1 /*
2 * Copyright 2005-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_escape.cpp.incl"
27
28 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
29 uint v = (targIdx << EdgeShift) + ((uint) et);
30 if (_edges == NULL) {
31 Arena *a = Compile::current()->comp_arena();
32 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
33 }
34 _edges->append_if_missing(v);
35 }
36
37 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
39
40 _edges->remove(v);
41 }
42
43 #ifndef PRODUCT
44 static const char *node_type_names[] = {
45 "UnknownType",
46 "JavaObject",
47 "LocalVar",
48 "Field"
49 };
50
51 static const char *esc_names[] = {
52 "UnknownEscape",
53 "NoEscape",
54 "ArgEscape",
55 "GlobalEscape"
56 };
57
58 static const char *edge_type_suffix[] = {
59 "?", // UnknownEdge
60 "P", // PointsToEdge
61 "D", // DeferredEdge
62 "F" // FieldEdge
63 };
64
65 void PointsToNode::dump(bool print_state) const {
66 NodeType nt = node_type();
67 tty->print("%s ", node_type_names[(int) nt]);
68 if (print_state) {
69 EscapeState es = escape_state();
70 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
71 }
72 tty->print("[[");
73 for (uint i = 0; i < edge_count(); i++) {
74 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
75 }
76 tty->print("]] ");
77 if (_node == NULL)
78 tty->print_cr("<null>");
79 else
80 _node->dump();
81 }
82 #endif
83
84 ConnectionGraph::ConnectionGraph(Compile * C) :
85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
86 _processed(C->comp_arena()),
87 _collecting(true),
88 _compile(C),
89 _node_map(C->comp_arena()) {
90
91 _phantom_object = C->top()->_idx,
92 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
93
94 // Add ConP(#NULL) and ConN(#NULL) nodes.
95 PhaseGVN* igvn = C->initial_gvn();
96 Node* oop_null = igvn->zerocon(T_OBJECT);
97 _oop_null = oop_null->_idx;
98 assert(_oop_null < C->unique(), "should be created already");
99 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
100
101 if (UseCompressedOops) {
102 Node* noop_null = igvn->zerocon(T_NARROWOOP);
103 _noop_null = noop_null->_idx;
104 assert(_noop_null < C->unique(), "should be created already");
105 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
106 }
107 }
108
109 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
110 PointsToNode *f = ptnode_adr(from_i);
111 PointsToNode *t = ptnode_adr(to_i);
112
113 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
114 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
115 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
116 f->add_edge(to_i, PointsToNode::PointsToEdge);
117 }
118
119 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
120 PointsToNode *f = ptnode_adr(from_i);
121 PointsToNode *t = ptnode_adr(to_i);
122
123 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
124 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
125 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
126 // don't add a self-referential edge, this can occur during removal of
127 // deferred edges
128 if (from_i != to_i)
129 f->add_edge(to_i, PointsToNode::DeferredEdge);
130 }
131
132 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
133 const Type *adr_type = phase->type(adr);
134 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
135 adr->in(AddPNode::Address)->is_Proj() &&
136 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
137 // We are computing a raw address for a store captured by an Initialize
138 // compute an appropriate address type. AddP cases #3 and #5 (see below).
139 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
140 assert(offs != Type::OffsetBot ||
141 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
142 "offset must be a constant or it is initialization of array");
143 return offs;
144 }
145 const TypePtr *t_ptr = adr_type->isa_ptr();
146 assert(t_ptr != NULL, "must be a pointer type");
147 return t_ptr->offset();
148 }
149
150 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
151 PointsToNode *f = ptnode_adr(from_i);
152 PointsToNode *t = ptnode_adr(to_i);
153
154 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
155 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
156 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
157 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
158 t->set_offset(offset);
159
160 f->add_edge(to_i, PointsToNode::FieldEdge);
161 }
162
163 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
164 PointsToNode *npt = ptnode_adr(ni);
165 PointsToNode::EscapeState old_es = npt->escape_state();
166 if (es > old_es)
167 npt->set_escape_state(es);
168 }
169
170 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
171 PointsToNode::EscapeState es, bool done) {
172 PointsToNode* ptadr = ptnode_adr(n->_idx);
173 ptadr->_node = n;
174 ptadr->set_node_type(nt);
175
176 // inline set_escape_state(idx, es);
177 PointsToNode::EscapeState old_es = ptadr->escape_state();
178 if (es > old_es)
179 ptadr->set_escape_state(es);
180
181 if (done)
182 _processed.set(n->_idx);
183 }
184
185 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n, PhaseTransform *phase) {
186 uint idx = n->_idx;
187 PointsToNode::EscapeState es;
188
189 // If we are still collecting or there were no non-escaping allocations
190 // we don't know the answer yet
191 if (_collecting)
192 return PointsToNode::UnknownEscape;
193
194 // if the node was created after the escape computation, return
195 // UnknownEscape
196 if (idx >= nodes_size())
197 return PointsToNode::UnknownEscape;
198
199 es = ptnode_adr(idx)->escape_state();
200
201 // if we have already computed a value, return it
202 if (es != PointsToNode::UnknownEscape &&
203 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
204 return es;
205
206 // PointsTo() calls n->uncast() which can return a new ideal node.
207 if (n->uncast()->_idx >= nodes_size())
208 return PointsToNode::UnknownEscape;
209
210 // compute max escape state of anything this node could point to
211 VectorSet ptset(Thread::current()->resource_area());
212 PointsTo(ptset, n, phase);
213 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) {
214 uint pt = i.elem;
215 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
216 if (pes > es)
217 es = pes;
218 }
219 // cache the computed escape state
220 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
221 ptnode_adr(idx)->set_escape_state(es);
222 return es;
223 }
224
225 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n, PhaseTransform *phase) {
226 VectorSet visited(Thread::current()->resource_area());
227 GrowableArray<uint> worklist;
228
229 #ifdef ASSERT
230 Node *orig_n = n;
231 #endif
232
233 n = n->uncast();
234 PointsToNode* npt = ptnode_adr(n->_idx);
235
236 // If we have a JavaObject, return just that object
237 if (npt->node_type() == PointsToNode::JavaObject) {
238 ptset.set(n->_idx);
239 return;
240 }
241 #ifdef ASSERT
242 if (npt->_node == NULL) {
243 if (orig_n != n)
244 orig_n->dump();
245 n->dump();
246 assert(npt->_node != NULL, "unregistered node");
247 }
248 #endif
249 worklist.push(n->_idx);
250 while(worklist.length() > 0) {
251 int ni = worklist.pop();
252 if (visited.test_set(ni))
253 continue;
254
255 PointsToNode* pn = ptnode_adr(ni);
256 // ensure that all inputs of a Phi have been processed
257 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
258
259 int edges_processed = 0;
260 uint e_cnt = pn->edge_count();
261 for (uint e = 0; e < e_cnt; e++) {
262 uint etgt = pn->edge_target(e);
263 PointsToNode::EdgeType et = pn->edge_type(e);
264 if (et == PointsToNode::PointsToEdge) {
265 ptset.set(etgt);
266 edges_processed++;
267 } else if (et == PointsToNode::DeferredEdge) {
268 worklist.push(etgt);
269 edges_processed++;
270 } else {
271 assert(false,"neither PointsToEdge or DeferredEdge");
272 }
273 }
274 if (edges_processed == 0) {
275 // no deferred or pointsto edges found. Assume the value was set
276 // outside this method. Add the phantom object to the pointsto set.
277 ptset.set(_phantom_object);
278 }
279 }
280 }
281
282 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
283 // This method is most expensive during ConnectionGraph construction.
284 // Reuse vectorSet and an additional growable array for deferred edges.
285 deferred_edges->clear();
286 visited->Clear();
287
288 visited->set(ni);
289 PointsToNode *ptn = ptnode_adr(ni);
290
291 // Mark current edges as visited and move deferred edges to separate array.
