Fri, 11 Mar 2011 07:50:51 -0800
7026631: field _klass is incorrectly set for dual type of TypeAryPtr::OOPS
Summary: add missing check this->dual() != TypeAryPtr::OOPS into TypeAryPtr::klass().
Reviewed-by: never
1 /*
2 * Copyright (c) 2005, 2011, 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 "libadt/vectset.hpp"
28 #include "memory/allocation.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/escape.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
37 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
39 if (_edges == NULL) {
40 Arena *a = Compile::current()->comp_arena();
41 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
42 }
43 _edges->append_if_missing(v);
44 }
46 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
47 uint v = (targIdx << EdgeShift) + ((uint) et);
49 _edges->remove(v);
50 }
52 #ifndef PRODUCT
53 static const char *node_type_names[] = {
54 "UnknownType",
55 "JavaObject",
56 "LocalVar",
57 "Field"
58 };
60 static const char *esc_names[] = {
61 "UnknownEscape",
62 "NoEscape",
63 "ArgEscape",
64 "GlobalEscape"
65 };
67 static const char *edge_type_suffix[] = {
68 "?", // UnknownEdge
69 "P", // PointsToEdge
70 "D", // DeferredEdge
71 "F" // FieldEdge
72 };
74 void PointsToNode::dump(bool print_state) const {
75 NodeType nt = node_type();
76 tty->print("%s ", node_type_names[(int) nt]);
77 if (print_state) {
78 EscapeState es = escape_state();
79 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
80 }
81 tty->print("[[");
82 for (uint i = 0; i < edge_count(); i++) {
83 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
84 }
85 tty->print("]] ");
86 if (_node == NULL)
87 tty->print_cr("<null>");
88 else
89 _node->dump();
90 }
91 #endif
93 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
94 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
95 _processed(C->comp_arena()),
96 pt_ptset(C->comp_arena()),
97 pt_visited(C->comp_arena()),
98 pt_worklist(C->comp_arena(), 4, 0, 0),
99 _collecting(true),
100 _progress(false),
101 _compile(C),
102 _igvn(igvn),
103 _node_map(C->comp_arena()) {
105 _phantom_object = C->top()->_idx,
106 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
108 // Add ConP(#NULL) and ConN(#NULL) nodes.
109 Node* oop_null = igvn->zerocon(T_OBJECT);
110 _oop_null = oop_null->_idx;
111 assert(_oop_null < C->unique(), "should be created already");
112 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
114 if (UseCompressedOops) {
115 Node* noop_null = igvn->zerocon(T_NARROWOOP);
116 _noop_null = noop_null->_idx;
117 assert(_noop_null < C->unique(), "should be created already");
118 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
119 }
120 }
122 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
123 PointsToNode *f = ptnode_adr(from_i);
124 PointsToNode *t = ptnode_adr(to_i);
126 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
127 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
128 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
129 add_edge(f, to_i, PointsToNode::PointsToEdge);
130 }
132 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
133 PointsToNode *f = ptnode_adr(from_i);
134 PointsToNode *t = ptnode_adr(to_i);
136 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
137 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
138 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
139 // don't add a self-referential edge, this can occur during removal of
140 // deferred edges
141 if (from_i != to_i)
142 add_edge(f, to_i, PointsToNode::DeferredEdge);
143 }
145 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
146 const Type *adr_type = phase->type(adr);
147 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
148 adr->in(AddPNode::Address)->is_Proj() &&
149 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
150 // We are computing a raw address for a store captured by an Initialize
151 // compute an appropriate address type. AddP cases #3 and #5 (see below).
152 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
153 assert(offs != Type::OffsetBot ||
154 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
155 "offset must be a constant or it is initialization of array");
156 return offs;
157 }
158 const TypePtr *t_ptr = adr_type->isa_ptr();
159 assert(t_ptr != NULL, "must be a pointer type");
160 return t_ptr->offset();
161 }
163 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
164 PointsToNode *f = ptnode_adr(from_i);
165 PointsToNode *t = ptnode_adr(to_i);
167 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
168 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
169 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
170 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
171 t->set_offset(offset);
173 add_edge(f, to_i, PointsToNode::FieldEdge);
174 }
176 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
177 PointsToNode *npt = ptnode_adr(ni);
178 PointsToNode::EscapeState old_es = npt->escape_state();
179 if (es > old_es)
180 npt->set_escape_state(es);
181 }
183 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
184 PointsToNode::EscapeState es, bool done) {
185 PointsToNode* ptadr = ptnode_adr(n->_idx);
186 ptadr->_node = n;
187 ptadr->set_node_type(nt);
189 // inline set_escape_state(idx, es);
190 PointsToNode::EscapeState old_es = ptadr->escape_state();
191 if (es > old_es)
192 ptadr->set_escape_state(es);
194 if (done)
195 _processed.set(n->_idx);
196 }
198 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
199 uint idx = n->_idx;
200 PointsToNode::EscapeState es;
202 // If we are still collecting or there were no non-escaping allocations
203 // we don't know the answer yet
204 if (_collecting)
205 return PointsToNode::UnknownEscape;
207 // if the node was created after the escape computation, return
208 // UnknownEscape
209 if (idx >= nodes_size())
210 return PointsToNode::UnknownEscape;
212 es = ptnode_adr(idx)->escape_state();
214 // if we have already computed a value, return it
215 if (es != PointsToNode::UnknownEscape &&
216 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
217 return es;
219 // PointsTo() calls n->uncast() which can return a new ideal node.
220 if (n->uncast()->_idx >= nodes_size())
221 return PointsToNode::UnknownEscape;
223 PointsToNode::EscapeState orig_es = es;
225 // compute max escape state of anything this node could point to
226 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) {
227 uint pt = i.elem;
228 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
229 if (pes > es)
230 es = pes;
231 }
232 if (orig_es != es) {
233 // cache the computed escape state
234 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
235 ptnode_adr(idx)->set_escape_state(es);
236 } // orig_es could be PointsToNode::UnknownEscape
237 return es;
238 }
240 VectorSet* ConnectionGraph::PointsTo(Node * n) {
241 pt_ptset.Reset();
242 pt_visited.Reset();
243 pt_worklist.clear();
245 #ifdef ASSERT
246 Node *orig_n = n;
247 #endif
249 n = n->uncast();
250 PointsToNode* npt = ptnode_adr(n->_idx);
252 // If we have a JavaObject, return just that object
253 if (npt->node_type() == PointsToNode::JavaObject) {
254 pt_ptset.set(n->_idx);
255 return &pt_ptset;
256 }
257 #ifdef ASSERT
258 if (npt->_node == NULL) {
259 if (orig_n != n)
260 orig_n->dump();
261 n->dump();
262 assert(npt->_node != NULL, "unregistered node");
263 }
264 #endif
265 pt_worklist.push(n->_idx);
266 while(pt_worklist.length() > 0) {
267 int ni = pt_worklist.pop();
268 if (pt_visited.test_set(ni))
269 continue;
271 PointsToNode* pn = ptnode_adr(ni);
272 // ensure that all inputs of a Phi have been processed
273 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
275 int edges_processed = 0;
276 uint e_cnt = pn->edge_count();
277 for (uint e = 0; e < e_cnt; e++) {
278 uint etgt = pn->edge_target(e);
279 PointsToNode::EdgeType et = pn->edge_type(e);
280 if (et == PointsToNode::PointsToEdge) {
281 pt_ptset.set(etgt);
282 edges_processed++;
283 } else if (et == PointsToNode::DeferredEdge) {
284 pt_worklist.push(etgt);
285 edges_processed++;
286 } else {
287 assert(false,"neither PointsToEdge or DeferredEdge");
288 }
289 }
290 if (edges_processed == 0) {
291 // no deferred or pointsto edges found. Assume the value was set
292 // outside this method. Add the phantom object to the pointsto set.
293 pt_ptset.set(_phantom_object);
294 }
295 }
296 return &pt_ptset;
297 }
299 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
300 // This method is most expensive during ConnectionGraph construction.
301 // Reuse vectorSet and an additional growable array for deferred edges.
302 deferred_edges->clear();
303 visited->Reset();
305 visited->set(ni);
306 PointsToNode *ptn = ptnode_adr(ni);
308 // Mark current edges as visited and move deferred edges to separate array.
309 for (uint i = 0; i < ptn->edge_count(); ) {
310 uint t = ptn->edge_target(i);
311 #ifdef ASSERT
312 assert(!visited->test_set(t), "expecting no duplications");
313 #else
314 visited->set(t);
315 #endif
316 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
317 ptn->remove_edge(t, PointsToNode::DeferredEdge);
318 deferred_edges->append(t);
319 } else {
320 i++;
321 }
322 }
323 for (int next = 0; next < deferred_edges->length(); ++next) {
324 uint t = deferred_edges->at(next);
325 PointsToNode *ptt = ptnode_adr(t);
326 uint e_cnt = ptt->edge_count();
327 for (uint e = 0; e < e_cnt; e++) {
328 uint etgt = ptt->edge_target(e);
329 if (visited->test_set(etgt))
330 continue;
332 PointsToNode::EdgeType et = ptt->edge_type(e);
333 if (et == PointsToNode::PointsToEdge) {
334 add_pointsto_edge(ni, etgt);
335 if(etgt == _phantom_object) {
336 // Special case - field set outside (globally escaping).
337 ptn->set_escape_state(PointsToNode::GlobalEscape);
338 }
339 } else if (et == PointsToNode::DeferredEdge) {
340 deferred_edges->append(etgt);
341 } else {
342 assert(false,"invalid connection graph");
343 }
344 }
345 }
346 }
349 // Add an edge to node given by "to_i" from any field of adr_i whose offset
350 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
351 // a pointsto edge is added if it is a JavaObject
353 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
354 PointsToNode* an = ptnode_adr(adr_i);
355 PointsToNode* to = ptnode_adr(to_i);
356 bool deferred = (to->node_type() == PointsToNode::LocalVar);
358 for (uint fe = 0; fe < an->edge_count(); fe++) {
359 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
360 int fi = an->edge_target(fe);
361 PointsToNode* pf = ptnode_adr(fi);
362 int po = pf->offset();
363 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
364 if (deferred)
365 add_deferred_edge(fi, to_i);
366 else
367 add_pointsto_edge(fi, to_i);
368 }
369 }
370 }
372 // Add a deferred edge from node given by "from_i" to any field of adr_i
373 // whose offset matches "offset".
