if (use->is_Arraycopy()) {
2961 continue;
2962 }
2963 if (use->is_Field()) {
2964 FieldNode* field = use->as_Field();
2965 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
2966 // 1. An object is not scalar replaceable if the field into which it is
2967 // stored has unknown offset (stored into unknown element of an array).
2968 if (field->offset() == Type::OffsetBot) {
2969 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored at unknown offset"));
2970 return;
2971 }
2972 for (BaseIterator i(field); i.has_next(); i.next()) {
2973 PointsToNode* base = i.get();
2974 // 2. An object is not scalar replaceable if the field into which it is
2975 // stored has multiple bases one of which is null.
2976 if ((base == null_obj) && (field->base_count() > 1)) {
2977 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with potentially null base"));
2978 return;
2979 }
2980 // 2.5. An object is not scalar replaceable if the field into which it is
2981 // stored has NSR base.
2982 if (!base->scalar_replaceable()) {
2983 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
2984 return;
2985 }
2986 }
2987 }
2988 assert(use->is_Field() || use->is_LocalVar(), "sanity");
2989 // 3. An object is not scalar replaceable if it is merged with other objects
2990 // and we can't remove the merge
2991 for (EdgeIterator j(use); j.has_next(); j.next()) {
2992 PointsToNode* ptn = j.get();
2993 if (ptn->is_JavaObject() && ptn != jobj) {
2994 Node* use_n = use->ideal_node();
2995
2996 // These other local vars may point to multiple objects through a Phi
2997 // In this case we skip them and see if we can reduce the Phi.
2998 if (use_n->is_CastPP() || use_n->is_CheckCastPP()) {
2999 use_n = use_n->in(1);
3000 }
3001
3002 // If it's already a candidate or confirmed reducible merge we can skip verification
3003 if (candidates.member(use_n) || reducible_merges.member(use_n)) {
3004 continue;
3005 }
3006
3007 if (use_n->is_Phi() && can_reduce_phi(use_n->as_Phi())) {
3008 candidates.push(use_n);
3009 } else {
3010 // Mark all objects as NSR if we can't remove the merge
3011 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA trace_merged_message(ptn)));
3012 set_not_scalar_replaceable(ptn NOT_PRODUCT(COMMA trace_merged_message(jobj)));
3013 }
3014 }
3015 }
3016 if (!jobj->scalar_replaceable()) {
3017 return;
3018 }
3019 }
3020
3021 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
3022 if (j.get()->is_Arraycopy()) {
3023 continue;
3024 }
3025
3026 // Non-escaping object node should point only to field nodes.
3027 FieldNode* field = j.get()->as_Field();
3028 int offset = field->as_Field()->offset();
3029
3030 // 4. An object is not scalar replaceable if it has a field with unknown
3031 // offset (array's element is accessed in loop).
3032 if (offset == Type::OffsetBot) {
3033 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "has field with unknown offset"));
3034 return;
3035 }
3036 // 5. Currently an object is not scalar replaceable if a LoadStore node
3037 // access its field since the field value is unknown after it.
3038 //
3039 Node* n = field->ideal_node();
3040
3041 // Test for an unsafe access that was parsed as maybe off heap
3042 // (with a CheckCastPP to raw memory).
3043 assert(n->is_AddP(), "expect an address computation");
3044 if (n->in(AddPNode::Base)->is_top() &&
3045 n->in(AddPNode::Address)->Opcode() == Op_CheckCastPP) {
3046 assert(n->in(AddPNode::Address)->bottom_type()->isa_rawptr(), "raw address so raw cast expected");
3047 assert(_igvn->type(n->in(AddPNode::Address)->in(1))->isa_oopptr(), "cast pattern at unsafe access expected");
3048 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used as base of mixed unsafe access"));
3049 return;
3050 }
3051
3052 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3053 Node* u = n->fast_out(i);
3054 if (u->is_LoadStore() || (u->is_Mem() && u->as_Mem()->is_mismatched_access())) {
3055 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is used in LoadStore or mismatched access"));
3056 return;
3057 }
3058 }
3059
3060 // 6. Or the address may point to more then one object. This may produce
3061 // the false positive result (set not scalar replaceable)
3062 // since the flow-insensitive escape analysis can't separate
3063 // the case when stores overwrite the field's value from the case
3064 // when stores happened on different control branches.
3065 //
3066 // Note: it will disable scalar replacement in some cases:
3067 //
3068 // Point p[] = new Point[1];
3069 // p[0] = new Point(); // Will be not scalar replaced
3070 //
3071 // but it will save us from incorrect optimizations in next cases:
3072 //
3073 // Point p[] = new Point[1];
3074 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
3075 //
3076 if (field->base_count() > 1 && candidates.size() == 0) {
3077 if (has_non_reducible_merge(field, reducible_merges)) {
3078 for (BaseIterator i(field); i.has_next(); i.next()) {
3079 PointsToNode* base = i.get();
3080 // Don't take into account LocalVar nodes which
3081 // may point to only one object which should be also
3082 // this field's base by now.
3083 if (base->is_JavaObject() && base != jobj) {
3084 // Mark all bases.
3085 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "may point to more than one object"));
3086 set_not_scalar_replaceable(base NOT_PRODUCT(COMMA "may point to more than one object"));
3087 }
3088 }
3089
3090 if (!jobj->scalar_replaceable()) {
3091 return;
3092 }
3093 }
3094 }
3095 }
3096
3097 // The candidate is truly a reducible merge only if none of the other
3098 // constraints ruled it as NSR. There could be multiple merges involving the
3099 // same jobj.
3100 assert(jobj->scalar_replaceable(), "sanity");
3101 for (uint i = 0; i < candidates.size(); i++ ) {
3102 Node* candidate = candidates.at(i);
3103 reducible_merges.push(candidate);
3104 }
3105 }
3106
3107 bool ConnectionGraph::has_non_reducible_merge(FieldNode* field, Unique_Node_List& reducible_merges) {
3108 for (BaseIterator i(field); i.has_next(); i.next()) {
3109 Node* base = i.get()->ideal_node();
3110 if (base->is_Phi() && !reducible_merges.member(base)) {
3111 return true;
3112 }
3113 }
3114 return false;
3115 }
3116
3117 // Propagate NSR (Not scalar replaceable) state.
3118 void ConnectionGraph::find_scalar_replaceable_allocs(GrowableArray<JavaObjectNode*>& jobj_worklist) {
3119 int jobj_length = jobj_worklist.length();
3120 bool found_nsr_alloc = true;
3121 while (found_nsr_alloc) {
3122 found_nsr_alloc = false;
3123 for (int next = 0; next < jobj_length; ++next) {
3124 JavaObjectNode* jobj = jobj_worklist.at(next);
3125 for (UseIterator i(jobj); (jobj->scalar_replaceable() && i.has_next()); i.next()) {
3126 PointsToNode* use = i.get();
3127 if (use->is_Field()) {
3128 FieldNode* field = use->as_Field();
3129 assert(field->is_oop() && field->scalar_replaceable(), "sanity");
3130 assert(field->offset() != Type::OffsetBot, "sanity");
3131 for (BaseIterator i(field); i.has_next(); i.next()) {
3132 PointsToNode* base = i.get();
3133 // An object is not scalar replaceable if the field into which
3134 // it is stored has NSR base.
3135 if ((base != null_obj) && !base->scalar_replaceable()) {
3136 set_not_scalar_replaceable(jobj NOT_PRODUCT(COMMA "is stored into field with NSR base"));
3137 found_nsr_alloc = true;
3138 break;
3139 }
3140 }
3141 }
3142 }
3143 }
3144 }
3145 }
3146
3147 #ifdef ASSERT
3148 void ConnectionGraph::verify_connection_graph(
3149 GrowableArray<PointsToNode*>& ptnodes_worklist,
3150 GrowableArray<JavaObjectNode*>& non_escaped_allocs_worklist,
3151 GrowableArray<JavaObjectNode*>& java_objects_worklist,
3152 GrowableArray<Node*>& addp_worklist) {
3153 // Verify that graph is complete - no new edges could be added.
3154 int java_objects_length = java_objects_worklist.length();
3155 int non_escaped_length = non_escaped_allocs_worklist.length();
3156 int new_edges = 0;
3157 for (int next = 0; next < java_objects_length; ++next) {
3158 JavaObjectNode* ptn = java_objects_worklist.at(next);
3159 new_edges += add_java_object_edges(ptn, true);
3160 }
3161 assert(new_edges == 0, "graph was not complete");
3162 // Verify that escape state is final.
3163 int length = non_escaped_allocs_worklist.length();
3164 find_non_escaped_objects(ptnodes_worklist, non_escaped_allocs_worklist);
3165 assert((non_escaped_length == non_escaped_allocs_worklist.length()) &&
3166 (non_escaped_length == length) &&
3167 (_worklist.length() == 0), "escape state was not final");
3168
3169 // Verify fields information.
3170 int addp_length = addp_worklist.length();
3171 for (int next = 0; next < addp_length; ++next ) {
3172 Node* n = addp_worklist.at(next);
3173 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
3174 if (field->is_oop()) {
3175 // Verify that field has all bases
3176 Node* base = get_addp_base(n);
3177 PointsToNode* ptn = ptnode_adr(base->_idx);
3178 if (ptn->is_JavaObject()) {
3179 assert(field->has_base(ptn->as_JavaObject()), "sanity");
3180 } else {
3181 assert(ptn->is_LocalVar(), "sanity");
3182 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3183 PointsToNode* e = i.get();
3184 if (e->is_JavaObject()) {
3185 assert(field->has_base(e->as_JavaObject()), "sanity");
3186 }
3187 }
3188 }
3189 // Verify that all fields have initializing values.
3190 if (field->edge_count() == 0) {
3191 tty->print_cr("----------field does not have references----------");
3192 field->dump();
3193 for (BaseIterator i(field); i.has_next(); i.next()) {
3194 PointsToNode* base = i.get();
3195 tty->print_cr("----------field has next base---------------------");
3196 base->dump();
3197 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
3198 tty->print_cr("----------base has fields-------------------------");
3199 for (EdgeIterator j(base); j.has_next(); j.next()) {
3200 j.get()->dump();
3201 }
3202 tty->print_cr("----------base has references---------------------");
3203 for (UseIterator j(base); j.has_next(); j.next()) {
3204 j.get()->dump();
3205 }
3206 }
3207 }
3208 for (UseIterator i(field); i.has_next(); i.next()) {
3209 i.get()->dump();
3210 }
3211 assert(field->edge_count() > 0, "sanity");
3212 }
3213 }
3214 }
3215 }
3216 #endif
3217
3218 // Optimize ideal graph.
3219 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
3220 GrowableArray<MemBarStoreStoreNode*>& storestore_worklist) {
3221 Compile* C = _compile;
3222 PhaseIterGVN* igvn = _igvn;
3223 if (EliminateLocks) {
3224 // Mark locks before changing ideal graph.
