Thu, 20 Feb 2014 11:05:12 +0100
8035394: PPC64: Make usage of intrinsic dsqrt depend on processor recognition.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2005, 2013, 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 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/allocation.hpp"
30 #include "opto/c2compiler.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/escape.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
38 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
39 _nodes(C->comp_arena(), C->unique(), C->unique(), NULL),
40 _collecting(true),
41 _verify(false),
42 _compile(C),
43 _igvn(igvn),
44 _node_map(C->comp_arena()) {
45 // Add unknown java object.
46 add_java_object(C->top(), PointsToNode::GlobalEscape);
47 phantom_obj = ptnode_adr(C->top()->_idx)->as_JavaObject();
48 // Add ConP(#NULL) and ConN(#NULL) nodes.
49 Node* oop_null = igvn->zerocon(T_OBJECT);
50 assert(oop_null->_idx < nodes_size(), "should be created already");
51 add_java_object(oop_null, PointsToNode::NoEscape);
52 null_obj = ptnode_adr(oop_null->_idx)->as_JavaObject();
53 if (UseCompressedOops) {
54 Node* noop_null = igvn->zerocon(T_NARROWOOP);
55 assert(noop_null->_idx < nodes_size(), "should be created already");
56 map_ideal_node(noop_null, null_obj);
57 }
58 _pcmp_neq = NULL; // Should be initialized
59 _pcmp_eq = NULL;
60 }
62 bool ConnectionGraph::has_candidates(Compile *C) {
63 // EA brings benefits only when the code has allocations and/or locks which
64 // are represented by ideal Macro nodes.
65 int cnt = C->macro_count();
66 for (int i = 0; i < cnt; i++) {
67 Node *n = C->macro_node(i);
68 if (n->is_Allocate())
69 return true;
70 if (n->is_Lock()) {
71 Node* obj = n->as_Lock()->obj_node()->uncast();
72 if (!(obj->is_Parm() || obj->is_Con()))
73 return true;
74 }
75 if (n->is_CallStaticJava() &&
76 n->as_CallStaticJava()->is_boxing_method()) {
77 return true;
78 }
79 }
80 return false;
81 }
83 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
84 Compile::TracePhase t2("escapeAnalysis", &Phase::_t_escapeAnalysis, true);
85 ResourceMark rm;
87 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
88 // to create space for them in ConnectionGraph::_nodes[].
89 Node* oop_null = igvn->zerocon(T_OBJECT);
90 Node* noop_null = igvn->zerocon(T_NARROWOOP);
91 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
92 // Perform escape analysis
93 if (congraph->compute_escape()) {
94 // There are non escaping objects.
95 C->set_congraph(congraph);
96 }
97 // Cleanup.
98 if (oop_null->outcnt() == 0)
99 igvn->hash_delete(oop_null);
100 if (noop_null->outcnt() == 0)
101 igvn->hash_delete(noop_null);
102 }
104 bool ConnectionGraph::compute_escape() {
105 Compile* C = _compile;
106 PhaseGVN* igvn = _igvn;
108 // Worklists used by EA.
109 Unique_Node_List delayed_worklist;
110 GrowableArray<Node*> alloc_worklist;
111 GrowableArray<Node*> ptr_cmp_worklist;
112 GrowableArray<Node*> storestore_worklist;
113 GrowableArray<PointsToNode*> ptnodes_worklist;
114 GrowableArray<JavaObjectNode*> java_objects_worklist;
115 GrowableArray<JavaObjectNode*> non_escaped_worklist;
116 GrowableArray<FieldNode*> oop_fields_worklist;
117 DEBUG_ONLY( GrowableArray<Node*> addp_worklist; )
119 { Compile::TracePhase t3("connectionGraph", &Phase::_t_connectionGraph, true);
121 // 1. Populate Connection Graph (CG) with PointsTo nodes.
122 ideal_nodes.map(C->live_nodes(), NULL); // preallocate space
123 // Initialize worklist
124 if (C->root() != NULL) {
125 ideal_nodes.push(C->root());
126 }
127 for( uint next = 0; next < ideal_nodes.size(); ++next ) {
128 Node* n = ideal_nodes.at(next);
129 // Create PointsTo nodes and add them to Connection Graph. Called
130 // only once per ideal node since ideal_nodes is Unique_Node list.
131 add_node_to_connection_graph(n, &delayed_worklist);
132 PointsToNode* ptn = ptnode_adr(n->_idx);
133 if (ptn != NULL) {
134 ptnodes_worklist.append(ptn);
135 if (ptn->is_JavaObject()) {
136 java_objects_worklist.append(ptn->as_JavaObject());
137 if ((n->is_Allocate() || n->is_CallStaticJava()) &&
138 (ptn->escape_state() < PointsToNode::GlobalEscape)) {
139 // Only allocations and java static calls results are interesting.
140 non_escaped_worklist.append(ptn->as_JavaObject());
141 }
142 } else if (ptn->is_Field() && ptn->as_Field()->is_oop()) {
143 oop_fields_worklist.append(ptn->as_Field());
144 }
145 }
146 if (n->is_MergeMem()) {
147 // Collect all MergeMem nodes to add memory slices for
148 // scalar replaceable objects in split_unique_types().
149 _mergemem_worklist.append(n->as_MergeMem());
150 } else if (OptimizePtrCompare && n->is_Cmp() &&
151 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
152 // Collect compare pointers nodes.
153 ptr_cmp_worklist.append(n);
154 } else if (n->is_MemBarStoreStore()) {
155 // Collect all MemBarStoreStore nodes so that depending on the
156 // escape status of the associated Allocate node some of them
157 // may be eliminated.
158 storestore_worklist.append(n);
159 } else if (n->is_MemBar() && (n->Opcode() == Op_MemBarRelease) &&
160 (n->req() > MemBarNode::Precedent)) {
161 record_for_optimizer(n);
162 #ifdef ASSERT
163 } else if (n->is_AddP()) {
164 // Collect address nodes for graph verification.
165 addp_worklist.append(n);
166 #endif
167 }
168 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
169 Node* m = n->fast_out(i); // Get user
170 ideal_nodes.push(m);
171 }
172 }
173 if (non_escaped_worklist.length() == 0) {
174 _collecting = false;
175 return false; // Nothing to do.
176 }
177 // Add final simple edges to graph.
178 while(delayed_worklist.size() > 0) {
179 Node* n = delayed_worklist.pop();
180 add_final_edges(n);
181 }
182 int ptnodes_length = ptnodes_worklist.length();
184 #ifdef ASSERT
185 if (VerifyConnectionGraph) {
186 // Verify that no new simple edges could be created and all
187 // local vars has edges.
188 _verify = true;
189 for (int next = 0; next < ptnodes_length; ++next) {
190 PointsToNode* ptn = ptnodes_worklist.at(next);
191 add_final_edges(ptn->ideal_node());
192 if (ptn->is_LocalVar() && ptn->edge_count() == 0) {
193 ptn->dump();
194 assert(ptn->as_LocalVar()->edge_count() > 0, "sanity");
195 }
196 }
197 _verify = false;
198 }
199 #endif
201 // 2. Finish Graph construction by propagating references to all
202 // java objects through graph.
203 if (!complete_connection_graph(ptnodes_worklist, non_escaped_worklist,
204 java_objects_worklist, oop_fields_worklist)) {
205 // All objects escaped or hit time or iterations limits.
206 _collecting = false;
207 return false;
208 }
210 // 3. Adjust scalar_replaceable state of nonescaping objects and push
211 // scalar replaceable allocations on alloc_worklist for processing
212 // in split_unique_types().
213 int non_escaped_length = non_escaped_worklist.length();
214 for (int next = 0; next < non_escaped_length; next++) {
215 JavaObjectNode* ptn = non_escaped_worklist.at(next);
216 bool noescape = (ptn->escape_state() == PointsToNode::NoEscape);
217 Node* n = ptn->ideal_node();
218 if (n->is_Allocate()) {
219 n->as_Allocate()->_is_non_escaping = noescape;
220 }
221 if (n->is_CallStaticJava()) {
222 n->as_CallStaticJava()->_is_non_escaping = noescape;
223 }
224 if (noescape && ptn->scalar_replaceable()) {
225 adjust_scalar_replaceable_state(ptn);
226 if (ptn->scalar_replaceable()) {
227 alloc_worklist.append(ptn->ideal_node());
228 }
229 }
230 }
232 #ifdef ASSERT
233 if (VerifyConnectionGraph) {
234 // Verify that graph is complete - no new edges could be added or needed.
235 verify_connection_graph(ptnodes_worklist, non_escaped_worklist,
236 java_objects_worklist, addp_worklist);
237 }
238 assert(C->unique() == nodes_size(), "no new ideal nodes should be added during ConnectionGraph build");
239 assert(null_obj->escape_state() == PointsToNode::NoEscape &&
240 null_obj->edge_count() == 0 &&
241 !null_obj->arraycopy_src() &&
242 !null_obj->arraycopy_dst(), "sanity");
243 #endif
245 _collecting = false;
247 } // TracePhase t3("connectionGraph")
249 // 4. Optimize ideal graph based on EA information.
250 bool has_non_escaping_obj = (non_escaped_worklist.length() > 0);
251 if (has_non_escaping_obj) {
252 optimize_ideal_graph(ptr_cmp_worklist, storestore_worklist);
253 }
255 #ifndef PRODUCT
256 if (PrintEscapeAnalysis) {
257 dump(ptnodes_worklist); // Dump ConnectionGraph
258 }
259 #endif
261 bool has_scalar_replaceable_candidates = (alloc_worklist.length() > 0);
262 #ifdef ASSERT
263 if (VerifyConnectionGraph) {
264 int alloc_length = alloc_worklist.length();
265 for (int next = 0; next < alloc_length; ++next) {
266 Node* n = alloc_worklist.at(next);
267 PointsToNode* ptn = ptnode_adr(n->_idx);
268 assert(ptn->escape_state() == PointsToNode::NoEscape && ptn->scalar_replaceable(), "sanity");
269 }
270 }
271 #endif
273 // 5. Separate memory graph for scalar replaceable allcations.
274 if (has_scalar_replaceable_candidates &&
275 C->AliasLevel() >= 3 && EliminateAllocations) {
276 // Now use the escape information to create unique types for
277 // scalar replaceable objects.
278 split_unique_types(alloc_worklist);
279 if (C->failing()) return false;
280 C->print_method(PHASE_AFTER_EA, 2);
282 #ifdef ASSERT
283 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
284 tty->print("=== No allocations eliminated for ");
285 C->method()->print_short_name();
286 if(!EliminateAllocations) {
287 tty->print(" since EliminateAllocations is off ===");
288 } else if(!has_scalar_replaceable_candidates) {
289 tty->print(" since there are no scalar replaceable candidates ===");
290 } else if(C->AliasLevel() < 3) {
291 tty->print(" since AliasLevel < 3 ===");
292 }
293 tty->cr();
294 #endif
295 }
296 return has_non_escaping_obj;
297 }
299 // Utility function for nodes that load an object
300 void ConnectionGraph::add_objload_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
301 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
302 // ThreadLocal has RawPtr type.
303 const Type* t = _igvn->type(n);
304 if (t->make_ptr() != NULL) {
305 Node* adr = n->in(MemNode::Address);
306 #ifdef ASSERT
307 if (!adr->is_AddP()) {
308 assert(_igvn->type(adr)->isa_rawptr(), "sanity");
309 } else {
310 assert((ptnode_adr(adr->_idx) == NULL ||
311 ptnode_adr(adr->_idx)->as_Field()->is_oop()), "sanity");
312 }
313 #endif
314 add_local_var_and_edge(n, PointsToNode::NoEscape,
315 adr, delayed_worklist);
316 }
317 }
319 // Populate Connection Graph with PointsTo nodes and create simple
320 // connection graph edges.
321 void ConnectionGraph::add_node_to_connection_graph(Node *n, Unique_Node_List *delayed_worklist) {
322 assert(!_verify, "this method sould not be called for verification");
323 PhaseGVN* igvn = _igvn;
324 uint n_idx = n->_idx;
325 PointsToNode* n_ptn = ptnode_adr(n_idx);
326 if (n_ptn != NULL)
327 return; // No need to redefine PointsTo node during first iteration.
329 if (n->is_Call()) {
330 // Arguments to allocation and locking don't escape.
331 if (n->is_AbstractLock()) {
332 // Put Lock and Unlock nodes on IGVN worklist to process them during
333 // first IGVN optimization when escape information is still available.
334 record_for_optimizer(n);
335 } else if (n->is_Allocate()) {
336 add_call_node(n->as_Call());
337 record_for_optimizer(n);
338 } else {
339 if (n->is_CallStaticJava()) {
340 const char* name = n->as_CallStaticJava()->_name;
341 if (name != NULL && strcmp(name, "uncommon_trap") == 0)
342 return; // Skip uncommon traps
343 }
344 // Don't mark as processed since call's arguments have to be processed.
345 delayed_worklist->push(n);
346 // Check if a call returns an object.
347 if ((n->as_Call()->returns_pointer() &&
348 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) ||
349 (n->is_CallStaticJava() &&
350 n->as_CallStaticJava()->is_boxing_method())) {
351 add_call_node(n->as_Call());
352 }
353 }
354 return;
355 }
356 // Put this check here to process call arguments since some call nodes
357 // point to phantom_obj.
358 if (n_ptn == phantom_obj || n_ptn == null_obj)
359 return; // Skip predefined nodes.
361 int opcode = n->Opcode();
362 switch (opcode) {
363 case Op_AddP: {
364 Node* base = get_addp_base(n);
365 PointsToNode* ptn_base = ptnode_adr(base->_idx);
366 // Field nodes are created for all field types. They are used in
367 // adjust_scalar_replaceable_state() and split_unique_types().
368 // Note, non-oop fields will have only base edges in Connection
369 // Graph because such fields are not used for oop loads and stores.
370 int offset = address_offset(n, igvn);
371 add_field(n, PointsToNode::NoEscape, offset);
372 if (ptn_base == NULL) {
373 delayed_worklist->push(n); // Process it later.
374 } else {
375 n_ptn = ptnode_adr(n_idx);
376 add_base(n_ptn->as_Field(), ptn_base);
377 }
378 break;
379 }
380 case Op_CastX2P: {
381 map_ideal_node(n, phantom_obj);
382 break;
383 }
384 case Op_CastPP:
385 case Op_CheckCastPP:
386 case Op_EncodeP:
387 case Op_DecodeN:
388 case Op_EncodePKlass:
389 case Op_DecodeNKlass: {
390 add_local_var_and_edge(n, PointsToNode::NoEscape,
391 n->in(1), delayed_worklist);
392 break;
393 }
394 case Op_CMoveP: {
395 add_local_var(n, PointsToNode::NoEscape);
396 // Do not add edges during first iteration because some could be
397 // not defined yet.
398 delayed_worklist->push(n);
399 break;
400 }
401 case Op_ConP:
402 case Op_ConN:
403 case Op_ConNKlass: {
404 // assume all oop constants globally escape except for null
405 PointsToNode::EscapeState es;
406 const Type* t = igvn->type(n);
407 if (t == TypePtr::NULL_PTR || t == TypeNarrowOop::NULL_PTR) {
408 es = PointsToNode::NoEscape;
409 } else {
410 es = PointsToNode::GlobalEscape;
411 }
412 add_java_object(n, es);
413 break;
414 }
415 case Op_CreateEx: {
416 // assume that all exception objects globally escape
417 add_java_object(n, PointsToNode::GlobalEscape);
418 break;
419 }
420 case Op_LoadKlass:
421 case Op_LoadNKlass: {
422 // Unknown class is loaded
423 map_ideal_node(n, phantom_obj);
424 break;
425 }
426 case Op_LoadP:
427 case Op_LoadN:
428 case Op_LoadPLocked: {
429 add_objload_to_connection_graph(n, delayed_worklist);
430 break;
431 }
432 case Op_Parm: {
433 map_ideal_node(n, phantom_obj);
434 break;
435 }
436 case Op_PartialSubtypeCheck: {
437 // Produces Null or notNull and is used in only in CmpP so
438 // phantom_obj could be used.