374 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
375 PointsToNode* an = ptnode_adr(adr_i);
376 for (uint fe = 0; fe < an->edge_count(); fe++) {
377 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
378 int fi = an->edge_target(fe);
379 PointsToNode* pf = ptnode_adr(fi);
380 int po = pf->offset();
381 if (pf->edge_count() == 0) {
382 // we have not seen any stores to this field, assume it was set outside this method
383 add_pointsto_edge(fi, _phantom_object);
384 }
385 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
386 add_deferred_edge(from_i, fi);
387 }
388 }
389 }
391 // Helper functions
393 static Node* get_addp_base(Node *addp) {
394 assert(addp->is_AddP(), "must be AddP");
395 //
396 // AddP cases for Base and Address inputs:
397 // case #1. Direct object's field reference:
398 // Allocate
399 // |
400 // Proj #5 ( oop result )
401 // |
402 // CheckCastPP (cast to instance type)
403 // | |
404 // AddP ( base == address )
405 //
406 // case #2. Indirect object's field reference:
407 // Phi
408 // |
409 // CastPP (cast to instance type)
410 // | |
411 // AddP ( base == address )
412 //
413 // case #3. Raw object's field reference for Initialize node:
414 // Allocate
415 // |
416 // Proj #5 ( oop result )
417 // top |
418 // \ |
419 // AddP ( base == top )
420 //
421 // case #4. Array's element reference:
422 // {CheckCastPP | CastPP}
423 // | | |
424 // | AddP ( array's element offset )
425 // | |
426 // AddP ( array's offset )
427 //
428 // case #5. Raw object's field reference for arraycopy stub call:
429 // The inline_native_clone() case when the arraycopy stub is called
430 // after the allocation before Initialize and CheckCastPP nodes.
431 // Allocate
432 // |
433 // Proj #5 ( oop result )
434 // | |
435 // AddP ( base == address )
436 //
437 // case #6. Constant Pool, ThreadLocal, CastX2P or
438 // Raw object's field reference:
439 // {ConP, ThreadLocal, CastX2P, raw Load}
440 // top |
441 // \ |
442 // AddP ( base == top )
443 //
444 // case #7. Klass's field reference.
445 // LoadKlass
446 // | |
447 // AddP ( base == address )
448 //
449 // case #8. narrow Klass's field reference.
450 // LoadNKlass
451 // |
452 // DecodeN
453 // | |
454 // AddP ( base == address )
455 //
456 Node *base = addp->in(AddPNode::Base)->uncast();
457 if (base->is_top()) { // The AddP case #3 and #6.
458 base = addp->in(AddPNode::Address)->uncast();
459 while (base->is_AddP()) {
460 // Case #6 (unsafe access) may have several chained AddP nodes.
461 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
462 base = base->in(AddPNode::Address)->uncast();
463 }
464 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
465 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
466 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
467 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
468 }
469 return base;
470 }
472 static Node* find_second_addp(Node* addp, Node* n) {
473 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
475 Node* addp2 = addp->raw_out(0);
476 if (addp->outcnt() == 1 && addp2->is_AddP() &&
477 addp2->in(AddPNode::Base) == n &&
478 addp2->in(AddPNode::Address) == addp) {
480 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
481 //
482 // Find array's offset to push it on worklist first and
483 // as result process an array's element offset first (pushed second)
484 // to avoid CastPP for the array's offset.
485 // Otherwise the inserted CastPP (LocalVar) will point to what
486 // the AddP (Field) points to. Which would be wrong since
487 // the algorithm expects the CastPP has the same point as
488 // as AddP's base CheckCastPP (LocalVar).
489 //
490 // ArrayAllocation
491 // |
492 // CheckCastPP
493 // |
494 // memProj (from ArrayAllocation CheckCastPP)
495 // | ||
496 // | || Int (element index)
497 // | || | ConI (log(element size))
498 // | || | /
499 // | || LShift
500 // | || /
501 // | AddP (array's element offset)
502 // | |
503 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
504 // | / /
505 // AddP (array's offset)
506 // |
507 // Load/Store (memory operation on array's element)
508 //
509 return addp2;
510 }
511 return NULL;
512 }
514 //
515 // Adjust the type and inputs of an AddP which computes the
516 // address of a field of an instance
517 //
518 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
519 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
520 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
521 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
522 if (t == NULL) {
523 // We are computing a raw address for a store captured by an Initialize
524 // compute an appropriate address type (cases #3 and #5).
525 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
526 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
527 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
528 assert(offs != Type::OffsetBot, "offset must be a constant");
529 t = base_t->add_offset(offs)->is_oopptr();
530 }
531 int inst_id = base_t->instance_id();
532 assert(!t->is_known_instance() || t->instance_id() == inst_id,
533 "old type must be non-instance or match new type");
535 // The type 't' could be subclass of 'base_t'.
536 // As result t->offset() could be large then base_t's size and it will
537 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
538 // constructor verifies correctness of the offset.
539 //
540 // It could happened on subclass's branch (from the type profiling
541 // inlining) which was not eliminated during parsing since the exactness
542 // of the allocation type was not propagated to the subclass type check.
543 //
544 // Or the type 't' could be not related to 'base_t' at all.
545 // It could happened when CHA type is different from MDO type on a dead path
546 // (for example, from instanceof check) which is not collapsed during parsing.
547 //
548 // Do nothing for such AddP node and don't process its users since
549 // this code branch will go away.
550 //
551 if (!t->is_known_instance() &&
552 !base_t->klass()->is_subtype_of(t->klass())) {
553 return false; // bail out
554 }
556 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
557 // Do NOT remove the next line: ensure a new alias index is allocated
558 // for the instance type. Note: C++ will not remove it since the call
559 // has side effect.
560 int alias_idx = _compile->get_alias_index(tinst);
561 igvn->set_type(addp, tinst);
562 // record the allocation in the node map
563 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
564 set_map(addp->_idx, get_map(base->_idx));
566 // Set addp's Base and Address to 'base'.
567 Node *abase = addp->in(AddPNode::Base);
568 Node *adr = addp->in(AddPNode::Address);
569 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
570 adr->in(0)->_idx == (uint)inst_id) {
571 // Skip AddP cases #3 and #5.
572 } else {
573 assert(!abase->is_top(), "sanity"); // AddP case #3
574 if (abase != base) {
575 igvn->hash_delete(addp);
576 addp->set_req(AddPNode::Base, base);
577 if (abase == adr) {
578 addp->set_req(AddPNode::Address, base);
579 } else {
580 // AddP case #4 (adr is array's element offset AddP node)
581 #ifdef ASSERT
582 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
583 assert(adr->is_AddP() && atype != NULL &&
584 atype->instance_id() == inst_id, "array's element offset should be processed first");
585 #endif
586 }
587 igvn->hash_insert(addp);
588 }
589 }
590 // Put on IGVN worklist since at least addp's type was changed above.
591 record_for_optimizer(addp);
592 return true;
593 }
595 //
596 // Create a new version of orig_phi if necessary. Returns either the newly
597 // created phi or an existing phi. Sets create_new to indicate wheter a new
598 // phi was created. Cache the last newly created phi in the node map.
599 //
600 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
601 Compile *C = _compile;
602 new_created = false;
603 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
604 // nothing to do if orig_phi is bottom memory or matches alias_idx
605 if (phi_alias_idx == alias_idx) {
606 return orig_phi;
607 }
608 // Have we recently created a Phi for this alias index?
609 PhiNode *result = get_map_phi(orig_phi->_idx);
610 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
611 return result;
612 }
613 // Previous check may fail when the same wide memory Phi was split into Phis
614 // for different memory slices. Search all Phis for this region.
615 if (result != NULL) {
616 Node* region = orig_phi->in(0);
617 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
618 Node* phi = region->fast_out(i);
619 if (phi->is_Phi() &&
620 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
621 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
622 return phi->as_Phi();
623 }
624 }
625 }
626 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
627 if (C->do_escape_analysis() == true && !C->failing()) {
628 // Retry compilation without escape analysis.
629 // If this is the first failure, the sentinel string will "stick"
630 // to the Compile object, and the C2Compiler will see it and retry.
631 C->record_failure(C2Compiler::retry_no_escape_analysis());
632 }
633 return NULL;
634 }
635 orig_phi_worklist.append_if_missing(orig_phi);
636 const TypePtr *atype = C->get_adr_type(alias_idx);
637 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
638 C->copy_node_notes_to(result, orig_phi);
639 igvn->set_type(result, result->bottom_type());
640 record_for_optimizer(result);
642 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
643 assert(pn == NULL || pn == orig_phi, "wrong node");
644 set_map(orig_phi->_idx, result);
645 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
647 new_created = true;
648 return result;
649 }
651 //
652 // Return a new version of Memory Phi "orig_phi" with the inputs having the
653 // specified alias index.
654 //
655 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
657 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
658 Compile *C = _compile;
659 bool new_phi_created;
660 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
661 if (!new_phi_created) {
662 return result;
663 }
665 GrowableArray<PhiNode *> phi_list;
666 GrowableArray<uint> cur_input;
668 PhiNode *phi = orig_phi;
669 uint idx = 1;
670 bool finished = false;
671 while(!finished) {
672 while (idx < phi->req()) {
673 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
674 if (mem != NULL && mem->is_Phi()) {
675 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
676 if (new_phi_created) {
677 // found an phi for which we created a new split, push current one on worklist and begin
678 // processing new one
679 phi_list.push(phi);
680 cur_input.push(idx);
681 phi = mem->as_Phi();
682 result = newphi;
683 idx = 1;
684 continue;
685 } else {
686 mem = newphi;
687 }
688 }
689 if (C->failing()) {
690 return NULL;
691 }
692 result->set_req(idx++, mem);
693 }
694 #ifdef ASSERT
695 // verify that the new Phi has an input for each input of the original
696 assert( phi->req() == result->req(), "must have same number of inputs.");
697 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
698 #endif
699 // Check if all new phi's inputs have specified alias index.
700 // Otherwise use old phi.
701 for (uint i = 1; i < phi->req(); i++) {
702 Node* in = result->in(i);
703 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
704 }
705 // we have finished processing a Phi, see if there are any more to do
706 finished = (phi_list.length() == 0 );
707 if (!finished) {
708 phi = phi_list.pop();
709 idx = cur_input.pop();
710 PhiNode *prev_result = get_map_phi(phi->_idx);
711 prev_result->set_req(idx++, result);
712 result = prev_result;
713 }
714 }
715 return result;
716 }
719 //
720 // The next methods are derived from methods in MemNode.
721 //
722 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
723 Node *mem = mmem;
724 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
725 // means an array I have not precisely typed yet. Do not do any
726 // alias stuff with it any time soon.
727 if( toop->base() != Type::AnyPtr &&
728 !(toop->klass() != NULL &&
729 toop->klass()->is_java_lang_Object() &&
730 toop->offset() == Type::OffsetBot) ) {
731 mem = mmem->memory_at(alias_idx);
732 // Update input if it is progress over what we have now
733 }
734 return mem;
735 }
737 //
738 // Move memory users to their memory slices.
739 //
740 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
741 Compile* C = _compile;
743 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
744 assert(tp != NULL, "ptr type");
745 int alias_idx = C->get_alias_index(tp);
746 int general_idx = C->get_general_index(alias_idx);
748 // Move users first
749 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
750 Node* use = n->fast_out(i);
751 if (use->is_MergeMem()) {
752 MergeMemNode* mmem = use->as_MergeMem();
753 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
754 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
755 continue; // Nothing to do
756 }
757 // Replace previous general reference to mem node.
758 uint orig_uniq = C->unique();
759 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
760 assert(orig_uniq == C->unique(), "no new nodes");
761 mmem->set_memory_at(general_idx, m);
762 --imax;
763 --i;
764 } else if (use->is_MemBar()) {
765 assert(!use->is_Initialize(), "initializing stores should not be moved");
766 if (use->req() > MemBarNode::Precedent &&
767 use->in(MemBarNode::Precedent) == n) {
768 // Don't move related membars.