3225 int cnt = C->macro_count();
3226 for (int i = 0; i < cnt; i++) {
3227 Node *n = C->macro_node(i);
3228 if (n->is_AbstractLock()) { // Lock and Unlock nodes
3229 AbstractLockNode* alock = n->as_AbstractLock();
3230 if (!alock->is_non_esc_obj()) {
3231 const Type* obj_type = igvn->type(alock->obj_node());
3232 if (can_eliminate_lock(alock) && !obj_type->is_inlinetypeptr()) {
3233 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
3234 // The lock could be marked eliminated by lock coarsening
3235 // code during first IGVN before EA. Replace coarsened flag
3236 // to eliminate all associated locks/unlocks.
3237 #ifdef ASSERT
3238 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc3");
3239 #endif
3240 alock->set_non_esc_obj();
3241 }
3242 }
3243 }
3244 }
3245 }
3246
3247 if (OptimizePtrCompare) {
3248 for (int i = 0; i < ptr_cmp_worklist.length(); i++) {
3249 Node *n = ptr_cmp_worklist.at(i);
3250 assert(n->Opcode() == Op_CmpN || n->Opcode() == Op_CmpP, "must be");
3251 const TypeInt* tcmp = optimize_ptr_compare(n->in(1), n->in(2));
3252 if (tcmp->singleton()) {
3253 Node* cmp = igvn->makecon(tcmp);
3254 #ifndef PRODUCT
3255 if (PrintOptimizePtrCompare) {
3256 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"));
3257 if (Verbose) {
3258 n->dump(1);
3259 }
3260 }
3261 #endif
3262 igvn->replace_node(n, cmp);
3263 }
3264 }
3265 }
3266
3267 // For MemBarStoreStore nodes added in library_call.cpp, check
3268 // escape status of associated AllocateNode and optimize out
3269 // MemBarStoreStore node if the allocated object never escapes.
3270 for (int i = 0; i < storestore_worklist.length(); i++) {
3271 Node* storestore = storestore_worklist.at(i);
3272 Node* alloc = storestore->in(MemBarNode::Precedent)->in(0);
3273 if (alloc->is_Allocate() && not_global_escape(alloc)) {
3274 if (alloc->in(AllocateNode::InlineType) != nullptr) {
3275 // Non-escaping inline type buffer allocations don't require a membar
3276 storestore->as_MemBar()->remove(_igvn);
3277 } else {
3278 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
3279 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
3280 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
3281 igvn->register_new_node_with_optimizer(mb);
3282 igvn->replace_node(storestore, mb);
3283 }
3284 }
3285 }
3286 }
3287
3288 // Optimize objects compare.
3289 const TypeInt* ConnectionGraph::optimize_ptr_compare(Node* left, Node* right) {
3290 assert(OptimizePtrCompare, "sanity");
3291 const TypeInt* EQ = TypeInt::CC_EQ; // [0] == ZERO
3292 const TypeInt* NE = TypeInt::CC_GT; // [1] == ONE
3293 const TypeInt* UNKNOWN = TypeInt::CC; // [-1, 0,1]
3294
3295 PointsToNode* ptn1 = ptnode_adr(left->_idx);
3296 PointsToNode* ptn2 = ptnode_adr(right->_idx);
3297 JavaObjectNode* jobj1 = unique_java_object(left);
3298 JavaObjectNode* jobj2 = unique_java_object(right);
3299
3300 // The use of this method during allocation merge reduction may cause 'left'
3301 // or 'right' be something (e.g., a Phi) that isn't in the connection graph or
3302 // that doesn't reference an unique java object.
3303 if (ptn1 == nullptr || ptn2 == nullptr ||
3304 jobj1 == nullptr || jobj2 == nullptr) {
3305 return UNKNOWN;
3306 }
3307
3308 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
3309 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
3310
3311 // Check simple cases first.
3312 if (jobj1 != nullptr) {
3313 if (jobj1->escape_state() == PointsToNode::NoEscape) {
3314 if (jobj1 == jobj2) {
3315 // Comparing the same not escaping object.
3316 return EQ;
3317 }
3318 Node* obj = jobj1->ideal_node();
3319 // Comparing not escaping allocation.
3320 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3321 !ptn2->points_to(jobj1)) {
3322 return NE; // This includes nullness check.
3323 }
3324 }
3325 }
3326 if (jobj2 != nullptr) {
3327 if (jobj2->escape_state() == PointsToNode::NoEscape) {
3328 Node* obj = jobj2->ideal_node();
3329 // Comparing not escaping allocation.
3330 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
3331 !ptn1->points_to(jobj2)) {
3332 return NE; // This includes nullness check.
3333 }
3334 }
3335 }
3336 if (jobj1 != nullptr && jobj1 != phantom_obj &&
3337 jobj2 != nullptr && jobj2 != phantom_obj &&
3338 jobj1->ideal_node()->is_Con() &&
3339 jobj2->ideal_node()->is_Con()) {
3340 // Klass or String constants compare. Need to be careful with
3341 // compressed pointers - compare types of ConN and ConP instead of nodes.
3342 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
3343 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
3344 if (t1->make_ptr() == t2->make_ptr()) {
3345 return EQ;
3346 } else {
3347 return NE;
3348 }
3349 }
3350 if (ptn1->meet(ptn2)) {
3351 return UNKNOWN; // Sets are not disjoint
3352 }
3353
3354 // Sets are disjoint.
3355 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
3356 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
3357 bool set1_has_null_ptr = ptn1->points_to(null_obj);
3358 bool set2_has_null_ptr = ptn2->points_to(null_obj);
3359 if ((set1_has_unknown_ptr && set2_has_null_ptr) ||
3360 (set2_has_unknown_ptr && set1_has_null_ptr)) {
3361 // Check nullness of unknown object.
3362 return UNKNOWN;
3363 }
3364
3365 // Disjointness by itself is not sufficient since
3366 // alias analysis is not complete for escaped objects.
3367 // Disjoint sets are definitely unrelated only when
3368 // at least one set has only not escaping allocations.
3369 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
3370 if (ptn1->non_escaping_allocation()) {
3371 return NE;
3372 }
3373 }
3374 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
3375 if (ptn2->non_escaping_allocation()) {
3376 return NE;
3377 }
3378 }
3379 return UNKNOWN;
3380 }
3381
3382 // Connection Graph construction functions.
3383
3384 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
3385 PointsToNode* ptadr = _nodes.at(n->_idx);
3386 if (ptadr != nullptr) {
3387 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
3388 return;
3389 }
3390 Compile* C = _compile;
3391 ptadr = new (C->comp_arena()) LocalVarNode(this, n, es);
3392 map_ideal_node(n, ptadr);
3393 }
3394
3395 PointsToNode* ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
3396 PointsToNode* ptadr = _nodes.at(n->_idx);
3397 if (ptadr != nullptr) {
3398 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
3399 return ptadr;
3400 }
3401 Compile* C = _compile;
3402 ptadr = new (C->comp_arena()) JavaObjectNode(this, n, es);
3403 map_ideal_node(n, ptadr);
3404 return ptadr;
3405 }
3406
3407 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
3408 PointsToNode* ptadr = _nodes.at(n->_idx);
3409 if (ptadr != nullptr) {
3410 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
3411 return;
3412 }
3413 bool unsafe = false;
3414 bool is_oop = is_oop_field(n, offset, &unsafe);
3415 if (unsafe) {
3416 es = PointsToNode::GlobalEscape;
3417 }
3418 Compile* C = _compile;
3419 FieldNode* field = new (C->comp_arena()) FieldNode(this, n, es, offset, is_oop);
3420 map_ideal_node(n, field);
3421 }
3422
3423 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
3424 PointsToNode* src, PointsToNode* dst) {
3425 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
3426 assert((src != null_obj) && (dst != null_obj), "not for ConP null");
3427 PointsToNode* ptadr = _nodes.at(n->_idx);
3428 if (ptadr != nullptr) {
3429 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
3430 return;
3431 }
3432 Compile* C = _compile;
3433 ptadr = new (C->comp_arena()) ArraycopyNode(this, n, es);
3434 map_ideal_node(n, ptadr);
3435 // Add edge from arraycopy node to source object.
3436 (void)add_edge(ptadr, src);
3437 src->set_arraycopy_src();
3438 // Add edge from destination object to arraycopy node.
3439 (void)add_edge(dst, ptadr);
3440 dst->set_arraycopy_dst();
3441 }
3442
3443 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
3444 const Type* adr_type = n->as_AddP()->bottom_type();
3445 int field_offset = adr_type->isa_aryptr() ? adr_type->isa_aryptr()->field_offset().get() : Type::OffsetBot;
3446 BasicType bt = T_INT;
3447 if (offset == Type::OffsetBot && field_offset == Type::OffsetBot) {
3448 // Check only oop fields.
3449 if (!adr_type->isa_aryptr() ||
3450 adr_type->isa_aryptr()->elem() == Type::BOTTOM ||
3451 adr_type->isa_aryptr()->elem()->make_oopptr() != nullptr) {
3452 // OffsetBot is used to reference array's element. Ignore first AddP.
3453 if (find_second_addp(n, n->in(AddPNode::Base)) == nullptr) {
3454 bt = T_OBJECT;
3455 }
3456 }
3457 } else if (offset != oopDesc::klass_offset_in_bytes()) {
3458 if (adr_type->isa_instptr()) {
3459 ciField* field = _compile->alias_type(adr_type->is_ptr())->field();
3460 if (field != nullptr) {
3461 bt = field->layout_type();
3462 } else {
3463 // Check for unsafe oop field access
3464 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3465 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3466 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3467 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3468 bt = T_OBJECT;
3469 (*unsafe) = true;
3470 }
3471 }
3472 } else if (adr_type->isa_aryptr()) {
3473 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3474 // Ignore array length load.
3475 } else if (find_second_addp(n, n->in(AddPNode::Base)) != nullptr) {
3476 // Ignore first AddP.
3477 } else {
3478 const Type* elemtype = adr_type->is_aryptr()->elem();
3479 if (adr_type->is_aryptr()->is_flat() && field_offset != Type::OffsetBot) {
3480 ciInlineKlass* vk = elemtype->inline_klass();
3481 field_offset += vk->first_field_offset();
3482 bt = vk->get_field_by_offset(field_offset, false)->layout_type();
3483 } else {
3484 bt = elemtype->array_element_basic_type();
3485 }
3486 }
3487 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
3488 // Allocation initialization, ThreadLocal field access, unsafe access
3489 if (n->has_out_with(Op_StoreP, Op_LoadP, Op_StoreN, Op_LoadN) ||
3490 n->has_out_with(Op_GetAndSetP, Op_GetAndSetN, Op_CompareAndExchangeP, Op_CompareAndExchangeN) ||
3491 n->has_out_with(Op_CompareAndSwapP, Op_CompareAndSwapN, Op_WeakCompareAndSwapP, Op_WeakCompareAndSwapN) ||
3492 BarrierSet::barrier_set()->barrier_set_c2()->escape_has_out_with_unsafe_object(n)) {
3493 bt = T_OBJECT;
3494 }
3495 }
3496 }
3497 // Note: T_NARROWOOP is not classed as a real reference type
3498 return (is_reference_type(bt) || bt == T_NARROWOOP);
3499 }
3500
3501 // Returns unique pointed java object or null.