439 map_ideal_node(n, phantom_obj); // Result is unknown
440 break;
441 }
442 case Op_Phi: {
443 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
444 // ThreadLocal has RawPtr type.
445 const Type* t = n->as_Phi()->type();
446 if (t->make_ptr() != NULL) {
447 add_local_var(n, PointsToNode::NoEscape);
448 // Do not add edges during first iteration because some could be
449 // not defined yet.
450 delayed_worklist->push(n);
451 }
452 break;
453 }
454 case Op_Proj: {
455 // we are only interested in the oop result projection from a call
456 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
457 n->in(0)->as_Call()->returns_pointer()) {
458 add_local_var_and_edge(n, PointsToNode::NoEscape,
459 n->in(0), delayed_worklist);
460 }
461 break;
462 }
463 case Op_Rethrow: // Exception object escapes
464 case Op_Return: {
465 if (n->req() > TypeFunc::Parms &&
466 igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
467 // Treat Return value as LocalVar with GlobalEscape escape state.
468 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
469 n->in(TypeFunc::Parms), delayed_worklist);
470 }
471 break;
472 }
473 case Op_GetAndSetP:
474 case Op_GetAndSetN: {
475 add_objload_to_connection_graph(n, delayed_worklist);
476 // fallthrough
477 }
478 case Op_StoreP:
479 case Op_StoreN:
480 case Op_StoreNKlass:
481 case Op_StorePConditional:
482 case Op_CompareAndSwapP:
483 case Op_CompareAndSwapN: {
484 Node* adr = n->in(MemNode::Address);
485 const Type *adr_type = igvn->type(adr);
486 adr_type = adr_type->make_ptr();
487 if (adr_type == NULL) {
488 break; // skip dead nodes
489 }
490 if (adr_type->isa_oopptr() ||
491 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
492 (adr_type == TypeRawPtr::NOTNULL &&
493 adr->in(AddPNode::Address)->is_Proj() &&
494 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
495 delayed_worklist->push(n); // Process it later.
496 #ifdef ASSERT
497 assert(adr->is_AddP(), "expecting an AddP");
498 if (adr_type == TypeRawPtr::NOTNULL) {
499 // Verify a raw address for a store captured by Initialize node.
500 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
501 assert(offs != Type::OffsetBot, "offset must be a constant");
502 }
503 #endif
504 } else {
505 // Ignore copy the displaced header to the BoxNode (OSR compilation).
506 if (adr->is_BoxLock())
507 break;
508 // Stored value escapes in unsafe access.
509 if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
510 // Pointer stores in G1 barriers looks like unsafe access.
511 // Ignore such stores to be able scalar replace non-escaping
512 // allocations.
513 if (UseG1GC && adr->is_AddP()) {
514 Node* base = get_addp_base(adr);
515 if (base->Opcode() == Op_LoadP &&
516 base->in(MemNode::Address)->is_AddP()) {
517 adr = base->in(MemNode::Address);
518 Node* tls = get_addp_base(adr);
519 if (tls->Opcode() == Op_ThreadLocal) {
520 int offs = (int)igvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
521 if (offs == in_bytes(JavaThread::satb_mark_queue_offset() +
522 PtrQueue::byte_offset_of_buf())) {
523 break; // G1 pre barier previous oop value store.
524 }
525 if (offs == in_bytes(JavaThread::dirty_card_queue_offset() +
526 PtrQueue::byte_offset_of_buf())) {
527 break; // G1 post barier card address store.
528 }
529 }
530 }
531 }
532 delayed_worklist->push(n); // Process unsafe access later.
533 break;
534 }
535 #ifdef ASSERT
536 n->dump(1);
537 assert(false, "not unsafe or G1 barrier raw StoreP");
538 #endif
539 }
540 break;
541 }
542 case Op_AryEq:
543 case Op_StrComp:
544 case Op_StrEquals:
545 case Op_StrIndexOf:
546 case Op_EncodeISOArray: {
547 add_local_var(n, PointsToNode::ArgEscape);
548 delayed_worklist->push(n); // Process it later.
549 break;
550 }
551 case Op_ThreadLocal: {
552 add_java_object(n, PointsToNode::ArgEscape);
553 break;
554 }
555 default:
556 ; // Do nothing for nodes not related to EA.
557 }
558 return;
559 }
561 #ifdef ASSERT
562 #define ELSE_FAIL(name) \
563 /* Should not be called for not pointer type. */ \
564 n->dump(1); \
565 assert(false, name); \
566 break;
567 #else
568 #define ELSE_FAIL(name) \
569 break;
570 #endif
572 // Add final simple edges to graph.
573 void ConnectionGraph::add_final_edges(Node *n) {
574 PointsToNode* n_ptn = ptnode_adr(n->_idx);
575 #ifdef ASSERT
576 if (_verify && n_ptn->is_JavaObject())
577 return; // This method does not change graph for JavaObject.
578 #endif
580 if (n->is_Call()) {
581 process_call_arguments(n->as_Call());
582 return;
583 }
584 assert(n->is_Store() || n->is_LoadStore() ||
585 (n_ptn != NULL) && (n_ptn->ideal_node() != NULL),
586 "node should be registered already");
587 int opcode = n->Opcode();
588 switch (opcode) {
589 case Op_AddP: {
590 Node* base = get_addp_base(n);
591 PointsToNode* ptn_base = ptnode_adr(base->_idx);
592 assert(ptn_base != NULL, "field's base should be registered");
593 add_base(n_ptn->as_Field(), ptn_base);
594 break;
595 }
596 case Op_CastPP:
597 case Op_CheckCastPP:
598 case Op_EncodeP:
599 case Op_DecodeN:
600 case Op_EncodePKlass:
601 case Op_DecodeNKlass: {
602 add_local_var_and_edge(n, PointsToNode::NoEscape,
603 n->in(1), NULL);
604 break;
605 }
606 case Op_CMoveP: {
607 for (uint i = CMoveNode::IfFalse; i < n->req(); i++) {
608 Node* in = n->in(i);
609 if (in == NULL)
610 continue; // ignore NULL
611 Node* uncast_in = in->uncast();
612 if (uncast_in->is_top() || uncast_in == n)
613 continue; // ignore top or inputs which go back this node
614 PointsToNode* ptn = ptnode_adr(in->_idx);
615 assert(ptn != NULL, "node should be registered");
616 add_edge(n_ptn, ptn);
617 }
618 break;
619 }
620 case Op_LoadP:
621 case Op_LoadN:
622 case Op_LoadPLocked: {
623 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
624 // ThreadLocal has RawPtr type.
625 const Type* t = _igvn->type(n);
626 if (t->make_ptr() != NULL) {
627 Node* adr = n->in(MemNode::Address);
628 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
629 break;
630 }
631 ELSE_FAIL("Op_LoadP");
632 }
633 case Op_Phi: {
634 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
635 // ThreadLocal has RawPtr type.
636 const Type* t = n->as_Phi()->type();
637 if (t->make_ptr() != NULL) {
638 for (uint i = 1; i < n->req(); i++) {
639 Node* in = n->in(i);
640 if (in == NULL)
641 continue; // ignore NULL
642 Node* uncast_in = in->uncast();
643 if (uncast_in->is_top() || uncast_in == n)
644 continue; // ignore top or inputs which go back this node
645 PointsToNode* ptn = ptnode_adr(in->_idx);
646 assert(ptn != NULL, "node should be registered");
647 add_edge(n_ptn, ptn);
648 }
649 break;
650 }
651 ELSE_FAIL("Op_Phi");
652 }
653 case Op_Proj: {
654 // we are only interested in the oop result projection from a call
655 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() &&
656 n->in(0)->as_Call()->returns_pointer()) {
657 add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(0), NULL);
658 break;
659 }
660 ELSE_FAIL("Op_Proj");
661 }
662 case Op_Rethrow: // Exception object escapes
663 case Op_Return: {
664 if (n->req() > TypeFunc::Parms &&
665 _igvn->type(n->in(TypeFunc::Parms))->isa_oopptr()) {
666 // Treat Return value as LocalVar with GlobalEscape escape state.
667 add_local_var_and_edge(n, PointsToNode::GlobalEscape,
668 n->in(TypeFunc::Parms), NULL);
669 break;
670 }
671 ELSE_FAIL("Op_Return");
672 }
673 case Op_StoreP:
674 case Op_StoreN:
675 case Op_StoreNKlass:
676 case Op_StorePConditional:
677 case Op_CompareAndSwapP:
678 case Op_CompareAndSwapN:
679 case Op_GetAndSetP:
680 case Op_GetAndSetN: {
681 Node* adr = n->in(MemNode::Address);
682 const Type *adr_type = _igvn->type(adr);
683 adr_type = adr_type->make_ptr();
684 #ifdef ASSERT
685 if (adr_type == NULL) {
686 n->dump(1);
687 assert(adr_type != NULL, "dead node should not be on list");
688 break;
689 }
690 #endif
691 if (opcode == Op_GetAndSetP || opcode == Op_GetAndSetN) {
692 add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
693 }
694 if (adr_type->isa_oopptr() ||
695 (opcode == Op_StoreP || opcode == Op_StoreN || opcode == Op_StoreNKlass) &&
696 (adr_type == TypeRawPtr::NOTNULL &&
697 adr->in(AddPNode::Address)->is_Proj() &&
698 adr->in(AddPNode::Address)->in(0)->is_Allocate())) {
699 // Point Address to Value
700 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
701 assert(adr_ptn != NULL &&
702 adr_ptn->as_Field()->is_oop(), "node should be registered");
703 Node *val = n->in(MemNode::ValueIn);
704 PointsToNode* ptn = ptnode_adr(val->_idx);
705 assert(ptn != NULL, "node should be registered");
706 add_edge(adr_ptn, ptn);
707 break;
708 } else if ((opcode == Op_StoreP) && (adr_type == TypeRawPtr::BOTTOM)) {
709 // Stored value escapes in unsafe access.
710 Node *val = n->in(MemNode::ValueIn);
711 PointsToNode* ptn = ptnode_adr(val->_idx);
712 assert(ptn != NULL, "node should be registered");
713 ptn->set_escape_state(PointsToNode::GlobalEscape);
714 // Add edge to object for unsafe access with offset.
715 PointsToNode* adr_ptn = ptnode_adr(adr->_idx);
716 assert(adr_ptn != NULL, "node should be registered");
717 if (adr_ptn->is_Field()) {
718 assert(adr_ptn->as_Field()->is_oop(), "should be oop field");
719 add_edge(adr_ptn, ptn);
720 }
721 break;
722 }
723 ELSE_FAIL("Op_StoreP");
724 }
725 case Op_AryEq:
726 case Op_StrComp:
727 case Op_StrEquals:
728 case Op_StrIndexOf:
729 case Op_EncodeISOArray: {
730 // char[] arrays passed to string intrinsic do not escape but
731 // they are not scalar replaceable. Adjust escape state for them.
732 // Start from in(2) edge since in(1) is memory edge.
733 for (uint i = 2; i < n->req(); i++) {
734 Node* adr = n->in(i);
735 const Type* at = _igvn->type(adr);
736 if (!adr->is_top() && at->isa_ptr()) {
737 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
738 at->isa_ptr() != NULL, "expecting a pointer");
739 if (adr->is_AddP()) {
740 adr = get_addp_base(adr);
741 }
742 PointsToNode* ptn = ptnode_adr(adr->_idx);
743 assert(ptn != NULL, "node should be registered");
744 add_edge(n_ptn, ptn);
745 }
746 }
747 break;
748 }
749 default: {
750 // This method should be called only for EA specific nodes which may
751 // miss some edges when they were created.
752 #ifdef ASSERT
753 n->dump(1);
754 #endif
755 guarantee(false, "unknown node");
756 }
757 }
758 return;
759 }
761 void ConnectionGraph::add_call_node(CallNode* call) {
762 assert(call->returns_pointer(), "only for call which returns pointer");
763 uint call_idx = call->_idx;
764 if (call->is_Allocate()) {
765 Node* k = call->in(AllocateNode::KlassNode);
766 const TypeKlassPtr* kt = k->bottom_type()->isa_klassptr();
767 assert(kt != NULL, "TypeKlassPtr required.");
768 ciKlass* cik = kt->klass();
769 PointsToNode::EscapeState es = PointsToNode::NoEscape;
770 bool scalar_replaceable = true;
771 if (call->is_AllocateArray()) {
772 if (!cik->is_array_klass()) { // StressReflectiveCode
773 es = PointsToNode::GlobalEscape;
774 } else {
775 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
776 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
777 // Not scalar replaceable if the length is not constant or too big.
778 scalar_replaceable = false;
779 }
780 }
781 } else { // Allocate instance
782 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
783 cik->is_subclass_of(_compile->env()->Reference_klass()) ||
784 !cik->is_instance_klass() || // StressReflectiveCode
785 cik->as_instance_klass()->has_finalizer()) {
786 es = PointsToNode::GlobalEscape;
787 }
788 }
789 add_java_object(call, es);
790 PointsToNode* ptn = ptnode_adr(call_idx);
791 if (!scalar_replaceable && ptn->scalar_replaceable()) {
792 ptn->set_scalar_replaceable(false);
793 }
794 } else if (call->is_CallStaticJava()) {
795 // Call nodes could be different types:
796 //
797 // 1. CallDynamicJavaNode (what happened during call is unknown):
798 //
799 // - mapped to GlobalEscape JavaObject node if oop is returned;
800 //
801 // - all oop arguments are escaping globally;
802 //
803 // 2. CallStaticJavaNode (execute bytecode analysis if possible):
804 //
805 // - the same as CallDynamicJavaNode if can't do bytecode analysis;
806 //
807 // - mapped to GlobalEscape JavaObject node if unknown oop is returned;
808 // - mapped to NoEscape JavaObject node if non-escaping object allocated
809 // during call is returned;
810 // - mapped to ArgEscape LocalVar node pointed to object arguments
811 // which are returned and does not escape during call;
812 //
813 // - oop arguments escaping status is defined by bytecode analysis;
814 //
815 // For a static call, we know exactly what method is being called.
816 // Use bytecode estimator to record whether the call's return value escapes.
817 ciMethod* meth = call->as_CallJava()->method();
818 if (meth == NULL) {
819 const char* name = call->as_CallStaticJava()->_name;
820 assert(strncmp(name, "_multianewarray", 15) == 0, "TODO: add failed case check");
821 // Returns a newly allocated unescaped object.
822 add_java_object(call, PointsToNode::NoEscape);
823 ptnode_adr(call_idx)->set_scalar_replaceable(false);
824 } else if (meth->is_boxing_method()) {
825 // Returns boxing object
826 PointsToNode::EscapeState es;
827 vmIntrinsics::ID intr = meth->intrinsic_id();
828 if (intr == vmIntrinsics::_floatValue || intr == vmIntrinsics::_doubleValue) {
829 // It does not escape if object is always allocated.
830 es = PointsToNode::NoEscape;
831 } else {
832 // It escapes globally if object could be loaded from cache.
833 es = PointsToNode::GlobalEscape;
834 }
835 add_java_object(call, es);
836 } else {
837 BCEscapeAnalyzer* call_analyzer = meth->get_bcea();
838 call_analyzer->copy_dependencies(_compile->dependencies());
839 if (call_analyzer->is_return_allocated()) {
840 // Returns a newly allocated unescaped object, simply
841 // update dependency information.
842 // Mark it as NoEscape so that objects referenced by
843 // it's fields will be marked as NoEscape at least.
844 add_java_object(call, PointsToNode::NoEscape);
845 ptnode_adr(call_idx)->set_scalar_replaceable(false);
846 } else {
847 // Determine whether any arguments are returned.
848 const TypeTuple* d = call->tf()->domain();
849 bool ret_arg = false;
850 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
851 if (d->field_at(i)->isa_ptr() != NULL &&
852 call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
853 ret_arg = true;
854 break;
855 }
856 }
857 if (ret_arg) {
858 add_local_var(call, PointsToNode::ArgEscape);
859 } else {
860 // Returns unknown object.
861 map_ideal_node(call, phantom_obj);
862 }
863 }
864 }
865 } else {
866 // An other type of call, assume the worst case:
867 // returned value is unknown and globally escapes.