769 record_for_optimizer(use);
770 continue;
771 }
772 tp = use->as_MemBar()->adr_type()->isa_ptr();
773 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
774 alias_idx == general_idx) {
775 continue; // Nothing to do
776 }
777 // Move to general memory slice.
778 uint orig_uniq = C->unique();
779 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
780 assert(orig_uniq == C->unique(), "no new nodes");
781 igvn->hash_delete(use);
782 imax -= use->replace_edge(n, m);
783 igvn->hash_insert(use);
784 record_for_optimizer(use);
785 --i;
786 #ifdef ASSERT
787 } else if (use->is_Mem()) {
788 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
789 // Don't move related cardmark.
790 continue;
791 }
792 // Memory nodes should have new memory input.
793 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
794 assert(tp != NULL, "ptr type");
795 int idx = C->get_alias_index(tp);
796 assert(get_map(use->_idx) != NULL || idx == alias_idx,
797 "Following memory nodes should have new memory input or be on the same memory slice");
798 } else if (use->is_Phi()) {
799 // Phi nodes should be split and moved already.
800 tp = use->as_Phi()->adr_type()->isa_ptr();
801 assert(tp != NULL, "ptr type");
802 int idx = C->get_alias_index(tp);
803 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
804 } else {
805 use->dump();
806 assert(false, "should not be here");
807 #endif
808 }
809 }
810 }
812 //
813 // Search memory chain of "mem" to find a MemNode whose address
814 // is the specified alias index.
815 //
816 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
817 if (orig_mem == NULL)
818 return orig_mem;
819 Compile* C = phase->C;
820 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
821 bool is_instance = (toop != NULL) && toop->is_known_instance();
822 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
823 Node *prev = NULL;
824 Node *result = orig_mem;
825 while (prev != result) {
826 prev = result;
827 if (result == start_mem)
828 break; // hit one of our sentinels
829 if (result->is_Mem()) {
830 const Type *at = phase->type(result->in(MemNode::Address));
831 if (at != Type::TOP) {
832 assert (at->isa_ptr() != NULL, "pointer type required.");
833 int idx = C->get_alias_index(at->is_ptr());
834 if (idx == alias_idx)
835 break;
836 }
837 result = result->in(MemNode::Memory);
838 }
839 if (!is_instance)
840 continue; // don't search further for non-instance types
841 // skip over a call which does not affect this memory slice
842 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
843 Node *proj_in = result->in(0);
844 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
845 break; // hit one of our sentinels
846 } else if (proj_in->is_Call()) {
847 CallNode *call = proj_in->as_Call();
848 if (!call->may_modify(toop, phase)) {
849 result = call->in(TypeFunc::Memory);
850 }
851 } else if (proj_in->is_Initialize()) {
852 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
853 // Stop if this is the initialization for the object instance which
854 // which contains this memory slice, otherwise skip over it.
855 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
856 result = proj_in->in(TypeFunc::Memory);
857 }
858 } else if (proj_in->is_MemBar()) {
859 result = proj_in->in(TypeFunc::Memory);
860 }
861 } else if (result->is_MergeMem()) {
862 MergeMemNode *mmem = result->as_MergeMem();
863 result = step_through_mergemem(mmem, alias_idx, toop);
864 if (result == mmem->base_memory()) {
865 // Didn't find instance memory, search through general slice recursively.
866 result = mmem->memory_at(C->get_general_index(alias_idx));
867 result = find_inst_mem(result, alias_idx, orig_phis, phase);
868 if (C->failing()) {
869 return NULL;
870 }
871 mmem->set_memory_at(alias_idx, result);
872 }
873 } else if (result->is_Phi() &&
874 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
875 Node *un = result->as_Phi()->unique_input(phase);
876 if (un != NULL) {
877 orig_phis.append_if_missing(result->as_Phi());
878 result = un;
879 } else {
880 break;
881 }
882 } else if (result->is_ClearArray()) {
883 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
884 // Can not bypass initialization of the instance
885 // we are looking for.
886 break;
887 }
888 // Otherwise skip it (the call updated 'result' value).
889 } else if (result->Opcode() == Op_SCMemProj) {
890 assert(result->in(0)->is_LoadStore(), "sanity");
891 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
892 if (at != Type::TOP) {
893 assert (at->isa_ptr() != NULL, "pointer type required.");
894 int idx = C->get_alias_index(at->is_ptr());
895 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
896 break;
897 }
898 result = result->in(0)->in(MemNode::Memory);
899 }
900 }
901 if (result->is_Phi()) {
902 PhiNode *mphi = result->as_Phi();
903 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
904 const TypePtr *t = mphi->adr_type();
905 if (C->get_alias_index(t) != alias_idx) {
906 // Create a new Phi with the specified alias index type.
907 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
908 } else if (!is_instance) {
909 // Push all non-instance Phis on the orig_phis worklist to update inputs
910 // during Phase 4 if needed.
911 orig_phis.append_if_missing(mphi);
912 }
913 }
914 // the result is either MemNode, PhiNode, InitializeNode.
915 return result;
916 }
918 //
919 // Convert the types of unescaped object to instance types where possible,
920 // propagate the new type information through the graph, and update memory
921 // edges and MergeMem inputs to reflect the new type.
922 //
923 // We start with allocations (and calls which may be allocations) on alloc_worklist.
924 // The processing is done in 4 phases:
925 //
926 // Phase 1: Process possible allocations from alloc_worklist. Create instance
927 // types for the CheckCastPP for allocations where possible.
928 // Propagate the the new types through users as follows:
929 // casts and Phi: push users on alloc_worklist
930 // AddP: cast Base and Address inputs to the instance type
931 // push any AddP users on alloc_worklist and push any memnode
932 // users onto memnode_worklist.
933 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
934 // search the Memory chain for a store with the appropriate type
935 // address type. If a Phi is found, create a new version with
936 // the appropriate memory slices from each of the Phi inputs.
937 // For stores, process the users as follows:
938 // MemNode: push on memnode_worklist
939 // MergeMem: push on mergemem_worklist
940 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
941 // moving the first node encountered of each instance type to the
942 // the input corresponding to its alias index.
943 // appropriate memory slice.
944 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
945 //
946 // In the following example, the CheckCastPP nodes are the cast of allocation
947 // results and the allocation of node 29 is unescaped and eligible to be an
948 // instance type.
949 //
950 // We start with:
951 //
952 // 7 Parm #memory
953 // 10 ConI "12"
954 // 19 CheckCastPP "Foo"
955 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
956 // 29 CheckCastPP "Foo"
957 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
958 //
959 // 40 StoreP 25 7 20 ... alias_index=4
960 // 50 StoreP 35 40 30 ... alias_index=4
961 // 60 StoreP 45 50 20 ... alias_index=4
962 // 70 LoadP _ 60 30 ... alias_index=4
963 // 80 Phi 75 50 60 Memory alias_index=4
964 // 90 LoadP _ 80 30 ... alias_index=4
965 // 100 LoadP _ 80 20 ... alias_index=4
966 //
967 //
968 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
969 // and creating a new alias index for node 30. This gives:
970 //
971 // 7 Parm #memory
972 // 10 ConI "12"
973 // 19 CheckCastPP "Foo"
974 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
975 // 29 CheckCastPP "Foo" iid=24
976 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
977 //
978 // 40 StoreP 25 7 20 ... alias_index=4
979 // 50 StoreP 35 40 30 ... alias_index=6
980 // 60 StoreP 45 50 20 ... alias_index=4
981 // 70 LoadP _ 60 30 ... alias_index=6
982 // 80 Phi 75 50 60 Memory alias_index=4
983 // 90 LoadP _ 80 30 ... alias_index=6
984 // 100 LoadP _ 80 20 ... alias_index=4
985 //
986 // In phase 2, new memory inputs are computed for the loads and stores,
987 // And a new version of the phi is created. In phase 4, the inputs to
988 // node 80 are updated and then the memory nodes are updated with the
989 // values computed in phase 2. This results in:
990 //
991 // 7 Parm #memory
992 // 10 ConI "12"
993 // 19 CheckCastPP "Foo"
994 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
995 // 29 CheckCastPP "Foo" iid=24
996 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
997 //
998 // 40 StoreP 25 7 20 ... alias_index=4
999 // 50 StoreP 35 7 30 ... alias_index=6
1000 // 60 StoreP 45 40 20 ... alias_index=4
1001 // 70 LoadP _ 50 30 ... alias_index=6
1002 // 80 Phi 75 40 60 Memory alias_index=4
1003 // 120 Phi 75 50 50 Memory alias_index=6
1004 // 90 LoadP _ 120 30 ... alias_index=6
1005 // 100 LoadP _ 80 20 ... alias_index=4
1006 //
1007 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1008 GrowableArray<Node *> memnode_worklist;
1009 GrowableArray<PhiNode *> orig_phis;
1011 PhaseIterGVN *igvn = _igvn;
1012 uint new_index_start = (uint) _compile->num_alias_types();
1013 Arena* arena = Thread::current()->resource_area();
1014 VectorSet visited(arena);
1017 // Phase 1: Process possible allocations from alloc_worklist.
1018 // Create instance types for the CheckCastPP for allocations where possible.
1019 //
1020 // (Note: don't forget to change the order of the second AddP node on
1021 // the alloc_worklist if the order of the worklist processing is changed,
1022 // see the comment in find_second_addp().)
1023 //
1024 while (alloc_worklist.length() != 0) {
1025 Node *n = alloc_worklist.pop();
1026 uint ni = n->_idx;
1027 const TypeOopPtr* tinst = NULL;
1028 if (n->is_Call()) {
1029 CallNode *alloc = n->as_Call();
1030 // copy escape information to call node
1031 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1032 PointsToNode::EscapeState es = escape_state(alloc);
1033 // We have an allocation or call which returns a Java object,
1034 // see if it is unescaped.
1035 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
1036 continue;
1038 // Find CheckCastPP for the allocate or for the return value of a call
1039 n = alloc->result_cast();
1040 if (n == NULL) { // No uses except Initialize node
1041 if (alloc->is_Allocate()) {
1042 // Set the scalar_replaceable flag for allocation
1043 // so it could be eliminated if it has no uses.
1044 alloc->as_Allocate()->_is_scalar_replaceable = true;
1045 }
1046 continue;
1047 }
1048 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1049 assert(!alloc->is_Allocate(), "allocation should have unique type");
1050 continue;
1051 }
1053 // The inline code for Object.clone() casts the allocation result to
1054 // java.lang.Object and then to the actual type of the allocated
1055 // object. Detect this case and use the second cast.
1056 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1057 // the allocation result is cast to java.lang.Object and then
1058 // to the actual Array type.
1059 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1060 && (alloc->is_AllocateArray() ||
1061 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1062 Node *cast2 = NULL;
1063 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1064 Node *use = n->fast_out(i);
1065 if (use->is_CheckCastPP()) {
1066 cast2 = use;
1067 break;
1068 }
1069 }
1070 if (cast2 != NULL) {
1071 n = cast2;
1072 } else {
1073 // Non-scalar replaceable if the allocation type is unknown statically
1074 // (reflection allocation), the object can't be restored during
1075 // deoptimization without precise type.
1076 continue;
1077 }
1078 }
1079 if (alloc->is_Allocate()) {
1080 // Set the scalar_replaceable flag for allocation
1081 // so it could be eliminated.