292 for (uint i = 0; i < ptn->edge_count(); ) {
293 uint t = ptn->edge_target(i);
294 #ifdef ASSERT
295 assert(!visited->test_set(t), "expecting no duplications");
296 #else
297 visited->set(t);
298 #endif
299 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
300 ptn->remove_edge(t, PointsToNode::DeferredEdge);
301 deferred_edges->append(t);
302 } else {
303 i++;
304 }
305 }
306 for (int next = 0; next < deferred_edges->length(); ++next) {
307 uint t = deferred_edges->at(next);
308 PointsToNode *ptt = ptnode_adr(t);
309 uint e_cnt = ptt->edge_count();
310 for (uint e = 0; e < e_cnt; e++) {
311 uint etgt = ptt->edge_target(e);
312 if (visited->test_set(etgt))
313 continue;
314
315 PointsToNode::EdgeType et = ptt->edge_type(e);
316 if (et == PointsToNode::PointsToEdge) {
317 add_pointsto_edge(ni, etgt);
318 if(etgt == _phantom_object) {
319 // Special case - field set outside (globally escaping).
320 ptn->set_escape_state(PointsToNode::GlobalEscape);
321 }
322 } else if (et == PointsToNode::DeferredEdge) {
323 deferred_edges->append(etgt);
324 } else {
325 assert(false,"invalid connection graph");
326 }
327 }
328 }
329 }
330
331
332 // Add an edge to node given by "to_i" from any field of adr_i whose offset
333 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
334 // a pointsto edge is added if it is a JavaObject
335
336 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
337 PointsToNode* an = ptnode_adr(adr_i);
338 PointsToNode* to = ptnode_adr(to_i);
339 bool deferred = (to->node_type() == PointsToNode::LocalVar);
340
341 for (uint fe = 0; fe < an->edge_count(); fe++) {
342 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
343 int fi = an->edge_target(fe);
344 PointsToNode* pf = ptnode_adr(fi);
345 int po = pf->offset();
346 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
347 if (deferred)
348 add_deferred_edge(fi, to_i);
349 else
350 add_pointsto_edge(fi, to_i);
351 }
352 }
353 }
354
355 // Add a deferred edge from node given by "from_i" to any field of adr_i
356 // whose offset matches "offset".
357 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
358 PointsToNode* an = ptnode_adr(adr_i);
359 for (uint fe = 0; fe < an->edge_count(); fe++) {
360 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
361 int fi = an->edge_target(fe);
362 PointsToNode* pf = ptnode_adr(fi);
363 int po = pf->offset();
364 if (pf->edge_count() == 0) {
365 // we have not seen any stores to this field, assume it was set outside this method
366 add_pointsto_edge(fi, _phantom_object);
367 }
368 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
369 add_deferred_edge(from_i, fi);
370 }
371 }
372 }
373
374 // Helper functions
375
376 static Node* get_addp_base(Node *addp) {
377 assert(addp->is_AddP(), "must be AddP");
378 //
379 // AddP cases for Base and Address inputs:
380 // case #1. Direct object's field reference:
381 // Allocate
382 // |
383 // Proj #5 ( oop result )
384 // |
385 // CheckCastPP (cast to instance type)
386 // | |
387 // AddP ( base == address )
388 //
389 // case #2. Indirect object's field reference:
390 // Phi
391 // |
392 // CastPP (cast to instance type)
393 // | |
394 // AddP ( base == address )
395 //
396 // case #3. Raw object's field reference for Initialize node:
397 // Allocate
398 // |
399 // Proj #5 ( oop result )
400 // top |
401 // \ |
402 // AddP ( base == top )
403 //
404 // case #4. Array's element reference:
405 // {CheckCastPP | CastPP}
406 // | | |
407 // | AddP ( array's element offset )
408 // | |
409 // AddP ( array's offset )
410 //
411 // case #5. Raw object's field reference for arraycopy stub call:
412 // The inline_native_clone() case when the arraycopy stub is called
413 // after the allocation before Initialize and CheckCastPP nodes.
414 // Allocate
415 // |
416 // Proj #5 ( oop result )
417 // | |
418 // AddP ( base == address )
419 //
420 // case #6. Constant Pool, ThreadLocal, CastX2P or
421 // Raw object's field reference:
422 // {ConP, ThreadLocal, CastX2P, raw Load}
423 // top |
424 // \ |
425 // AddP ( base == top )
426 //
427 // case #7. Klass's field reference.
428 // LoadKlass
429 // | |
430 // AddP ( base == address )
431 //
432 // case #8. narrow Klass's field reference.
433 // LoadNKlass
434 // |
435 // DecodeN
436 // | |
437 // AddP ( base == address )
438 //
439 Node *base = addp->in(AddPNode::Base)->uncast();
440 if (base->is_top()) { // The AddP case #3 and #6.
441 base = addp->in(AddPNode::Address)->uncast();
442 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
443 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
444 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
445 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
446 }
447 return base;
448 }
449
450 static Node* find_second_addp(Node* addp, Node* n) {
451 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
452
453 Node* addp2 = addp->raw_out(0);
454 if (addp->outcnt() == 1 && addp2->is_AddP() &&
455 addp2->in(AddPNode::Base) == n &&
456 addp2->in(AddPNode::Address) == addp) {
457
458 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
459 //
460 // Find array's offset to push it on worklist first and
461 // as result process an array's element offset first (pushed second)
462 // to avoid CastPP for the array's offset.
463 // Otherwise the inserted CastPP (LocalVar) will point to what
464 // the AddP (Field) points to. Which would be wrong since
465 // the algorithm expects the CastPP has the same point as
466 // as AddP's base CheckCastPP (LocalVar).
467 //
468 // ArrayAllocation
469 // |
470 // CheckCastPP
471 // |
472 // memProj (from ArrayAllocation CheckCastPP)
473 // | ||
474 // | || Int (element index)
475 // | || | ConI (log(element size))
476 // | || | /
477 // | || LShift
478 // | || /
479 // | AddP (array's element offset)
480 // | |
481 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
482 // | / /
483 // AddP (array's offset)
484 // |
485 // Load/Store (memory operation on array's element)
486 //
487 return addp2;
488 }
489 return NULL;
490 }
491
492 //
493 // Adjust the type and inputs of an AddP which computes the
494 // address of a field of an instance
495 //
496 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
497 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
498 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
499 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
500 if (t == NULL) {
501 // We are computing a raw address for a store captured by an Initialize
502 // compute an appropriate address type (cases #3 and #5).
503 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
504 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
505 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
506 assert(offs != Type::OffsetBot, "offset must be a constant");
507 t = base_t->add_offset(offs)->is_oopptr();
508 }
509 int inst_id = base_t->instance_id();
510 assert(!t->is_known_instance() || t->instance_id() == inst_id,
511 "old type must be non-instance or match new type");
512
513 // The type 't' could be subclass of 'base_t'.
514 // As result t->offset() could be large then base_t's size and it will
515 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
516 // constructor verifies correctness of the offset.
517 //
518 // It could happend on subclass's branch (from the type profiling
519 // inlining) which was not eliminated during parsing since the exactness
520 // of the allocation type was not propagated to the subclass type check.
521 //
522 // Do nothing for such AddP node and don't process its users since
523 // this code branch will go away.
524 //
525 if (!t->is_known_instance() &&
526 !t->klass()->equals(base_t->klass()) &&
527 t->klass()->is_subtype_of(base_t->klass())) {
528 return false; // bail out
529 }
530
531 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
532 // Do NOT remove the next call: ensure an new alias index is allocated
533 // for the instance type
534 int alias_idx = _compile->get_alias_index(tinst);
535 igvn->set_type(addp, tinst);
536 // record the allocation in the node map
537 set_map(addp->_idx, get_map(base->_idx));
538
539 // Set addp's Base and Address to 'base'.
540 Node *abase = addp->in(AddPNode::Base);
541 Node *adr = addp->in(AddPNode::Address);
542 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
543 adr->in(0)->_idx == (uint)inst_id) {
544 // Skip AddP cases #3 and #5.