868 assert(call->Opcode() == Op_CallDynamicJava, "add failed case check");
869 map_ideal_node(call, phantom_obj);
870 }
871 }
873 void ConnectionGraph::process_call_arguments(CallNode *call) {
874 bool is_arraycopy = false;
875 switch (call->Opcode()) {
876 #ifdef ASSERT
877 case Op_Allocate:
878 case Op_AllocateArray:
879 case Op_Lock:
880 case Op_Unlock:
881 assert(false, "should be done already");
882 break;
883 #endif
884 case Op_CallLeafNoFP:
885 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
886 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
887 // fall through
888 case Op_CallLeaf: {
889 // Stub calls, objects do not escape but they are not scale replaceable.
890 // Adjust escape state for outgoing arguments.
891 const TypeTuple * d = call->tf()->domain();
892 bool src_has_oops = false;
893 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
894 const Type* at = d->field_at(i);
895 Node *arg = call->in(i);
896 const Type *aat = _igvn->type(arg);
897 if (arg->is_top() || !at->isa_ptr() || !aat->isa_ptr())
898 continue;
899 if (arg->is_AddP()) {
900 //
901 // The inline_native_clone() case when the arraycopy stub is called
902 // after the allocation before Initialize and CheckCastPP nodes.
903 // Or normal arraycopy for object arrays case.
904 //
905 // Set AddP's base (Allocate) as not scalar replaceable since
906 // pointer to the base (with offset) is passed as argument.
907 //
908 arg = get_addp_base(arg);
909 }
910 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
911 assert(arg_ptn != NULL, "should be registered");
912 PointsToNode::EscapeState arg_esc = arg_ptn->escape_state();
913 if (is_arraycopy || arg_esc < PointsToNode::ArgEscape) {
914 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
915 aat->isa_ptr() != NULL, "expecting an Ptr");
916 bool arg_has_oops = aat->isa_oopptr() &&
917 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
918 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
919 if (i == TypeFunc::Parms) {
920 src_has_oops = arg_has_oops;
921 }
922 //
923 // src or dst could be j.l.Object when other is basic type array:
924 //
925 // arraycopy(char[],0,Object*,0,size);
926 // arraycopy(Object*,0,char[],0,size);
927 //
928 // Don't add edges in such cases.
929 //
930 bool arg_is_arraycopy_dest = src_has_oops && is_arraycopy &&
931 arg_has_oops && (i > TypeFunc::Parms);
932 #ifdef ASSERT
933 if (!(is_arraycopy ||
934 (call->as_CallLeaf()->_name != NULL &&
935 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
936 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ||
937 strcmp(call->as_CallLeaf()->_name, "updateBytesCRC32") == 0 ||
938 strcmp(call->as_CallLeaf()->_name, "aescrypt_encryptBlock") == 0 ||
939 strcmp(call->as_CallLeaf()->_name, "aescrypt_decryptBlock") == 0 ||
940 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_encryptAESCrypt") == 0 ||
941 strcmp(call->as_CallLeaf()->_name, "cipherBlockChaining_decryptAESCrypt") == 0)
942 ))) {
943 call->dump();
944 fatal(err_msg_res("EA unexpected CallLeaf %s", call->as_CallLeaf()->_name));
945 }
946 #endif
947 // Always process arraycopy's destination object since
948 // we need to add all possible edges to references in
949 // source object.
950 if (arg_esc >= PointsToNode::ArgEscape &&
951 !arg_is_arraycopy_dest) {
952 continue;
953 }
954 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
955 if (arg_is_arraycopy_dest) {
956 Node* src = call->in(TypeFunc::Parms);
957 if (src->is_AddP()) {
958 src = get_addp_base(src);
959 }
960 PointsToNode* src_ptn = ptnode_adr(src->_idx);
961 assert(src_ptn != NULL, "should be registered");
962 if (arg_ptn != src_ptn) {
963 // Special arraycopy edge:
964 // A destination object's field can't have the source object
965 // as base since objects escape states are not related.
966 // Only escape state of destination object's fields affects
967 // escape state of fields in source object.
968 add_arraycopy(call, PointsToNode::ArgEscape, src_ptn, arg_ptn);
969 }
970 }
971 }
972 }
973 break;
974 }
975 case Op_CallStaticJava: {
976 // For a static call, we know exactly what method is being called.
977 // Use bytecode estimator to record the call's escape affects
978 #ifdef ASSERT
979 const char* name = call->as_CallStaticJava()->_name;
980 assert((name == NULL || strcmp(name, "uncommon_trap") != 0), "normal calls only");
981 #endif
982 ciMethod* meth = call->as_CallJava()->method();
983 if ((meth != NULL) && meth->is_boxing_method()) {
984 break; // Boxing methods do not modify any oops.
985 }
986 BCEscapeAnalyzer* call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
987 // fall-through if not a Java method or no analyzer information
988 if (call_analyzer != NULL) {
989 PointsToNode* call_ptn = ptnode_adr(call->_idx);
990 const TypeTuple* d = call->tf()->domain();
991 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
992 const Type* at = d->field_at(i);
993 int k = i - TypeFunc::Parms;
994 Node* arg = call->in(i);
995 PointsToNode* arg_ptn = ptnode_adr(arg->_idx);
996 if (at->isa_ptr() != NULL &&
997 call_analyzer->is_arg_returned(k)) {
998 // The call returns arguments.
999 if (call_ptn != NULL) { // Is call's result used?
1000 assert(call_ptn->is_LocalVar(), "node should be registered");
1001 assert(arg_ptn != NULL, "node should be registered");
1002 add_edge(call_ptn, arg_ptn);
1003 }
1004 }
1005 if (at->isa_oopptr() != NULL &&
1006 arg_ptn->escape_state() < PointsToNode::GlobalEscape) {
1007 if (!call_analyzer->is_arg_stack(k)) {
1008 // The argument global escapes
1009 set_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1010 } else {
1011 set_escape_state(arg_ptn, PointsToNode::ArgEscape);
1012 if (!call_analyzer->is_arg_local(k)) {
1013 // The argument itself doesn't escape, but any fields might
1014 set_fields_escape_state(arg_ptn, PointsToNode::GlobalEscape);
1015 }
1016 }
1017 }
1018 }
1019 if (call_ptn != NULL && call_ptn->is_LocalVar()) {
1020 // The call returns arguments.
1021 assert(call_ptn->edge_count() > 0, "sanity");
1022 if (!call_analyzer->is_return_local()) {
1023 // Returns also unknown object.
1024 add_edge(call_ptn, phantom_obj);
1025 }
1026 }
1027 break;
1028 }
1029 }
1030 default: {
1031 // Fall-through here if not a Java method or no analyzer information
1032 // or some other type of call, assume the worst case: all arguments
1033 // globally escape.
1034 const TypeTuple* d = call->tf()->domain();
1035 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1036 const Type* at = d->field_at(i);
1037 if (at->isa_oopptr() != NULL) {
1038 Node* arg = call->in(i);
1039 if (arg->is_AddP()) {
1040 arg = get_addp_base(arg);
1041 }
1042 assert(ptnode_adr(arg->_idx) != NULL, "should be defined already");
1043 set_escape_state(ptnode_adr(arg->_idx), PointsToNode::GlobalEscape);
1044 }
1045 }
1046 }
1047 }
1048 }
1051 // Finish Graph construction.
1052 bool ConnectionGraph::complete_connection_graph(
1053 GrowableArray<PointsToNode*>& ptnodes_worklist,
1054 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1055 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1056 GrowableArray<FieldNode*>& oop_fields_worklist) {
1057 // Normally only 1-3 passes needed to build Connection Graph depending
1058 // on graph complexity. Observed 8 passes in jvm2008 compiler.compiler.
1059 // Set limit to 20 to catch situation when something did go wrong and
1060 // bailout Escape Analysis.
1061 // Also limit build time to 30 sec (60 in debug VM).
1062 #define CG_BUILD_ITER_LIMIT 20
1063 #ifdef ASSERT
1064 #define CG_BUILD_TIME_LIMIT 60.0
1065 #else
1066 #define CG_BUILD_TIME_LIMIT 30.0
1067 #endif
1069 // Propagate GlobalEscape and ArgEscape escape states and check that
1070 // we still have non-escaping objects. The method pushs on _worklist
1071 // Field nodes which reference phantom_object.
1072 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1073 return false; // Nothing to do.
1074 }
1075 // Now propagate references to all JavaObject nodes.
1076 int java_objects_length = java_objects_worklist.length();
1077 elapsedTimer time;
1078 int new_edges = 1;
1079 int iterations = 0;
1080 do {
1081 while ((new_edges > 0) &&
1082 (iterations++ < CG_BUILD_ITER_LIMIT) &&
1083 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1084 time.start();
1085 new_edges = 0;
1086 // Propagate references to phantom_object for nodes pushed on _worklist
1087 // by find_non_escaped_objects() and find_field_value().
1088 new_edges += add_java_object_edges(phantom_obj, false);
1089 for (int next = 0; next < java_objects_length; ++next) {
1090 JavaObjectNode* ptn = java_objects_worklist.at(next);
1091 new_edges += add_java_object_edges(ptn, true);
1092 }
1093 if (new_edges > 0) {
1094 // Update escape states on each iteration if graph was updated.
1095 if (!find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist)) {
1096 return false; // Nothing to do.
1097 }
1098 }
1099 time.stop();
1100 }
1101 if ((iterations < CG_BUILD_ITER_LIMIT) &&
1102 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1103 time.start();
1104 // Find fields which have unknown value.
1105 int fields_length = oop_fields_worklist.length();
1106 for (int next = 0; next < fields_length; next++) {
1107 FieldNode* field = oop_fields_worklist.at(next);
1108 if (field->edge_count() == 0) {
1109 new_edges += find_field_value(field);
1110 // This code may added new edges to phantom_object.
1111 // Need an other cycle to propagate references to phantom_object.
1112 }
1113 }
1114 time.stop();
1115 } else {
1116 new_edges = 0; // Bailout
1117 }
1118 } while (new_edges > 0);
1120 // Bailout if passed limits.
1121 if ((iterations >= CG_BUILD_ITER_LIMIT) ||
1122 (time.seconds() >= CG_BUILD_TIME_LIMIT)) {
1123 Compile* C = _compile;
1124 if (C->log() != NULL) {
1125 C->log()->begin_elem("connectionGraph_bailout reason='reached ");
1126 C->log()->text("%s", (iterations >= CG_BUILD_ITER_LIMIT) ? "iterations" : "time");
1127 C->log()->end_elem(" limit'");
1128 }
1129 assert(ExitEscapeAnalysisOnTimeout, err_msg_res("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1130 time.seconds(), iterations, nodes_size(), ptnodes_worklist.length()));
1131 // Possible infinite build_connection_graph loop,
1132 // bailout (no changes to ideal graph were made).
1133 return false;
1134 }
1135 #ifdef ASSERT
1136 if (Verbose && PrintEscapeAnalysis) {
1137 tty->print_cr("EA: %d iterations to build connection graph with %d nodes and worklist size %d",
1138 iterations, nodes_size(), ptnodes_worklist.length());
1139 }
1140 #endif
1142 #undef CG_BUILD_ITER_LIMIT
1143 #undef CG_BUILD_TIME_LIMIT
1145 // Find fields initialized by NULL for non-escaping Allocations.
1146 int non_escaped_length = non_escaped_worklist.length();
1147 for (int next = 0; next < non_escaped_length; next++) {
1148 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1149 PointsToNode::EscapeState es = ptn->escape_state();
1150 assert(es <= PointsToNode::ArgEscape, "sanity");
1151 if (es == PointsToNode::NoEscape) {
1152 if (find_init_values(ptn, null_obj, _igvn) > 0) {
1153 // Adding references to NULL object does not change escape states
1154 // since it does not escape. Also no fields are added to NULL object.
1155 add_java_object_edges(null_obj, false);
1156 }
1157 }
1158 Node* n = ptn->ideal_node();
1159 if (n->is_Allocate()) {
1160 // The object allocated by this Allocate node will never be
1161 // seen by an other thread. Mark it so that when it is
1162 // expanded no MemBarStoreStore is added.
1163 InitializeNode* ini = n->as_Allocate()->initialization();
1164 if (ini != NULL)
1165 ini->set_does_not_escape();
1166 }
1167 }
1168 return true; // Finished graph construction.
1169 }
1171 // Propagate GlobalEscape and ArgEscape escape states to all nodes
1172 // and check that we still have non-escaping java objects.
1173 bool ConnectionGraph::find_non_escaped_objects(GrowableArray<PointsToNode*>& ptnodes_worklist,
1174 GrowableArray<JavaObjectNode*>& non_escaped_worklist) {
1175 GrowableArray<PointsToNode*> escape_worklist;
1176 // First, put all nodes with GlobalEscape and ArgEscape states on worklist.
1177 int ptnodes_length = ptnodes_worklist.length();
1178 for (int next = 0; next < ptnodes_length; ++next) {
1179 PointsToNode* ptn = ptnodes_worklist.at(next);
1180 if (ptn->escape_state() >= PointsToNode::ArgEscape ||
1181 ptn->fields_escape_state() >= PointsToNode::ArgEscape) {
1182 escape_worklist.push(ptn);
1183 }
1184 }
1185 // Set escape states to referenced nodes (edges list).
1186 while (escape_worklist.length() > 0) {
1187 PointsToNode* ptn = escape_worklist.pop();
1188 PointsToNode::EscapeState es = ptn->escape_state();
1189 PointsToNode::EscapeState field_es = ptn->fields_escape_state();
1190 if (ptn->is_Field() && ptn->as_Field()->is_oop() &&
1191 es >= PointsToNode::ArgEscape) {
1192 // GlobalEscape or ArgEscape state of field means it has unknown value.
1193 if (add_edge(ptn, phantom_obj)) {
1194 // New edge was added
1195 add_field_uses_to_worklist(ptn->as_Field());
1196 }
1197 }
1198 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1199 PointsToNode* e = i.get();
1200 if (e->is_Arraycopy()) {
1201 assert(ptn->arraycopy_dst(), "sanity");
1202 // Propagate only fields escape state through arraycopy edge.
1203 if (e->fields_escape_state() < field_es) {
1204 set_fields_escape_state(e, field_es);
1205 escape_worklist.push(e);
1206 }
1207 } else if (es >= field_es) {
1208 // fields_escape_state is also set to 'es' if it is less than 'es'.
1209 if (e->escape_state() < es) {
1210 set_escape_state(e, es);
1211 escape_worklist.push(e);
1212 }
1213 } else {
1214 // Propagate field escape state.
1215 bool es_changed = false;
1216 if (e->fields_escape_state() < field_es) {
1217 set_fields_escape_state(e, field_es);
1218 es_changed = true;
1219 }
1220 if ((e->escape_state() < field_es) &&
1221 e->is_Field() && ptn->is_JavaObject() &&
1222 e->as_Field()->is_oop()) {
1223 // Change escape state of referenced fileds.
1224 set_escape_state(e, field_es);
1225 es_changed = true;;
1226 } else if (e->escape_state() < es) {
1227 set_escape_state(e, es);
1228 es_changed = true;;
1229 }
1230 if (es_changed) {
1231 escape_worklist.push(e);
1232 }
1233 }
1234 }
1235 }
1236 // Remove escaped objects from non_escaped list.
1237 for (int next = non_escaped_worklist.length()-1; next >= 0 ; --next) {
1238 JavaObjectNode* ptn = non_escaped_worklist.at(next);
1239 if (ptn->escape_state() >= PointsToNode::GlobalEscape) {
1240 non_escaped_worklist.delete_at(next);
1241 }
1242 if (ptn->escape_state() == PointsToNode::NoEscape) {
1243 // Find fields in non-escaped allocations which have unknown value.
1244 find_init_values(ptn, phantom_obj, NULL);
1245 }
1246 }
1247 return (non_escaped_worklist.length() > 0);
1248 }
1250 // Add all references to JavaObject node by walking over all uses.
1251 int ConnectionGraph::add_java_object_edges(JavaObjectNode* jobj, bool populate_worklist) {
1252 int new_edges = 0;
1253 if (populate_worklist) {
1254 // Populate _worklist by uses of jobj's uses.