3502 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) const {
3503 // If the node was created after the escape computation we can't answer.
3504 uint idx = n->_idx;
3505 if (idx >= nodes_size()) {
3506 return nullptr;
3507 }
3508 PointsToNode* ptn = ptnode_adr(idx);
3509 if (ptn == nullptr) {
3510 return nullptr;
3511 }
3512 if (ptn->is_JavaObject()) {
3513 return ptn->as_JavaObject();
3514 }
3515 assert(ptn->is_LocalVar(), "sanity");
3516 // Check all java objects it points to.
3517 JavaObjectNode* jobj = nullptr;
3518 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3519 PointsToNode* e = i.get();
3520 if (e->is_JavaObject()) {
3521 if (jobj == nullptr) {
3522 jobj = e->as_JavaObject();
3523 } else if (jobj != e) {
3524 return nullptr;
3525 }
3526 }
3527 }
3528 return jobj;
3529 }
3530
3531 // Return true if this node points only to non-escaping allocations.
3532 bool PointsToNode::non_escaping_allocation() {
3533 if (is_JavaObject()) {
3534 Node* n = ideal_node();
3535 if (n->is_Allocate() || n->is_CallStaticJava()) {
3536 return (escape_state() == PointsToNode::NoEscape);
3537 } else {
3538 return false;
3539 }
3540 }
3541 assert(is_LocalVar(), "sanity");
3542 // Check all java objects it points to.
3543 for (EdgeIterator i(this); i.has_next(); i.next()) {
3544 PointsToNode* e = i.get();
3545 if (e->is_JavaObject()) {
3546 Node* n = e->ideal_node();
3547 if ((e->escape_state() != PointsToNode::NoEscape) ||
3548 !(n->is_Allocate() || n->is_CallStaticJava())) {
3549 return false;
3550 }
3551 }
3552 }
3553 return true;
3554 }
3555
3556 // Return true if we know the node does not escape globally.
3557 bool ConnectionGraph::not_global_escape(Node *n) {
3558 assert(!_collecting, "should not call during graph construction");
3559 // If the node was created after the escape computation we can't answer.
3560 uint idx = n->_idx;
3561 if (idx >= nodes_size()) {
3562 return false;
3563 }
3564 PointsToNode* ptn = ptnode_adr(idx);
3565 if (ptn == nullptr) {
3566 return false; // not in congraph (e.g. ConI)
3567 }
3568 PointsToNode::EscapeState es = ptn->escape_state();
3569 // If we have already computed a value, return it.
3570 if (es >= PointsToNode::GlobalEscape) {
3571 return false;
3572 }
3573 if (ptn->is_JavaObject()) {
3574 return true; // (es < PointsToNode::GlobalEscape);
3575 }
3576 assert(ptn->is_LocalVar(), "sanity");
3577 // Check all java objects it points to.
3578 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
3579 if (i.get()->escape_state() >= PointsToNode::GlobalEscape) {
3580 return false;
3581 }
3582 }
3583 return true;
3584 }
3585
3586 // Return true if locked object does not escape globally
3587 // and locked code region (identified by BoxLockNode) is balanced:
3588 // all compiled code paths have corresponding Lock/Unlock pairs.
3589 bool ConnectionGraph::can_eliminate_lock(AbstractLockNode* alock) {
3590 if (alock->is_balanced() && not_global_escape(alock->obj_node())) {
3591 if (EliminateNestedLocks) {
3592 // We can mark whole locking region as Local only when only
3593 // one object is used for locking.
3594 alock->box_node()->as_BoxLock()->set_local();
3595 }
3596 return true;
3597 }
3598 return false;
3599 }
3600
3601 // Helper functions
3602
3603 // Return true if this node points to specified node or nodes it points to.
3604 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
3605 if (is_JavaObject()) {
3606 return (this == ptn);
3607 }
3608 assert(is_LocalVar() || is_Field(), "sanity");
3609 for (EdgeIterator i(this); i.has_next(); i.next()) {
3610 if (i.get() == ptn) {
3611 return true;
3612 }
3613 }
3614 return false;
3615 }
3616
3617 // Return true if one node points to an other.
3618 bool PointsToNode::meet(PointsToNode* ptn) {
3619 if (this == ptn) {
3620 return true;
3621 } else if (ptn->is_JavaObject()) {
3622 return this->points_to(ptn->as_JavaObject());
3623 } else if (this->is_JavaObject()) {
3624 return ptn->points_to(this->as_JavaObject());
3625 }
3626 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
3627 int ptn_count = ptn->edge_count();
3628 for (EdgeIterator i(this); i.has_next(); i.next()) {
3629 PointsToNode* this_e = i.get();
3630 for (int j = 0; j < ptn_count; j++) {
3631 if (this_e == ptn->edge(j)) {
3632 return true;
3633 }
3634 }
3635 }
3636 return false;
3637 }
3638
3639 #ifdef ASSERT
3640 // Return true if bases point to this java object.
3641 bool FieldNode::has_base(JavaObjectNode* jobj) const {
3642 for (BaseIterator i(this); i.has_next(); i.next()) {
3643 if (i.get() == jobj) {
3644 return true;
3645 }
3646 }
3647 return false;
3648 }
3649 #endif
3650
3651 bool ConnectionGraph::is_captured_store_address(Node* addp) {
3652 // Handle simple case first.
3653 assert(_igvn->type(addp)->isa_oopptr() == nullptr, "should be raw access");
3654 if (addp->in(AddPNode::Address)->is_Proj() && addp->in(AddPNode::Address)->in(0)->is_Allocate()) {
3655 return true;
3656 } else if (addp->in(AddPNode::Address)->is_Phi()) {
3657 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
3658 Node* addp_use = addp->fast_out(i);
3659 if (addp_use->is_Store()) {
3660 for (DUIterator_Fast jmax, j = addp_use->fast_outs(jmax); j < jmax; j++) {
3661 if (addp_use->fast_out(j)->is_Initialize()) {
3662 return true;
3663 }
3664 }
3665 }
3666 }
3667 }
3668 return false;
3669 }
3670
3671 int ConnectionGraph::address_offset(Node* adr, PhaseValues* phase) {
3672 const Type *adr_type = phase->type(adr);
3673 if (adr->is_AddP() && adr_type->isa_oopptr() == nullptr && is_captured_store_address(adr)) {
3674 // We are computing a raw address for a store captured by an Initialize
3675 // compute an appropriate address type. AddP cases #3 and #5 (see below).
3676 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
3677 assert(offs != Type::OffsetBot ||
3678 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
3679 "offset must be a constant or it is initialization of array");
3680 return offs;
3681 }
3682 return adr_type->is_ptr()->flat_offset();
3683 }
3684
3685 Node* ConnectionGraph::get_addp_base(Node *addp) {
3686 assert(addp->is_AddP(), "must be AddP");
3687 //
3688 // AddP cases for Base and Address inputs:
3689 // case #1. Direct object's field reference:
3690 // Allocate
3691 // |
3692 // Proj #5 ( oop result )
3693 // |
3694 // CheckCastPP (cast to instance type)
3695 // | |
3696 // AddP ( base == address )
3697 //
3698 // case #2. Indirect object's field reference:
3699 // Phi
3700 // |
3701 // CastPP (cast to instance type)
3702 // | |
3703 // AddP ( base == address )
3704 //
3705 // case #3. Raw object's field reference for Initialize node:
3706 // Allocate
3707 // |
3708 // Proj #5 ( oop result )
3709 // top |
3710 // \ |
3711 // AddP ( base == top )
3712 //
3713 // case #4. Array's element reference:
3714 // {CheckCastPP | CastPP}
3715 // | | |
3716 // | AddP ( array's element offset )
3717 // | |
3718 // AddP ( array's offset )
3719 //
3720 // case #5. Raw object's field reference for arraycopy stub call:
3721 // The inline_native_clone() case when the arraycopy stub is called
3722 // after the allocation before Initialize and CheckCastPP nodes.
3723 // Allocate
3724 // |
3725 // Proj #5 ( oop result )
3726 // | |
3727 // AddP ( base == address )
3728 //
3729 // case #6. Constant Pool, ThreadLocal, CastX2P or
3730 // Raw object's field reference:
3731 // {ConP, ThreadLocal, CastX2P, raw Load}
3732 // top |
3733 // \ |
3734 // AddP ( base == top )
3735 //
3736 // case #7. Klass's field reference.
3737 // LoadKlass
3738 // | |
3739 // AddP ( base == address )
3740 //
3741 // case #8. narrow Klass's field reference.
3742 // LoadNKlass
3743 // |
3744 // DecodeN
3745 // | |
3746 // AddP ( base == address )
3747 //
3748 // case #9. Mixed unsafe access
3749 // {instance}
3750 // |
3751 // CheckCastPP (raw)
3752 // top |
3753 // \ |
3754 // AddP ( base == top )
3755 //
3756 Node *base = addp->in(AddPNode::Base);
3757 if (base->uncast()->is_top()) { // The AddP case #3 and #6 and #9.
3758 base = addp->in(AddPNode::Address);
3759 while (base->is_AddP()) {
3760 // Case #6 (unsafe access) may have several chained AddP nodes.
3761 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
3762 base = base->in(AddPNode::Address);
3763 }
3764 if (base->Opcode() == Op_CheckCastPP &&
3765 base->bottom_type()->isa_rawptr() &&
3766 _igvn->type(base->in(1))->isa_oopptr()) {
3767 base = base->in(1); // Case #9
3768 } else {
3769 Node* uncast_base = base->uncast();
3770 int opcode = uncast_base->Opcode();
3771 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
3772 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
3773 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != nullptr)) ||
3774 is_captured_store_address(addp), "sanity");
3775 }
3776 }
3777 return base;
3778 }
3779
3780 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
3781 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
3782 Node* addp2 = addp->raw_out(0);
3783 if (addp->outcnt() == 1 && addp2->is_AddP() &&
3784 addp2->in(AddPNode::Base) == n &&
3785 addp2->in(AddPNode::Address) == addp) {
3786 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
3787 //
3788 // Find array's offset to push it on worklist first and
3789 // as result process an array's element offset first (pushed second)
3790 // to avoid CastPP for the array's offset.
3791 // Otherwise the inserted CastPP (LocalVar) will point to what
3792 // the AddP (Field) points to. Which would be wrong since
3793 // the algorithm expects the CastPP has the same point as
3794 // as AddP's base CheckCastPP (LocalVar).
3795 //
3796 // ArrayAllocation
3797 // |
3798 // CheckCastPP
3799 // |
3800 // memProj (from ArrayAllocation CheckCastPP)
3801 // | ||
3802 // | || Int (element index)
3803 // | || | ConI (log(element size))
3804 // | || | /
3805 // | || LShift
3806 // | || /
3807 // | AddP (array's element offset)
3808 // | |
3809 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
3810 // | / /
3811 // AddP (array's offset)
3812 // |
3813 // Load/Store (memory operation on array's element)
3814 //
3815 return addp2;
3816 }
3817 return nullptr;
3818 }
3819
3820 //
3821 // Adjust the type and inputs of an AddP which computes the
3822 // address of a field of an instance
3823 //
3824 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
3825 PhaseGVN* igvn = _igvn;
3826 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
3827 assert(base_t != nullptr && base_t->is_known_instance(), "expecting instance oopptr");
3828 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
3829 if (t == nullptr) {
3830 // We are computing a raw address for a store captured by an Initialize
3831 // compute an appropriate address type (cases #3 and #5).
