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