1255 for (UseIterator i(jobj); i.has_next(); i.next()) {
1256 PointsToNode* use = i.get();
1257 if (use->is_Arraycopy())
1258 continue;
1259 add_uses_to_worklist(use);
1260 if (use->is_Field() && use->as_Field()->is_oop()) {
1261 // Put on worklist all field's uses (loads) and
1262 // related field nodes (same base and offset).
1263 add_field_uses_to_worklist(use->as_Field());
1264 }
1265 }
1266 }
1267 while(_worklist.length() > 0) {
1268 PointsToNode* use = _worklist.pop();
1269 if (PointsToNode::is_base_use(use)) {
1270 // Add reference from jobj to field and from field to jobj (field's base).
1271 use = PointsToNode::get_use_node(use)->as_Field();
1272 if (add_base(use->as_Field(), jobj)) {
1273 new_edges++;
1274 }
1275 continue;
1276 }
1277 assert(!use->is_JavaObject(), "sanity");
1278 if (use->is_Arraycopy()) {
1279 if (jobj == null_obj) // NULL object does not have field edges
1280 continue;
1281 // Added edge from Arraycopy node to arraycopy's source java object
1282 if (add_edge(use, jobj)) {
1283 jobj->set_arraycopy_src();
1284 new_edges++;
1285 }
1286 // and stop here.
1287 continue;
1288 }
1289 if (!add_edge(use, jobj))
1290 continue; // No new edge added, there was such edge already.
1291 new_edges++;
1292 if (use->is_LocalVar()) {
1293 add_uses_to_worklist(use);
1294 if (use->arraycopy_dst()) {
1295 for (EdgeIterator i(use); i.has_next(); i.next()) {
1296 PointsToNode* e = i.get();
1297 if (e->is_Arraycopy()) {
1298 if (jobj == null_obj) // NULL object does not have field edges
1299 continue;
1300 // Add edge from arraycopy's destination java object to Arraycopy node.
1301 if (add_edge(jobj, e)) {
1302 new_edges++;
1303 jobj->set_arraycopy_dst();
1304 }
1305 }
1306 }
1307 }
1308 } else {
1309 // Added new edge to stored in field values.
1310 // Put on worklist all field's uses (loads) and
1311 // related field nodes (same base and offset).
1312 add_field_uses_to_worklist(use->as_Field());
1313 }
1314 }
1315 return new_edges;
1316 }
1318 // Put on worklist all related field nodes.
1319 void ConnectionGraph::add_field_uses_to_worklist(FieldNode* field) {
1320 assert(field->is_oop(), "sanity");
1321 int offset = field->offset();
1322 add_uses_to_worklist(field);
1323 // Loop over all bases of this field and push on worklist Field nodes
1324 // with the same offset and base (since they may reference the same field).
1325 for (BaseIterator i(field); i.has_next(); i.next()) {
1326 PointsToNode* base = i.get();
1327 add_fields_to_worklist(field, base);
1328 // Check if the base was source object of arraycopy and go over arraycopy's
1329 // destination objects since values stored to a field of source object are
1330 // accessable by uses (loads) of fields of destination objects.
1331 if (base->arraycopy_src()) {
1332 for (UseIterator j(base); j.has_next(); j.next()) {
1333 PointsToNode* arycp = j.get();
1334 if (arycp->is_Arraycopy()) {
1335 for (UseIterator k(arycp); k.has_next(); k.next()) {
1336 PointsToNode* abase = k.get();
1337 if (abase->arraycopy_dst() && abase != base) {
1338 // Look for the same arracopy reference.
1339 add_fields_to_worklist(field, abase);
1340 }
1341 }
1342 }
1343 }
1344 }
1345 }
1346 }
1348 // Put on worklist all related field nodes.
1349 void ConnectionGraph::add_fields_to_worklist(FieldNode* field, PointsToNode* base) {
1350 int offset = field->offset();
1351 if (base->is_LocalVar()) {
1352 for (UseIterator j(base); j.has_next(); j.next()) {
1353 PointsToNode* f = j.get();
1354 if (PointsToNode::is_base_use(f)) { // Field
1355 f = PointsToNode::get_use_node(f);
1356 if (f == field || !f->as_Field()->is_oop())
1357 continue;
1358 int offs = f->as_Field()->offset();
1359 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1360 add_to_worklist(f);
1361 }
1362 }
1363 }
1364 } else {
1365 assert(base->is_JavaObject(), "sanity");
1366 if (// Skip phantom_object since it is only used to indicate that
1367 // this field's content globally escapes.
1368 (base != phantom_obj) &&
1369 // NULL object node does not have fields.
1370 (base != null_obj)) {
1371 for (EdgeIterator i(base); i.has_next(); i.next()) {
1372 PointsToNode* f = i.get();
1373 // Skip arraycopy edge since store to destination object field
1374 // does not update value in source object field.
1375 if (f->is_Arraycopy()) {
1376 assert(base->arraycopy_dst(), "sanity");
1377 continue;
1378 }
1379 if (f == field || !f->as_Field()->is_oop())
1380 continue;
1381 int offs = f->as_Field()->offset();
1382 if (offs == offset || offset == Type::OffsetBot || offs == Type::OffsetBot) {
1383 add_to_worklist(f);
1384 }
1385 }
1386 }
1387 }
1388 }
1390 // Find fields which have unknown value.
1391 int ConnectionGraph::find_field_value(FieldNode* field) {
1392 // Escaped fields should have init value already.
1393 assert(field->escape_state() == PointsToNode::NoEscape, "sanity");
1394 int new_edges = 0;
1395 for (BaseIterator i(field); i.has_next(); i.next()) {
1396 PointsToNode* base = i.get();
1397 if (base->is_JavaObject()) {
1398 // Skip Allocate's fields which will be processed later.
1399 if (base->ideal_node()->is_Allocate())
1400 return 0;
1401 assert(base == null_obj, "only NULL ptr base expected here");
1402 }
1403 }
1404 if (add_edge(field, phantom_obj)) {
1405 // New edge was added
1406 new_edges++;
1407 add_field_uses_to_worklist(field);
1408 }
1409 return new_edges;
1410 }
1412 // Find fields initializing values for allocations.
1413 int ConnectionGraph::find_init_values(JavaObjectNode* pta, PointsToNode* init_val, PhaseTransform* phase) {
1414 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1415 int new_edges = 0;
1416 Node* alloc = pta->ideal_node();
1417 if (init_val == phantom_obj) {
1418 // Do nothing for Allocate nodes since its fields values are "known".
1419 if (alloc->is_Allocate())
1420 return 0;
1421 assert(alloc->as_CallStaticJava(), "sanity");
1422 #ifdef ASSERT
1423 if (alloc->as_CallStaticJava()->method() == NULL) {
1424 const char* name = alloc->as_CallStaticJava()->_name;
1425 assert(strncmp(name, "_multianewarray", 15) == 0, "sanity");
1426 }
1427 #endif
1428 // Non-escaped allocation returned from Java or runtime call have
1429 // unknown values in fields.
1430 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1431 PointsToNode* field = i.get();
1432 if (field->is_Field() && field->as_Field()->is_oop()) {
1433 if (add_edge(field, phantom_obj)) {
1434 // New edge was added
1435 new_edges++;
1436 add_field_uses_to_worklist(field->as_Field());
1437 }
1438 }
1439 }
1440 return new_edges;
1441 }
1442 assert(init_val == null_obj, "sanity");
1443 // Do nothing for Call nodes since its fields values are unknown.
1444 if (!alloc->is_Allocate())
1445 return 0;
1447 InitializeNode* ini = alloc->as_Allocate()->initialization();
1448 Compile* C = _compile;
1449 bool visited_bottom_offset = false;
1450 GrowableArray<int> offsets_worklist;
1452 // Check if an oop field's initializing value is recorded and add
1453 // a corresponding NULL if field's value if it is not recorded.
1454 // Connection Graph does not record a default initialization by NULL
1455 // captured by Initialize node.
1456 //
1457 for (EdgeIterator i(pta); i.has_next(); i.next()) {
1458 PointsToNode* field = i.get(); // Field (AddP)
1459 if (!field->is_Field() || !field->as_Field()->is_oop())
1460 continue; // Not oop field
1461 int offset = field->as_Field()->offset();
1462 if (offset == Type::OffsetBot) {
1463 if (!visited_bottom_offset) {
1464 // OffsetBot is used to reference array's element,
1465 // always add reference to NULL to all Field nodes since we don't
1466 // known which element is referenced.
1467 if (add_edge(field, null_obj)) {
1468 // New edge was added
1469 new_edges++;
1470 add_field_uses_to_worklist(field->as_Field());
1471 visited_bottom_offset = true;
1472 }
1473 }
1474 } else {
1475 // Check only oop fields.
1476 const Type* adr_type = field->ideal_node()->as_AddP()->bottom_type();
1477 if (adr_type->isa_rawptr()) {
1478 #ifdef ASSERT
1479 // Raw pointers are used for initializing stores so skip it
1480 // since it should be recorded already
1481 Node* base = get_addp_base(field->ideal_node());
1482 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1483 (base->in(0) == alloc),"unexpected pointer type");
1484 #endif
1485 continue;
1486 }
1487 if (!offsets_worklist.contains(offset)) {
1488 offsets_worklist.append(offset);
1489 Node* value = NULL;
1490 if (ini != NULL) {
1491 // StoreP::memory_type() == T_ADDRESS
1492 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_ADDRESS;
1493 Node* store = ini->find_captured_store(offset, type2aelembytes(ft, true), phase);
1494 // Make sure initializing store has the same type as this AddP.
1495 // This AddP may reference non existing field because it is on a
1496 // dead branch of bimorphic call which is not eliminated yet.
1497 if (store != NULL && store->is_Store() &&
1498 store->as_Store()->memory_type() == ft) {
1499 value = store->in(MemNode::ValueIn);
1500 #ifdef ASSERT
1501 if (VerifyConnectionGraph) {
1502 // Verify that AddP already points to all objects the value points to.
1503 PointsToNode* val = ptnode_adr(value->_idx);
1504 assert((val != NULL), "should be processed already");
1505 PointsToNode* missed_obj = NULL;
1506 if (val->is_JavaObject()) {
1507 if (!field->points_to(val->as_JavaObject())) {
1508 missed_obj = val;
1509 }
1510 } else {
1511 if (!val->is_LocalVar() || (val->edge_count() == 0)) {
1512 tty->print_cr("----------init store has invalid value -----");
1513 store->dump();
1514 val->dump();
1515 assert(val->is_LocalVar() && (val->edge_count() > 0), "should be processed already");
1516 }
1517 for (EdgeIterator j(val); j.has_next(); j.next()) {
1518 PointsToNode* obj = j.get();
1519 if (obj->is_JavaObject()) {
1520 if (!field->points_to(obj->as_JavaObject())) {
1521 missed_obj = obj;
1522 break;
1523 }
1524 }
1525 }
1526 }
1527 if (missed_obj != NULL) {
1528 tty->print_cr("----------field---------------------------------");
1529 field->dump();
1530 tty->print_cr("----------missed referernce to object-----------");
1531 missed_obj->dump();
1532 tty->print_cr("----------object referernced by init store -----");
1533 store->dump();
1534 val->dump();
1535 assert(!field->points_to(missed_obj->as_JavaObject()), "missed JavaObject reference");
1536 }
1537 }
1538 #endif
1539 } else {
1540 // There could be initializing stores which follow allocation.
1541 // For example, a volatile field store is not collected
1542 // by Initialize node.
1543 //
1544 // Need to check for dependent loads to separate such stores from
1545 // stores which follow loads. For now, add initial value NULL so
1546 // that compare pointers optimization works correctly.
1547 }
1548 }
1549 if (value == NULL) {
1550 // A field's initializing value was not recorded. Add NULL.
1551 if (add_edge(field, null_obj)) {
1552 // New edge was added
1553 new_edges++;
1554 add_field_uses_to_worklist(field->as_Field());
1555 }
1556 }
1557 }
1558 }
1559 }
1560 return new_edges;
1561 }
1563 // Adjust scalar_replaceable state after Connection Graph is built.
1564 void ConnectionGraph::adjust_scalar_replaceable_state(JavaObjectNode* jobj) {
1565 // Search for non-escaping objects which are not scalar replaceable
1566 // and mark them to propagate the state to referenced objects.
1568 // 1. An object is not scalar replaceable if the field into which it is
1569 // stored has unknown offset (stored into unknown element of an array).
1570 //
1571 for (UseIterator i(jobj); i.has_next(); i.next()) {
1572 PointsToNode* use = i.get();
1573 assert(!use->is_Arraycopy(), "sanity");
1574 if (use->is_Field()) {
1575 FieldNode* field = use->as_Field();
1576 assert(field->is_oop() && field->scalar_replaceable() &&
1577 field->fields_escape_state() == PointsToNode::NoEscape, "sanity");
1578 if (field->offset() == Type::OffsetBot) {
1579 jobj->set_scalar_replaceable(false);
1580 return;
1581 }
1582 // 2. An object is not scalar replaceable if the field into which it is
1583 // stored has multiple bases one of which is null.
1584 if (field->base_count() > 1) {
1585 for (BaseIterator i(field); i.has_next(); i.next()) {
1586 PointsToNode* base = i.get();
1587 if (base == null_obj) {
1588 jobj->set_scalar_replaceable(false);
1589 return;
1590 }
1591 }
1592 }
1593 }
1594 assert(use->is_Field() || use->is_LocalVar(), "sanity");
1595 // 3. An object is not scalar replaceable if it is merged with other objects.
1596 for (EdgeIterator j(use); j.has_next(); j.next()) {
1597 PointsToNode* ptn = j.get();
1598 if (ptn->is_JavaObject() && ptn != jobj) {
1599 // Mark all objects.
1600 jobj->set_scalar_replaceable(false);
1601 ptn->set_scalar_replaceable(false);
1602 }
1603 }
1604 if (!jobj->scalar_replaceable()) {
1605 return;
1606 }
1607 }
1609 for (EdgeIterator j(jobj); j.has_next(); j.next()) {
1610 // Non-escaping object node should point only to field nodes.
1611 FieldNode* field = j.get()->as_Field();
1612 int offset = field->as_Field()->offset();
1614 // 4. An object is not scalar replaceable if it has a field with unknown
1615 // offset (array's element is accessed in loop).
1616 if (offset == Type::OffsetBot) {
1617 jobj->set_scalar_replaceable(false);
1618 return;
1619 }
1620 // 5. Currently an object is not scalar replaceable if a LoadStore node
1621 // access its field since the field value is unknown after it.
1622 //
1623 Node* n = field->ideal_node();
1624 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1625 if (n->fast_out(i)->is_LoadStore()) {
1626 jobj->set_scalar_replaceable(false);
1627 return;
1628 }
1629 }
1631 // 6. Or the address may point to more then one object. This may produce
1632 // the false positive result (set not scalar replaceable)
1633 // since the flow-insensitive escape analysis can't separate
1634 // the case when stores overwrite the field's value from the case
1635 // when stores happened on different control branches.
1636 //
1637 // Note: it will disable scalar replacement in some cases:
1638 //
1639 // Point p[] = new Point[1];
1640 // p[0] = new Point(); // Will be not scalar replaced
1641 //
1642 // but it will save us from incorrect optimizations in next cases:
1643 //
1644 // Point p[] = new Point[1];
1645 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1646 //
1647 if (field->base_count() > 1) {
1648 for (BaseIterator i(field); i.has_next(); i.next()) {
1649 PointsToNode* base = i.get();
1650 // Don't take into account LocalVar nodes which
1651 // may point to only one object which should be also
1652 // this field's base by now.
1653 if (base->is_JavaObject() && base != jobj) {
1654 // Mark all bases.
1655 jobj->set_scalar_replaceable(false);
1656 base->set_scalar_replaceable(false);
1657 }
1658 }
1659 }
1660 }
1661 }
1663 #ifdef ASSERT
1664 void ConnectionGraph::verify_connection_graph(
1665 GrowableArray<PointsToNode*>& ptnodes_worklist,
1666 GrowableArray<JavaObjectNode*>& non_escaped_worklist,
1667 GrowableArray<JavaObjectNode*>& java_objects_worklist,
1668 GrowableArray<Node*>& addp_worklist) {
1669 // Verify that graph is complete - no new edges could be added.