3832 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
3833 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
3834 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
3835 assert(offs != Type::OffsetBot, "offset must be a constant");
3836 if (base_t->isa_aryptr() != nullptr) {
3837 // In the case of a flat inline type array, each field has its
3838 // own slice so we need to extract the field being accessed from
3839 // the address computation
3840 t = base_t->isa_aryptr()->add_field_offset_and_offset(offs)->is_oopptr();
3841 } else {
3842 t = base_t->add_offset(offs)->is_oopptr();
3843 }
3844 }
3845 int inst_id = base_t->instance_id();
3846 assert(!t->is_known_instance() || t->instance_id() == inst_id,
3847 "old type must be non-instance or match new type");
3848
3849 // The type 't' could be subclass of 'base_t'.
3850 // As result t->offset() could be large then base_t's size and it will
3851 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
3852 // constructor verifies correctness of the offset.
3853 //
3854 // It could happened on subclass's branch (from the type profiling
3855 // inlining) which was not eliminated during parsing since the exactness
3856 // of the allocation type was not propagated to the subclass type check.
3857 //
3858 // Or the type 't' could be not related to 'base_t' at all.
3859 // It could happen when CHA type is different from MDO type on a dead path
3860 // (for example, from instanceof check) which is not collapsed during parsing.
3861 //
3862 // Do nothing for such AddP node and don't process its users since
3863 // this code branch will go away.
3864 //
3865 if (!t->is_known_instance() &&
3866 !base_t->maybe_java_subtype_of(t)) {
3867 return false; // bail out
3868 }
3869 const TypePtr* tinst = base_t->add_offset(t->offset());
3870 if (tinst->isa_aryptr() && t->isa_aryptr()) {
3871 // In the case of a flat inline type array, each field has its
3872 // own slice so we need to keep track of the field being accessed.
3873 tinst = tinst->is_aryptr()->with_field_offset(t->is_aryptr()->field_offset().get());
3874 // Keep array properties (not flat/null-free)
3875 tinst = tinst->is_aryptr()->update_properties(t->is_aryptr());
3876 if (tinst == nullptr) {
3877 return false; // Skip dead path with inconsistent properties
3878 }
3879 }
3880
3881 // Do NOT remove the next line: ensure a new alias index is allocated
3882 // for the instance type. Note: C++ will not remove it since the call
3883 // has side effect.
3884 int alias_idx = _compile->get_alias_index(tinst);
3885 igvn->set_type(addp, tinst);
3886 // record the allocation in the node map
3887 set_map(addp, get_map(base->_idx));
3888 // Set addp's Base and Address to 'base'.
3889 Node *abase = addp->in(AddPNode::Base);
3890 Node *adr = addp->in(AddPNode::Address);
3891 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
3892 adr->in(0)->_idx == (uint)inst_id) {
3893 // Skip AddP cases #3 and #5.
3894 } else {
3895 assert(!abase->is_top(), "sanity"); // AddP case #3
3896 if (abase != base) {
3897 igvn->hash_delete(addp);
3898 addp->set_req(AddPNode::Base, base);
3899 if (abase == adr) {
3900 addp->set_req(AddPNode::Address, base);
3901 } else {
3902 // AddP case #4 (adr is array's element offset AddP node)
3903 #ifdef ASSERT
3904 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
3905 assert(adr->is_AddP() && atype != nullptr &&
3906 atype->instance_id() == inst_id, "array's element offset should be processed first");
3907 #endif
3908 }
3909 igvn->hash_insert(addp);
3910 }
3911 }
3912 // Put on IGVN worklist since at least addp's type was changed above.
3913 record_for_optimizer(addp);
3914 return true;
3915 }
3916
3917 //
3918 // Create a new version of orig_phi if necessary. Returns either the newly
3919 // created phi or an existing phi. Sets create_new to indicate whether a new
3920 // phi was created. Cache the last newly created phi in the node map.
3921 //
3922 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
3923 Compile *C = _compile;
3924 PhaseGVN* igvn = _igvn;
3925 new_created = false;
3926 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
3927 // nothing to do if orig_phi is bottom memory or matches alias_idx
3928 if (phi_alias_idx == alias_idx) {
3929 return orig_phi;
3930 }
3931 // Have we recently created a Phi for this alias index?
3932 PhiNode *result = get_map_phi(orig_phi->_idx);
3933 if (result != nullptr && C->get_alias_index(result->adr_type()) == alias_idx) {
3934 return result;
3935 }
3936 // Previous check may fail when the same wide memory Phi was split into Phis
3937 // for different memory slices. Search all Phis for this region.
3938 if (result != nullptr) {
3939 Node* region = orig_phi->in(0);
3940 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3941 Node* phi = region->fast_out(i);
3942 if (phi->is_Phi() &&
3943 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
3944 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
3945 return phi->as_Phi();
3946 }
3947 }
3948 }
3949 if (C->live_nodes() + 2*NodeLimitFudgeFactor > C->max_node_limit()) {
3950 if (C->do_escape_analysis() == true && !C->failing()) {
3951 // Retry compilation without escape analysis.
3952 // If this is the first failure, the sentinel string will "stick"
3953 // to the Compile object, and the C2Compiler will see it and retry.
3954 C->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
3955 }
3956 return nullptr;
3957 }
3958 orig_phi_worklist.append_if_missing(orig_phi);
3959 const TypePtr *atype = C->get_adr_type(alias_idx);
3960 result = PhiNode::make(orig_phi->in(0), nullptr, Type::MEMORY, atype);
3961 C->copy_node_notes_to(result, orig_phi);
3962 igvn->set_type(result, result->bottom_type());
3963 record_for_optimizer(result);
3964 set_map(orig_phi, result);
3965 new_created = true;
3966 return result;
3967 }
3968
3969 //
3970 // Return a new version of Memory Phi "orig_phi" with the inputs having the
3971 // specified alias index.
3972 //
3973 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, uint rec_depth) {
3974 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
3975 Compile *C = _compile;
3976 PhaseGVN* igvn = _igvn;
3977 bool new_phi_created;
3978 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
3979 if (!new_phi_created) {
3980 return result;
3981 }
3982 GrowableArray<PhiNode *> phi_list;
3983 GrowableArray<uint> cur_input;
3984 PhiNode *phi = orig_phi;
3985 uint idx = 1;
3986 bool finished = false;
3987 while(!finished) {
3988 while (idx < phi->req()) {
3989 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, rec_depth + 1);
3990 if (mem != nullptr && mem->is_Phi()) {
3991 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
3992 if (new_phi_created) {
3993 // found an phi for which we created a new split, push current one on worklist and begin
3994 // processing new one
3995 phi_list.push(phi);
3996 cur_input.push(idx);
3997 phi = mem->as_Phi();
3998 result = newphi;
3999 idx = 1;
4000 continue;
4001 } else {
4002 mem = newphi;
4003 }
4004 }
4005 if (C->failing()) {
4006 return nullptr;
4007 }
4008 result->set_req(idx++, mem);
4009 }
4010 #ifdef ASSERT
4011 // verify that the new Phi has an input for each input of the original
4012 assert( phi->req() == result->req(), "must have same number of inputs.");
4013 assert( result->in(0) != nullptr && result->in(0) == phi->in(0), "regions must match");
4014 #endif
4015 // Check if all new phi's inputs have specified alias index.
4016 // Otherwise use old phi.
4017 for (uint i = 1; i < phi->req(); i++) {
4018 Node* in = result->in(i);
4019 assert((phi->in(i) == nullptr) == (in == nullptr), "inputs must correspond.");
4020 }
4021 // we have finished processing a Phi, see if there are any more to do
4022 finished = (phi_list.length() == 0 );
4023 if (!finished) {
4024 phi = phi_list.pop();
4025 idx = cur_input.pop();
4026 PhiNode *prev_result = get_map_phi(phi->_idx);
4027 prev_result->set_req(idx++, result);
4028 result = prev_result;
4029 }
4030 }
4031 return result;
4032 }
4033
4034 //
4035 // The next methods are derived from methods in MemNode.
4036 //
4037 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
4038 Node *mem = mmem;
4039 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
4040 // means an array I have not precisely typed yet. Do not do any
4041 // alias stuff with it any time soon.
4042 if (toop->base() != Type::AnyPtr &&
4043 !(toop->isa_instptr() &&
4044 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
4045 toop->offset() == Type::OffsetBot)) {
4046 mem = mmem->memory_at(alias_idx);
4047 // Update input if it is progress over what we have now
4048 }
4049 return mem;
4050 }
4051
4052 //
4053 // Move memory users to their memory slices.
4054 //
4055 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
4056 Compile* C = _compile;
4057 PhaseGVN* igvn = _igvn;
4058 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
4059 assert(tp != nullptr, "ptr type");
4060 int alias_idx = C->get_alias_index(tp);
4061 int general_idx = C->get_general_index(alias_idx);
4062
4063 // Move users first
4064 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4065 Node* use = n->fast_out(i);
4066 if (use->is_MergeMem()) {
4067 MergeMemNode* mmem = use->as_MergeMem();
4068 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
4069 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
4070 continue; // Nothing to do
4071 }
4072 // Replace previous general reference to mem node.
4073 uint orig_uniq = C->unique();
4074 Node* m = find_inst_mem(n, general_idx, orig_phis);
4075 assert(orig_uniq == C->unique(), "no new nodes");
4076 mmem->set_memory_at(general_idx, m);
4077 --imax;
4078 --i;
4079 } else if (use->is_MemBar()) {
4080 assert(!use->is_Initialize(), "initializing stores should not be moved");
4081 if (use->req() > MemBarNode::Precedent &&
4082 use->in(MemBarNode::Precedent) == n) {
4083 // Don't move related membars.
4084 record_for_optimizer(use);
4085 continue;
4086 }
4087 tp = use->as_MemBar()->adr_type()->isa_ptr();
4088 if ((tp != nullptr && C->get_alias_index(tp) == alias_idx) ||
4089 alias_idx == general_idx) {
4090 continue; // Nothing to do
4091 }
4092 // Move to general memory slice.
4093 uint orig_uniq = C->unique();
4094 Node* m = find_inst_mem(n, general_idx, orig_phis);
4095 assert(orig_uniq == C->unique(), "no new nodes");
4096 igvn->hash_delete(use);
4097 imax -= use->replace_edge(n, m, igvn);
4098 igvn->hash_insert(use);
4099 record_for_optimizer(use);
4100 --i;
4101 #ifdef ASSERT
4102 } else if (use->is_Mem()) {
4103 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
4104 // Don't move related cardmark.
4105 continue;
4106 }
4107 // Memory nodes should have new memory input.