1082 alloc->as_Allocate()->_is_scalar_replaceable = true;
1083 }
1084 set_escape_state(n->_idx, es);
1085 // in order for an object to be scalar-replaceable, it must be:
1086 // - a direct allocation (not a call returning an object)
1087 // - non-escaping
1088 // - eligible to be a unique type
1089 // - not determined to be ineligible by escape analysis
1090 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1091 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1092 set_map(alloc->_idx, n);
1093 set_map(n->_idx, alloc);
1094 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1095 if (t == NULL)
1096 continue; // not a TypeInstPtr
1097 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1098 igvn->hash_delete(n);
1099 igvn->set_type(n, tinst);
1100 n->raise_bottom_type(tinst);
1101 igvn->hash_insert(n);
1102 record_for_optimizer(n);
1103 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
1104 (t->isa_instptr() || t->isa_aryptr())) {
1106 // First, put on the worklist all Field edges from Connection Graph
1107 // which is more accurate then putting immediate users from Ideal Graph.
1108 for (uint e = 0; e < ptn->edge_count(); e++) {
1109 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1110 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1111 "only AddP nodes are Field edges in CG");
1112 if (use->outcnt() > 0) { // Don't process dead nodes
1113 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1114 if (addp2 != NULL) {
1115 assert(alloc->is_AllocateArray(),"array allocation was expected");
1116 alloc_worklist.append_if_missing(addp2);
1117 }
1118 alloc_worklist.append_if_missing(use);
1119 }
1120 }
1122 // An allocation may have an Initialize which has raw stores. Scan
1123 // the users of the raw allocation result and push AddP users
1124 // on alloc_worklist.
1125 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1126 assert (raw_result != NULL, "must have an allocation result");
1127 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1128 Node *use = raw_result->fast_out(i);
1129 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1130 Node* addp2 = find_second_addp(use, raw_result);
1131 if (addp2 != NULL) {
1132 assert(alloc->is_AllocateArray(),"array allocation was expected");
1133 alloc_worklist.append_if_missing(addp2);
1134 }
1135 alloc_worklist.append_if_missing(use);
1136 } else if (use->is_MemBar()) {
1137 memnode_worklist.append_if_missing(use);
1138 }
1139 }
1140 }
1141 } else if (n->is_AddP()) {
1142 VectorSet* ptset = PointsTo(get_addp_base(n));
1143 assert(ptset->Size() == 1, "AddP address is unique");
1144 uint elem = ptset->getelem(); // Allocation node's index
1145 if (elem == _phantom_object) {
1146 assert(false, "escaped allocation");
1147 continue; // Assume the value was set outside this method.
1148 }
1149 Node *base = get_map(elem); // CheckCastPP node
1150 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1151 tinst = igvn->type(base)->isa_oopptr();
1152 } else if (n->is_Phi() ||
1153 n->is_CheckCastPP() ||
1154 n->is_EncodeP() ||
1155 n->is_DecodeN() ||
1156 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1157 if (visited.test_set(n->_idx)) {
1158 assert(n->is_Phi(), "loops only through Phi's");
1159 continue; // already processed
1160 }
1161 VectorSet* ptset = PointsTo(n);
1162 if (ptset->Size() == 1) {
1163 uint elem = ptset->getelem(); // Allocation node's index
1164 if (elem == _phantom_object) {
1165 assert(false, "escaped allocation");
1166 continue; // Assume the value was set outside this method.
1167 }
1168 Node *val = get_map(elem); // CheckCastPP node
1169 TypeNode *tn = n->as_Type();
1170 tinst = igvn->type(val)->isa_oopptr();
1171 assert(tinst != NULL && tinst->is_known_instance() &&
1172 (uint)tinst->instance_id() == elem , "instance type expected.");
1174 const Type *tn_type = igvn->type(tn);
1175 const TypeOopPtr *tn_t;
1176 if (tn_type->isa_narrowoop()) {
1177 tn_t = tn_type->make_ptr()->isa_oopptr();
1178 } else {
1179 tn_t = tn_type->isa_oopptr();
1180 }
1182 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1183 if (tn_type->isa_narrowoop()) {
1184 tn_type = tinst->make_narrowoop();
1185 } else {
1186 tn_type = tinst;
1187 }
1188 igvn->hash_delete(tn);
1189 igvn->set_type(tn, tn_type);
1190 tn->set_type(tn_type);
1191 igvn->hash_insert(tn);
1192 record_for_optimizer(n);
1193 } else {
1194 assert(tn_type == TypePtr::NULL_PTR ||
1195 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1196 "unexpected type");
1197 continue; // Skip dead path with different type
1198 }
1199 }
1200 } else {
1201 debug_only(n->dump();)
1202 assert(false, "EA: unexpected node");
1203 continue;
1204 }
1205 // push allocation's users on appropriate worklist
1206 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1207 Node *use = n->fast_out(i);
1208 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1209 // Load/store to instance's field
1210 memnode_worklist.append_if_missing(use);
1211 } else if (use->is_MemBar()) {
1212 memnode_worklist.append_if_missing(use);
1213 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1214 Node* addp2 = find_second_addp(use, n);
1215 if (addp2 != NULL) {
1216 alloc_worklist.append_if_missing(addp2);
1217 }
1218 alloc_worklist.append_if_missing(use);
1219 } else if (use->is_Phi() ||
1220 use->is_CheckCastPP() ||
1221 use->is_EncodeP() ||
1222 use->is_DecodeN() ||
1223 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1224 alloc_worklist.append_if_missing(use);
1225 #ifdef ASSERT
1226 } else if (use->is_Mem()) {
1227 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1228 } else if (use->is_MergeMem()) {
1229 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1230 } else if (use->is_SafePoint()) {
1231 // Look for MergeMem nodes for calls which reference unique allocation
1232 // (through CheckCastPP nodes) even for debug info.
1233 Node* m = use->in(TypeFunc::Memory);
1234 if (m->is_MergeMem()) {
1235 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1236 }
1237 } else {
1238 uint op = use->Opcode();
1239 if (!(op == Op_CmpP || op == Op_Conv2B ||
1240 op == Op_CastP2X || op == Op_StoreCM ||
1241 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1242 op == Op_StrEquals || op == Op_StrIndexOf)) {
1243 n->dump();
1244 use->dump();
1245 assert(false, "EA: missing allocation reference path");
1246 }
1247 #endif
1248 }
1249 }
1251 }
1252 // New alias types were created in split_AddP().
1253 uint new_index_end = (uint) _compile->num_alias_types();
1255 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1256 // compute new values for Memory inputs (the Memory inputs are not
1257 // actually updated until phase 4.)
1258 if (memnode_worklist.length() == 0)
1259 return; // nothing to do
1261 while (memnode_worklist.length() != 0) {
1262 Node *n = memnode_worklist.pop();
1263 if (visited.test_set(n->_idx))
1264 continue;
1265 if (n->is_Phi() || n->is_ClearArray()) {
1266 // we don't need to do anything, but the users must be pushed
1267 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1268 // we don't need to do anything, but the users must be pushed
1269 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1270 if (n == NULL)
1271 continue;
1272 } else {
1273 assert(n->is_Mem(), "memory node required.");
1274 Node *addr = n->in(MemNode::Address);
1275 const Type *addr_t = igvn->type(addr);
1276 if (addr_t == Type::TOP)
1277 continue;
1278 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1279 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1280 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1281 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1282 if (_compile->failing()) {
1283 return;
1284 }
1285 if (mem != n->in(MemNode::Memory)) {
1286 // We delay the memory edge update since we need old one in
1287 // MergeMem code below when instances memory slices are separated.
1288 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1289 assert(pn == NULL || pn == n, "wrong node");
1290 set_map(n->_idx, mem);
1291 ptnode_adr(n->_idx)->_node = n;
1292 }
1293 if (n->is_Load()) {
1294 continue; // don't push users
1295 } else if (n->is_LoadStore()) {
1296 // get the memory projection
1297 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1298 Node *use = n->fast_out(i);
1299 if (use->Opcode() == Op_SCMemProj) {
1300 n = use;
1301 break;
1302 }
1303 }
1304 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1305 }
1306 }
1307 // push user on appropriate worklist
1308 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1309 Node *use = n->fast_out(i);
1310 if (use->is_Phi() || use->is_ClearArray()) {
1311 memnode_worklist.append_if_missing(use);
1312 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1313 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1314 continue;
1315 memnode_worklist.append_if_missing(use);
1316 } else if (use->is_MemBar()) {
1317 memnode_worklist.append_if_missing(use);
1318 #ifdef ASSERT
1319 } else if(use->is_Mem()) {
1320 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1321 } else if (use->is_MergeMem()) {
1322 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1323 } else {
1324 uint op = use->Opcode();
1325 if (!(op == Op_StoreCM ||
1326 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1327 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1328 op == Op_AryEq || op == Op_StrComp ||
1329 op == Op_StrEquals || op == Op_StrIndexOf)) {
1330 n->dump();
1331 use->dump();
1332 assert(false, "EA: missing memory path");
1333 }
1334 #endif
1335 }
1336 }
1337 }
1339 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1340 // Walk each memory slice moving the first node encountered of each
1341 // instance type to the the input corresponding to its alias index.
1342 uint length = _mergemem_worklist.length();
1343 for( uint next = 0; next < length; ++next ) {
1344 MergeMemNode* nmm = _mergemem_worklist.at(next);
1345 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1346 // Note: we don't want to use MergeMemStream here because we only want to
1347 // scan inputs which exist at the start, not ones we add during processing.
1348 // Note 2: MergeMem may already contains instance memory slices added
1349 // during find_inst_mem() call when memory nodes were processed above.
1350 igvn->hash_delete(nmm);
1351 uint nslices = nmm->req();
1352 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1353 Node* mem = nmm->in(i);
1354 Node* cur = NULL;
1355 if (mem == NULL || mem->is_top())
1356 continue;
1357 // First, update mergemem by moving memory nodes to corresponding slices
1358 // if their type became more precise since this mergemem was created.
1359 while (mem->is_Mem()) {
1360 const Type *at = igvn->type(mem->in(MemNode::Address));
1361 if (at != Type::TOP) {
1362 assert (at->isa_ptr() != NULL, "pointer type required.");
1363 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1364 if (idx == i) {
1365 if (cur == NULL)
1366 cur = mem;
1367 } else {
1368 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1369 nmm->set_memory_at(idx, mem);
1370 }
1371 }
1372 }
1373 mem = mem->in(MemNode::Memory);
1374 }
1375 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1376 // Find any instance of the current type if we haven't encountered
1377 // already a memory slice of the instance along the memory chain.
1378 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1379 if((uint)_compile->get_general_index(ni) == i) {
1380 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1381 if (nmm->is_empty_memory(m)) {
1382 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1383 if (_compile->failing()) {
1384 return;
1385 }
1386 nmm->set_memory_at(ni, result);
1387 }
1388 }
1389 }
1390 }
1391 // Find the rest of instances values
1392 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1393 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1394 Node* result = step_through_mergemem(nmm, ni, tinst);
1395 if (result == nmm->base_memory()) {
1396 // Didn't find instance memory, search through general slice recursively.