1670 int java_objects_length = java_objects_worklist.length();
1671 int non_escaped_length = non_escaped_worklist.length();
1672 int new_edges = 0;
1673 for (int next = 0; next < java_objects_length; ++next) {
1674 JavaObjectNode* ptn = java_objects_worklist.at(next);
1675 new_edges += add_java_object_edges(ptn, true);
1676 }
1677 assert(new_edges == 0, "graph was not complete");
1678 // Verify that escape state is final.
1679 int length = non_escaped_worklist.length();
1680 find_non_escaped_objects(ptnodes_worklist, non_escaped_worklist);
1681 assert((non_escaped_length == non_escaped_worklist.length()) &&
1682 (non_escaped_length == length) &&
1683 (_worklist.length() == 0), "escape state was not final");
1685 // Verify fields information.
1686 int addp_length = addp_worklist.length();
1687 for (int next = 0; next < addp_length; ++next ) {
1688 Node* n = addp_worklist.at(next);
1689 FieldNode* field = ptnode_adr(n->_idx)->as_Field();
1690 if (field->is_oop()) {
1691 // Verify that field has all bases
1692 Node* base = get_addp_base(n);
1693 PointsToNode* ptn = ptnode_adr(base->_idx);
1694 if (ptn->is_JavaObject()) {
1695 assert(field->has_base(ptn->as_JavaObject()), "sanity");
1696 } else {
1697 assert(ptn->is_LocalVar(), "sanity");
1698 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
1699 PointsToNode* e = i.get();
1700 if (e->is_JavaObject()) {
1701 assert(field->has_base(e->as_JavaObject()), "sanity");
1702 }
1703 }
1704 }
1705 // Verify that all fields have initializing values.
1706 if (field->edge_count() == 0) {
1707 tty->print_cr("----------field does not have references----------");
1708 field->dump();
1709 for (BaseIterator i(field); i.has_next(); i.next()) {
1710 PointsToNode* base = i.get();
1711 tty->print_cr("----------field has next base---------------------");
1712 base->dump();
1713 if (base->is_JavaObject() && (base != phantom_obj) && (base != null_obj)) {
1714 tty->print_cr("----------base has fields-------------------------");
1715 for (EdgeIterator j(base); j.has_next(); j.next()) {
1716 j.get()->dump();
1717 }
1718 tty->print_cr("----------base has references---------------------");
1719 for (UseIterator j(base); j.has_next(); j.next()) {
1720 j.get()->dump();
1721 }
1722 }
1723 }
1724 for (UseIterator i(field); i.has_next(); i.next()) {
1725 i.get()->dump();
1726 }
1727 assert(field->edge_count() > 0, "sanity");
1728 }
1729 }
1730 }
1731 }
1732 #endif
1734 // Optimize ideal graph.
1735 void ConnectionGraph::optimize_ideal_graph(GrowableArray<Node*>& ptr_cmp_worklist,
1736 GrowableArray<Node*>& storestore_worklist) {
1737 Compile* C = _compile;
1738 PhaseIterGVN* igvn = _igvn;
1739 if (EliminateLocks) {
1740 // Mark locks before changing ideal graph.
1741 int cnt = C->macro_count();
1742 for( int i=0; i < cnt; i++ ) {
1743 Node *n = C->macro_node(i);
1744 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1745 AbstractLockNode* alock = n->as_AbstractLock();
1746 if (!alock->is_non_esc_obj()) {
1747 if (not_global_escape(alock->obj_node())) {
1748 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1749 // The lock could be marked eliminated by lock coarsening
1750 // code during first IGVN before EA. Replace coarsened flag
1751 // to eliminate all associated locks/unlocks.
1752 alock->set_non_esc_obj();
1753 }
1754 }
1755 }
1756 }
1757 }
1759 if (OptimizePtrCompare) {
1760 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1761 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1762 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1763 // Optimize objects compare.
1764 while (ptr_cmp_worklist.length() != 0) {
1765 Node *n = ptr_cmp_worklist.pop();
1766 Node *res = optimize_ptr_compare(n);
1767 if (res != NULL) {
1768 #ifndef PRODUCT
1769 if (PrintOptimizePtrCompare) {
1770 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, (res == _pcmp_eq ? "EQ" : "NotEQ"));
1771 if (Verbose) {
1772 n->dump(1);
1773 }
1774 }
1775 #endif
1776 igvn->replace_node(n, res);
1777 }
1778 }
1779 // cleanup
1780 if (_pcmp_neq->outcnt() == 0)
1781 igvn->hash_delete(_pcmp_neq);
1782 if (_pcmp_eq->outcnt() == 0)
1783 igvn->hash_delete(_pcmp_eq);
1784 }
1786 // For MemBarStoreStore nodes added in library_call.cpp, check
1787 // escape status of associated AllocateNode and optimize out
1788 // MemBarStoreStore node if the allocated object never escapes.
1789 while (storestore_worklist.length() != 0) {
1790 Node *n = storestore_worklist.pop();
1791 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1792 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1793 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1794 if (not_global_escape(alloc)) {
1795 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1796 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1797 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1798 igvn->register_new_node_with_optimizer(mb);
1799 igvn->replace_node(storestore, mb);
1800 }
1801 }
1802 }
1804 // Optimize objects compare.
1805 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
1806 assert(OptimizePtrCompare, "sanity");
1807 PointsToNode* ptn1 = ptnode_adr(n->in(1)->_idx);
1808 PointsToNode* ptn2 = ptnode_adr(n->in(2)->_idx);
1809 JavaObjectNode* jobj1 = unique_java_object(n->in(1));
1810 JavaObjectNode* jobj2 = unique_java_object(n->in(2));
1811 assert(ptn1->is_JavaObject() || ptn1->is_LocalVar(), "sanity");
1812 assert(ptn2->is_JavaObject() || ptn2->is_LocalVar(), "sanity");
1814 // Check simple cases first.
1815 if (jobj1 != NULL) {
1816 if (jobj1->escape_state() == PointsToNode::NoEscape) {
1817 if (jobj1 == jobj2) {
1818 // Comparing the same not escaping object.
1819 return _pcmp_eq;
1820 }
1821 Node* obj = jobj1->ideal_node();
1822 // Comparing not escaping allocation.
1823 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1824 !ptn2->points_to(jobj1)) {
1825 return _pcmp_neq; // This includes nullness check.
1826 }
1827 }
1828 }
1829 if (jobj2 != NULL) {
1830 if (jobj2->escape_state() == PointsToNode::NoEscape) {
1831 Node* obj = jobj2->ideal_node();
1832 // Comparing not escaping allocation.
1833 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
1834 !ptn1->points_to(jobj2)) {
1835 return _pcmp_neq; // This includes nullness check.
1836 }
1837 }
1838 }
1839 if (jobj1 != NULL && jobj1 != phantom_obj &&
1840 jobj2 != NULL && jobj2 != phantom_obj &&
1841 jobj1->ideal_node()->is_Con() &&
1842 jobj2->ideal_node()->is_Con()) {
1843 // Klass or String constants compare. Need to be careful with
1844 // compressed pointers - compare types of ConN and ConP instead of nodes.
1845 const Type* t1 = jobj1->ideal_node()->get_ptr_type();
1846 const Type* t2 = jobj2->ideal_node()->get_ptr_type();
1847 if (t1->make_ptr() == t2->make_ptr()) {
1848 return _pcmp_eq;
1849 } else {
1850 return _pcmp_neq;
1851 }
1852 }
1853 if (ptn1->meet(ptn2)) {
1854 return NULL; // Sets are not disjoint
1855 }
1857 // Sets are disjoint.
1858 bool set1_has_unknown_ptr = ptn1->points_to(phantom_obj);
1859 bool set2_has_unknown_ptr = ptn2->points_to(phantom_obj);
1860 bool set1_has_null_ptr = ptn1->points_to(null_obj);
1861 bool set2_has_null_ptr = ptn2->points_to(null_obj);
1862 if (set1_has_unknown_ptr && set2_has_null_ptr ||
1863 set2_has_unknown_ptr && set1_has_null_ptr) {
1864 // Check nullness of unknown object.
1865 return NULL;
1866 }
1868 // Disjointness by itself is not sufficient since
1869 // alias analysis is not complete for escaped objects.
1870 // Disjoint sets are definitely unrelated only when
1871 // at least one set has only not escaping allocations.
1872 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
1873 if (ptn1->non_escaping_allocation()) {
1874 return _pcmp_neq;
1875 }
1876 }
1877 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
1878 if (ptn2->non_escaping_allocation()) {
1879 return _pcmp_neq;
1880 }
1881 }
1882 return NULL;
1883 }
1885 // Connection Graph constuction functions.
1887 void ConnectionGraph::add_local_var(Node *n, PointsToNode::EscapeState es) {
1888 PointsToNode* ptadr = _nodes.at(n->_idx);
1889 if (ptadr != NULL) {
1890 assert(ptadr->is_LocalVar() && ptadr->ideal_node() == n, "sanity");
1891 return;
1892 }
1893 Compile* C = _compile;
1894 ptadr = new (C->comp_arena()) LocalVarNode(C, n, es);
1895 _nodes.at_put(n->_idx, ptadr);
1896 }
1898 void ConnectionGraph::add_java_object(Node *n, PointsToNode::EscapeState es) {
1899 PointsToNode* ptadr = _nodes.at(n->_idx);
1900 if (ptadr != NULL) {
1901 assert(ptadr->is_JavaObject() && ptadr->ideal_node() == n, "sanity");
1902 return;
1903 }
1904 Compile* C = _compile;
1905 ptadr = new (C->comp_arena()) JavaObjectNode(C, n, es);
1906 _nodes.at_put(n->_idx, ptadr);
1907 }
1909 void ConnectionGraph::add_field(Node *n, PointsToNode::EscapeState es, int offset) {
1910 PointsToNode* ptadr = _nodes.at(n->_idx);
1911 if (ptadr != NULL) {
1912 assert(ptadr->is_Field() && ptadr->ideal_node() == n, "sanity");
1913 return;
1914 }
1915 bool unsafe = false;
1916 bool is_oop = is_oop_field(n, offset, &unsafe);
1917 if (unsafe) {
1918 es = PointsToNode::GlobalEscape;
1919 }
1920 Compile* C = _compile;
1921 FieldNode* field = new (C->comp_arena()) FieldNode(C, n, es, offset, is_oop);
1922 _nodes.at_put(n->_idx, field);
1923 }
1925 void ConnectionGraph::add_arraycopy(Node *n, PointsToNode::EscapeState es,
1926 PointsToNode* src, PointsToNode* dst) {
1927 assert(!src->is_Field() && !dst->is_Field(), "only for JavaObject and LocalVar");
1928 assert((src != null_obj) && (dst != null_obj), "not for ConP NULL");
1929 PointsToNode* ptadr = _nodes.at(n->_idx);
1930 if (ptadr != NULL) {
1931 assert(ptadr->is_Arraycopy() && ptadr->ideal_node() == n, "sanity");
1932 return;
1933 }
1934 Compile* C = _compile;
1935 ptadr = new (C->comp_arena()) ArraycopyNode(C, n, es);
1936 _nodes.at_put(n->_idx, ptadr);
1937 // Add edge from arraycopy node to source object.
1938 (void)add_edge(ptadr, src);
1939 src->set_arraycopy_src();
1940 // Add edge from destination object to arraycopy node.
1941 (void)add_edge(dst, ptadr);
1942 dst->set_arraycopy_dst();
1943 }
1945 bool ConnectionGraph::is_oop_field(Node* n, int offset, bool* unsafe) {
1946 const Type* adr_type = n->as_AddP()->bottom_type();
1947 BasicType bt = T_INT;
1948 if (offset == Type::OffsetBot) {
1949 // Check only oop fields.
1950 if (!adr_type->isa_aryptr() ||
1951 (adr_type->isa_aryptr()->klass() == NULL) ||
1952 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1953 // OffsetBot is used to reference array's element. Ignore first AddP.
1954 if (find_second_addp(n, n->in(AddPNode::Base)) == NULL) {
1955 bt = T_OBJECT;
1956 }
1957 }
1958 } else if (offset != oopDesc::klass_offset_in_bytes()) {
1959 if (adr_type->isa_instptr()) {
1960 ciField* field = _compile->alias_type(adr_type->isa_instptr())->field();
1961 if (field != NULL) {
1962 bt = field->layout_type();
1963 } else {
1964 // Check for unsafe oop field access
1965 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1966 int opcode = n->fast_out(i)->Opcode();
1967 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1968 opcode == Op_StoreN || opcode == Op_LoadN) {
1969 bt = T_OBJECT;
1970 (*unsafe) = true;
1971 break;
1972 }
1973 }
1974 }
1975 } else if (adr_type->isa_aryptr()) {
1976 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1977 // Ignore array length load.
1978 } else if (find_second_addp(n, n->in(AddPNode::Base)) != NULL) {
1979 // Ignore first AddP.
1980 } else {
1981 const Type* elemtype = adr_type->isa_aryptr()->elem();
1982 bt = elemtype->array_element_basic_type();
1983 }
1984 } else if (adr_type->isa_rawptr() || adr_type->isa_klassptr()) {
1985 // Allocation initialization, ThreadLocal field access, unsafe access
1986 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1987 int opcode = n->fast_out(i)->Opcode();
1988 if (opcode == Op_StoreP || opcode == Op_LoadP ||
1989 opcode == Op_StoreN || opcode == Op_LoadN) {
1990 bt = T_OBJECT;
1991 break;
1992 }
1993 }
1994 }
1995 }
1996 return (bt == T_OBJECT || bt == T_NARROWOOP || bt == T_ARRAY);
1997 }
1999 // Returns unique pointed java object or NULL.
2000 JavaObjectNode* ConnectionGraph::unique_java_object(Node *n) {
2001 assert(!_collecting, "should not call when contructed graph");
2002 // If the node was created after the escape computation we can't answer.
2003 uint idx = n->_idx;
2004 if (idx >= nodes_size()) {
2005 return NULL;
2006 }
2007 PointsToNode* ptn = ptnode_adr(idx);
2008 if (ptn->is_JavaObject()) {
2009 return ptn->as_JavaObject();
2010 }
2011 assert(ptn->is_LocalVar(), "sanity");
2012 // Check all java objects it points to.
2013 JavaObjectNode* jobj = NULL;
2014 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2015 PointsToNode* e = i.get();
2016 if (e->is_JavaObject()) {
2017 if (jobj == NULL) {
2018 jobj = e->as_JavaObject();
2019 } else if (jobj != e) {
2020 return NULL;
2021 }
2022 }
2023 }
2024 return jobj;
2025 }
2027 // Return true if this node points only to non-escaping allocations.
2028 bool PointsToNode::non_escaping_allocation() {
2029 if (is_JavaObject()) {
2030 Node* n = ideal_node();
2031 if (n->is_Allocate() || n->is_CallStaticJava()) {
2032 return (escape_state() == PointsToNode::NoEscape);
2033 } else {
2034 return false;
2035 }
2036 }
2037 assert(is_LocalVar(), "sanity");
2038 // Check all java objects it points to.
2039 for (EdgeIterator i(this); i.has_next(); i.next()) {
2040 PointsToNode* e = i.get();
2041 if (e->is_JavaObject()) {
2042 Node* n = e->ideal_node();
2043 if ((e->escape_state() != PointsToNode::NoEscape) ||
2044 !(n->is_Allocate() || n->is_CallStaticJava())) {
2045 return false;
2046 }
2047 }
2048 }
2049 return true;
2050 }
2052 // Return true if we know the node does not escape globally.
2053 bool ConnectionGraph::not_global_escape(Node *n) {
2054 assert(!_collecting, "should not call during graph construction");
2055 // If the node was created after the escape computation we can't answer.
2056 uint idx = n->_idx;
2057 if (idx >= nodes_size()) {
2058 return false;
2059 }
2060 PointsToNode* ptn = ptnode_adr(idx);
2061 PointsToNode::EscapeState es = ptn->escape_state();
2062 // If we have already computed a value, return it.