4108 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
4109 assert(tp != nullptr, "ptr type");
4110 int idx = C->get_alias_index(tp);
4111 assert(get_map(use->_idx) != nullptr || idx == alias_idx,
4112 "Following memory nodes should have new memory input or be on the same memory slice");
4113 } else if (use->is_Phi()) {
4114 // Phi nodes should be split and moved already.
4115 tp = use->as_Phi()->adr_type()->isa_ptr();
4116 assert(tp != nullptr, "ptr type");
4117 int idx = C->get_alias_index(tp);
4118 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
4119 } else {
4120 use->dump();
4121 assert(false, "should not be here");
4122 #endif
4123 }
4124 }
4125 }
4126
4127 //
4128 // Search memory chain of "mem" to find a MemNode whose address
4129 // is the specified alias index.
4130 //
4131 #define FIND_INST_MEM_RECURSION_DEPTH_LIMIT 1000
4132 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, uint rec_depth) {
4133 if (rec_depth > FIND_INST_MEM_RECURSION_DEPTH_LIMIT) {
4134 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4135 return nullptr;
4136 }
4137 if (orig_mem == nullptr) {
4138 return orig_mem;
4139 }
4140 Compile* C = _compile;
4141 PhaseGVN* igvn = _igvn;
4142 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
4143 bool is_instance = (toop != nullptr) && toop->is_known_instance();
4144 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
4145 Node *prev = nullptr;
4146 Node *result = orig_mem;
4147 while (prev != result) {
4148 prev = result;
4149 if (result == start_mem) {
4150 break; // hit one of our sentinels
4151 }
4152 if (result->is_Mem()) {
4153 const Type *at = igvn->type(result->in(MemNode::Address));
4154 if (at == Type::TOP) {
4155 break; // Dead
4156 }
4157 assert (at->isa_ptr() != nullptr, "pointer type required.");
4158 int idx = C->get_alias_index(at->is_ptr());
4159 if (idx == alias_idx) {
4160 break; // Found
4161 }
4162 if (!is_instance && (at->isa_oopptr() == nullptr ||
4163 !at->is_oopptr()->is_known_instance())) {
4164 break; // Do not skip store to general memory slice.
4165 }
4166 result = result->in(MemNode::Memory);
4167 }
4168 if (!is_instance) {
4169 continue; // don't search further for non-instance types
4170 }
4171 // skip over a call which does not affect this memory slice
4172 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
4173 Node *proj_in = result->in(0);
4174 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
4175 break; // hit one of our sentinels
4176 } else if (proj_in->is_Call()) {
4177 // ArrayCopy node processed here as well
4178 CallNode *call = proj_in->as_Call();
4179 if (!call->may_modify(toop, igvn)) {
4180 result = call->in(TypeFunc::Memory);
4181 }
4182 } else if (proj_in->is_Initialize()) {
4183 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
4184 // Stop if this is the initialization for the object instance which
4185 // which contains this memory slice, otherwise skip over it.
4186 if (alloc == nullptr || alloc->_idx != (uint)toop->instance_id()) {
4187 result = proj_in->in(TypeFunc::Memory);
4188 }
4189 } else if (proj_in->is_MemBar()) {
4190 // Check if there is an array copy for a clone
4191 // Step over GC barrier when ReduceInitialCardMarks is disabled
4192 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4193 Node* control_proj_ac = bs->step_over_gc_barrier(proj_in->in(0));
4194
4195 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
4196 // Stop if it is a clone
4197 ArrayCopyNode* ac = control_proj_ac->in(0)->as_ArrayCopy();
4198 if (ac->may_modify(toop, igvn)) {
4199 break;
4200 }
4201 }
4202 result = proj_in->in(TypeFunc::Memory);
4203 }
4204 } else if (result->is_MergeMem()) {
4205 MergeMemNode *mmem = result->as_MergeMem();
4206 result = step_through_mergemem(mmem, alias_idx, toop);
4207 if (result == mmem->base_memory()) {
4208 // Didn't find instance memory, search through general slice recursively.
4209 result = mmem->memory_at(C->get_general_index(alias_idx));
4210 result = find_inst_mem(result, alias_idx, orig_phis, rec_depth + 1);
4211 if (C->failing()) {
4212 return nullptr;
4213 }
4214 mmem->set_memory_at(alias_idx, result);
4215 }
4216 } else if (result->is_Phi() &&
4217 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
4218 Node *un = result->as_Phi()->unique_input(igvn);
4219 if (un != nullptr) {
4220 orig_phis.append_if_missing(result->as_Phi());
4221 result = un;
4222 } else {
4223 break;
4224 }
4225 } else if (result->is_ClearArray()) {
4226 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
4227 // Can not bypass initialization of the instance
4228 // we are looking for.
4229 break;
4230 }
4231 // Otherwise skip it (the call updated 'result' value).
4232 } else if (result->Opcode() == Op_SCMemProj) {
4233 Node* mem = result->in(0);
4234 Node* adr = nullptr;
4235 if (mem->is_LoadStore()) {
4236 adr = mem->in(MemNode::Address);
4237 } else {
4238 assert(mem->Opcode() == Op_EncodeISOArray ||
4239 mem->Opcode() == Op_StrCompressedCopy, "sanity");
4240 adr = mem->in(3); // Memory edge corresponds to destination array
4241 }
4242 const Type *at = igvn->type(adr);
4243 if (at != Type::TOP) {
4244 assert(at->isa_ptr() != nullptr, "pointer type required.");
4245 int idx = C->get_alias_index(at->is_ptr());
4246 if (idx == alias_idx) {
4247 // Assert in debug mode
4248 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
4249 break; // In product mode return SCMemProj node
4250 }
4251 }
4252 result = mem->in(MemNode::Memory);
4253 } else if (result->Opcode() == Op_StrInflatedCopy) {
4254 Node* adr = result->in(3); // Memory edge corresponds to destination array
4255 const Type *at = igvn->type(adr);
4256 if (at != Type::TOP) {
4257 assert(at->isa_ptr() != nullptr, "pointer type required.");
4258 int idx = C->get_alias_index(at->is_ptr());
4259 if (idx == alias_idx) {
4260 // Assert in debug mode
4261 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
4262 break; // In product mode return SCMemProj node
4263 }
4264 }
4265 result = result->in(MemNode::Memory);
4266 }
4267 }
4268 if (result->is_Phi()) {
4269 PhiNode *mphi = result->as_Phi();
4270 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
4271 const TypePtr *t = mphi->adr_type();
4272 if (!is_instance) {
4273 // Push all non-instance Phis on the orig_phis worklist to update inputs
4274 // during Phase 4 if needed.
4275 orig_phis.append_if_missing(mphi);
4276 } else if (C->get_alias_index(t) != alias_idx) {
4277 // Create a new Phi with the specified alias index type.
4278 result = split_memory_phi(mphi, alias_idx, orig_phis, rec_depth + 1);
4279 }
4280 }
4281 // the result is either MemNode, PhiNode, InitializeNode.
4282 return result;
4283 }
4284
4285 //
4286 // Convert the types of non-escaped object to instance types where possible,
4287 // propagate the new type information through the graph, and update memory
4288 // edges and MergeMem inputs to reflect the new type.
4289 //
4290 // We start with allocations (and calls which may be allocations) on alloc_worklist.
4291 // The processing is done in 4 phases:
4292 //
4293 // Phase 1: Process possible allocations from alloc_worklist. Create instance
4294 // types for the CheckCastPP for allocations where possible.
4295 // Propagate the new types through users as follows:
4296 // casts and Phi: push users on alloc_worklist
4297 // AddP: cast Base and Address inputs to the instance type
4298 // push any AddP users on alloc_worklist and push any memnode
4299 // users onto memnode_worklist.
4300 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4301 // search the Memory chain for a store with the appropriate type
4302 // address type. If a Phi is found, create a new version with
4303 // the appropriate memory slices from each of the Phi inputs.
4304 // For stores, process the users as follows:
4305 // MemNode: push on memnode_worklist
4306 // MergeMem: push on mergemem_worklist
4307 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
4308 // moving the first node encountered of each instance type to the
4309 // the input corresponding to its alias index.
4310 // appropriate memory slice.
4311 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
4312 //
4313 // In the following example, the CheckCastPP nodes are the cast of allocation
4314 // results and the allocation of node 29 is non-escaped and eligible to be an
4315 // instance type.
4316 //
4317 // We start with:
4318 //
4319 // 7 Parm #memory
4320 // 10 ConI "12"
4321 // 19 CheckCastPP "Foo"
4322 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4323 // 29 CheckCastPP "Foo"
4324 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
4325 //
4326 // 40 StoreP 25 7 20 ... alias_index=4
4327 // 50 StoreP 35 40 30 ... alias_index=4
4328 // 60 StoreP 45 50 20 ... alias_index=4
4329 // 70 LoadP _ 60 30 ... alias_index=4
4330 // 80 Phi 75 50 60 Memory alias_index=4
4331 // 90 LoadP _ 80 30 ... alias_index=4
4332 // 100 LoadP _ 80 20 ... alias_index=4
4333 //
4334 //
4335 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
4336 // and creating a new alias index for node 30. This gives:
4337 //
4338 // 7 Parm #memory
4339 // 10 ConI "12"
4340 // 19 CheckCastPP "Foo"
4341 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4342 // 29 CheckCastPP "Foo" iid=24
4343 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4344 //
4345 // 40 StoreP 25 7 20 ... alias_index=4
4346 // 50 StoreP 35 40 30 ... alias_index=6
4347 // 60 StoreP 45 50 20 ... alias_index=4
4348 // 70 LoadP _ 60 30 ... alias_index=6
4349 // 80 Phi 75 50 60 Memory alias_index=4
4350 // 90 LoadP _ 80 30 ... alias_index=6
4351 // 100 LoadP _ 80 20 ... alias_index=4
4352 //
4353 // In phase 2, new memory inputs are computed for the loads and stores,
4354 // And a new version of the phi is created. In phase 4, the inputs to
4355 // node 80 are updated and then the memory nodes are updated with the
4356 // values computed in phase 2. This results in:
4357 //
4358 // 7 Parm #memory
4359 // 10 ConI "12"
4360 // 19 CheckCastPP "Foo"
4361 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
4362 // 29 CheckCastPP "Foo" iid=24
4363 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
4364 //
4365 // 40 StoreP 25 7 20 ... alias_index=4
4366 // 50 StoreP 35 7 30 ... alias_index=6
4367 // 60 StoreP 45 40 20 ... alias_index=4
4368 // 70 LoadP _ 50 30 ... alias_index=6
4369 // 80 Phi 75 40 60 Memory alias_index=4
4370 // 120 Phi 75 50 50 Memory alias_index=6
4371 // 90 LoadP _ 120 30 ... alias_index=6
4372 // 100 LoadP _ 80 20 ... alias_index=4
4373 //
4374 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist,
4375 GrowableArray<ArrayCopyNode*> &arraycopy_worklist,
4376 GrowableArray<MergeMemNode*> &mergemem_worklist,
4377 Unique_Node_List &reducible_merges) {
4378 DEBUG_ONLY(Unique_Node_List reduced_merges;)
4379 GrowableArray<Node *> memnode_worklist;
4380 GrowableArray<PhiNode *> orig_phis;
4381 PhaseIterGVN *igvn = _igvn;
4382 uint new_index_start = (uint) _compile->num_alias_types();
4383 VectorSet visited;
4384 ideal_nodes.clear(); // Reset for use with set_map/get_map.