1397 result = nmm->memory_at(_compile->get_general_index(ni));
1398 result = find_inst_mem(result, ni, orig_phis, igvn);
1399 if (_compile->failing()) {
1400 return;
1401 }
1402 nmm->set_memory_at(ni, result);
1403 }
1404 }
1405 igvn->hash_insert(nmm);
1406 record_for_optimizer(nmm);
1407 }
1409 // Phase 4: Update the inputs of non-instance memory Phis and
1410 // the Memory input of memnodes
1411 // First update the inputs of any non-instance Phi's from
1412 // which we split out an instance Phi. Note we don't have
1413 // to recursively process Phi's encounted on the input memory
1414 // chains as is done in split_memory_phi() since they will
1415 // also be processed here.
1416 for (int j = 0; j < orig_phis.length(); j++) {
1417 PhiNode *phi = orig_phis.at(j);
1418 int alias_idx = _compile->get_alias_index(phi->adr_type());
1419 igvn->hash_delete(phi);
1420 for (uint i = 1; i < phi->req(); i++) {
1421 Node *mem = phi->in(i);
1422 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1423 if (_compile->failing()) {
1424 return;
1425 }
1426 if (mem != new_mem) {
1427 phi->set_req(i, new_mem);
1428 }
1429 }
1430 igvn->hash_insert(phi);
1431 record_for_optimizer(phi);
1432 }
1434 // Update the memory inputs of MemNodes with the value we computed
1435 // in Phase 2 and move stores memory users to corresponding memory slices.
1436 #ifdef ASSERT
1437 visited.Reset();
1438 Node_Stack old_mems(arena, _compile->unique() >> 2);
1439 #endif
1440 for (uint i = 0; i < nodes_size(); i++) {
1441 Node *nmem = get_map(i);
1442 if (nmem != NULL) {
1443 Node *n = ptnode_adr(i)->_node;
1444 assert(n != NULL, "sanity");
1445 if (n->is_Mem()) {
1446 #ifdef ASSERT
1447 Node* old_mem = n->in(MemNode::Memory);
1448 if (!visited.test_set(old_mem->_idx)) {
1449 old_mems.push(old_mem, old_mem->outcnt());
1450 }
1451 #endif
1452 assert(n->in(MemNode::Memory) != nmem, "sanity");
1453 if (!n->is_Load()) {
1454 // Move memory users of a store first.
1455 move_inst_mem(n, orig_phis, igvn);
1456 }
1457 // Now update memory input
1458 igvn->hash_delete(n);
1459 n->set_req(MemNode::Memory, nmem);
1460 igvn->hash_insert(n);
1461 record_for_optimizer(n);
1462 } else {
1463 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1464 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1465 }
1466 }
1467 }
1468 #ifdef ASSERT
1469 // Verify that memory was split correctly
1470 while (old_mems.is_nonempty()) {
1471 Node* old_mem = old_mems.node();
1472 uint old_cnt = old_mems.index();
1473 old_mems.pop();
1474 assert(old_cnt = old_mem->outcnt(), "old mem could be lost");
1475 }
1476 #endif
1477 }
1479 bool ConnectionGraph::has_candidates(Compile *C) {
1480 // EA brings benefits only when the code has allocations and/or locks which
1481 // are represented by ideal Macro nodes.
1482 int cnt = C->macro_count();
1483 for( int i=0; i < cnt; i++ ) {
1484 Node *n = C->macro_node(i);
1485 if ( n->is_Allocate() )
1486 return true;
1487 if( n->is_Lock() ) {
1488 Node* obj = n->as_Lock()->obj_node()->uncast();
1489 if( !(obj->is_Parm() || obj->is_Con()) )
1490 return true;
1491 }
1492 }
1493 return false;
1494 }
1496 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1497 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1498 // to create space for them in ConnectionGraph::_nodes[].
1499 Node* oop_null = igvn->zerocon(T_OBJECT);
1500 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1502 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1503 // Perform escape analysis
1504 if (congraph->compute_escape()) {
1505 // There are non escaping objects.
1506 C->set_congraph(congraph);
1507 }
1509 // Cleanup.
1510 if (oop_null->outcnt() == 0)
1511 igvn->hash_delete(oop_null);
1512 if (noop_null->outcnt() == 0)
1513 igvn->hash_delete(noop_null);
1514 }
1516 bool ConnectionGraph::compute_escape() {
1517 Compile* C = _compile;
1519 // 1. Populate Connection Graph (CG) with Ideal nodes.
1521 Unique_Node_List worklist_init;
1522 worklist_init.map(C->unique(), NULL); // preallocate space
1524 // Initialize worklist
1525 if (C->root() != NULL) {
1526 worklist_init.push(C->root());
1527 }
1529 GrowableArray<int> cg_worklist;
1530 PhaseGVN* igvn = _igvn;
1531 bool has_allocations = false;
1533 // Push all useful nodes onto CG list and set their type.
1534 for( uint next = 0; next < worklist_init.size(); ++next ) {
1535 Node* n = worklist_init.at(next);
1536 record_for_escape_analysis(n, igvn);
1537 // Only allocations and java static calls results are checked
1538 // for an escape status. See process_call_result() below.
1539 if (n->is_Allocate() || n->is_CallStaticJava() &&
1540 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1541 has_allocations = true;
1542 }
1543 if(n->is_AddP()) {
1544 // Collect address nodes. Use them during stage 3 below
1545 // to build initial connection graph field edges.
1546 cg_worklist.append(n->_idx);
1547 } else if (n->is_MergeMem()) {
1548 // Collect all MergeMem nodes to add memory slices for
1549 // scalar replaceable objects in split_unique_types().
1550 _mergemem_worklist.append(n->as_MergeMem());
1551 }
1552 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1553 Node* m = n->fast_out(i); // Get user
1554 worklist_init.push(m);
1555 }
1556 }
1558 if (!has_allocations) {
1559 _collecting = false;
1560 return false; // Nothing to do.
1561 }
1563 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1564 uint delayed_size = _delayed_worklist.size();
1565 for( uint next = 0; next < delayed_size; ++next ) {
1566 Node* n = _delayed_worklist.at(next);
1567 build_connection_graph(n, igvn);
1568 }
1570 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1571 // to reduce number of iterations during stage 4 below.
1572 uint cg_length = cg_worklist.length();
1573 for( uint next = 0; next < cg_length; ++next ) {
1574 int ni = cg_worklist.at(next);
1575 Node* n = ptnode_adr(ni)->_node;
1576 Node* base = get_addp_base(n);
1577 if (base->is_Proj())
1578 base = base->in(0);
1579 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1580 if (nt == PointsToNode::JavaObject) {
1581 build_connection_graph(n, igvn);
1582 }
1583 }
1585 cg_worklist.clear();
1586 cg_worklist.append(_phantom_object);
1587 GrowableArray<uint> worklist;
1589 // 4. Build Connection Graph which need
1590 // to walk the connection graph.
1591 _progress = false;
1592 for (uint ni = 0; ni < nodes_size(); ni++) {
1593 PointsToNode* ptn = ptnode_adr(ni);
1594 Node *n = ptn->_node;
1595 if (n != NULL) { // Call, AddP, LoadP, StoreP
1596 build_connection_graph(n, igvn);
1597 if (ptn->node_type() != PointsToNode::UnknownType)
1598 cg_worklist.append(n->_idx); // Collect CG nodes
1599 if (!_processed.test(n->_idx))
1600 worklist.append(n->_idx); // Collect C/A/L/S nodes
1601 }
1602 }
1604 // After IGVN user nodes may have smaller _idx than
1605 // their inputs so they will be processed first in
1606 // previous loop. Because of that not all Graph
1607 // edges will be created. Walk over interesting
1608 // nodes again until no new edges are created.
1609 //
1610 // Normally only 1-3 passes needed to build
1611 // Connection Graph depending on graph complexity.
1612 // Observed 8 passes in jvm2008 compiler.compiler.
1613 // Set limit to 20 to catch situation when something
1614 // did go wrong and recompile the method without EA.
1616 #define CG_BUILD_ITER_LIMIT 20
1618 uint length = worklist.length();
1619 int iterations = 0;
1620 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) {
1621 _progress = false;
1622 for( uint next = 0; next < length; ++next ) {
1623 int ni = worklist.at(next);
1624 PointsToNode* ptn = ptnode_adr(ni);
1625 Node* n = ptn->_node;
1626 assert(n != NULL, "should be known node");
1627 build_connection_graph(n, igvn);
1628 }
1629 }
1630 if (iterations >= CG_BUILD_ITER_LIMIT) {
1631 assert(iterations < CG_BUILD_ITER_LIMIT,
1632 err_msg("infinite EA connection graph build with %d nodes and worklist size %d",
1633 nodes_size(), length));
1634 // Possible infinite build_connection_graph loop,
1635 // retry compilation without escape analysis.
1636 C->record_failure(C2Compiler::retry_no_escape_analysis());
1637 _collecting = false;
1638 return false;
1639 }
1640 #undef CG_BUILD_ITER_LIMIT
1642 Arena* arena = Thread::current()->resource_area();
1643 VectorSet visited(arena);
1644 worklist.clear();
1646 // 5. Remove deferred edges from the graph and adjust
1647 // escape state of nonescaping objects.
1648 cg_length = cg_worklist.length();
1649 for( uint next = 0; next < cg_length; ++next ) {
1650 int ni = cg_worklist.at(next);
1651 PointsToNode* ptn = ptnode_adr(ni);
1652 PointsToNode::NodeType nt = ptn->node_type();
1653 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1654 remove_deferred(ni, &worklist, &visited);
1655 Node *n = ptn->_node;
1656 if (n->is_AddP()) {
1657 // Search for objects which are not scalar replaceable
1658 // and adjust their escape state.
1659 adjust_escape_state(ni, igvn);
1660 }
1661 }
1662 }
1664 // 6. Propagate escape states.
1665 worklist.clear();
1666 bool has_non_escaping_obj = false;
1668 // push all GlobalEscape nodes on the worklist
1669 for( uint next = 0; next < cg_length; ++next ) {
1670 int nk = cg_worklist.at(next);
1671 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1672 worklist.push(nk);
1673 }
1674 // mark all nodes reachable from GlobalEscape nodes
1675 while(worklist.length() > 0) {
1676 PointsToNode* ptn = ptnode_adr(worklist.pop());
1677 uint e_cnt = ptn->edge_count();
1678 for (uint ei = 0; ei < e_cnt; ei++) {
1679 uint npi = ptn->edge_target(ei);
1680 PointsToNode *np = ptnode_adr(npi);
1681 if (np->escape_state() < PointsToNode::GlobalEscape) {
1682 np->set_escape_state(PointsToNode::GlobalEscape);
1683 worklist.push(npi);
1684 }
1685 }
1686 }
1688 // push all ArgEscape nodes on the worklist
1689 for( uint next = 0; next < cg_length; ++next ) {
1690 int nk = cg_worklist.at(next);
1691 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1692 worklist.push(nk);
1693 }
1694 // mark all nodes reachable from ArgEscape nodes
1695 while(worklist.length() > 0) {
1696 PointsToNode* ptn = ptnode_adr(worklist.pop());
1697 if (ptn->node_type() == PointsToNode::JavaObject)
1698 has_non_escaping_obj = true; // Non GlobalEscape
1699 uint e_cnt = ptn->edge_count();
1700 for (uint ei = 0; ei < e_cnt; ei++) {
1701 uint npi = ptn->edge_target(ei);
1702 PointsToNode *np = ptnode_adr(npi);
1703 if (np->escape_state() < PointsToNode::ArgEscape) {
1704 np->set_escape_state(PointsToNode::ArgEscape);
1705 worklist.push(npi);
1706 }
1707 }
1708 }
1710 GrowableArray<Node*> alloc_worklist;
1712 // push all NoEscape nodes on the worklist
1713 for( uint next = 0; next < cg_length; ++next ) {
1714 int nk = cg_worklist.at(next);
1715 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1716 worklist.push(nk);
1717 }
1718 // mark all nodes reachable from NoEscape nodes
1719 while(worklist.length() > 0) {
1720 PointsToNode* ptn = ptnode_adr(worklist.pop());
1721 if (ptn->node_type() == PointsToNode::JavaObject)
1722 has_non_escaping_obj = true; // Non GlobalEscape
1723 Node* n = ptn->_node;
1724 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1725 // Push scalar replaceable allocations on alloc_worklist
1726 // for processing in split_unique_types().