545 } else {
546 assert(!abase->is_top(), "sanity"); // AddP case #3
547 if (abase != base) {
548 igvn->hash_delete(addp);
549 addp->set_req(AddPNode::Base, base);
550 if (abase == adr) {
551 addp->set_req(AddPNode::Address, base);
552 } else {
553 // AddP case #4 (adr is array's element offset AddP node)
554 #ifdef ASSERT
555 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
556 assert(adr->is_AddP() && atype != NULL &&
557 atype->instance_id() == inst_id, "array's element offset should be processed first");
558 #endif
559 }
560 igvn->hash_insert(addp);
561 }
562 }
563 // Put on IGVN worklist since at least addp's type was changed above.
564 record_for_optimizer(addp);
565 return true;
566 }
567
568 //
569 // Create a new version of orig_phi if necessary. Returns either the newly
570 // created phi or an existing phi. Sets create_new to indicate wheter a new
571 // phi was created. Cache the last newly created phi in the node map.
572 //
573 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
574 Compile *C = _compile;
575 new_created = false;
576 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
577 // nothing to do if orig_phi is bottom memory or matches alias_idx
578 if (phi_alias_idx == alias_idx) {
579 return orig_phi;
580 }
581 // have we already created a Phi for this alias index?
582 PhiNode *result = get_map_phi(orig_phi->_idx);
583 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
584 return result;
585 }
586 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
587 if (C->do_escape_analysis() == true && !C->failing()) {
588 // Retry compilation without escape analysis.
589 // If this is the first failure, the sentinel string will "stick"
590 // to the Compile object, and the C2Compiler will see it and retry.
591 C->record_failure(C2Compiler::retry_no_escape_analysis());
592 }
593 return NULL;
594 }
595 orig_phi_worklist.append_if_missing(orig_phi);
596 const TypePtr *atype = C->get_adr_type(alias_idx);
597 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
598 set_map_phi(orig_phi->_idx, result);
599 igvn->set_type(result, result->bottom_type());
600 record_for_optimizer(result);
601 new_created = true;
602 return result;
603 }
604
605 //
606 // Return a new version of Memory Phi "orig_phi" with the inputs having the
607 // specified alias index.
608 //
609 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
610
611 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
612 Compile *C = _compile;
613 bool new_phi_created;
614 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
615 if (!new_phi_created) {
616 return result;
617 }
618
619 GrowableArray<PhiNode *> phi_list;
620 GrowableArray<uint> cur_input;
621
622 PhiNode *phi = orig_phi;
623 uint idx = 1;
624 bool finished = false;
625 while(!finished) {
626 while (idx < phi->req()) {
627 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
628 if (mem != NULL && mem->is_Phi()) {
629 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
630 if (new_phi_created) {
631 // found an phi for which we created a new split, push current one on worklist and begin
632 // processing new one
633 phi_list.push(phi);
634 cur_input.push(idx);
635 phi = mem->as_Phi();
636 result = newphi;
637 idx = 1;
638 continue;
639 } else {
640 mem = newphi;
641 }
642 }
643 if (C->failing()) {
644 return NULL;
645 }
646 result->set_req(idx++, mem);
647 }
648 #ifdef ASSERT
649 // verify that the new Phi has an input for each input of the original
650 assert( phi->req() == result->req(), "must have same number of inputs.");
651 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
652 #endif
653 // Check if all new phi's inputs have specified alias index.
654 // Otherwise use old phi.
655 for (uint i = 1; i < phi->req(); i++) {
656 Node* in = result->in(i);
657 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
658 }
659 // we have finished processing a Phi, see if there are any more to do
660 finished = (phi_list.length() == 0 );
661 if (!finished) {
662 phi = phi_list.pop();
663 idx = cur_input.pop();
664 PhiNode *prev_result = get_map_phi(phi->_idx);
665 prev_result->set_req(idx++, result);
666 result = prev_result;
667 }
668 }
669 return result;
670 }
671
672
673 //
674 // The next methods are derived from methods in MemNode.
675 //
676 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) {
677 Node *mem = mmem;
678 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
679 // means an array I have not precisely typed yet. Do not do any
680 // alias stuff with it any time soon.
681 if( tinst->base() != Type::AnyPtr &&
682 !(tinst->klass()->is_java_lang_Object() &&
683 tinst->offset() == Type::OffsetBot) ) {
684 mem = mmem->memory_at(alias_idx);
685 // Update input if it is progress over what we have now
686 }
687 return mem;
688 }
689
690 //
691 // Search memory chain of "mem" to find a MemNode whose address
692 // is the specified alias index.
693 //
694 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
695 if (orig_mem == NULL)
696 return orig_mem;
697 Compile* C = phase->C;
698 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr();
699 bool is_instance = (tinst != NULL) && tinst->is_known_instance();
700 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
701 Node *prev = NULL;
702 Node *result = orig_mem;
703 while (prev != result) {
704 prev = result;
705 if (result == start_mem)
706 break; // hit one of our sentinals
707 if (result->is_Mem()) {
708 const Type *at = phase->type(result->in(MemNode::Address));
709 if (at != Type::TOP) {
710 assert (at->isa_ptr() != NULL, "pointer type required.");
711 int idx = C->get_alias_index(at->is_ptr());
712 if (idx == alias_idx)
713 break;
714 }
715 result = result->in(MemNode::Memory);
716 }
717 if (!is_instance)
718 continue; // don't search further for non-instance types
719 // skip over a call which does not affect this memory slice
720 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
721 Node *proj_in = result->in(0);
722 if (proj_in->is_Allocate() && proj_in->_idx == (uint)tinst->instance_id()) {
723 break; // hit one of our sentinals
724 } else if (proj_in->is_Call()) {
725 CallNode *call = proj_in->as_Call();
726 if (!call->may_modify(tinst, phase)) {
727 result = call->in(TypeFunc::Memory);
728 }
729 } else if (proj_in->is_Initialize()) {
730 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
731 // Stop if this is the initialization for the object instance which
732 // which contains this memory slice, otherwise skip over it.
733 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) {
734 result = proj_in->in(TypeFunc::Memory);
735 }
736 } else if (proj_in->is_MemBar()) {
737 result = proj_in->in(TypeFunc::Memory);
738 }
739 } else if (result->is_MergeMem()) {
740 MergeMemNode *mmem = result->as_MergeMem();
741 result = step_through_mergemem(mmem, alias_idx, tinst);
742 if (result == mmem->base_memory()) {
743 // Didn't find instance memory, search through general slice recursively.
744 result = mmem->memory_at(C->get_general_index(alias_idx));
745 result = find_inst_mem(result, alias_idx, orig_phis, phase);
746 if (C->failing()) {
747 return NULL;
748 }
749 mmem->set_memory_at(alias_idx, result);
750 }
751 } else if (result->is_Phi() &&
752 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
753 Node *un = result->as_Phi()->unique_input(phase);
754 if (un != NULL) {
755 result = un;
756 } else {
757 break;
758 }
759 }
760 }
761 if (result->is_Phi()) {
762 PhiNode *mphi = result->as_Phi();
763 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
764 const TypePtr *t = mphi->adr_type();
765 if (C->get_alias_index(t) != alias_idx) {
766 // Create a new Phi with the specified alias index type.
767 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
768 } else if (!is_instance) {
769 // Push all non-instance Phis on the orig_phis worklist to update inputs
770 // during Phase 4 if needed.
771 orig_phis.append_if_missing(mphi);
772 }
773 }
774 // the result is either MemNode, PhiNode, InitializeNode.
775 return result;
776 }
777
778
779 //
780 // Convert the types of unescaped object to instance types where possible,
781 // propagate the new type information through the graph, and update memory
782 // edges and MergeMem inputs to reflect the new type.
783 //
784 // We start with allocations (and calls which may be allocations) on alloc_worklist.
785 // The processing is done in 4 phases:
786 //
787 // Phase 1: Process possible allocations from alloc_worklist. Create instance
788 // types for the CheckCastPP for allocations where possible.
789 // Propagate the the new types through users as follows:
790 // casts and Phi: push users on alloc_worklist
791 // AddP: cast Base and Address inputs to the instance type
792 // push any AddP users on alloc_worklist and push any memnode
793 // users onto memnode_worklist.
794 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
795 // search the Memory chain for a store with the appropriate type
796 // address type. If a Phi is found, create a new version with
797 // the approriate memory slices from each of the Phi inputs.