4385 uint unique_old = _compile->unique();
4386
4387 // Phase 1: Process possible allocations from alloc_worklist.
4388 // Create instance types for the CheckCastPP for allocations where possible.
4389 //
4390 // (Note: don't forget to change the order of the second AddP node on
4391 // the alloc_worklist if the order of the worklist processing is changed,
4392 // see the comment in find_second_addp().)
4393 //
4394 while (alloc_worklist.length() != 0) {
4395 Node *n = alloc_worklist.pop();
4396 uint ni = n->_idx;
4397 if (n->is_Call()) {
4398 CallNode *alloc = n->as_Call();
4399 // copy escape information to call node
4400 PointsToNode* ptn = ptnode_adr(alloc->_idx);
4401 PointsToNode::EscapeState es = ptn->escape_state();
4402 // We have an allocation or call which returns a Java object,
4403 // see if it is non-escaped.
4404 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable()) {
4405 continue;
4406 }
4407 // Find CheckCastPP for the allocate or for the return value of a call
4408 n = alloc->result_cast();
4409 if (n == nullptr) { // No uses except Initialize node
4410 if (alloc->is_Allocate()) {
4411 // Set the scalar_replaceable flag for allocation
4412 // so it could be eliminated if it has no uses.
4413 alloc->as_Allocate()->_is_scalar_replaceable = true;
4414 }
4415 continue;
4416 }
4417 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
4418 // we could reach here for allocate case if one init is associated with many allocs.
4419 if (alloc->is_Allocate()) {
4420 alloc->as_Allocate()->_is_scalar_replaceable = false;
4421 }
4422 continue;
4423 }
4424
4425 // The inline code for Object.clone() casts the allocation result to
4426 // java.lang.Object and then to the actual type of the allocated
4427 // object. Detect this case and use the second cast.
4428 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
4429 // the allocation result is cast to java.lang.Object and then
4430 // to the actual Array type.
4431 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
4432 && (alloc->is_AllocateArray() ||
4433 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeInstKlassPtr::OBJECT)) {
4434 Node *cast2 = nullptr;
4435 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4436 Node *use = n->fast_out(i);
4437 if (use->is_CheckCastPP()) {
4438 cast2 = use;
4439 break;
4440 }
4441 }
4442 if (cast2 != nullptr) {
4443 n = cast2;
4444 } else {
4445 // Non-scalar replaceable if the allocation type is unknown statically
4446 // (reflection allocation), the object can't be restored during
4447 // deoptimization without precise type.
4448 continue;
4449 }
4450 }
4451
4452 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
4453 if (t == nullptr) {
4454 continue; // not a TypeOopPtr
4455 }
4456 if (!t->klass_is_exact()) {
4457 continue; // not an unique type
4458 }
4459 if (alloc->is_Allocate()) {
4460 // Set the scalar_replaceable flag for allocation
4461 // so it could be eliminated.
4462 alloc->as_Allocate()->_is_scalar_replaceable = true;
4463 }
4464 set_escape_state(ptnode_adr(n->_idx), es NOT_PRODUCT(COMMA trace_propagate_message(ptn))); // CheckCastPP escape state
4465 // in order for an object to be scalar-replaceable, it must be:
4466 // - a direct allocation (not a call returning an object)
4467 // - non-escaping
4468 // - eligible to be a unique type
4469 // - not determined to be ineligible by escape analysis
4470 set_map(alloc, n);
4471 set_map(n, alloc);
4472 const TypeOopPtr* tinst = t->cast_to_instance_id(ni);
4473 igvn->hash_delete(n);
4474 igvn->set_type(n, tinst);
4475 n->raise_bottom_type(tinst);
4476 igvn->hash_insert(n);
4477 record_for_optimizer(n);
4478 // Allocate an alias index for the header fields. Accesses to
4479 // the header emitted during macro expansion wouldn't have
4480 // correct memory state otherwise.
4481 _compile->get_alias_index(tinst->add_offset(oopDesc::mark_offset_in_bytes()));
4482 _compile->get_alias_index(tinst->add_offset(oopDesc::klass_offset_in_bytes()));
4483 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
4484
4485 // First, put on the worklist all Field edges from Connection Graph
4486 // which is more accurate than putting immediate users from Ideal Graph.
4487 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
4488 PointsToNode* tgt = e.get();
4489 if (tgt->is_Arraycopy()) {
4490 continue;
4491 }
4492 Node* use = tgt->ideal_node();
4493 assert(tgt->is_Field() && use->is_AddP(),
4494 "only AddP nodes are Field edges in CG");
4495 if (use->outcnt() > 0) { // Don't process dead nodes
4496 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
4497 if (addp2 != nullptr) {
4498 assert(alloc->is_AllocateArray(),"array allocation was expected");
4499 alloc_worklist.append_if_missing(addp2);
4500 }
4501 alloc_worklist.append_if_missing(use);
4502 }
4503 }
4504
4505 // An allocation may have an Initialize which has raw stores. Scan
4506 // the users of the raw allocation result and push AddP users
4507 // on alloc_worklist.
4508 Node *raw_result = alloc->proj_out_or_null(TypeFunc::Parms);
4509 assert (raw_result != nullptr, "must have an allocation result");
4510 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
4511 Node *use = raw_result->fast_out(i);
4512 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
4513 Node* addp2 = find_second_addp(use, raw_result);
4514 if (addp2 != nullptr) {
4515 assert(alloc->is_AllocateArray(),"array allocation was expected");
4516 alloc_worklist.append_if_missing(addp2);
4517 }
4518 alloc_worklist.append_if_missing(use);
4519 } else if (use->is_MemBar()) {
4520 memnode_worklist.append_if_missing(use);
4521 }
4522 }
4523 }
4524 } else if (n->is_AddP()) {
4525 if (has_reducible_merge_base(n->as_AddP(), reducible_merges)) {
4526 // This AddP will go away when we reduce the the Phi
4527 continue;
4528 }
4529 Node* addp_base = get_addp_base(n);
4530 JavaObjectNode* jobj = unique_java_object(addp_base);
4531 if (jobj == nullptr || jobj == phantom_obj) {
4532 #ifdef ASSERT
4533 ptnode_adr(get_addp_base(n)->_idx)->dump();
4534 ptnode_adr(n->_idx)->dump();
4535 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4536 #endif
4537 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4538 return;
4539 }
4540 Node *base = get_map(jobj->idx()); // CheckCastPP node
4541 if (!split_AddP(n, base)) continue; // wrong type from dead path
4542 } else if (n->is_Phi() ||
4543 n->is_CheckCastPP() ||
4544 n->is_EncodeP() ||
4545 n->is_DecodeN() ||
4546 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
4547 if (visited.test_set(n->_idx)) {
4548 assert(n->is_Phi(), "loops only through Phi's");
4549 continue; // already processed
4550 }
4551 // Reducible Phi's will be removed from the graph after split_unique_types
4552 // finishes. For now we just try to split out the SR inputs of the merge.
4553 Node* parent = n->in(1);
4554 if (reducible_merges.member(n)) {
4555 reduce_phi(n->as_Phi(), alloc_worklist, memnode_worklist);
4556 #ifdef ASSERT
4557 if (VerifyReduceAllocationMerges) {
4558 reduced_merges.push(n);
4559 }
4560 #endif
4561 continue;
4562 } else if (reducible_merges.member(parent)) {
4563 // 'n' is an user of a reducible merge (a Phi). It will be simplified as
4564 // part of reduce_merge.
4565 continue;
4566 }
4567 JavaObjectNode* jobj = unique_java_object(n);
4568 if (jobj == nullptr || jobj == phantom_obj) {
4569 #ifdef ASSERT
4570 ptnode_adr(n->_idx)->dump();
4571 assert(jobj != nullptr && jobj != phantom_obj, "escaped allocation");
4572 #endif
4573 _compile->record_failure(_invocation > 0 ? C2Compiler::retry_no_iterative_escape_analysis() : C2Compiler::retry_no_escape_analysis());
4574 return;
4575 } else {
4576 Node *val = get_map(jobj->idx()); // CheckCastPP node
4577 TypeNode *tn = n->as_Type();
4578 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
4579 assert(tinst != nullptr && tinst->is_known_instance() &&
4580 tinst->instance_id() == jobj->idx() , "instance type expected.");
4581
4582 const Type *tn_type = igvn->type(tn);
4583 const TypeOopPtr *tn_t;
4584 if (tn_type->isa_narrowoop()) {
4585 tn_t = tn_type->make_ptr()->isa_oopptr();
4586 } else {
4587 tn_t = tn_type->isa_oopptr();
4588 }
4589 if (tn_t != nullptr && tinst->maybe_java_subtype_of(tn_t)) {
4590 if (tn_t->isa_aryptr()) {
4591 // Keep array properties (not flat/null-free)
4592 tinst = tinst->is_aryptr()->update_properties(tn_t->is_aryptr());
4593 if (tinst == nullptr) {
4594 continue; // Skip dead path with inconsistent properties
4595 }
4596 }
4597 if (tn_type->isa_narrowoop()) {
4598 tn_type = tinst->make_narrowoop();
4599 } else {
4600 tn_type = tinst;
4601 }
4602 igvn->hash_delete(tn);
4603 igvn->set_type(tn, tn_type);
4604 tn->set_type(tn_type);
4605 igvn->hash_insert(tn);
4606 record_for_optimizer(n);
4607 } else {
4608 assert(tn_type == TypePtr::NULL_PTR ||
4609 (tn_t != nullptr && !tinst->maybe_java_subtype_of(tn_t)),
4610 "unexpected type");
4611 continue; // Skip dead path with different type
4612 }
4613 }
4614 } else {
4615 debug_only(n->dump();)
4616 assert(false, "EA: unexpected node");
4617 continue;
4618 }
4619 // push allocation's users on appropriate worklist
4620 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4621 Node *use = n->fast_out(i);
4622 if (use->is_Mem() && use->in(MemNode::Address) == n) {
4623 // Load/store to instance's field
4624 memnode_worklist.append_if_missing(use);
4625 } else if (use->is_MemBar()) {
4626 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4627 memnode_worklist.append_if_missing(use);
4628 }
4629 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
4630 Node* addp2 = find_second_addp(use, n);
4631 if (addp2 != nullptr) {
4632 alloc_worklist.append_if_missing(addp2);
4633 }
4634 alloc_worklist.append_if_missing(use);
4635 } else if (use->is_Phi() ||
4636 use->is_CheckCastPP() ||
4637 use->is_EncodeNarrowPtr() ||
4638 use->is_DecodeNarrowPtr() ||
4639 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
4640 alloc_worklist.append_if_missing(use);
4641 #ifdef ASSERT
4642 } else if (use->is_Mem()) {
4643 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
4644 } else if (use->is_MergeMem()) {
4645 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4646 } else if (use->is_SafePoint()) {
4647 // Look for MergeMem nodes for calls which reference unique allocation
4648 // (through CheckCastPP nodes) even for debug info.