2063 if (es >= PointsToNode::GlobalEscape)
2064 return false;
2065 if (ptn->is_JavaObject()) {
2066 return true; // (es < PointsToNode::GlobalEscape);
2067 }
2068 assert(ptn->is_LocalVar(), "sanity");
2069 // Check all java objects it points to.
2070 for (EdgeIterator i(ptn); i.has_next(); i.next()) {
2071 if (i.get()->escape_state() >= PointsToNode::GlobalEscape)
2072 return false;
2073 }
2074 return true;
2075 }
2078 // Helper functions
2080 // Return true if this node points to specified node or nodes it points to.
2081 bool PointsToNode::points_to(JavaObjectNode* ptn) const {
2082 if (is_JavaObject()) {
2083 return (this == ptn);
2084 }
2085 assert(is_LocalVar() || is_Field(), "sanity");
2086 for (EdgeIterator i(this); i.has_next(); i.next()) {
2087 if (i.get() == ptn)
2088 return true;
2089 }
2090 return false;
2091 }
2093 // Return true if one node points to an other.
2094 bool PointsToNode::meet(PointsToNode* ptn) {
2095 if (this == ptn) {
2096 return true;
2097 } else if (ptn->is_JavaObject()) {
2098 return this->points_to(ptn->as_JavaObject());
2099 } else if (this->is_JavaObject()) {
2100 return ptn->points_to(this->as_JavaObject());
2101 }
2102 assert(this->is_LocalVar() && ptn->is_LocalVar(), "sanity");
2103 int ptn_count = ptn->edge_count();
2104 for (EdgeIterator i(this); i.has_next(); i.next()) {
2105 PointsToNode* this_e = i.get();
2106 for (int j = 0; j < ptn_count; j++) {
2107 if (this_e == ptn->edge(j))
2108 return true;
2109 }
2110 }
2111 return false;
2112 }
2114 #ifdef ASSERT
2115 // Return true if bases point to this java object.
2116 bool FieldNode::has_base(JavaObjectNode* jobj) const {
2117 for (BaseIterator i(this); i.has_next(); i.next()) {
2118 if (i.get() == jobj)
2119 return true;
2120 }
2121 return false;
2122 }
2123 #endif
2125 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
2126 const Type *adr_type = phase->type(adr);
2127 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
2128 adr->in(AddPNode::Address)->is_Proj() &&
2129 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2130 // We are computing a raw address for a store captured by an Initialize
2131 // compute an appropriate address type. AddP cases #3 and #5 (see below).
2132 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2133 assert(offs != Type::OffsetBot ||
2134 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
2135 "offset must be a constant or it is initialization of array");
2136 return offs;
2137 }
2138 const TypePtr *t_ptr = adr_type->isa_ptr();
2139 assert(t_ptr != NULL, "must be a pointer type");
2140 return t_ptr->offset();
2141 }
2143 Node* ConnectionGraph::get_addp_base(Node *addp) {
2144 assert(addp->is_AddP(), "must be AddP");
2145 //
2146 // AddP cases for Base and Address inputs:
2147 // case #1. Direct object's field reference:
2148 // Allocate
2149 // |
2150 // Proj #5 ( oop result )
2151 // |
2152 // CheckCastPP (cast to instance type)
2153 // | |
2154 // AddP ( base == address )
2155 //
2156 // case #2. Indirect object's field reference:
2157 // Phi
2158 // |
2159 // CastPP (cast to instance type)
2160 // | |
2161 // AddP ( base == address )
2162 //
2163 // case #3. Raw object's field reference for Initialize node:
2164 // Allocate
2165 // |
2166 // Proj #5 ( oop result )
2167 // top |
2168 // \ |
2169 // AddP ( base == top )
2170 //
2171 // case #4. Array's element reference:
2172 // {CheckCastPP | CastPP}
2173 // | | |
2174 // | AddP ( array's element offset )
2175 // | |
2176 // AddP ( array's offset )
2177 //
2178 // case #5. Raw object's field reference for arraycopy stub call:
2179 // The inline_native_clone() case when the arraycopy stub is called
2180 // after the allocation before Initialize and CheckCastPP nodes.
2181 // Allocate
2182 // |
2183 // Proj #5 ( oop result )
2184 // | |
2185 // AddP ( base == address )
2186 //
2187 // case #6. Constant Pool, ThreadLocal, CastX2P or
2188 // Raw object's field reference:
2189 // {ConP, ThreadLocal, CastX2P, raw Load}
2190 // top |
2191 // \ |
2192 // AddP ( base == top )
2193 //
2194 // case #7. Klass's field reference.
2195 // LoadKlass
2196 // | |
2197 // AddP ( base == address )
2198 //
2199 // case #8. narrow Klass's field reference.
2200 // LoadNKlass
2201 // |
2202 // DecodeN
2203 // | |
2204 // AddP ( base == address )
2205 //
2206 Node *base = addp->in(AddPNode::Base);
2207 if (base->uncast()->is_top()) { // The AddP case #3 and #6.
2208 base = addp->in(AddPNode::Address);
2209 while (base->is_AddP()) {
2210 // Case #6 (unsafe access) may have several chained AddP nodes.
2211 assert(base->in(AddPNode::Base)->uncast()->is_top(), "expected unsafe access address only");
2212 base = base->in(AddPNode::Address);
2213 }
2214 Node* uncast_base = base->uncast();
2215 int opcode = uncast_base->Opcode();
2216 assert(opcode == Op_ConP || opcode == Op_ThreadLocal ||
2217 opcode == Op_CastX2P || uncast_base->is_DecodeNarrowPtr() ||
2218 (uncast_base->is_Mem() && (uncast_base->bottom_type()->isa_rawptr() != NULL)) ||
2219 (uncast_base->is_Proj() && uncast_base->in(0)->is_Allocate()), "sanity");
2220 }
2221 return base;
2222 }
2224 Node* ConnectionGraph::find_second_addp(Node* addp, Node* n) {
2225 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
2226 Node* addp2 = addp->raw_out(0);
2227 if (addp->outcnt() == 1 && addp2->is_AddP() &&
2228 addp2->in(AddPNode::Base) == n &&
2229 addp2->in(AddPNode::Address) == addp) {
2230 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
2231 //
2232 // Find array's offset to push it on worklist first and
2233 // as result process an array's element offset first (pushed second)
2234 // to avoid CastPP for the array's offset.
2235 // Otherwise the inserted CastPP (LocalVar) will point to what
2236 // the AddP (Field) points to. Which would be wrong since
2237 // the algorithm expects the CastPP has the same point as
2238 // as AddP's base CheckCastPP (LocalVar).
2239 //
2240 // ArrayAllocation
2241 // |
2242 // CheckCastPP
2243 // |
2244 // memProj (from ArrayAllocation CheckCastPP)
2245 // | ||
2246 // | || Int (element index)
2247 // | || | ConI (log(element size))
2248 // | || | /
2249 // | || LShift
2250 // | || /
2251 // | AddP (array's element offset)
2252 // | |
2253 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
2254 // | / /
2255 // AddP (array's offset)
2256 // |
2257 // Load/Store (memory operation on array's element)
2258 //
2259 return addp2;
2260 }
2261 return NULL;
2262 }
2264 //
2265 // Adjust the type and inputs of an AddP which computes the
2266 // address of a field of an instance
2267 //
2268 bool ConnectionGraph::split_AddP(Node *addp, Node *base) {
2269 PhaseGVN* igvn = _igvn;
2270 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
2271 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
2272 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
2273 if (t == NULL) {
2274 // We are computing a raw address for a store captured by an Initialize
2275 // compute an appropriate address type (cases #3 and #5).
2276 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
2277 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
2278 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
2279 assert(offs != Type::OffsetBot, "offset must be a constant");
2280 t = base_t->add_offset(offs)->is_oopptr();
2281 }
2282 int inst_id = base_t->instance_id();
2283 assert(!t->is_known_instance() || t->instance_id() == inst_id,
2284 "old type must be non-instance or match new type");
2286 // The type 't' could be subclass of 'base_t'.
2287 // As result t->offset() could be large then base_t's size and it will
2288 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
2289 // constructor verifies correctness of the offset.
2290 //
2291 // It could happened on subclass's branch (from the type profiling
2292 // inlining) which was not eliminated during parsing since the exactness
2293 // of the allocation type was not propagated to the subclass type check.
2294 //
2295 // Or the type 't' could be not related to 'base_t' at all.
2296 // It could happened when CHA type is different from MDO type on a dead path
2297 // (for example, from instanceof check) which is not collapsed during parsing.
2298 //
2299 // Do nothing for such AddP node and don't process its users since
2300 // this code branch will go away.
2301 //
2302 if (!t->is_known_instance() &&
2303 !base_t->klass()->is_subtype_of(t->klass())) {
2304 return false; // bail out
2305 }
2306 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
2307 // Do NOT remove the next line: ensure a new alias index is allocated
2308 // for the instance type. Note: C++ will not remove it since the call
2309 // has side effect.
2310 int alias_idx = _compile->get_alias_index(tinst);
2311 igvn->set_type(addp, tinst);
2312 // record the allocation in the node map
2313 set_map(addp, get_map(base->_idx));
2314 // Set addp's Base and Address to 'base'.
2315 Node *abase = addp->in(AddPNode::Base);
2316 Node *adr = addp->in(AddPNode::Address);
2317 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
2318 adr->in(0)->_idx == (uint)inst_id) {
2319 // Skip AddP cases #3 and #5.
2320 } else {
2321 assert(!abase->is_top(), "sanity"); // AddP case #3
2322 if (abase != base) {
2323 igvn->hash_delete(addp);
2324 addp->set_req(AddPNode::Base, base);
2325 if (abase == adr) {
2326 addp->set_req(AddPNode::Address, base);
2327 } else {
2328 // AddP case #4 (adr is array's element offset AddP node)
2329 #ifdef ASSERT
2330 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
2331 assert(adr->is_AddP() && atype != NULL &&
2332 atype->instance_id() == inst_id, "array's element offset should be processed first");
2333 #endif
2334 }
2335 igvn->hash_insert(addp);
2336 }
2337 }
2338 // Put on IGVN worklist since at least addp's type was changed above.
2339 record_for_optimizer(addp);
2340 return true;
2341 }
2343 //
2344 // Create a new version of orig_phi if necessary. Returns either the newly
2345 // created phi or an existing phi. Sets create_new to indicate whether a new
2346 // phi was created. Cache the last newly created phi in the node map.
2347 //
2348 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, bool &new_created) {
2349 Compile *C = _compile;
2350 PhaseGVN* igvn = _igvn;
2351 new_created = false;
2352 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
2353 // nothing to do if orig_phi is bottom memory or matches alias_idx
2354 if (phi_alias_idx == alias_idx) {
2355 return orig_phi;
2356 }
2357 // Have we recently created a Phi for this alias index?
2358 PhiNode *result = get_map_phi(orig_phi->_idx);
2359 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
2360 return result;
2361 }
2362 // Previous check may fail when the same wide memory Phi was split into Phis
2363 // for different memory slices. Search all Phis for this region.
2364 if (result != NULL) {
2365 Node* region = orig_phi->in(0);
2366 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2367 Node* phi = region->fast_out(i);
2368 if (phi->is_Phi() &&
2369 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
2370 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
2371 return phi->as_Phi();
2372 }
2373 }
2374 }
2375 if ((int) (C->live_nodes() + 2*NodeLimitFudgeFactor) > MaxNodeLimit) {
2376 if (C->do_escape_analysis() == true && !C->failing()) {
2377 // Retry compilation without escape analysis.
2378 // If this is the first failure, the sentinel string will "stick"
2379 // to the Compile object, and the C2Compiler will see it and retry.
2380 C->record_failure(C2Compiler::retry_no_escape_analysis());
2381 }
2382 return NULL;
2383 }
2384 orig_phi_worklist.append_if_missing(orig_phi);
2385 const TypePtr *atype = C->get_adr_type(alias_idx);
2386 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
2387 C->copy_node_notes_to(result, orig_phi);
2388 igvn->set_type(result, result->bottom_type());
2389 record_for_optimizer(result);
2390 set_map(orig_phi, result);
2391 new_created = true;
2392 return result;
2393 }
2395 //
2396 // Return a new version of Memory Phi "orig_phi" with the inputs having the
2397 // specified alias index.
2398 //
2399 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist) {
2400 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
2401 Compile *C = _compile;
2402 PhaseGVN* igvn = _igvn;
2403 bool new_phi_created;
2404 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, new_phi_created);
2405 if (!new_phi_created) {
2406 return result;
2407 }
2408 GrowableArray<PhiNode *> phi_list;
2409 GrowableArray<uint> cur_input;
2410 PhiNode *phi = orig_phi;
2411 uint idx = 1;
2412 bool finished = false;
2413 while(!finished) {
2414 while (idx < phi->req()) {
2415 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist);
2416 if (mem != NULL && mem->is_Phi()) {
2417 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, new_phi_created);
2418 if (new_phi_created) {
2419 // found an phi for which we created a new split, push current one on worklist and begin
2420 // processing new one
2421 phi_list.push(phi);
2422 cur_input.push(idx);
2423 phi = mem->as_Phi();
2424 result = newphi;
2425 idx = 1;
2426 continue;
2427 } else {
2428 mem = newphi;
2429 }
2430 }
2431 if (C->failing()) {
2432 return NULL;
2433 }
2434 result->set_req(idx++, mem);
2435 }
2436 #ifdef ASSERT
2437 // verify that the new Phi has an input for each input of the original
2438 assert( phi->req() == result->req(), "must have same number of inputs.");
2439 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
2440 #endif
2441 // Check if all new phi's inputs have specified alias index.
2442 // Otherwise use old phi.
2443 for (uint i = 1; i < phi->req(); i++) {
2444 Node* in = result->in(i);
2445 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
2446 }
2447 // we have finished processing a Phi, see if there are any more to do
2448 finished = (phi_list.length() == 0 );
2449 if (!finished) {
2450 phi = phi_list.pop();
2451 idx = cur_input.pop();
2452 PhiNode *prev_result = get_map_phi(phi->_idx);
2453 prev_result->set_req(idx++, result);
2454 result = prev_result;
2455 }
2456 }
2457 return result;
2458 }
2460 //
2461 // The next methods are derived from methods in MemNode.
2462 //
2463 Node* ConnectionGraph::step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
2464 Node *mem = mmem;
2465 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
2466 // means an array I have not precisely typed yet. Do not do any
2467 // alias stuff with it any time soon.
2468 if (toop->base() != Type::AnyPtr &&
2469 !(toop->klass() != NULL &&
2470 toop->klass()->is_java_lang_Object() &&
2471 toop->offset() == Type::OffsetBot)) {
2472 mem = mmem->memory_at(alias_idx);
2473 // Update input if it is progress over what we have now
2474 }
2475 return mem;
2476 }
2478 //
2479 // Move memory users to their memory slices.
2480 //
2481 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis) {
2482 Compile* C = _compile;
2483 PhaseGVN* igvn = _igvn;
2484 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
2485 assert(tp != NULL, "ptr type");
2486 int alias_idx = C->get_alias_index(tp);
2487 int general_idx = C->get_general_index(alias_idx);
2489 // Move users first
2490 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2491 Node* use = n->fast_out(i);
2492 if (use->is_MergeMem()) {
2493 MergeMemNode* mmem = use->as_MergeMem();
2494 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
2495 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
2496 continue; // Nothing to do
2497 }
2498 // Replace previous general reference to mem node.
2499 uint orig_uniq = C->unique();
2500 Node* m = find_inst_mem(n, general_idx, orig_phis);
2501 assert(orig_uniq == C->unique(), "no new nodes");
2502 mmem->set_memory_at(general_idx, m);
2503 --imax;
2504 --i;
2505 } else if (use->is_MemBar()) {
2506 assert(!use->is_Initialize(), "initializing stores should not be moved");
2507 if (use->req() > MemBarNode::Precedent &&
2508 use->in(MemBarNode::Precedent) == n) {
2509 // Don't move related membars.
2510 record_for_optimizer(use);
2511 continue;
2512 }
2513 tp = use->as_MemBar()->adr_type()->isa_ptr();
2514 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
2515 alias_idx == general_idx) {
2516 continue; // Nothing to do
2517 }
2518 // Move to general memory slice.