798 // For stores, process the users as follows:
799 // MemNode: push on memnode_worklist
800 // MergeMem: push on mergemem_worklist
801 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
802 // moving the first node encountered of each instance type to the
803 // the input corresponding to its alias index.
804 // appropriate memory slice.
805 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
806 //
807 // In the following example, the CheckCastPP nodes are the cast of allocation
808 // results and the allocation of node 29 is unescaped and eligible to be an
809 // instance type.
810 //
811 // We start with:
812 //
813 // 7 Parm #memory
814 // 10 ConI "12"
815 // 19 CheckCastPP "Foo"
816 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
817 // 29 CheckCastPP "Foo"
818 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
819 //
820 // 40 StoreP 25 7 20 ... alias_index=4
821 // 50 StoreP 35 40 30 ... alias_index=4
822 // 60 StoreP 45 50 20 ... alias_index=4
823 // 70 LoadP _ 60 30 ... alias_index=4
824 // 80 Phi 75 50 60 Memory alias_index=4
825 // 90 LoadP _ 80 30 ... alias_index=4
826 // 100 LoadP _ 80 20 ... alias_index=4
827 //
828 //
829 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
830 // and creating a new alias index for node 30. This gives:
831 //
832 // 7 Parm #memory
833 // 10 ConI "12"
834 // 19 CheckCastPP "Foo"
835 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
836 // 29 CheckCastPP "Foo" iid=24
837 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
838 //
839 // 40 StoreP 25 7 20 ... alias_index=4
840 // 50 StoreP 35 40 30 ... alias_index=6
841 // 60 StoreP 45 50 20 ... alias_index=4
842 // 70 LoadP _ 60 30 ... alias_index=6
843 // 80 Phi 75 50 60 Memory alias_index=4
844 // 90 LoadP _ 80 30 ... alias_index=6
845 // 100 LoadP _ 80 20 ... alias_index=4
846 //
847 // In phase 2, new memory inputs are computed for the loads and stores,
848 // And a new version of the phi is created. In phase 4, the inputs to
849 // node 80 are updated and then the memory nodes are updated with the
850 // values computed in phase 2. This results in:
851 //
852 // 7 Parm #memory
853 // 10 ConI "12"
854 // 19 CheckCastPP "Foo"
855 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
856 // 29 CheckCastPP "Foo" iid=24
857 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
858 //
859 // 40 StoreP 25 7 20 ... alias_index=4
860 // 50 StoreP 35 7 30 ... alias_index=6
861 // 60 StoreP 45 40 20 ... alias_index=4
862 // 70 LoadP _ 50 30 ... alias_index=6
863 // 80 Phi 75 40 60 Memory alias_index=4
864 // 120 Phi 75 50 50 Memory alias_index=6
865 // 90 LoadP _ 120 30 ... alias_index=6
866 // 100 LoadP _ 80 20 ... alias_index=4
867 //
868 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
869 GrowableArray<Node *> memnode_worklist;
870 GrowableArray<Node *> mergemem_worklist;
871 GrowableArray<PhiNode *> orig_phis;
872 PhaseGVN *igvn = _compile->initial_gvn();
873 uint new_index_start = (uint) _compile->num_alias_types();
874 VectorSet visited(Thread::current()->resource_area());
875 VectorSet ptset(Thread::current()->resource_area());
876
877
878 // Phase 1: Process possible allocations from alloc_worklist.
879 // Create instance types for the CheckCastPP for allocations where possible.
880 //
881 // (Note: don't forget to change the order of the second AddP node on
882 // the alloc_worklist if the order of the worklist processing is changed,
883 // see the comment in find_second_addp().)
884 //
885 while (alloc_worklist.length() != 0) {
886 Node *n = alloc_worklist.pop();
887 uint ni = n->_idx;
888 const TypeOopPtr* tinst = NULL;
889 if (n->is_Call()) {
890 CallNode *alloc = n->as_Call();
891 // copy escape information to call node
892 PointsToNode* ptn = ptnode_adr(alloc->_idx);
893 PointsToNode::EscapeState es = escape_state(alloc, igvn);
894 // We have an allocation or call which returns a Java object,
895 // see if it is unescaped.
896 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
897 continue;
898 if (alloc->is_Allocate()) {
899 // Set the scalar_replaceable flag before the next check.
900 alloc->as_Allocate()->_is_scalar_replaceable = true;
901 }
902 // find CheckCastPP of call return value
903 n = alloc->result_cast();
904 if (n == NULL || // No uses accept Initialize or
905 !n->is_CheckCastPP()) // not unique CheckCastPP.
906 continue;
907 // The inline code for Object.clone() casts the allocation result to
908 // java.lang.Object and then to the actual type of the allocated
909 // object. Detect this case and use the second cast.
910 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
911 // the allocation result is cast to java.lang.Object and then
912 // to the actual Array type.
913 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
914 && (alloc->is_AllocateArray() ||
915 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
916 Node *cast2 = NULL;
917 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
918 Node *use = n->fast_out(i);
919 if (use->is_CheckCastPP()) {
920 cast2 = use;
921 break;
922 }
923 }
924 if (cast2 != NULL) {
925 n = cast2;
926 } else {
927 continue;
928 }
929 }
930 set_escape_state(n->_idx, es);
931 // in order for an object to be scalar-replaceable, it must be:
932 // - a direct allocation (not a call returning an object)
933 // - non-escaping
934 // - eligible to be a unique type
935 // - not determined to be ineligible by escape analysis
936 set_map(alloc->_idx, n);
937 set_map(n->_idx, alloc);
938 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
939 if (t == NULL)
940 continue; // not a TypeInstPtr
941 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
942 igvn->hash_delete(n);
943 igvn->set_type(n, tinst);
944 n->raise_bottom_type(tinst);
945 igvn->hash_insert(n);
946 record_for_optimizer(n);
947 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
948 (t->isa_instptr() || t->isa_aryptr())) {
949
950 // First, put on the worklist all Field edges from Connection Graph
951 // which is more accurate then putting immediate users from Ideal Graph.
952 for (uint e = 0; e < ptn->edge_count(); e++) {
953 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
954 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
955 "only AddP nodes are Field edges in CG");
956 if (use->outcnt() > 0) { // Don't process dead nodes
957 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
958 if (addp2 != NULL) {
959 assert(alloc->is_AllocateArray(),"array allocation was expected");
960 alloc_worklist.append_if_missing(addp2);
961 }
962 alloc_worklist.append_if_missing(use);
963 }
964 }
965
966 // An allocation may have an Initialize which has raw stores. Scan
967 // the users of the raw allocation result and push AddP users
968 // on alloc_worklist.
969 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
970 assert (raw_result != NULL, "must have an allocation result");
971 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
972 Node *use = raw_result->fast_out(i);
973 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
974 Node* addp2 = find_second_addp(use, raw_result);
975 if (addp2 != NULL) {
976 assert(alloc->is_AllocateArray(),"array allocation was expected");
977 alloc_worklist.append_if_missing(addp2);
978 }
979 alloc_worklist.append_if_missing(use);
980 } else if (use->is_Initialize()) {
981 memnode_worklist.append_if_missing(use);
982 }
983 }
984 }
985 } else if (n->is_AddP()) {
986 ptset.Clear();
987 PointsTo(ptset, get_addp_base(n), igvn);
988 assert(ptset.Size() == 1, "AddP address is unique");
989 uint elem = ptset.getelem(); // Allocation node's index
990 if (elem == _phantom_object)
991 continue; // Assume the value was set outside this method.
992 Node *base = get_map(elem); // CheckCastPP node
993 if (!split_AddP(n, base, igvn)) continue; // wrong type
994 tinst = igvn->type(base)->isa_oopptr();
995 } else if (n->is_Phi() ||
996 n->is_CheckCastPP() ||
997 n->is_EncodeP() ||
998 n->is_DecodeN() ||
999 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1000 if (visited.test_set(n->_idx)) {
1001 assert(n->is_Phi(), "loops only through Phi's");
1002 continue; // already processed
1003 }
1004 ptset.Clear();
1005 PointsTo(ptset, n, igvn);
1006 if (ptset.Size() == 1) {
1007 uint elem = ptset.getelem(); // Allocation node's index
1008 if (elem == _phantom_object)
1009 continue; // Assume the value was set outside this method.