1727 alloc_worklist.append(n);
1728 }
1729 uint e_cnt = ptn->edge_count();
1730 for (uint ei = 0; ei < e_cnt; ei++) {
1731 uint npi = ptn->edge_target(ei);
1732 PointsToNode *np = ptnode_adr(npi);
1733 if (np->escape_state() < PointsToNode::NoEscape) {
1734 np->set_escape_state(PointsToNode::NoEscape);
1735 worklist.push(npi);
1736 }
1737 }
1738 }
1740 _collecting = false;
1741 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1743 #ifndef PRODUCT
1744 if (PrintEscapeAnalysis) {
1745 dump(); // Dump ConnectionGraph
1746 }
1747 #endif
1749 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1750 if ( has_scalar_replaceable_candidates &&
1751 C->AliasLevel() >= 3 && EliminateAllocations ) {
1753 // Now use the escape information to create unique types for
1754 // scalar replaceable objects.
1755 split_unique_types(alloc_worklist);
1757 if (C->failing()) return false;
1759 C->print_method("After Escape Analysis", 2);
1761 #ifdef ASSERT
1762 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1763 tty->print("=== No allocations eliminated for ");
1764 C->method()->print_short_name();
1765 if(!EliminateAllocations) {
1766 tty->print(" since EliminateAllocations is off ===");
1767 } else if(!has_scalar_replaceable_candidates) {
1768 tty->print(" since there are no scalar replaceable candidates ===");
1769 } else if(C->AliasLevel() < 3) {
1770 tty->print(" since AliasLevel < 3 ===");
1771 }
1772 tty->cr();
1773 #endif
1774 }
1775 return has_non_escaping_obj;
1776 }
1778 // Adjust escape state after Connection Graph is built.
1779 void ConnectionGraph::adjust_escape_state(int nidx, PhaseTransform* phase) {
1780 PointsToNode* ptn = ptnode_adr(nidx);
1781 Node* n = ptn->_node;
1782 assert(n->is_AddP(), "Should be called for AddP nodes only");
1783 // Search for objects which are not scalar replaceable.
1784 // Mark their escape state as ArgEscape to propagate the state
1785 // to referenced objects.
1786 // Note: currently there are no difference in compiler optimizations
1787 // for ArgEscape objects and NoEscape objects which are not
1788 // scalar replaceable.
1790 Compile* C = _compile;
1792 int offset = ptn->offset();
1793 Node* base = get_addp_base(n);
1794 VectorSet* ptset = PointsTo(base);
1795 int ptset_size = ptset->Size();
1797 // Check if a oop field's initializing value is recorded and add
1798 // a corresponding NULL field's value if it is not recorded.
1799 // Connection Graph does not record a default initialization by NULL
1800 // captured by Initialize node.
1801 //
1802 // Note: it will disable scalar replacement in some cases:
1803 //
1804 // Point p[] = new Point[1];
1805 // p[0] = new Point(); // Will be not scalar replaced
1806 //
1807 // but it will save us from incorrect optimizations in next cases:
1808 //
1809 // Point p[] = new Point[1];
1810 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1811 //
1812 // Do a simple control flow analysis to distinguish above cases.
1813 //
1814 if (offset != Type::OffsetBot && ptset_size == 1) {
1815 uint elem = ptset->getelem(); // Allocation node's index
1816 // It does not matter if it is not Allocation node since
1817 // only non-escaping allocations are scalar replaced.
1818 if (ptnode_adr(elem)->_node->is_Allocate() &&
1819 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1820 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1821 InitializeNode* ini = alloc->initialization();
1823 // Check only oop fields.
1824 const Type* adr_type = n->as_AddP()->bottom_type();
1825 BasicType basic_field_type = T_INT;
1826 if (adr_type->isa_instptr()) {
1827 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1828 if (field != NULL) {
1829 basic_field_type = field->layout_type();
1830 } else {
1831 // Ignore non field load (for example, klass load)
1832 }
1833 } else if (adr_type->isa_aryptr()) {
1834 const Type* elemtype = adr_type->isa_aryptr()->elem();
1835 basic_field_type = elemtype->array_element_basic_type();
1836 } else {
1837 // Raw pointers are used for initializing stores so skip it.
1838 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1839 (base->in(0) == alloc),"unexpected pointer type");
1840 }
1841 if (basic_field_type == T_OBJECT ||
1842 basic_field_type == T_NARROWOOP ||
1843 basic_field_type == T_ARRAY) {
1844 Node* value = NULL;
1845 if (ini != NULL) {
1846 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1847 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1848 if (store != NULL && store->is_Store()) {
1849 value = store->in(MemNode::ValueIn);
1850 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1851 // Check for a store which follows allocation without branches.
1852 // For example, a volatile field store is not collected
1853 // by Initialize node. TODO: it would be nice to use idom() here.
1854 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1855 store = n->fast_out(i);
1856 if (store->is_Store() && store->in(0) != NULL) {
1857 Node* ctrl = store->in(0);
1858 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1859 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1860 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1861 ctrl = ctrl->in(0);
1862 }
1863 if (ctrl == ini || ctrl == alloc) {
1864 value = store->in(MemNode::ValueIn);
1865 break;
1866 }
1867 }
1868 }
1869 }
1870 }
1871 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1872 // A field's initializing value was not recorded. Add NULL.
1873 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1874 add_pointsto_edge(nidx, null_idx);
1875 }
1876 }
1877 }
1878 }
1880 // An object is not scalar replaceable if the field which may point
1881 // to it has unknown offset (unknown element of an array of objects).
1882 //
1883 if (offset == Type::OffsetBot) {
1884 uint e_cnt = ptn->edge_count();
1885 for (uint ei = 0; ei < e_cnt; ei++) {
1886 uint npi = ptn->edge_target(ei);
1887 set_escape_state(npi, PointsToNode::ArgEscape);
1888 ptnode_adr(npi)->_scalar_replaceable = false;
1889 }
1890 }
1892 // Currently an object is not scalar replaceable if a LoadStore node
1893 // access its field since the field value is unknown after it.
1894 //
1895 bool has_LoadStore = false;
1896 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1897 Node *use = n->fast_out(i);
1898 if (use->is_LoadStore()) {
1899 has_LoadStore = true;
1900 break;
1901 }
1902 }
1903 // An object is not scalar replaceable if the address points
1904 // to unknown field (unknown element for arrays, offset is OffsetBot).
1905 //
1906 // Or the address may point to more then one object. This may produce
1907 // the false positive result (set scalar_replaceable to false)
1908 // since the flow-insensitive escape analysis can't separate
1909 // the case when stores overwrite the field's value from the case
1910 // when stores happened on different control branches.
1911 //
1912 if (ptset_size > 1 || ptset_size != 0 &&
1913 (has_LoadStore || offset == Type::OffsetBot)) {
1914 for( VectorSetI j(ptset); j.test(); ++j ) {
1915 set_escape_state(j.elem, PointsToNode::ArgEscape);
1916 ptnode_adr(j.elem)->_scalar_replaceable = false;
1917 }
1918 }
1919 }
1921 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1923 switch (call->Opcode()) {
1924 #ifdef ASSERT
1925 case Op_Allocate:
1926 case Op_AllocateArray:
1927 case Op_Lock:
1928 case Op_Unlock:
1929 assert(false, "should be done already");
1930 break;
1931 #endif
1932 case Op_CallLeaf:
1933 case Op_CallLeafNoFP:
1934 {
1935 // Stub calls, objects do not escape but they are not scale replaceable.
1936 // Adjust escape state for outgoing arguments.
1937 const TypeTuple * d = call->tf()->domain();
1938 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1939 const Type* at = d->field_at(i);
1940 Node *arg = call->in(i)->uncast();
1941 const Type *aat = phase->type(arg);
1942 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
1943 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
1945 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1946 aat->isa_ptr() != NULL, "expecting an Ptr");
1947 #ifdef ASSERT
1948 if (!(call->Opcode() == Op_CallLeafNoFP &&
1949 call->as_CallLeaf()->_name != NULL &&
1950 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
1951 call->as_CallLeaf()->_name != NULL &&
1952 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
1953 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
1954 ) {
1955 call->dump();
1956 assert(false, "EA: unexpected CallLeaf");
1957 }
1958 #endif
1959 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1960 if (arg->is_AddP()) {
1961 //
1962 // The inline_native_clone() case when the arraycopy stub is called
1963 // after the allocation before Initialize and CheckCastPP nodes.
1964 //
1965 // Set AddP's base (Allocate) as not scalar replaceable since
1966 // pointer to the base (with offset) is passed as argument.
1967 //
1968 arg = get_addp_base(arg);
1969 }
1970 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
1971 uint pt = j.elem;
1972 set_escape_state(pt, PointsToNode::ArgEscape);
1973 }
1974 }
1975 }
1976 break;
1977 }
1979 case Op_CallStaticJava:
1980 // For a static call, we know exactly what method is being called.
1981 // Use bytecode estimator to record the call's escape affects
1982 {
1983 ciMethod *meth = call->as_CallJava()->method();
1984 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1985 // fall-through if not a Java method or no analyzer information
1986 if (call_analyzer != NULL) {
1987 const TypeTuple * d = call->tf()->domain();
1988 bool copy_dependencies = false;
1989 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1990 const Type* at = d->field_at(i);
1991 int k = i - TypeFunc::Parms;
1992 Node *arg = call->in(i)->uncast();
1994 if (at->isa_oopptr() != NULL &&
1995 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
1997 bool global_escapes = false;
1998 bool fields_escapes = false;
1999 if (!call_analyzer->is_arg_stack(k)) {
2000 // The argument global escapes, mark everything it could point to
2001 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2002 global_escapes = true;
2003 } else {
2004 if (!call_analyzer->is_arg_local(k)) {
2005 // The argument itself doesn't escape, but any fields might
2006 fields_escapes = true;
2007 }
2008 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2009 copy_dependencies = true;
2010 }
2012 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2013 uint pt = j.elem;
2014 if (global_escapes) {
2015 //The argument global escapes, mark everything it could point to
2016 set_escape_state(pt, PointsToNode::GlobalEscape);
2017 } else {
2018 if (fields_escapes) {
2019 // The argument itself doesn't escape, but any fields might
2020 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2021 }
2022 set_escape_state(pt, PointsToNode::ArgEscape);
2023 }
2024 }
2025 }
2026 }
2027 if (copy_dependencies)
2028 call_analyzer->copy_dependencies(_compile->dependencies());
2029 break;
2030 }
2031 }
2033 default:
2034 // Fall-through here if not a Java method or no analyzer information
2035 // or some other type of call, assume the worst case: all arguments
2036 // globally escape.