2519 uint orig_uniq = C->unique();
2520 Node* m = find_inst_mem(n, general_idx, orig_phis);
2521 assert(orig_uniq == C->unique(), "no new nodes");
2522 igvn->hash_delete(use);
2523 imax -= use->replace_edge(n, m);
2524 igvn->hash_insert(use);
2525 record_for_optimizer(use);
2526 --i;
2527 #ifdef ASSERT
2528 } else if (use->is_Mem()) {
2529 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
2530 // Don't move related cardmark.
2531 continue;
2532 }
2533 // Memory nodes should have new memory input.
2534 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
2535 assert(tp != NULL, "ptr type");
2536 int idx = C->get_alias_index(tp);
2537 assert(get_map(use->_idx) != NULL || idx == alias_idx,
2538 "Following memory nodes should have new memory input or be on the same memory slice");
2539 } else if (use->is_Phi()) {
2540 // Phi nodes should be split and moved already.
2541 tp = use->as_Phi()->adr_type()->isa_ptr();
2542 assert(tp != NULL, "ptr type");
2543 int idx = C->get_alias_index(tp);
2544 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
2545 } else {
2546 use->dump();
2547 assert(false, "should not be here");
2548 #endif
2549 }
2550 }
2551 }
2553 //
2554 // Search memory chain of "mem" to find a MemNode whose address
2555 // is the specified alias index.
2556 //
2557 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis) {
2558 if (orig_mem == NULL)
2559 return orig_mem;
2560 Compile* C = _compile;
2561 PhaseGVN* igvn = _igvn;
2562 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
2563 bool is_instance = (toop != NULL) && toop->is_known_instance();
2564 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
2565 Node *prev = NULL;
2566 Node *result = orig_mem;
2567 while (prev != result) {
2568 prev = result;
2569 if (result == start_mem)
2570 break; // hit one of our sentinels
2571 if (result->is_Mem()) {
2572 const Type *at = igvn->type(result->in(MemNode::Address));
2573 if (at == Type::TOP)
2574 break; // Dead
2575 assert (at->isa_ptr() != NULL, "pointer type required.");
2576 int idx = C->get_alias_index(at->is_ptr());
2577 if (idx == alias_idx)
2578 break; // Found
2579 if (!is_instance && (at->isa_oopptr() == NULL ||
2580 !at->is_oopptr()->is_known_instance())) {
2581 break; // Do not skip store to general memory slice.
2582 }
2583 result = result->in(MemNode::Memory);
2584 }
2585 if (!is_instance)
2586 continue; // don't search further for non-instance types
2587 // skip over a call which does not affect this memory slice
2588 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
2589 Node *proj_in = result->in(0);
2590 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
2591 break; // hit one of our sentinels
2592 } else if (proj_in->is_Call()) {
2593 CallNode *call = proj_in->as_Call();
2594 if (!call->may_modify(toop, igvn)) {
2595 result = call->in(TypeFunc::Memory);
2596 }
2597 } else if (proj_in->is_Initialize()) {
2598 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
2599 // Stop if this is the initialization for the object instance which
2600 // which contains this memory slice, otherwise skip over it.
2601 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
2602 result = proj_in->in(TypeFunc::Memory);
2603 }
2604 } else if (proj_in->is_MemBar()) {
2605 result = proj_in->in(TypeFunc::Memory);
2606 }
2607 } else if (result->is_MergeMem()) {
2608 MergeMemNode *mmem = result->as_MergeMem();
2609 result = step_through_mergemem(mmem, alias_idx, toop);
2610 if (result == mmem->base_memory()) {
2611 // Didn't find instance memory, search through general slice recursively.
2612 result = mmem->memory_at(C->get_general_index(alias_idx));
2613 result = find_inst_mem(result, alias_idx, orig_phis);
2614 if (C->failing()) {
2615 return NULL;
2616 }
2617 mmem->set_memory_at(alias_idx, result);
2618 }
2619 } else if (result->is_Phi() &&
2620 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
2621 Node *un = result->as_Phi()->unique_input(igvn);
2622 if (un != NULL) {
2623 orig_phis.append_if_missing(result->as_Phi());
2624 result = un;
2625 } else {
2626 break;
2627 }
2628 } else if (result->is_ClearArray()) {
2629 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), igvn)) {
2630 // Can not bypass initialization of the instance
2631 // we are looking for.
2632 break;
2633 }
2634 // Otherwise skip it (the call updated 'result' value).
2635 } else if (result->Opcode() == Op_SCMemProj) {
2636 Node* mem = result->in(0);
2637 Node* adr = NULL;
2638 if (mem->is_LoadStore()) {
2639 adr = mem->in(MemNode::Address);
2640 } else {
2641 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
2642 adr = mem->in(3); // Memory edge corresponds to destination array
2643 }
2644 const Type *at = igvn->type(adr);
2645 if (at != Type::TOP) {
2646 assert (at->isa_ptr() != NULL, "pointer type required.");
2647 int idx = C->get_alias_index(at->is_ptr());
2648 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
2649 break;
2650 }
2651 result = mem->in(MemNode::Memory);
2652 }
2653 }
2654 if (result->is_Phi()) {
2655 PhiNode *mphi = result->as_Phi();
2656 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
2657 const TypePtr *t = mphi->adr_type();
2658 if (!is_instance) {
2659 // Push all non-instance Phis on the orig_phis worklist to update inputs
2660 // during Phase 4 if needed.
2661 orig_phis.append_if_missing(mphi);
2662 } else if (C->get_alias_index(t) != alias_idx) {
2663 // Create a new Phi with the specified alias index type.
2664 result = split_memory_phi(mphi, alias_idx, orig_phis);
2665 }
2666 }
2667 // the result is either MemNode, PhiNode, InitializeNode.
2668 return result;
2669 }
2671 //
2672 // Convert the types of unescaped object to instance types where possible,
2673 // propagate the new type information through the graph, and update memory
2674 // edges and MergeMem inputs to reflect the new type.
2675 //
2676 // We start with allocations (and calls which may be allocations) on alloc_worklist.
2677 // The processing is done in 4 phases:
2678 //
2679 // Phase 1: Process possible allocations from alloc_worklist. Create instance
2680 // types for the CheckCastPP for allocations where possible.
2681 // Propagate the the new types through users as follows:
2682 // casts and Phi: push users on alloc_worklist
2683 // AddP: cast Base and Address inputs to the instance type
2684 // push any AddP users on alloc_worklist and push any memnode
2685 // users onto memnode_worklist.
2686 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
2687 // search the Memory chain for a store with the appropriate type
2688 // address type. If a Phi is found, create a new version with
2689 // the appropriate memory slices from each of the Phi inputs.
2690 // For stores, process the users as follows:
2691 // MemNode: push on memnode_worklist
2692 // MergeMem: push on mergemem_worklist
2693 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
2694 // moving the first node encountered of each instance type to the
2695 // the input corresponding to its alias index.
2696 // appropriate memory slice.
2697 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
2698 //
2699 // In the following example, the CheckCastPP nodes are the cast of allocation
2700 // results and the allocation of node 29 is unescaped and eligible to be an
2701 // instance type.
2702 //
2703 // We start with:
2704 //
2705 // 7 Parm #memory
2706 // 10 ConI "12"
2707 // 19 CheckCastPP "Foo"
2708 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2709 // 29 CheckCastPP "Foo"
2710 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
2711 //
2712 // 40 StoreP 25 7 20 ... alias_index=4
2713 // 50 StoreP 35 40 30 ... alias_index=4
2714 // 60 StoreP 45 50 20 ... alias_index=4
2715 // 70 LoadP _ 60 30 ... alias_index=4
2716 // 80 Phi 75 50 60 Memory alias_index=4
2717 // 90 LoadP _ 80 30 ... alias_index=4
2718 // 100 LoadP _ 80 20 ... alias_index=4
2719 //
2720 //
2721 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
2722 // and creating a new alias index for node 30. This gives:
2723 //
2724 // 7 Parm #memory
2725 // 10 ConI "12"
2726 // 19 CheckCastPP "Foo"
2727 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2728 // 29 CheckCastPP "Foo" iid=24
2729 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2730 //
2731 // 40 StoreP 25 7 20 ... alias_index=4
2732 // 50 StoreP 35 40 30 ... alias_index=6
2733 // 60 StoreP 45 50 20 ... alias_index=4
2734 // 70 LoadP _ 60 30 ... alias_index=6
2735 // 80 Phi 75 50 60 Memory alias_index=4
2736 // 90 LoadP _ 80 30 ... alias_index=6
2737 // 100 LoadP _ 80 20 ... alias_index=4
2738 //
2739 // In phase 2, new memory inputs are computed for the loads and stores,
2740 // And a new version of the phi is created. In phase 4, the inputs to
2741 // node 80 are updated and then the memory nodes are updated with the
2742 // values computed in phase 2. This results in:
2743 //
2744 // 7 Parm #memory
2745 // 10 ConI "12"
2746 // 19 CheckCastPP "Foo"
2747 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
2748 // 29 CheckCastPP "Foo" iid=24
2749 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
2750 //
2751 // 40 StoreP 25 7 20 ... alias_index=4
2752 // 50 StoreP 35 7 30 ... alias_index=6
2753 // 60 StoreP 45 40 20 ... alias_index=4
2754 // 70 LoadP _ 50 30 ... alias_index=6
2755 // 80 Phi 75 40 60 Memory alias_index=4
2756 // 120 Phi 75 50 50 Memory alias_index=6
2757 // 90 LoadP _ 120 30 ... alias_index=6
2758 // 100 LoadP _ 80 20 ... alias_index=4
2759 //
2760 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
2761 GrowableArray<Node *> memnode_worklist;
2762 GrowableArray<PhiNode *> orig_phis;
2763 PhaseIterGVN *igvn = _igvn;
2764 uint new_index_start = (uint) _compile->num_alias_types();
2765 Arena* arena = Thread::current()->resource_area();
2766 VectorSet visited(arena);
2767 ideal_nodes.clear(); // Reset for use with set_map/get_map.
2768 uint unique_old = _compile->unique();
2770 // Phase 1: Process possible allocations from alloc_worklist.
2771 // Create instance types for the CheckCastPP for allocations where possible.
2772 //
2773 // (Note: don't forget to change the order of the second AddP node on
2774 // the alloc_worklist if the order of the worklist processing is changed,
2775 // see the comment in find_second_addp().)
2776 //
2777 while (alloc_worklist.length() != 0) {
2778 Node *n = alloc_worklist.pop();
2779 uint ni = n->_idx;
2780 if (n->is_Call()) {
2781 CallNode *alloc = n->as_Call();
2782 // copy escape information to call node
2783 PointsToNode* ptn = ptnode_adr(alloc->_idx);
2784 PointsToNode::EscapeState es = ptn->escape_state();
2785 // We have an allocation or call which returns a Java object,
2786 // see if it is unescaped.
2787 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
2788 continue;
2789 // Find CheckCastPP for the allocate or for the return value of a call
2790 n = alloc->result_cast();
2791 if (n == NULL) { // No uses except Initialize node
2792 if (alloc->is_Allocate()) {
2793 // Set the scalar_replaceable flag for allocation
2794 // so it could be eliminated if it has no uses.
2795 alloc->as_Allocate()->_is_scalar_replaceable = true;
2796 }
2797 if (alloc->is_CallStaticJava()) {
2798 // Set the scalar_replaceable flag for boxing method
2799 // so it could be eliminated if it has no uses.
2800 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2801 }
2802 continue;
2803 }
2804 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
2805 assert(!alloc->is_Allocate(), "allocation should have unique type");
2806 continue;
2807 }
2809 // The inline code for Object.clone() casts the allocation result to
2810 // java.lang.Object and then to the actual type of the allocated
2811 // object. Detect this case and use the second cast.
2812 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
2813 // the allocation result is cast to java.lang.Object and then
2814 // to the actual Array type.
2815 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
2816 && (alloc->is_AllocateArray() ||
2817 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
2818 Node *cast2 = NULL;
2819 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2820 Node *use = n->fast_out(i);
2821 if (use->is_CheckCastPP()) {
2822 cast2 = use;
2823 break;
2824 }
2825 }
2826 if (cast2 != NULL) {
2827 n = cast2;
2828 } else {
2829 // Non-scalar replaceable if the allocation type is unknown statically
2830 // (reflection allocation), the object can't be restored during
2831 // deoptimization without precise type.
2832 continue;
2833 }
2834 }
2835 if (alloc->is_Allocate()) {
2836 // Set the scalar_replaceable flag for allocation
2837 // so it could be eliminated.
2838 alloc->as_Allocate()->_is_scalar_replaceable = true;
2839 }
2840 if (alloc->is_CallStaticJava()) {
2841 // Set the scalar_replaceable flag for boxing method
2842 // so it could be eliminated.
2843 alloc->as_CallStaticJava()->_is_scalar_replaceable = true;
2844 }
2845 set_escape_state(ptnode_adr(n->_idx), es); // CheckCastPP escape state
2846 // in order for an object to be scalar-replaceable, it must be:
2847 // - a direct allocation (not a call returning an object)
2848 // - non-escaping
2849 // - eligible to be a unique type
2850 // - not determined to be ineligible by escape analysis
2851 set_map(alloc, n);
2852 set_map(n, alloc);
2853 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
2854 if (t == NULL)
2855 continue; // not a TypeOopPtr
2856 const TypeOopPtr* tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
2857 igvn->hash_delete(n);
2858 igvn->set_type(n, tinst);
2859 n->raise_bottom_type(tinst);
2860 igvn->hash_insert(n);
2861 record_for_optimizer(n);
2862 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
2864 // First, put on the worklist all Field edges from Connection Graph
2865 // which is more accurate then putting immediate users from Ideal Graph.
2866 for (EdgeIterator e(ptn); e.has_next(); e.next()) {
2867 PointsToNode* tgt = e.get();
2868 Node* use = tgt->ideal_node();
2869 assert(tgt->is_Field() && use->is_AddP(),
2870 "only AddP nodes are Field edges in CG");
2871 if (use->outcnt() > 0) { // Don't process dead nodes
2872 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
2873 if (addp2 != NULL) {
2874 assert(alloc->is_AllocateArray(),"array allocation was expected");
2875 alloc_worklist.append_if_missing(addp2);
2876 }
2877 alloc_worklist.append_if_missing(use);
2878 }
2879 }
2881 // An allocation may have an Initialize which has raw stores. Scan
2882 // the users of the raw allocation result and push AddP users
2883 // on alloc_worklist.