1010 Node *val = get_map(elem); // CheckCastPP node
1011 TypeNode *tn = n->as_Type();
1012 tinst = igvn->type(val)->isa_oopptr();
1013 assert(tinst != NULL && tinst->is_known_instance() &&
1014 (uint)tinst->instance_id() == elem , "instance type expected.");
1015
1016 const Type *tn_type = igvn->type(tn);
1017 const TypeOopPtr *tn_t;
1018 if (tn_type->isa_narrowoop()) {
1019 tn_t = tn_type->make_ptr()->isa_oopptr();
1020 } else {
1021 tn_t = tn_type->isa_oopptr();
1022 }
1023
1024 if (tn_t != NULL &&
1025 tinst->cast_to_instance_id(TypeOopPtr::InstanceBot)->higher_equal(tn_t)) {
1026 if (tn_type->isa_narrowoop()) {
1027 tn_type = tinst->make_narrowoop();
1028 } else {
1029 tn_type = tinst;
1030 }
1031 igvn->hash_delete(tn);
1032 igvn->set_type(tn, tn_type);
1033 tn->set_type(tn_type);
1034 igvn->hash_insert(tn);
1035 record_for_optimizer(n);
1036 } else {
1037 continue; // wrong type
1038 }
1039 }
1040 } else {
1041 continue;
1042 }
1043 // push users on appropriate worklist
1044 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1045 Node *use = n->fast_out(i);
1046 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1047 memnode_worklist.append_if_missing(use);
1048 } else if (use->is_Initialize()) {
1049 memnode_worklist.append_if_missing(use);
1050 } else if (use->is_MergeMem()) {
1051 mergemem_worklist.append_if_missing(use);
1052 } else if (use->is_SafePoint() && tinst != NULL) {
1053 // Look for MergeMem nodes for calls which reference unique allocation
1054 // (through CheckCastPP nodes) even for debug info.
1055 Node* m = use->in(TypeFunc::Memory);
1056 uint iid = tinst->instance_id();
1057 while (m->is_Proj() && m->in(0)->is_SafePoint() &&
1058 m->in(0) != use && !m->in(0)->_idx != iid) {
1059 m = m->in(0)->in(TypeFunc::Memory);
1060 }
1061 if (m->is_MergeMem()) {
1062 mergemem_worklist.append_if_missing(m);
1063 }
1064 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1065 Node* addp2 = find_second_addp(use, n);
1066 if (addp2 != NULL) {
1067 alloc_worklist.append_if_missing(addp2);
1068 }
1069 alloc_worklist.append_if_missing(use);
1070 } else if (use->is_Phi() ||
1071 use->is_CheckCastPP() ||
1072 use->is_EncodeP() ||
1073 use->is_DecodeN() ||
1074 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1075 alloc_worklist.append_if_missing(use);
1076 }
1077 }
1078
1079 }
1080 // New alias types were created in split_AddP().
1081 uint new_index_end = (uint) _compile->num_alias_types();
1082
1083 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1084 // compute new values for Memory inputs (the Memory inputs are not
1085 // actually updated until phase 4.)
1086 if (memnode_worklist.length() == 0)
1087 return; // nothing to do
1088
1089 while (memnode_worklist.length() != 0) {
1090 Node *n = memnode_worklist.pop();
1091 if (visited.test_set(n->_idx))
1092 continue;
1093 if (n->is_Phi()) {
1094 assert(n->as_Phi()->adr_type() != TypePtr::BOTTOM, "narrow memory slice required");
1095 // we don't need to do anything, but the users must be pushed if we haven't processed
1096 // this Phi before
1097 } else if (n->is_Initialize()) {
1098 // we don't need to do anything, but the users of the memory projection must be pushed
1099 n = n->as_Initialize()->proj_out(TypeFunc::Memory);
1100 if (n == NULL)
1101 continue;
1102 } else {
1103 assert(n->is_Mem(), "memory node required.");
1104 Node *addr = n->in(MemNode::Address);
1105 assert(addr->is_AddP(), "AddP required");
1106 const Type *addr_t = igvn->type(addr);
1107 if (addr_t == Type::TOP)
1108 continue;
1109 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1110 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1111 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1112 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1113 if (_compile->failing()) {
1114 return;
1115 }
1116 if (mem != n->in(MemNode::Memory)) {
1117 set_map(n->_idx, mem);
1118 ptnode_adr(n->_idx)->_node = n;
1119 }
1120 if (n->is_Load()) {
1121 continue; // don't push users
1122 } else if (n->is_LoadStore()) {
1123 // get the memory projection
1124 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1125 Node *use = n->fast_out(i);
1126 if (use->Opcode() == Op_SCMemProj) {
1127 n = use;
1128 break;
1129 }
1130 }
1131 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1132 }
1133 }
1134 // push user on appropriate worklist
1135 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1136 Node *use = n->fast_out(i);
1137 if (use->is_Phi()) {
1138 memnode_worklist.append_if_missing(use);
1139 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1140 memnode_worklist.append_if_missing(use);
1141 } else if (use->is_Initialize()) {
1142 memnode_worklist.append_if_missing(use);
1143 } else if (use->is_MergeMem()) {
1144 mergemem_worklist.append_if_missing(use);
1145 }
1146 }
1147 }
1148
1149 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1150 // Walk each memory moving the first node encountered of each
1151 // instance type to the the input corresponding to its alias index.
1152 while (mergemem_worklist.length() != 0) {
1153 Node *n = mergemem_worklist.pop();
1154 assert(n->is_MergeMem(), "MergeMem node required.");
1155 if (visited.test_set(n->_idx))
1156 continue;
1157 MergeMemNode *nmm = n->as_MergeMem();
1158 // Note: we don't want to use MergeMemStream here because we only want to
1159 // scan inputs which exist at the start, not ones we add during processing.
1160 uint nslices = nmm->req();
1161 igvn->hash_delete(nmm);
1162 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1163 Node* mem = nmm->in(i);
1164 Node* cur = NULL;
1165 if (mem == NULL || mem->is_top())
1166 continue;
1167 while (mem->is_Mem()) {
1168 const Type *at = igvn->type(mem->in(MemNode::Address));
1169 if (at != Type::TOP) {
1170 assert (at->isa_ptr() != NULL, "pointer type required.");
1171 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1172 if (idx == i) {
1173 if (cur == NULL)
1174 cur = mem;
1175 } else {
1176 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1177 nmm->set_memory_at(idx, mem);
1178 }
1179 }
1180 }
1181 mem = mem->in(MemNode::Memory);
1182 }
1183 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1184 // Find any instance of the current type if we haven't encountered
1185 // a value of the instance along the chain.
1186 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1187 if((uint)_compile->get_general_index(ni) == i) {
1188 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1189 if (nmm->is_empty_memory(m)) {
1190 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1191 if (_compile->failing()) {
1192 return;
1193 }
1194 nmm->set_memory_at(ni, result);
1195 }
1196 }
1197 }
1198 }
1199 // Find the rest of instances values
1200 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1201 const TypeOopPtr *tinst = igvn->C->get_adr_type(ni)->isa_oopptr();
1202 Node* result = step_through_mergemem(nmm, ni, tinst);
1203 if (result == nmm->base_memory()) {
1204 // Didn't find instance memory, search through general slice recursively.
1205 result = nmm->memory_at(igvn->C->get_general_index(ni));
1206 result = find_inst_mem(result, ni, orig_phis, igvn);
1207 if (_compile->failing()) {
1208 return;
1209 }
1210 nmm->set_memory_at(ni, result);
1211 }
1212 }
1213 igvn->hash_insert(nmm);
1214 record_for_optimizer(nmm);
1215
1216 // Propagate new memory slices to following MergeMem nodes.