2037 {
2038 // adjust escape state for outgoing arguments
2039 const TypeTuple * d = call->tf()->domain();
2040 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2041 const Type* at = d->field_at(i);
2042 if (at->isa_oopptr() != NULL) {
2043 Node *arg = call->in(i)->uncast();
2044 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2045 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2046 uint pt = j.elem;
2047 set_escape_state(pt, PointsToNode::GlobalEscape);
2048 }
2049 }
2050 }
2051 }
2052 }
2053 }
2054 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2055 CallNode *call = resproj->in(0)->as_Call();
2056 uint call_idx = call->_idx;
2057 uint resproj_idx = resproj->_idx;
2059 switch (call->Opcode()) {
2060 case Op_Allocate:
2061 {
2062 Node *k = call->in(AllocateNode::KlassNode);
2063 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2064 assert(kt != NULL, "TypeKlassPtr required.");
2065 ciKlass* cik = kt->klass();
2067 PointsToNode::EscapeState es;
2068 uint edge_to;
2069 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2070 !cik->is_instance_klass() || // StressReflectiveCode
2071 cik->as_instance_klass()->has_finalizer()) {
2072 es = PointsToNode::GlobalEscape;
2073 edge_to = _phantom_object; // Could not be worse
2074 } else {
2075 es = PointsToNode::NoEscape;
2076 edge_to = call_idx;
2077 }
2078 set_escape_state(call_idx, es);
2079 add_pointsto_edge(resproj_idx, edge_to);
2080 _processed.set(resproj_idx);
2081 break;
2082 }
2084 case Op_AllocateArray:
2085 {
2087 Node *k = call->in(AllocateNode::KlassNode);
2088 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2089 assert(kt != NULL, "TypeKlassPtr required.");
2090 ciKlass* cik = kt->klass();
2092 PointsToNode::EscapeState es;
2093 uint edge_to;
2094 if (!cik->is_array_klass()) { // StressReflectiveCode
2095 es = PointsToNode::GlobalEscape;
2096 edge_to = _phantom_object;
2097 } else {
2098 es = PointsToNode::NoEscape;
2099 edge_to = call_idx;
2100 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2101 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2102 // Not scalar replaceable if the length is not constant or too big.
2103 ptnode_adr(call_idx)->_scalar_replaceable = false;
2104 }
2105 }
2106 set_escape_state(call_idx, es);
2107 add_pointsto_edge(resproj_idx, edge_to);
2108 _processed.set(resproj_idx);
2109 break;
2110 }
2112 case Op_CallStaticJava:
2113 // For a static call, we know exactly what method is being called.
2114 // Use bytecode estimator to record whether the call's return value escapes
2115 {
2116 bool done = true;
2117 const TypeTuple *r = call->tf()->range();
2118 const Type* ret_type = NULL;
2120 if (r->cnt() > TypeFunc::Parms)
2121 ret_type = r->field_at(TypeFunc::Parms);
2123 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2124 // _multianewarray functions return a TypeRawPtr.
2125 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2126 _processed.set(resproj_idx);
2127 break; // doesn't return a pointer type
2128 }
2129 ciMethod *meth = call->as_CallJava()->method();
2130 const TypeTuple * d = call->tf()->domain();
2131 if (meth == NULL) {
2132 // not a Java method, assume global escape
2133 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2134 add_pointsto_edge(resproj_idx, _phantom_object);
2135 } else {
2136 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2137 bool copy_dependencies = false;
2139 if (call_analyzer->is_return_allocated()) {
2140 // Returns a newly allocated unescaped object, simply
2141 // update dependency information.
2142 // Mark it as NoEscape so that objects referenced by
2143 // it's fields will be marked as NoEscape at least.
2144 set_escape_state(call_idx, PointsToNode::NoEscape);
2145 add_pointsto_edge(resproj_idx, call_idx);
2146 copy_dependencies = true;
2147 } else if (call_analyzer->is_return_local()) {
2148 // determine whether any arguments are returned
2149 set_escape_state(call_idx, PointsToNode::NoEscape);
2150 bool ret_arg = false;
2151 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2152 const Type* at = d->field_at(i);
2154 if (at->isa_oopptr() != NULL) {
2155 Node *arg = call->in(i)->uncast();
2157 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2158 ret_arg = true;
2159 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2160 if (arg_esp->node_type() == PointsToNode::UnknownType)
2161 done = false;
2162 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2163 add_pointsto_edge(resproj_idx, arg->_idx);
2164 else
2165 add_deferred_edge(resproj_idx, arg->_idx);
2166 arg_esp->_hidden_alias = true;
2167 }
2168 }
2169 }
2170 if (done && !ret_arg) {
2171 // Returns unknown object.
2172 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2173 add_pointsto_edge(resproj_idx, _phantom_object);
2174 }
2175 copy_dependencies = true;
2176 } else {
2177 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2178 add_pointsto_edge(resproj_idx, _phantom_object);
2179 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2180 const Type* at = d->field_at(i);
2181 if (at->isa_oopptr() != NULL) {
2182 Node *arg = call->in(i)->uncast();
2183 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2184 arg_esp->_hidden_alias = true;
2185 }
2186 }
2187 }
2188 if (copy_dependencies)
2189 call_analyzer->copy_dependencies(_compile->dependencies());
2190 }
2191 if (done)
2192 _processed.set(resproj_idx);
2193 break;
2194 }
2196 default:
2197 // Some other type of call, assume the worst case that the
2198 // returned value, if any, globally escapes.
2199 {
2200 const TypeTuple *r = call->tf()->range();
2201 if (r->cnt() > TypeFunc::Parms) {
2202 const Type* ret_type = r->field_at(TypeFunc::Parms);
2204 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2205 // _multianewarray functions return a TypeRawPtr.
2206 if (ret_type->isa_ptr() != NULL) {
2207 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2208 add_pointsto_edge(resproj_idx, _phantom_object);
2209 }
2210 }
2211 _processed.set(resproj_idx);
2212 }
2213 }
2214 }
2216 // Populate Connection Graph with Ideal nodes and create simple
2217 // connection graph edges (do not need to check the node_type of inputs
2218 // or to call PointsTo() to walk the connection graph).
2219 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2220 if (_processed.test(n->_idx))
2221 return; // No need to redefine node's state.
2223 if (n->is_Call()) {
2224 // Arguments to allocation and locking don't escape.
2225 if (n->is_Allocate()) {
2226 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2227 record_for_optimizer(n);
2228 } else if (n->is_Lock() || n->is_Unlock()) {
2229 // Put Lock and Unlock nodes on IGVN worklist to process them during
2230 // the first IGVN optimization when escape information is still available.
2231 record_for_optimizer(n);
2232 _processed.set(n->_idx);
2233 } else {
2234 // Don't mark as processed since call's arguments have to be processed.
2235 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2236 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2238 // Check if a call returns an object.
2239 const TypeTuple *r = n->as_Call()->tf()->range();
2240 if (r->cnt() > TypeFunc::Parms &&
2241 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2242 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2243 nt = PointsToNode::JavaObject;
2244 if (!n->is_CallStaticJava()) {
2245 // Since the called mathod is statically unknown assume
2246 // the worst case that the returned value globally escapes.
2247 es = PointsToNode::GlobalEscape;
2248 }
2249 }
2250 add_node(n, nt, es, false);
2251 }
2252 return;
2253 }
2255 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2256 // ThreadLocal has RawPrt type.
2257 switch (n->Opcode()) {
2258 case Op_AddP:
2259 {
2260 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2261 break;
2262 }
2263 case Op_CastX2P:
2264 { // "Unsafe" memory access.
2265 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2266 break;
2267 }
2268 case Op_CastPP:
2269 case Op_CheckCastPP:
2270 case Op_EncodeP:
2271 case Op_DecodeN:
2272 {
2273 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2274 int ti = n->in(1)->_idx;
2275 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2276 if (nt == PointsToNode::UnknownType) {
2277 _delayed_worklist.push(n); // Process it later.
2278 break;
2279 } else if (nt == PointsToNode::JavaObject) {
2280 add_pointsto_edge(n->_idx, ti);
2281 } else {
2282 add_deferred_edge(n->_idx, ti);
2283 }
2284 _processed.set(n->_idx);
2285 break;
2286 }
2287 case Op_ConP:
2288 {
2289 // assume all pointer constants globally escape except for null
2290 PointsToNode::EscapeState es;
2291 if (phase->type(n) == TypePtr::NULL_PTR)
2292 es = PointsToNode::NoEscape;
2293 else
2294 es = PointsToNode::GlobalEscape;
2296 add_node(n, PointsToNode::JavaObject, es, true);
2297 break;
2298 }
2299 case Op_ConN:
2300 {
2301 // assume all narrow oop constants globally escape except for null
2302 PointsToNode::EscapeState es;
2303 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2304 es = PointsToNode::NoEscape;
2305 else
2306 es = PointsToNode::GlobalEscape;
2308 add_node(n, PointsToNode::JavaObject, es, true);
2309 break;
2310 }
2311 case Op_CreateEx:
2312 {
2313 // assume that all exception objects globally escape
2314 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2315 break;
2316 }
2317 case Op_LoadKlass:
2318 case Op_LoadNKlass:
2319 {
2320 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2321 break;
2322 }
2323 case Op_LoadP:
2324 case Op_LoadN:
2325 {
2326 const Type *t = phase->type(n);
2327 if (t->make_ptr() == NULL) {
2328 _processed.set(n->_idx);
2329 return;
2330 }
2331 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2332 break;
2333 }
2334 case Op_Parm:
2335 {
2336 _processed.set(n->_idx); // No need to redefine it state.
2337 uint con = n->as_Proj()->_con;
2338 if (con < TypeFunc::Parms)
2339 return;
2340 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2341 if (t->isa_ptr() == NULL)
2342 return;
2343 // We have to assume all input parameters globally escape
2344 // (Note: passing 'false' since _processed is already set).
2345 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2346 break;
2347 }
2348 case Op_Phi:
2349 {
2350 const Type *t = n->as_Phi()->type();
2351 if (t->make_ptr() == NULL) {
2352 // nothing to do if not an oop or narrow oop
2353 _processed.set(n->_idx);
2354 return;
2355 }
2356 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2357 uint i;
2358 for (i = 1; i < n->req() ; i++) {
2359 Node* in = n->in(i);
2360 if (in == NULL)
2361 continue; // ignore NULL
2362 in = in->uncast();
2363 if (in->is_top() || in == n)
2364 continue; // ignore top or inputs which go back this node
2365 int ti = in->_idx;
2366 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2367 if (nt == PointsToNode::UnknownType) {
2368 break;
2369 } else if (nt == PointsToNode::JavaObject) {
2370 add_pointsto_edge(n->_idx, ti);
2371 } else {
2372 add_deferred_edge(n->_idx, ti);
2373 }
2374 }
2375 if (i >= n->req())
2376 _processed.set(n->_idx);
2377 else
2378 _delayed_worklist.push(n);
2379 break;
2380 }
2381 case Op_Proj:
2382 {
2383 // we are only interested in the oop result projection from a call
2384 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2385 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2386 assert(r->cnt() > TypeFunc::Parms, "sanity");
2387 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2388 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2389 int ti = n->in(0)->_idx;
2390 // The call may not be registered yet (since not all its inputs are registered)
2391 // if this is the projection from backbranch edge of Phi.