4649 Node* m = use->in(TypeFunc::Memory);
4650 if (m->is_MergeMem()) {
4651 assert(mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4652 }
4653 } else if (use->Opcode() == Op_EncodeISOArray) {
4654 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4655 // EncodeISOArray overwrites destination array
4656 memnode_worklist.append_if_missing(use);
4657 }
4658 } else if (use->Opcode() == Op_Return) {
4659 // Allocation is referenced by field of returned inline type
4660 assert(_compile->tf()->returns_inline_type_as_fields(), "EA: unexpected reference by ReturnNode");
4661 } else {
4662 uint op = use->Opcode();
4663 if ((op == Op_StrCompressedCopy || op == Op_StrInflatedCopy) &&
4664 (use->in(MemNode::Memory) == n)) {
4665 // They overwrite memory edge corresponding to destination array,
4666 memnode_worklist.append_if_missing(use);
4667 } else if (!(op == Op_CmpP || op == Op_Conv2B ||
4668 op == Op_CastP2X || op == Op_StoreCM ||
4669 op == Op_FastLock || op == Op_AryEq ||
4670 op == Op_StrComp || op == Op_CountPositives ||
4671 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy ||
4672 op == Op_StrEquals || op == Op_VectorizedHashCode ||
4673 op == Op_StrIndexOf || op == Op_StrIndexOfChar ||
4674 op == Op_SubTypeCheck || op == Op_InlineType || op == Op_FlatArrayCheck ||
4675 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use))) {
4676 n->dump();
4677 use->dump();
4678 assert(false, "EA: missing allocation reference path");
4679 }
4680 #endif
4681 }
4682 }
4683
4684 }
4685
4686 #ifdef ASSERT
4687 if (VerifyReduceAllocationMerges) {
4688 for (uint i = 0; i < reducible_merges.size(); i++) {
4689 Node* phi = reducible_merges.at(i);
4690
4691 if (!reduced_merges.member(phi)) {
4692 phi->dump(2);
4693 phi->dump(-2);
4694 assert(false, "This reducible merge wasn't reduced.");
4695 }
4696
4697 // At this point reducible Phis shouldn't have AddP users anymore; only SafePoints or Casts.
4698 for (DUIterator_Fast jmax, j = phi->fast_outs(jmax); j < jmax; j++) {
4699 Node* use = phi->fast_out(j);
4700 if (!use->is_SafePoint() && !use->is_CastPP()) {
4701 phi->dump(2);
4702 phi->dump(-2);
4703 assert(false, "Unexpected user of reducible Phi -> %d:%s:%d", use->_idx, use->Name(), use->outcnt());
4704 }
4705 }
4706 }
4707 }
4708 #endif
4709
4710 // Go over all ArrayCopy nodes and if one of the inputs has a unique
4711 // type, record it in the ArrayCopy node so we know what memory this
4712 // node uses/modified.
4713 for (int next = 0; next < arraycopy_worklist.length(); next++) {
4714 ArrayCopyNode* ac = arraycopy_worklist.at(next);
4715 Node* dest = ac->in(ArrayCopyNode::Dest);
4716 if (dest->is_AddP()) {
4717 dest = get_addp_base(dest);
4718 }
4719 JavaObjectNode* jobj = unique_java_object(dest);
4720 if (jobj != nullptr) {
4721 Node *base = get_map(jobj->idx());
4722 if (base != nullptr) {
4723 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4724 ac->_dest_type = base_t;
4725 }
4726 }
4727 Node* src = ac->in(ArrayCopyNode::Src);
4728 if (src->is_AddP()) {
4729 src = get_addp_base(src);
4730 }
4731 jobj = unique_java_object(src);
4732 if (jobj != nullptr) {
4733 Node* base = get_map(jobj->idx());
4734 if (base != nullptr) {
4735 const TypeOopPtr *base_t = _igvn->type(base)->isa_oopptr();
4736 ac->_src_type = base_t;
4737 }
4738 }
4739 }
4740
4741 // New alias types were created in split_AddP().
4742 uint new_index_end = (uint) _compile->num_alias_types();
4743
4744 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
4745 // compute new values for Memory inputs (the Memory inputs are not
4746 // actually updated until phase 4.)
4747 if (memnode_worklist.length() == 0)
4748 return; // nothing to do
4749 while (memnode_worklist.length() != 0) {
4750 Node *n = memnode_worklist.pop();
4751 if (visited.test_set(n->_idx)) {
4752 continue;
4753 }
4754 if (n->is_Phi() || n->is_ClearArray()) {
4755 // we don't need to do anything, but the users must be pushed
4756 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
4757 // we don't need to do anything, but the users must be pushed
4758 n = n->as_MemBar()->proj_out_or_null(TypeFunc::Memory);
4759 if (n == nullptr) {
4760 continue;
4761 }
4762 } else if (n->is_CallLeaf()) {
4763 // Runtime calls with narrow memory input (no MergeMem node)
4764 // get the memory projection
4765 n = n->as_Call()->proj_out_or_null(TypeFunc::Memory);
4766 if (n == nullptr) {
4767 continue;
4768 }
4769 } else if (n->Opcode() == Op_StrCompressedCopy ||
4770 n->Opcode() == Op_EncodeISOArray) {
4771 // get the memory projection
4772 n = n->find_out_with(Op_SCMemProj);
4773 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4774 } else if (n->is_CallLeaf() && n->as_CallLeaf()->_name != nullptr &&
4775 strcmp(n->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
4776 n = n->as_CallLeaf()->proj_out(TypeFunc::Memory);
4777 } else {
4778 assert(n->is_Mem(), "memory node required.");
4779 Node *addr = n->in(MemNode::Address);
4780 const Type *addr_t = igvn->type(addr);
4781 if (addr_t == Type::TOP) {
4782 continue;
4783 }
4784 assert (addr_t->isa_ptr() != nullptr, "pointer type required.");
4785 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
4786 assert ((uint)alias_idx < new_index_end, "wrong alias index");
4787 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
4788 if (_compile->failing()) {
4789 return;
4790 }
4791 if (mem != n->in(MemNode::Memory)) {
4792 // We delay the memory edge update since we need old one in
4793 // MergeMem code below when instances memory slices are separated.
4794 set_map(n, mem);
4795 }
4796 if (n->is_Load()) {
4797 continue; // don't push users
4798 } else if (n->is_LoadStore()) {
4799 // get the memory projection
4800 n = n->find_out_with(Op_SCMemProj);
4801 assert(n != nullptr && n->Opcode() == Op_SCMemProj, "memory projection required");
4802 }
4803 }
4804 // push user on appropriate worklist
4805 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4806 Node *use = n->fast_out(i);
4807 if (use->is_Phi() || use->is_ClearArray()) {
4808 memnode_worklist.append_if_missing(use);
4809 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
4810 if (use->Opcode() == Op_StoreCM) { // Ignore cardmark stores
4811 continue;
4812 }
4813 memnode_worklist.append_if_missing(use);
4814 } else if (use->is_MemBar() || use->is_CallLeaf()) {
4815 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
4816 memnode_worklist.append_if_missing(use);
4817 }
4818 #ifdef ASSERT
4819 } else if (use->is_Mem()) {
4820 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
4821 } else if (use->is_MergeMem()) {
4822 assert(mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
4823 } else if (use->Opcode() == Op_EncodeISOArray) {
4824 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
4825 // EncodeISOArray overwrites destination array
4826 memnode_worklist.append_if_missing(use);
4827 }
4828 } else if (use->is_CallLeaf() && use->as_CallLeaf()->_name != nullptr &&
4829 strcmp(use->as_CallLeaf()->_name, "store_unknown_inline") == 0) {
4830 // store_unknown_inline overwrites destination array
4831 memnode_worklist.append_if_missing(use);
4832 } else {
4833 uint op = use->Opcode();
4834 if ((use->in(MemNode::Memory) == n) &&
4835 (op == Op_StrCompressedCopy || op == Op_StrInflatedCopy)) {
4836 // They overwrite memory edge corresponding to destination array,
4837 memnode_worklist.append_if_missing(use);
4838 } else if (!(BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(use) ||
4839 op == Op_AryEq || op == Op_StrComp || op == Op_CountPositives ||
4840 op == Op_StrCompressedCopy || op == Op_StrInflatedCopy || op == Op_VectorizedHashCode ||
4841 op == Op_StrEquals || op == Op_StrIndexOf || op == Op_StrIndexOfChar || op == Op_FlatArrayCheck)) {
4842 n->dump();
4843 use->dump();
4844 assert(false, "EA: missing memory path");
4845 }
4846 #endif
4847 }
4848 }
4849 }
4850
4851 // Phase 3: Process MergeMem nodes from mergemem_worklist.
4852 // Walk each memory slice moving the first node encountered of each
4853 // instance type to the input corresponding to its alias index.
4854 uint length = mergemem_worklist.length();
4855 for( uint next = 0; next < length; ++next ) {
4856 MergeMemNode* nmm = mergemem_worklist.at(next);
4857 assert(!visited.test_set(nmm->_idx), "should not be visited before");
4858 // Note: we don't want to use MergeMemStream here because we only want to
4859 // scan inputs which exist at the start, not ones we add during processing.
4860 // Note 2: MergeMem may already contains instance memory slices added
4861 // during find_inst_mem() call when memory nodes were processed above.
4862 igvn->hash_delete(nmm);
4863 uint nslices = MIN2(nmm->req(), new_index_start);
4864 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
4865 Node* mem = nmm->in(i);
4866 Node* cur = nullptr;
4867 if (mem == nullptr || mem->is_top()) {
4868 continue;
4869 }
4870 // First, update mergemem by moving memory nodes to corresponding slices
4871 // if their type became more precise since this mergemem was created.
4872 while (mem->is_Mem()) {
4873 const Type *at = igvn->type(mem->in(MemNode::Address));
4874 if (at != Type::TOP) {
4875 assert (at->isa_ptr() != nullptr, "pointer type required.");
4876 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
4877 if (idx == i) {
4878 if (cur == nullptr) {
4879 cur = mem;
4880 }
4881 } else {
4882 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
4883 nmm->set_memory_at(idx, mem);
4884 }
4885 }
4886 }
4887 mem = mem->in(MemNode::Memory);
4888 }
4889 nmm->set_memory_at(i, (cur != nullptr) ? cur : mem);
4890 // Find any instance of the current type if we haven't encountered
4891 // already a memory slice of the instance along the memory chain.
4892 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4893 if((uint)_compile->get_general_index(ni) == i) {
4894 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
4895 if (nmm->is_empty_memory(m)) {
4896 Node* result = find_inst_mem(mem, ni, orig_phis);
4897 if (_compile->failing()) {
4898 return;
4899 }
4900 nmm->set_memory_at(ni, result);
4901 }
4902 }
4903 }
4904 }
4905 // Find the rest of instances values
4906 for (uint ni = new_index_start; ni < new_index_end; ni++) {
4907 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
4908 Node* result = step_through_mergemem(nmm, ni, tinst);
4909 if (result == nmm->base_memory()) {
4910 // Didn't find instance memory, search through general slice recursively.