1217 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1218 Node *use = n->fast_out(i);
1219 if (use->is_Call()) {
1220 CallNode* in = use->as_Call();
1221 if (in->proj_out(TypeFunc::Memory) != NULL) {
1222 Node* m = in->proj_out(TypeFunc::Memory);
1223 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
1224 Node* mm = m->fast_out(j);
1225 if (mm->is_MergeMem()) {
1226 mergemem_worklist.append_if_missing(mm);
1227 }
1228 }
1229 }
1230 if (use->is_Allocate()) {
1231 use = use->as_Allocate()->initialization();
1232 if (use == NULL) {
1233 continue;
1234 }
1235 }
1236 }
1237 if (use->is_Initialize()) {
1238 InitializeNode* in = use->as_Initialize();
1239 if (in->proj_out(TypeFunc::Memory) != NULL) {
1240 Node* m = in->proj_out(TypeFunc::Memory);
1241 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
1242 Node* mm = m->fast_out(j);
1243 if (mm->is_MergeMem()) {
1244 mergemem_worklist.append_if_missing(mm);
1245 }
1246 }
1247 }
1248 }
1249 }
1250 }
1251
1252 // Phase 4: Update the inputs of non-instance memory Phis and
1253 // the Memory input of memnodes
1254 // First update the inputs of any non-instance Phi's from
1255 // which we split out an instance Phi. Note we don't have
1256 // to recursively process Phi's encounted on the input memory
1257 // chains as is done in split_memory_phi() since they will
1258 // also be processed here.
1259 for (int j = 0; j < orig_phis.length(); j++) {
1260 PhiNode *phi = orig_phis.at(j);
1261 int alias_idx = _compile->get_alias_index(phi->adr_type());
1262 igvn->hash_delete(phi);
1263 for (uint i = 1; i < phi->req(); i++) {
1264 Node *mem = phi->in(i);
1265 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1266 if (_compile->failing()) {
1267 return;
1268 }
1269 if (mem != new_mem) {
1270 phi->set_req(i, new_mem);
1271 }
1272 }
1273 igvn->hash_insert(phi);
1274 record_for_optimizer(phi);
1275 }
1276
1277 // Update the memory inputs of MemNodes with the value we computed
1278 // in Phase 2.
1279 for (uint i = 0; i < nodes_size(); i++) {
1280 Node *nmem = get_map(i);
1281 if (nmem != NULL) {
1282 Node *n = ptnode_adr(i)->_node;
1283 if (n != NULL && n->is_Mem()) {
1284 igvn->hash_delete(n);
1285 n->set_req(MemNode::Memory, nmem);
1286 igvn->hash_insert(n);
1287 record_for_optimizer(n);
1288 }
1289 }
1290 }
1291 }
1292
1293 bool ConnectionGraph::has_candidates(Compile *C) {
1294 // EA brings benefits only when the code has allocations and/or locks which
1295 // are represented by ideal Macro nodes.
1296 int cnt = C->macro_count();
1297 for( int i=0; i < cnt; i++ ) {
1298 Node *n = C->macro_node(i);
1299 if ( n->is_Allocate() )
1300 return true;
1301 if( n->is_Lock() ) {
1302 Node* obj = n->as_Lock()->obj_node()->uncast();
1303 if( !(obj->is_Parm() || obj->is_Con()) )
1304 return true;
1305 }
1306 }
1307 return false;
1308 }
1309
1310 bool ConnectionGraph::compute_escape() {
1311 Compile* C = _compile;
1312
1313 // 1. Populate Connection Graph (CG) with Ideal nodes.
1314
1315 Unique_Node_List worklist_init;
1316 worklist_init.map(C->unique(), NULL); // preallocate space
1317
1318 // Initialize worklist
1319 if (C->root() != NULL) {
1320 worklist_init.push(C->root());
1321 }
1322
1323 GrowableArray<int> cg_worklist;
1324 PhaseGVN* igvn = C->initial_gvn();
1325 bool has_allocations = false;
1326
1327 // Push all useful nodes onto CG list and set their type.
1328 for( uint next = 0; next < worklist_init.size(); ++next ) {
1329 Node* n = worklist_init.at(next);
1330 record_for_escape_analysis(n, igvn);
1331 // Only allocations and java static calls results are checked
1332 // for an escape status. See process_call_result() below.
1333 if (n->is_Allocate() || n->is_CallStaticJava() &&
1334 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1335 has_allocations = true;
1336 }
1337 if(n->is_AddP())
1338 cg_worklist.append(n->_idx);
1339 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1340 Node* m = n->fast_out(i); // Get user
1341 worklist_init.push(m);
1342 }
1343 }
1344
1345 if (!has_allocations) {
1346 _collecting = false;
1347 return false; // Nothing to do.
1348 }
1349
1350 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1351 uint delayed_size = _delayed_worklist.size();
1352 for( uint next = 0; next < delayed_size; ++next ) {
1353 Node* n = _delayed_worklist.at(next);
1354 build_connection_graph(n, igvn);
1355 }
1356
1357 // 3. Pass to create fields edges (Allocate -F-> AddP).
1358 uint cg_length = cg_worklist.length();
1359 for( uint next = 0; next < cg_length; ++next ) {
1360 int ni = cg_worklist.at(next);
1361 build_connection_graph(ptnode_adr(ni)->_node, igvn);
1362 }
1363
1364 cg_worklist.clear();
1365 cg_worklist.append(_phantom_object);
1366
1367 // 4. Build Connection Graph which need
1368 // to walk the connection graph.
1369 for (uint ni = 0; ni < nodes_size(); ni++) {
1370 PointsToNode* ptn = ptnode_adr(ni);
1371 Node *n = ptn->_node;
1372 if (n != NULL) { // Call, AddP, LoadP, StoreP
1373 build_connection_graph(n, igvn);
1374 if (ptn->node_type() != PointsToNode::UnknownType)
1375 cg_worklist.append(n->_idx); // Collect CG nodes
1376 }
1377 }
1378
1379 VectorSet ptset(Thread::current()->resource_area());
1380 GrowableArray<uint> deferred_edges;
1381 VectorSet visited(Thread::current()->resource_area());
1382
1383 // 5. Remove deferred edges from the graph and collect
1384 // information needed for type splitting.
1385 cg_length = cg_worklist.length();
1386 for( uint next = 0; next < cg_length; ++next ) {
1387 int ni = cg_worklist.at(next);
1388 PointsToNode* ptn = ptnode_adr(ni);
1389 PointsToNode::NodeType nt = ptn->node_type();
1390 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1391 remove_deferred(ni, &deferred_edges, &visited);
1392 Node *n = ptn->_node;
1393 if (n->is_AddP()) {
1394 // Search for objects which are not scalar replaceable.
1395 // Mark their escape state as ArgEscape to propagate the state
1396 // to referenced objects.
1397 // Note: currently there are no difference in compiler optimizations
1398 // for ArgEscape objects and NoEscape objects which are not
1399 // scalar replaceable.
1400
1401 int offset = ptn->offset();
1402 Node *base = get_addp_base(n);
1403 ptset.Clear();
1404 PointsTo(ptset, base, igvn);
1405 int ptset_size = ptset.Size();
1406
1407 // Check if a field's initializing value is recorded and add
1408 // a corresponding NULL field's value if it is not recorded.
1409 // Connection Graph does not record a default initialization by NULL
1410 // captured by Initialize node.
1411 //
1412 // Note: it will disable scalar replacement in some cases:
1413 //
1414 // Point p[] = new Point[1];
1415 // p[0] = new Point(); // Will be not scalar replaced
1416 //
1417 // but it will save us from incorrect optimizations in next cases:
1418 //
1419 // Point p[] = new Point[1];
1420 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1421 //
1422 // Without a control flow analysis we can't distinguish above cases.
1423 //
1424 if (offset != Type::OffsetBot && ptset_size == 1) {
1425 uint elem = ptset.getelem(); // Allocation node's index
1426 // It does not matter if it is not Allocation node since
1427 // only non-escaping allocations are scalar replaced.
1428 if (ptnode_adr(elem)->_node->is_Allocate() &&
1429 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1430 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1431 InitializeNode* ini = alloc->initialization();
1432 Node* value = NULL;
1433 if (ini != NULL) {
1434 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1435 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), igvn);
1436 if (store != NULL && store->is_Store())
1437 value = store->in(MemNode::ValueIn);
1438 }
1439 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1440 // A field's initializing value was not recorded. Add NULL.