2884 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
2885 assert (raw_result != NULL, "must have an allocation result");
2886 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
2887 Node *use = raw_result->fast_out(i);
2888 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
2889 Node* addp2 = find_second_addp(use, raw_result);
2890 if (addp2 != NULL) {
2891 assert(alloc->is_AllocateArray(),"array allocation was expected");
2892 alloc_worklist.append_if_missing(addp2);
2893 }
2894 alloc_worklist.append_if_missing(use);
2895 } else if (use->is_MemBar()) {
2896 memnode_worklist.append_if_missing(use);
2897 }
2898 }
2899 }
2900 } else if (n->is_AddP()) {
2901 JavaObjectNode* jobj = unique_java_object(get_addp_base(n));
2902 if (jobj == NULL || jobj == phantom_obj) {
2903 #ifdef ASSERT
2904 ptnode_adr(get_addp_base(n)->_idx)->dump();
2905 ptnode_adr(n->_idx)->dump();
2906 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2907 #endif
2908 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2909 return;
2910 }
2911 Node *base = get_map(jobj->idx()); // CheckCastPP node
2912 if (!split_AddP(n, base)) continue; // wrong type from dead path
2913 } else if (n->is_Phi() ||
2914 n->is_CheckCastPP() ||
2915 n->is_EncodeP() ||
2916 n->is_DecodeN() ||
2917 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
2918 if (visited.test_set(n->_idx)) {
2919 assert(n->is_Phi(), "loops only through Phi's");
2920 continue; // already processed
2921 }
2922 JavaObjectNode* jobj = unique_java_object(n);
2923 if (jobj == NULL || jobj == phantom_obj) {
2924 #ifdef ASSERT
2925 ptnode_adr(n->_idx)->dump();
2926 assert(jobj != NULL && jobj != phantom_obj, "escaped allocation");
2927 #endif
2928 _compile->record_failure(C2Compiler::retry_no_escape_analysis());
2929 return;
2930 } else {
2931 Node *val = get_map(jobj->idx()); // CheckCastPP node
2932 TypeNode *tn = n->as_Type();
2933 const TypeOopPtr* tinst = igvn->type(val)->isa_oopptr();
2934 assert(tinst != NULL && tinst->is_known_instance() &&
2935 tinst->instance_id() == jobj->idx() , "instance type expected.");
2937 const Type *tn_type = igvn->type(tn);
2938 const TypeOopPtr *tn_t;
2939 if (tn_type->isa_narrowoop()) {
2940 tn_t = tn_type->make_ptr()->isa_oopptr();
2941 } else {
2942 tn_t = tn_type->isa_oopptr();
2943 }
2944 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
2945 if (tn_type->isa_narrowoop()) {
2946 tn_type = tinst->make_narrowoop();
2947 } else {
2948 tn_type = tinst;
2949 }
2950 igvn->hash_delete(tn);
2951 igvn->set_type(tn, tn_type);
2952 tn->set_type(tn_type);
2953 igvn->hash_insert(tn);
2954 record_for_optimizer(n);
2955 } else {
2956 assert(tn_type == TypePtr::NULL_PTR ||
2957 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
2958 "unexpected type");
2959 continue; // Skip dead path with different type
2960 }
2961 }
2962 } else {
2963 debug_only(n->dump();)
2964 assert(false, "EA: unexpected node");
2965 continue;
2966 }
2967 // push allocation's users on appropriate worklist
2968 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2969 Node *use = n->fast_out(i);
2970 if(use->is_Mem() && use->in(MemNode::Address) == n) {
2971 // Load/store to instance's field
2972 memnode_worklist.append_if_missing(use);
2973 } else if (use->is_MemBar()) {
2974 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
2975 memnode_worklist.append_if_missing(use);
2976 }
2977 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
2978 Node* addp2 = find_second_addp(use, n);
2979 if (addp2 != NULL) {
2980 alloc_worklist.append_if_missing(addp2);
2981 }
2982 alloc_worklist.append_if_missing(use);
2983 } else if (use->is_Phi() ||
2984 use->is_CheckCastPP() ||
2985 use->is_EncodeNarrowPtr() ||
2986 use->is_DecodeNarrowPtr() ||
2987 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
2988 alloc_worklist.append_if_missing(use);
2989 #ifdef ASSERT
2990 } else if (use->is_Mem()) {
2991 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
2992 } else if (use->is_MergeMem()) {
2993 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
2994 } else if (use->is_SafePoint()) {
2995 // Look for MergeMem nodes for calls which reference unique allocation
2996 // (through CheckCastPP nodes) even for debug info.
2997 Node* m = use->in(TypeFunc::Memory);
2998 if (m->is_MergeMem()) {
2999 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3000 }
3001 } else if (use->Opcode() == Op_EncodeISOArray) {
3002 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3003 // EncodeISOArray overwrites destination array
3004 memnode_worklist.append_if_missing(use);
3005 }
3006 } else {
3007 uint op = use->Opcode();
3008 if (!(op == Op_CmpP || op == Op_Conv2B ||
3009 op == Op_CastP2X || op == Op_StoreCM ||
3010 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
3011 op == Op_StrEquals || op == Op_StrIndexOf)) {
3012 n->dump();
3013 use->dump();
3014 assert(false, "EA: missing allocation reference path");
3015 }
3016 #endif
3017 }
3018 }
3020 }
3021 // New alias types were created in split_AddP().
3022 uint new_index_end = (uint) _compile->num_alias_types();
3023 assert(unique_old == _compile->unique(), "there should be no new ideal nodes after Phase 1");
3025 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
3026 // compute new values for Memory inputs (the Memory inputs are not
3027 // actually updated until phase 4.)
3028 if (memnode_worklist.length() == 0)
3029 return; // nothing to do
3030 while (memnode_worklist.length() != 0) {
3031 Node *n = memnode_worklist.pop();
3032 if (visited.test_set(n->_idx))
3033 continue;
3034 if (n->is_Phi() || n->is_ClearArray()) {
3035 // we don't need to do anything, but the users must be pushed
3036 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
3037 // we don't need to do anything, but the users must be pushed
3038 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
3039 if (n == NULL)
3040 continue;
3041 } else if (n->Opcode() == Op_EncodeISOArray) {
3042 // get the memory projection
3043 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3044 Node *use = n->fast_out(i);
3045 if (use->Opcode() == Op_SCMemProj) {
3046 n = use;
3047 break;
3048 }
3049 }
3050 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3051 } else {
3052 assert(n->is_Mem(), "memory node required.");
3053 Node *addr = n->in(MemNode::Address);
3054 const Type *addr_t = igvn->type(addr);
3055 if (addr_t == Type::TOP)
3056 continue;
3057 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
3058 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
3059 assert ((uint)alias_idx < new_index_end, "wrong alias index");
3060 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis);
3061 if (_compile->failing()) {
3062 return;
3063 }
3064 if (mem != n->in(MemNode::Memory)) {
3065 // We delay the memory edge update since we need old one in
3066 // MergeMem code below when instances memory slices are separated.
3067 set_map(n, mem);
3068 }
3069 if (n->is_Load()) {
3070 continue; // don't push users
3071 } else if (n->is_LoadStore()) {
3072 // get the memory projection
3073 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3074 Node *use = n->fast_out(i);
3075 if (use->Opcode() == Op_SCMemProj) {
3076 n = use;
3077 break;
3078 }
3079 }
3080 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
3081 }
3082 }
3083 // push user on appropriate worklist
3084 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3085 Node *use = n->fast_out(i);
3086 if (use->is_Phi() || use->is_ClearArray()) {
3087 memnode_worklist.append_if_missing(use);
3088 } else if (use->is_Mem() && use->in(MemNode::Memory) == n) {
3089 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
3090 continue;
3091 memnode_worklist.append_if_missing(use);
3092 } else if (use->is_MemBar()) {
3093 if (use->in(TypeFunc::Memory) == n) { // Ignore precedent edge
3094 memnode_worklist.append_if_missing(use);
3095 }
3096 #ifdef ASSERT
3097 } else if(use->is_Mem()) {
3098 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
3099 } else if (use->is_MergeMem()) {
3100 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
3101 } else if (use->Opcode() == Op_EncodeISOArray) {
3102 if (use->in(MemNode::Memory) == n || use->in(3) == n) {
3103 // EncodeISOArray overwrites destination array
3104 memnode_worklist.append_if_missing(use);
3105 }
3106 } else {
3107 uint op = use->Opcode();
3108 if (!(op == Op_StoreCM ||
3109 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
3110 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
3111 op == Op_AryEq || op == Op_StrComp ||
3112 op == Op_StrEquals || op == Op_StrIndexOf)) {
3113 n->dump();
3114 use->dump();
3115 assert(false, "EA: missing memory path");
3116 }
3117 #endif
3118 }
3119 }
3120 }
3122 // Phase 3: Process MergeMem nodes from mergemem_worklist.
3123 // Walk each memory slice moving the first node encountered of each
3124 // instance type to the the input corresponding to its alias index.
3125 uint length = _mergemem_worklist.length();
3126 for( uint next = 0; next < length; ++next ) {
3127 MergeMemNode* nmm = _mergemem_worklist.at(next);
3128 assert(!visited.test_set(nmm->_idx), "should not be visited before");
3129 // Note: we don't want to use MergeMemStream here because we only want to
3130 // scan inputs which exist at the start, not ones we add during processing.
3131 // Note 2: MergeMem may already contains instance memory slices added
3132 // during find_inst_mem() call when memory nodes were processed above.
3133 igvn->hash_delete(nmm);
3134 uint nslices = nmm->req();
3135 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
3136 Node* mem = nmm->in(i);
3137 Node* cur = NULL;
3138 if (mem == NULL || mem->is_top())
3139 continue;
3140 // First, update mergemem by moving memory nodes to corresponding slices
3141 // if their type became more precise since this mergemem was created.
3142 while (mem->is_Mem()) {
3143 const Type *at = igvn->type(mem->in(MemNode::Address));
3144 if (at != Type::TOP) {
3145 assert (at->isa_ptr() != NULL, "pointer type required.");
3146 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
3147 if (idx == i) {
3148 if (cur == NULL)
3149 cur = mem;
3150 } else {
3151 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
3152 nmm->set_memory_at(idx, mem);
3153 }
3154 }
3155 }
3156 mem = mem->in(MemNode::Memory);
3157 }
3158 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
3159 // Find any instance of the current type if we haven't encountered
3160 // already a memory slice of the instance along the memory chain.
3161 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3162 if((uint)_compile->get_general_index(ni) == i) {
3163 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
3164 if (nmm->is_empty_memory(m)) {
3165 Node* result = find_inst_mem(mem, ni, orig_phis);
3166 if (_compile->failing()) {
3167 return;
3168 }
3169 nmm->set_memory_at(ni, result);
3170 }
3171 }
3172 }
3173 }
3174 // Find the rest of instances values
3175 for (uint ni = new_index_start; ni < new_index_end; ni++) {
3176 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
3177 Node* result = step_through_mergemem(nmm, ni, tinst);
3178 if (result == nmm->base_memory()) {
3179 // Didn't find instance memory, search through general slice recursively.
3180 result = nmm->memory_at(_compile->get_general_index(ni));
3181 result = find_inst_mem(result, ni, orig_phis);
3182 if (_compile->failing()) {
3183 return;
3184 }
3185 nmm->set_memory_at(ni, result);
3186 }
3187 }
3188 igvn->hash_insert(nmm);
3189 record_for_optimizer(nmm);
3190 }
3192 // Phase 4: Update the inputs of non-instance memory Phis and
3193 // the Memory input of memnodes
3194 // First update the inputs of any non-instance Phi's from
3195 // which we split out an instance Phi. Note we don't have
3196 // to recursively process Phi's encounted on the input memory
3197 // chains as is done in split_memory_phi() since they will
3198 // also be processed here.
3199 for (int j = 0; j < orig_phis.length(); j++) {
3200 PhiNode *phi = orig_phis.at(j);
3201 int alias_idx = _compile->get_alias_index(phi->adr_type());
3202 igvn->hash_delete(phi);
3203 for (uint i = 1; i < phi->req(); i++) {
3204 Node *mem = phi->in(i);
3205 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis);
3206 if (_compile->failing()) {
3207 return;
3208 }
3209 if (mem != new_mem) {
3210 phi->set_req(i, new_mem);
3211 }
3212 }
3213 igvn->hash_insert(phi);
3214 record_for_optimizer(phi);
3215 }
3217 // Update the memory inputs of MemNodes with the value we computed
3218 // in Phase 2 and move stores memory users to corresponding memory slices.
3219 // Disable memory split verification code until the fix for 6984348.
3220 // Currently it produces false negative results since it does not cover all cases.
3221 #if 0 // ifdef ASSERT
3222 visited.Reset();
3223 Node_Stack old_mems(arena, _compile->unique() >> 2);
3224 #endif
3225 for (uint i = 0; i < ideal_nodes.size(); i++) {
3226 Node* n = ideal_nodes.at(i);
3227 Node* nmem = get_map(n->_idx);
3228 assert(nmem != NULL, "sanity");
3229 if (n->is_Mem()) {
3230 #if 0 // ifdef ASSERT
3231 Node* old_mem = n->in(MemNode::Memory);
3232 if (!visited.test_set(old_mem->_idx)) {
3233 old_mems.push(old_mem, old_mem->outcnt());
3234 }
3235 #endif
3236 assert(n->in(MemNode::Memory) != nmem, "sanity");
3237 if (!n->is_Load()) {
3238 // Move memory users of a store first.
3239 move_inst_mem(n, orig_phis);
3240 }
3241 // Now update memory input
3242 igvn->hash_delete(n);
3243 n->set_req(MemNode::Memory, nmem);
3244 igvn->hash_insert(n);
3245 record_for_optimizer(n);
3246 } else {
3247 assert(n->is_Allocate() || n->is_CheckCastPP() ||
3248 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
3249 }
3250 }
3251 #if 0 // ifdef ASSERT
3252 // Verify that memory was split correctly
3253 while (old_mems.is_nonempty()) {
3254 Node* old_mem = old_mems.node();
3255 uint old_cnt = old_mems.index();
3256 old_mems.pop();
3257 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
3258 }
3259 #endif
3260 }
3262 #ifndef PRODUCT
3263 static const char *node_type_names[] = {
3264 "UnknownType",
3265 "JavaObject",
3266 "LocalVar",
3267 "Field",
3268 "Arraycopy"
3269 };
3271 static const char *esc_names[] = {
3272 "UnknownEscape",
3273 "NoEscape",
3274 "ArgEscape",
3275 "GlobalEscape"
3276 };
3278 void PointsToNode::dump(bool print_state) const {
3279 NodeType nt = node_type();
3280 tty->print("%s ", node_type_names[(int) nt]);
3281 if (print_state) {
3282 EscapeState es = escape_state();
3283 EscapeState fields_es = fields_escape_state();
3284 tty->print("%s(%s) ", esc_names[(int)es], esc_names[(int)fields_es]);
3285 if (nt == PointsToNode::JavaObject && !this->scalar_replaceable())
3286 tty->print("NSR ");
3287 }
3288 if (is_Field()) {
3289 FieldNode* f = (FieldNode*)this;
3290 if (f->is_oop())
3291 tty->print("oop ");
3292 if (f->offset() > 0)
3293 tty->print("+%d ", f->offset());
3294 tty->print("(");
3295 for (BaseIterator i(f); i.has_next(); i.next()) {
3296 PointsToNode* b = i.get();
3297 tty->print(" %d%s", b->idx(),(b->is_JavaObject() ? "P" : ""));
3298 }
3299 tty->print(" )");
3300 }
3301 tty->print("[");
3302 for (EdgeIterator i(this); i.has_next(); i.next()) {
3303 PointsToNode* e = i.get();
3304 tty->print(" %d%s%s", e->idx(),(e->is_JavaObject() ? "P" : (e->is_Field() ? "F" : "")), e->is_Arraycopy() ? "cp" : "");
3305 }
3306 tty->print(" [");
3307 for (UseIterator i(this); i.has_next(); i.next()) {
3308 PointsToNode* u = i.get();
3309 bool is_base = false;
3310 if (PointsToNode::is_base_use(u)) {
3311 is_base = true;
3312 u = PointsToNode::get_use_node(u)->as_Field();
3313 }
3314 tty->print(" %d%s%s", u->idx(), is_base ? "b" : "", u->is_Arraycopy() ? "cp" : "");
3315 }
3316 tty->print(" ]] ");
3317 if (_node == NULL)
3318 tty->print_cr("<null>");
3319 else
3320 _node->dump();
3321 }
3323 void ConnectionGraph::dump(GrowableArray<PointsToNode*>& ptnodes_worklist) {
3324 bool first = true;
3325 int ptnodes_length = ptnodes_worklist.length();
3326 for (int i = 0; i < ptnodes_length; i++) {
3327 PointsToNode *ptn = ptnodes_worklist.at(i);
3328 if (ptn == NULL || !ptn->is_JavaObject())
3329 continue;
3330 PointsToNode::EscapeState es = ptn->escape_state();
3331 if ((es != PointsToNode::NoEscape) && !Verbose) {
3332 continue;
3333 }
3334 Node* n = ptn->ideal_node();
3335 if (n->is_Allocate() || (n->is_CallStaticJava() &&
3336 n->as_CallStaticJava()->is_boxing_method())) {
3337 if (first) {
3338 tty->cr();
3339 tty->print("======== Connection graph for ");
3340 _compile->method()->print_short_name();
3341 tty->cr();
3342 first = false;
3343 }
3344 ptn->dump();
3345 // Print all locals and fields which reference this allocation
3346 for (UseIterator j(ptn); j.has_next(); j.next()) {
3347 PointsToNode* use = j.get();
3348 if (use->is_LocalVar()) {
3349 use->dump(Verbose);
3350 } else if (Verbose) {
3351 use->dump();
3352 }
3353 }
3354 tty->cr();
3355 }
3356 }
3357 }
3358 #endif