4911 result = nmm->memory_at(_compile->get_general_index(ni));
4912 result = find_inst_mem(result, ni, orig_phis);
4913 if (_compile->failing()) {
4914 return;
4915 }
4916 nmm->set_memory_at(ni, result);
4917 }
4918 }
4919
4920 // If we have crossed the 3/4 point of max node limit it's too risky
4921 // to continue with EA/SR because we might hit the max node limit.
4922 if (_compile->live_nodes() >= _compile->max_node_limit() * 0.75) {
4923 if (_compile->do_reduce_allocation_merges()) {
4924 _compile->record_failure(C2Compiler::retry_no_reduce_allocation_merges());
4925 } else if (_invocation > 0) {
4926 _compile->record_failure(C2Compiler::retry_no_iterative_escape_analysis());
4927 } else {
4928 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
4929 }
4930 return;
4931 }
4932
4933 igvn->hash_insert(nmm);
4934 record_for_optimizer(nmm);
4935 }
4936
4937 // Phase 4: Update the inputs of non-instance memory Phis and
4938 // the Memory input of memnodes
4939 // First update the inputs of any non-instance Phi's from
4940 // which we split out an instance Phi. Note we don't have
4941 // to recursively process Phi's encountered on the input memory
4942 // chains as is done in split_memory_phi() since they will
4943 // also be processed here.
4944 for (int j = 0; j < orig_phis.length(); j++) {
4945 PhiNode *phi = orig_phis.at(j);
4946 int alias_idx = _compile->get_alias_index(phi->adr_type());
4947 igvn->hash_delete(phi);
4948 for (uint i = 1; i < phi->req(); i++) {
4949 Node *mem = phi->in(i);
4950 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
4951 if (_compile->failing()) {
4952 return;
4953 }
4954 if (mem != new_mem) {
4955 phi->set_req(i, new_mem);
4956 }
4957 }
4958 igvn->hash_insert(phi);
4959 record_for_optimizer(phi);
4960 }
4961
4962 // Update the memory inputs of MemNodes with the value we computed
4963 // in Phase 2 and move stores memory users to corresponding memory slices.
4964 // Disable memory split verification code until the fix for 6984348.
4965 // Currently it produces false negative results since it does not cover all cases.
4966 #if 0 // ifdef ASSERT
4967 visited.Reset();
4968 Node_Stack old_mems(arena, _compile->unique() >> 2);
4969 #endif
4970 for (uint i = 0; i < ideal_nodes.size(); i++) {
4971 Node* n = ideal_nodes.at(i);
4972 Node* nmem = get_map(n->_idx);
4973 assert(nmem != nullptr, "sanity");
4974 if (n->is_Mem()) {
4975 #if 0 // ifdef ASSERT
4976 Node* old_mem = n->in(MemNode::Memory);
4977 if (!visited.test_set(old_mem->_idx)) {
4978 old_mems.push(old_mem, old_mem->outcnt());
4979 }
4980 #endif
4981 assert(n->in(MemNode::Memory) != nmem, "sanity");
4982 if (!n->is_Load()) {
4983 // Move memory users of a store first.
4984 move_inst_mem(n, orig_phis);
4985 }
4986 // Now update memory input
4987 igvn->hash_delete(n);
4988 n->set_req(MemNode::Memory, nmem);
4989 igvn->hash_insert(n);
4990 record_for_optimizer(n);
4991 } else {
4992 assert(n->is_Allocate() || n->is_CheckCastPP() ||
4993 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
4994 }
4995 }
4996 #if 0 // ifdef ASSERT
4997 // Verify that memory was split correctly
4998 while (old_mems.is_nonempty()) {
4999 Node* old_mem = old_mems.node();
5000 uint old_cnt = old_mems.index();
5001 old_mems.pop();
5002 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
5003 }
5004 #endif
5005 }
5006
5007 #ifndef PRODUCT
5008 int ConnectionGraph::_no_escape_counter = 0;
5009 int ConnectionGraph::_arg_escape_counter = 0;
5010 int ConnectionGraph::_global_escape_counter = 0;
5011
5012 static const char *node_type_names[] = {
5013 "UnknownType",
5014 "JavaObject",
5015 "LocalVar",
5016 "Field",
5017 "Arraycopy"
5018 };
5019
5020 static const char *esc_names[] = {
5021 "UnknownEscape",
5022 "NoEscape",
5023 "ArgEscape",
5024 "GlobalEscape"
5025 };
5026
5027 void PointsToNode::dump_header(bool print_state, outputStream* out) const {
5028 NodeType nt = node_type();
5029 out->print("%s(%d) ", node_type_names[(int) nt], _pidx);
5030 if (print_state) {
5031 EscapeState es = escape_state();
5032 EscapeState fields_es = fields_escape_state();
5033 out->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
5034 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable()) {
5035 out->print("NSR ");
5036 }
5037 }
5038 }
5039
5040 void PointsToNode::dump(bool print_state, outputStream* out, bool newline) const {
5041 dump_header(print_state, out);
5042 if (is_Field()) {
5043 FieldNode* f = (FieldNode*)this;
5044 if (f->is_oop()) {
5045 out->print("oop ");
5046 }
5047 if (f->offset() > 0) {
5048 out->print("+%d ", f->offset());
5049 }
5050 out->print("(");
5051 for (BaseIterator i(f); i.has_next(); i.next()) {
5052 PointsToNode* b = i.get();
5053 out->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
5054 }
5055 out->print(" )");
5056 }
5057 out->print("[");
5058 for (EdgeIterator i(this); i.has_next(); i.next()) {
5059 PointsToNode* e = i.get();
5060 out->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
5061 }
5062 out->print(" [");
5063 for (UseIterator i(this); i.has_next(); i.next()) {
5064 PointsToNode* u = i.get();
5065 bool is_base = false;
5066 if (PointsToNode::is_base_use(u)) {
5067 is_base = true;
5068 u = PointsToNode::get_use_node(u)->as_Field();
5069 }
5070 out->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
5071 }
5072 out->print(" ]] ");
5073 if (_node == nullptr) {
5074 out->print("<null>%s", newline ? "\n" : "");
5075 } else {
5076 _node->dump(newline ? "\n" : "", false, out);
5077 }
5078 }
5079
5080 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
5081 bool first = true;
5082 int ptnodes_length = ptnodes_worklist.length();
5083 for (int i = 0; i < ptnodes_length; i++) {
5084 PointsToNode *ptn = ptnodes_worklist.at(i);
5085 if (ptn == nullptr || !ptn->is_JavaObject()) {
5086 continue;
5087 }
5088 PointsToNode::EscapeState es = ptn->escape_state();
5089 if ((es != PointsToNode::NoEscape) && !Verbose) {
5090 continue;
5091 }
5092 Node* n = ptn->ideal_node();
5093 if (n->is_Allocate() || (n->is_CallStaticJava() &&
5094 n->as_CallStaticJava()->is_boxing_method())) {
5095 if (first) {
5096 tty->cr();
5097 tty->print("======== Connection graph for ");
5098 _compile->method()->print_short_name();
5099 tty->cr();
5100 tty->print_cr("invocation #%d: %d iterations and %f sec to build connection graph with %d nodes and worklist size %d",
5101 _invocation, _build_iterations, _build_time, nodes_size(), ptnodes_worklist.length());
5102 tty->cr();
5103 first = false;
5104 }
5105 ptn->dump();
5106 // Print all locals and fields which reference this allocation
5107 for (UseIterator j(ptn); j.has_next(); j.next()) {
5108 PointsToNode* use = j.get();
5109 if (use->is_LocalVar()) {
5110 use->dump(Verbose);
5111 } else if (Verbose) {
5112 use->dump();
5113 }
5114 }
5115 tty->cr();
5116 }
5117 }
5118 }
5119
5120 void ConnectionGraph::print_statistics() {
5121 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));
5122 }
5123
5124 void ConnectionGraph::escape_state_statistics(GrowableArray<JavaObjectNode*>& java_objects_worklist) {
5125 if (!PrintOptoStatistics || (_invocation > 0)) { // Collect data only for the first invocation
5126 return;
5127 }
5128 for (int next = 0; next < java_objects_worklist.length(); ++next) {
5129 JavaObjectNode* ptn = java_objects_worklist.at(next);
5130 if (ptn->ideal_node()->is_Allocate()) {
5131 if (ptn->escape_state() == PointsToNode::NoEscape) {
5132 Atomic::inc(&ConnectionGraph::_no_escape_counter);
5133 } else if (ptn->escape_state() == PointsToNode::ArgEscape) {
5134 Atomic::inc(&ConnectionGraph::_arg_escape_counter);
5135 } else if (ptn->escape_state() == PointsToNode::GlobalEscape) {
5136 Atomic::inc(&ConnectionGraph::_global_escape_counter);
5137 } else {
5138 assert(false, "Unexpected Escape State");
5139 }
5140 }
5141 }
5142 }
5143
5144 void ConnectionGraph::trace_es_update_helper(PointsToNode* ptn, PointsToNode::EscapeState es, bool fields, const char* reason) const {
5145 if (_compile->directive()->TraceEscapeAnalysisOption) {
5146 assert(ptn != nullptr, "should not be null");
5147 assert(reason != nullptr, "should not be null");
5148 ptn->dump_header(true);
5149 PointsToNode::EscapeState new_es = fields ? ptn->escape_state() : es;
5150 PointsToNode::EscapeState new_fields_es = fields ? es : ptn->fields_escape_state();
5151 tty->print_cr("-> %s(%s) %s", esc_names[(int)new_es], esc_names[(int)new_fields_es], reason);
5152 }
5153 }
5154
5155 const char* ConnectionGraph::trace_propagate_message(PointsToNode* from) const {
5156 if (_compile->directive()->TraceEscapeAnalysisOption) {
5157 stringStream ss;
5158 ss.print("propagated from: ");
5159 from->dump(true, &ss, false);
5160 return ss.as_string();
5161 } else {
5162 return nullptr;
5163 }
5164 }
5165
5166 const char* ConnectionGraph::trace_arg_escape_message(CallNode* call) const {
5167 if (_compile->directive()->TraceEscapeAnalysisOption) {
5168 stringStream ss;
5169 ss.print("escapes as arg to:");
5170 call->dump("", false, &ss);
5171 return ss.as_string();
5172 } else {
5173 return nullptr;
5174 }
5175 }
5176
5177 const char* ConnectionGraph::trace_merged_message(PointsToNode* other) const {
5178 if (_compile->directive()->TraceEscapeAnalysisOption) {
5179 stringStream ss;
5180 ss.print("is merged with other object: ");
5181 other->dump_header(true, &ss);
5182 return ss.as_string();
5183 } else {
5184 return nullptr;
5185 }
5186 }
5187
5188 #endif
5189
5190 void ConnectionGraph::record_for_optimizer(Node *n) {
5191 _igvn->_worklist.push(n);
5192 _igvn->add_users_to_worklist(n);
5193 }