1441 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1442 add_pointsto_edge(ni, null_idx);
1443 }
1444 }
1445 }
1446
1447 // An object is not scalar replaceable if the field which may point
1448 // to it has unknown offset (unknown element of an array of objects).
1449 //
1450 if (offset == Type::OffsetBot) {
1451 uint e_cnt = ptn->edge_count();
1452 for (uint ei = 0; ei < e_cnt; ei++) {
1453 uint npi = ptn->edge_target(ei);
1454 set_escape_state(npi, PointsToNode::ArgEscape);
1455 ptnode_adr(npi)->_scalar_replaceable = false;
1456 }
1457 }
1458
1459 // Currently an object is not scalar replaceable if a LoadStore node
1460 // access its field since the field value is unknown after it.
1461 //
1462 bool has_LoadStore = false;
1463 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1464 Node *use = n->fast_out(i);
1465 if (use->is_LoadStore()) {
1466 has_LoadStore = true;
1467 break;
1468 }
1469 }
1470 // An object is not scalar replaceable if the address points
1471 // to unknown field (unknown element for arrays, offset is OffsetBot).
1472 //
1473 // Or the address may point to more then one object. This may produce
1474 // the false positive result (set scalar_replaceable to false)
1475 // since the flow-insensitive escape analysis can't separate
1476 // the case when stores overwrite the field's value from the case
1477 // when stores happened on different control branches.
1478 //
1479 if (ptset_size > 1 || ptset_size != 0 &&
1480 (has_LoadStore || offset == Type::OffsetBot)) {
1481 for( VectorSetI j(&ptset); j.test(); ++j ) {
1482 set_escape_state(j.elem, PointsToNode::ArgEscape);
1483 ptnode_adr(j.elem)->_scalar_replaceable = false;
1484 }
1485 }
1486 }
1487 }
1488 }
1489
1490 // 6. Propagate escape states.
1491 GrowableArray<int> worklist;
1492 bool has_non_escaping_obj = false;
1493
1494 // push all GlobalEscape nodes on the worklist
1495 for( uint next = 0; next < cg_length; ++next ) {
1496 int nk = cg_worklist.at(next);
1497 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1498 worklist.push(nk);
1499 }
1500 // mark all nodes reachable from GlobalEscape nodes
1501 while(worklist.length() > 0) {
1502 PointsToNode* ptn = ptnode_adr(worklist.pop());
1503 uint e_cnt = ptn->edge_count();
1504 for (uint ei = 0; ei < e_cnt; ei++) {
1505 uint npi = ptn->edge_target(ei);
1506 PointsToNode *np = ptnode_adr(npi);
1507 if (np->escape_state() < PointsToNode::GlobalEscape) {
1508 np->set_escape_state(PointsToNode::GlobalEscape);
1509 worklist.push(npi);
1510 }
1511 }
1512 }
1513
1514 // push all ArgEscape nodes on the worklist
1515 for( uint next = 0; next < cg_length; ++next ) {
1516 int nk = cg_worklist.at(next);
1517 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1518 worklist.push(nk);
1519 }
1520 // mark all nodes reachable from ArgEscape nodes
1521 while(worklist.length() > 0) {
1522 PointsToNode* ptn = ptnode_adr(worklist.pop());
1523 if (ptn->node_type() == PointsToNode::JavaObject)
1524 has_non_escaping_obj = true; // Non GlobalEscape
1525 uint e_cnt = ptn->edge_count();
1526 for (uint ei = 0; ei < e_cnt; ei++) {
1527 uint npi = ptn->edge_target(ei);
1528 PointsToNode *np = ptnode_adr(npi);
1529 if (np->escape_state() < PointsToNode::ArgEscape) {
1530 np->set_escape_state(PointsToNode::ArgEscape);
1531 worklist.push(npi);
1532 }
1533 }
1534 }
1535
1536 GrowableArray<Node*> alloc_worklist;
1537
1538 // push all NoEscape nodes on the worklist
1539 for( uint next = 0; next < cg_length; ++next ) {
1540 int nk = cg_worklist.at(next);
1541 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1542 worklist.push(nk);
1543 }
1544 // mark all nodes reachable from NoEscape nodes
1545 while(worklist.length() > 0) {
1546 PointsToNode* ptn = ptnode_adr(worklist.pop());
1547 if (ptn->node_type() == PointsToNode::JavaObject)
1548 has_non_escaping_obj = true; // Non GlobalEscape
1549 Node* n = ptn->_node;
1550 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1551 // Push scalar replaceable alocations on alloc_worklist
1552 // for processing in split_unique_types().
1553 alloc_worklist.append(n);
1554 }
1555 uint e_cnt = ptn->edge_count();
1556 for (uint ei = 0; ei < e_cnt; ei++) {
1557 uint npi = ptn->edge_target(ei);
1558 PointsToNode *np = ptnode_adr(npi);
1559 if (np->escape_state() < PointsToNode::NoEscape) {
1560 np->set_escape_state(PointsToNode::NoEscape);
1561 worklist.push(npi);
1562 }
1563 }
1564 }
1565
1566 _collecting = false;
1567 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1568
1569 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1570 if ( has_scalar_replaceable_candidates &&
1571 C->AliasLevel() >= 3 && EliminateAllocations ) {
1572
1573 // Now use the escape information to create unique types for
1574 // scalar replaceable objects.
1575 split_unique_types(alloc_worklist);
1576
1577 if (C->failing()) return false;
1578
1579 // Clean up after split unique types.
1580 ResourceMark rm;
1581 PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn());
1582
1583 C->print_method("After Escape Analysis", 2);
1584
1585 #ifdef ASSERT
1586 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1587 tty->print("=== No allocations eliminated for ");
1588 C->method()->print_short_name();
1589 if(!EliminateAllocations) {
1590 tty->print(" since EliminateAllocations is off ===");
1591 } else if(!has_scalar_replaceable_candidates) {
1592 tty->print(" since there are no scalar replaceable candidates ===");
1593 } else if(C->AliasLevel() < 3) {
1594 tty->print(" since AliasLevel < 3 ===");
1595 }
1596 tty->cr();
1597 #endif
1598 }
1599 return has_non_escaping_obj;
1600 }
1601
1602 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1603
1604 switch (call->Opcode()) {
1605 #ifdef ASSERT
1606 case Op_Allocate:
1607 case Op_AllocateArray:
1608 case Op_Lock:
1609 case Op_Unlock:
1610 assert(false, "should be done already");
1611 break;
1612 #endif
1613 case Op_CallLeafNoFP:
1614 {
1615 // Stub calls, objects do not escape but they are not scale replaceable.
1616 // Adjust escape state for outgoing arguments.
1617 const TypeTuple * d = call->tf()->domain();
1618 VectorSet ptset(Thread::current()->resource_area());
1619 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1620 const Type* at = d->field_at(i);
1621 Node *arg = call->in(i)->uncast();
1622 const Type *aat = phase->type(arg);
1623 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr()) {
1624 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1625 aat->isa_ptr() != NULL, "expecting an Ptr");
1626 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1627 if (arg->is_AddP()) {
1628 //
1629 // The inline_native_clone() case when the arraycopy stub is called
1630 // after the allocation before Initialize and CheckCastPP nodes.
1631 //
1632 // Set AddP's base (Allocate) as not scalar replaceable since
1633 // pointer to the base (with offset) is passed as argument.
1634 //
1635 arg = get_addp_base(arg);
1636 }
1637 ptset.Clear();
1638 PointsTo(ptset, arg, phase);
1639 for( VectorSetI j(&ptset); j.test(); ++j ) {
1640 uint pt = j.elem;
1641 set_escape_state(pt, PointsToNode::ArgEscape);
1642 }
1643 }
1644 }
1645 break;
1646 }
1647
1648 case Op_CallStaticJava:
1649 // For a static call, we know exactly what method is being called.