2392 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2393 process_call_result(n->as_Proj(), phase);
2394 }
2395 if (!_processed.test(n->_idx)) {
2396 // The call's result may need to be processed later if the call
2397 // returns it's argument and the argument is not processed yet.
2398 _delayed_worklist.push(n);
2399 }
2400 break;
2401 }
2402 }
2403 _processed.set(n->_idx);
2404 break;
2405 }
2406 case Op_Return:
2407 {
2408 if( n->req() > TypeFunc::Parms &&
2409 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2410 // Treat Return value as LocalVar with GlobalEscape escape state.
2411 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2412 int ti = n->in(TypeFunc::Parms)->_idx;
2413 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2414 if (nt == PointsToNode::UnknownType) {
2415 _delayed_worklist.push(n); // Process it later.
2416 break;
2417 } else if (nt == PointsToNode::JavaObject) {
2418 add_pointsto_edge(n->_idx, ti);
2419 } else {
2420 add_deferred_edge(n->_idx, ti);
2421 }
2422 }
2423 _processed.set(n->_idx);
2424 break;
2425 }
2426 case Op_StoreP:
2427 case Op_StoreN:
2428 {
2429 const Type *adr_type = phase->type(n->in(MemNode::Address));
2430 adr_type = adr_type->make_ptr();
2431 if (adr_type->isa_oopptr()) {
2432 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2433 } else {
2434 Node* adr = n->in(MemNode::Address);
2435 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2436 adr->in(AddPNode::Address)->is_Proj() &&
2437 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2438 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2439 // We are computing a raw address for a store captured
2440 // by an Initialize compute an appropriate address type.
2441 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2442 assert(offs != Type::OffsetBot, "offset must be a constant");
2443 } else {
2444 _processed.set(n->_idx);
2445 return;
2446 }
2447 }
2448 break;
2449 }
2450 case Op_StorePConditional:
2451 case Op_CompareAndSwapP:
2452 case Op_CompareAndSwapN:
2453 {
2454 const Type *adr_type = phase->type(n->in(MemNode::Address));
2455 adr_type = adr_type->make_ptr();
2456 if (adr_type->isa_oopptr()) {
2457 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2458 } else {
2459 _processed.set(n->_idx);
2460 return;
2461 }
2462 break;
2463 }
2464 case Op_AryEq:
2465 case Op_StrComp:
2466 case Op_StrEquals:
2467 case Op_StrIndexOf:
2468 {
2469 // char[] arrays passed to string intrinsics are not scalar replaceable.
2470 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2471 break;
2472 }
2473 case Op_ThreadLocal:
2474 {
2475 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2476 break;
2477 }
2478 default:
2479 ;
2480 // nothing to do
2481 }
2482 return;
2483 }
2485 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2486 uint n_idx = n->_idx;
2487 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2489 // Don't set processed bit for AddP, LoadP, StoreP since
2490 // they may need more then one pass to process.
2491 // Also don't mark as processed Call nodes since their
2492 // arguments may need more then one pass to process.
2493 if (_processed.test(n_idx))
2494 return; // No need to redefine node's state.
2496 if (n->is_Call()) {
2497 CallNode *call = n->as_Call();
2498 process_call_arguments(call, phase);
2499 return;
2500 }
2502 switch (n->Opcode()) {
2503 case Op_AddP:
2504 {
2505 Node *base = get_addp_base(n);
2506 // Create a field edge to this node from everything base could point to.
2507 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) {
2508 uint pt = i.elem;
2509 add_field_edge(pt, n_idx, address_offset(n, phase));
2510 }
2511 break;
2512 }
2513 case Op_CastX2P:
2514 {
2515 assert(false, "Op_CastX2P");
2516 break;
2517 }
2518 case Op_CastPP:
2519 case Op_CheckCastPP:
2520 case Op_EncodeP:
2521 case Op_DecodeN:
2522 {
2523 int ti = n->in(1)->_idx;
2524 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2525 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2526 add_pointsto_edge(n_idx, ti);
2527 } else {
2528 add_deferred_edge(n_idx, ti);
2529 }
2530 _processed.set(n_idx);
2531 break;
2532 }
2533 case Op_ConP:
2534 {
2535 assert(false, "Op_ConP");
2536 break;
2537 }
2538 case Op_ConN:
2539 {
2540 assert(false, "Op_ConN");
2541 break;
2542 }
2543 case Op_CreateEx:
2544 {
2545 assert(false, "Op_CreateEx");
2546 break;
2547 }
2548 case Op_LoadKlass:
2549 case Op_LoadNKlass:
2550 {
2551 assert(false, "Op_LoadKlass");
2552 break;
2553 }
2554 case Op_LoadP:
2555 case Op_LoadN:
2556 {
2557 const Type *t = phase->type(n);
2558 #ifdef ASSERT
2559 if (t->make_ptr() == NULL)
2560 assert(false, "Op_LoadP");
2561 #endif
2563 Node* adr = n->in(MemNode::Address)->uncast();
2564 Node* adr_base;
2565 if (adr->is_AddP()) {
2566 adr_base = get_addp_base(adr);
2567 } else {
2568 adr_base = adr;
2569 }
2571 // For everything "adr_base" could point to, create a deferred edge from
2572 // this node to each field with the same offset.
2573 int offset = address_offset(adr, phase);
2574 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2575 uint pt = i.elem;
2576 add_deferred_edge_to_fields(n_idx, pt, offset);
2577 }
2578 break;
2579 }
2580 case Op_Parm:
2581 {
2582 assert(false, "Op_Parm");
2583 break;
2584 }
2585 case Op_Phi:
2586 {
2587 #ifdef ASSERT
2588 const Type *t = n->as_Phi()->type();
2589 if (t->make_ptr() == NULL)
2590 assert(false, "Op_Phi");
2591 #endif
2592 for (uint i = 1; i < n->req() ; i++) {
2593 Node* in = n->in(i);
2594 if (in == NULL)
2595 continue; // ignore NULL
2596 in = in->uncast();
2597 if (in->is_top() || in == n)
2598 continue; // ignore top or inputs which go back this node
2599 int ti = in->_idx;
2600 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2601 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2602 if (nt == PointsToNode::JavaObject) {
2603 add_pointsto_edge(n_idx, ti);
2604 } else {
2605 add_deferred_edge(n_idx, ti);
2606 }
2607 }
2608 _processed.set(n_idx);
2609 break;
2610 }
2611 case Op_Proj:
2612 {
2613 // we are only interested in the oop result projection from a call
2614 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2615 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2616 "all nodes should be registered");
2617 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2618 assert(r->cnt() > TypeFunc::Parms, "sanity");
2619 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2620 process_call_result(n->as_Proj(), phase);
2621 assert(_processed.test(n_idx), "all call results should be processed");
2622 break;
2623 }
2624 }
2625 assert(false, "Op_Proj");
2626 break;
2627 }
2628 case Op_Return:
2629 {
2630 #ifdef ASSERT
2631 if( n->req() <= TypeFunc::Parms ||
2632 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2633 assert(false, "Op_Return");
2634 }
2635 #endif
2636 int ti = n->in(TypeFunc::Parms)->_idx;
2637 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2638 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2639 add_pointsto_edge(n_idx, ti);
2640 } else {
2641 add_deferred_edge(n_idx, ti);
2642 }
2643 _processed.set(n_idx);
2644 break;
2645 }
2646 case Op_StoreP:
2647 case Op_StoreN:
2648 case Op_StorePConditional:
2649 case Op_CompareAndSwapP:
2650 case Op_CompareAndSwapN:
2651 {
2652 Node *adr = n->in(MemNode::Address);
2653 const Type *adr_type = phase->type(adr)->make_ptr();
2654 #ifdef ASSERT
2655 if (!adr_type->isa_oopptr())
2656 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2657 #endif
2659 assert(adr->is_AddP(), "expecting an AddP");
2660 Node *adr_base = get_addp_base(adr);
2661 Node *val = n->in(MemNode::ValueIn)->uncast();
2662 // For everything "adr_base" could point to, create a deferred edge
2663 // to "val" from each field with the same offset.
2664 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2665 uint pt = i.elem;
2666 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2667 }
2668 break;
2669 }
2670 case Op_AryEq:
2671 case Op_StrComp:
2672 case Op_StrEquals:
2673 case Op_StrIndexOf:
2674 {
2675 // char[] arrays passed to string intrinsic do not escape but
2676 // they are not scalar replaceable. Adjust escape state for them.
2677 // Start from in(2) edge since in(1) is memory edge.
2678 for (uint i = 2; i < n->req(); i++) {
2679 Node* adr = n->in(i)->uncast();
2680 const Type *at = phase->type(adr);
2681 if (!adr->is_top() && at->isa_ptr()) {
2682 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2683 at->isa_ptr() != NULL, "expecting an Ptr");
2684 if (adr->is_AddP()) {
2685 adr = get_addp_base(adr);
2686 }
2687 // Mark as ArgEscape everything "adr" could point to.
2688 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2689 }
2690 }
2691 _processed.set(n_idx);
2692 break;
2693 }
2694 case Op_ThreadLocal:
2695 {
2696 assert(false, "Op_ThreadLocal");
2697 break;
2698 }
2699 default:
2700 // This method should be called only for EA specific nodes.
2701 ShouldNotReachHere();
2702 }
2703 }
2705 #ifndef PRODUCT
2706 void ConnectionGraph::dump() {
2707 bool first = true;
2709 uint size = nodes_size();
2710 for (uint ni = 0; ni < size; ni++) {
2711 PointsToNode *ptn = ptnode_adr(ni);
2712 PointsToNode::NodeType ptn_type = ptn->node_type();
2714 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2715 continue;
2716 PointsToNode::EscapeState es = escape_state(ptn->_node);
2717 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2718 if (first) {
2719 tty->cr();
2720 tty->print("======== Connection graph for ");
2721 _compile->method()->print_short_name();
2722 tty->cr();
2723 first = false;
2724 }
2725 tty->print("%6d ", ni);
2726 ptn->dump();
2727 // Print all locals which reference this allocation
2728 for (uint li = ni; li < size; li++) {
2729 PointsToNode *ptn_loc = ptnode_adr(li);
2730 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2731 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2732 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2733 ptnode_adr(li)->dump(false);
2734 }
2735 }
2736 if (Verbose) {
2737 // Print all fields which reference this allocation
2738 for (uint i = 0; i < ptn->edge_count(); i++) {
2739 uint ei = ptn->edge_target(i);
2740 ptnode_adr(ei)->dump(false);
2741 }
2742 }
2743 tty->cr();
2744 }
2745 }
2746 }
2747 #endif