1650 // Use bytecode estimator to record the call's escape affects
1651 {
1652 ciMethod *meth = call->as_CallJava()->method();
1653 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1654 // fall-through if not a Java method or no analyzer information
1655 if (call_analyzer != NULL) {
1656 const TypeTuple * d = call->tf()->domain();
1657 VectorSet ptset(Thread::current()->resource_area());
1658 bool copy_dependencies = false;
1659 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1660 const Type* at = d->field_at(i);
1661 int k = i - TypeFunc::Parms;
1662
1663 if (at->isa_oopptr() != NULL) {
1664 Node *arg = call->in(i)->uncast();
1665
1666 bool global_escapes = false;
1667 bool fields_escapes = false;
1668 if (!call_analyzer->is_arg_stack(k)) {
1669 // The argument global escapes, mark everything it could point to
1670 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1671 global_escapes = true;
1672 } else {
1673 if (!call_analyzer->is_arg_local(k)) {
1674 // The argument itself doesn't escape, but any fields might
1675 fields_escapes = true;
1676 }
1677 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1678 copy_dependencies = true;
1679 }
1680
1681 ptset.Clear();
1682 PointsTo(ptset, arg, phase);
1683 for( VectorSetI j(&ptset); j.test(); ++j ) {
1684 uint pt = j.elem;
1685 if (global_escapes) {
1686 //The argument global escapes, mark everything it could point to
1687 set_escape_state(pt, PointsToNode::GlobalEscape);
1688 } else {
1689 if (fields_escapes) {
1690 // The argument itself doesn't escape, but any fields might
1691 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
1692 }
1693 set_escape_state(pt, PointsToNode::ArgEscape);
1694 }
1695 }
1696 }
1697 }
1698 if (copy_dependencies)
1699 call_analyzer->copy_dependencies(_compile->dependencies());
1700 break;
1701 }
1702 }
1703
1704 default:
1705 // Fall-through here if not a Java method or no analyzer information
1706 // or some other type of call, assume the worst case: all arguments
1707 // globally escape.
1708 {
1709 // adjust escape state for outgoing arguments
1710 const TypeTuple * d = call->tf()->domain();
1711 VectorSet ptset(Thread::current()->resource_area());
1712 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1713 const Type* at = d->field_at(i);
1714 if (at->isa_oopptr() != NULL) {
1715 Node *arg = call->in(i)->uncast();
1716 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1717 ptset.Clear();
1718 PointsTo(ptset, arg, phase);
1719 for( VectorSetI j(&ptset); j.test(); ++j ) {
1720 uint pt = j.elem;
1721 set_escape_state(pt, PointsToNode::GlobalEscape);
1722 }
1723 }
1724 }
1725 }
1726 }
1727 }
1728 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
1729 CallNode *call = resproj->in(0)->as_Call();
1730 uint call_idx = call->_idx;
1731 uint resproj_idx = resproj->_idx;
1732
1733 switch (call->Opcode()) {
1734 case Op_Allocate:
1735 {
1736 Node *k = call->in(AllocateNode::KlassNode);
1737 const TypeKlassPtr *kt;
1738 if (k->Opcode() == Op_LoadKlass) {
1739 kt = k->as_Load()->type()->isa_klassptr();
1740 } else {
1741 // Also works for DecodeN(LoadNKlass).
1742 kt = k->as_Type()->type()->isa_klassptr();
1743 }
1744 assert(kt != NULL, "TypeKlassPtr required.");
1745 ciKlass* cik = kt->klass();
1746 ciInstanceKlass* ciik = cik->as_instance_klass();
1747
1748 PointsToNode::EscapeState es;
1749 uint edge_to;
1750 if (cik->is_subclass_of(_compile->env()->Thread_klass()) || ciik->has_finalizer()) {
1751 es = PointsToNode::GlobalEscape;
1752 edge_to = _phantom_object; // Could not be worse
1753 } else {
1754 es = PointsToNode::NoEscape;
1755 edge_to = call_idx;
1756 }
1757 set_escape_state(call_idx, es);
1758 add_pointsto_edge(resproj_idx, edge_to);
1759 _processed.set(resproj_idx);
1760 break;
1761 }
1762
1763 case Op_AllocateArray:
1764 {
1765 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
1766 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
1767 // Not scalar replaceable if the length is not constant or too big.
1768 ptnode_adr(call_idx)->_scalar_replaceable = false;
1769 }
1770 set_escape_state(call_idx, PointsToNode::NoEscape);
1771 add_pointsto_edge(resproj_idx, call_idx);
1772 _processed.set(resproj_idx);
1773 break;
1774 }
1775
1776 case Op_CallStaticJava:
1777 // For a static call, we know exactly what method is being called.
1778 // Use bytecode estimator to record whether the call's return value escapes
1779 {
1780 bool done = true;
1781 const TypeTuple *r = call->tf()->range();
1782 const Type* ret_type = NULL;
1783
1784 if (r->cnt() > TypeFunc::Parms)
1785 ret_type = r->field_at(TypeFunc::Parms);
1786
1787 // Note: we use isa_ptr() instead of isa_oopptr() here because the
1788 // _multianewarray functions return a TypeRawPtr.
1789 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
1790 _processed.set(resproj_idx);
1791 break; // doesn't return a pointer type
1792 }
1793 ciMethod *meth = call->as_CallJava()->method();
1794 const TypeTuple * d = call->tf()->domain();
1795 if (meth == NULL) {
1796 // not a Java method, assume global escape
1797 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1798 add_pointsto_edge(resproj_idx, _phantom_object);
1799 } else {
1800 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
1801 bool copy_dependencies = false;
1802
1803 if (call_analyzer->is_return_allocated()) {
1804 // Returns a newly allocated unescaped object, simply
1805 // update dependency information.
1806 // Mark it as NoEscape so that objects referenced by
1807 // it's fields will be marked as NoEscape at least.
1808 set_escape_state(call_idx, PointsToNode::NoEscape);
1809 add_pointsto_edge(resproj_idx, call_idx);
1810 copy_dependencies = true;
1811 } else if (call_analyzer->is_return_local()) {
1812 // determine whether any arguments are returned
1813 set_escape_state(call_idx, PointsToNode::NoEscape);
1814 bool ret_arg = false;
1815 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1816 const Type* at = d->field_at(i);
1817
1818 if (at->isa_oopptr() != NULL) {
1819 Node *arg = call->in(i)->uncast();
1820
1821 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
1822 ret_arg = true;
1823 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
1824 if (arg_esp->node_type() == PointsToNode::UnknownType)
1825 done = false;
1826 else if (arg_esp->node_type() == PointsToNode::JavaObject)
1827 add_pointsto_edge(resproj_idx, arg->_idx);
1828 else
1829 add_deferred_edge(resproj_idx, arg->_idx);
1830 arg_esp->_hidden_alias = true;
1831 }
1832 }
1833 }
1834 if (done && !ret_arg) {
1835 // Returns unknown object.
1836 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1837 add_pointsto_edge(resproj_idx, _phantom_object);
1838 }
1839 copy_dependencies = true;
1840 } else {
1841 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1842 add_pointsto_edge(resproj_idx, _phantom_object);
1843 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1844 const Type* at = d->field_at(i);
1845 if (at->isa_oopptr() != NULL) {
1846 Node *arg = call->in(i)->uncast();
1847 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
1848 arg_esp->_hidden_alias = true;
1849 }
1850 }
1851 }
1852 if (copy_dependencies)
1853 call_analyzer->copy_dependencies(_compile->dependencies());
1854 }
1855 if (done)
1856 _processed.set(resproj_idx);
1857 break;
1858 }
1859
1860 default:
1861 // Some other type of call, assume the worst case that the
1862 // returned value, if any, globally escapes.
1863 {
1864 const TypeTuple *r = call->tf()->range();
1865 if (r->cnt() > TypeFunc::Parms) {
1866 const Type* ret_type = r->field_at(TypeFunc::Parms);
1867
1868 // Note: we use isa_ptr() instead of isa_oopptr() here because the
1869 // _multianewarray functions return a TypeRawPtr.
1870 if (ret_type->isa_ptr() != NULL) {
1871 set_escape_state(call_idx, PointsToNode::GlobalEscape);
1872 add_pointsto_edge(resproj_idx, _phantom_object);
1873 }
1874 }
1875 _processed.set(resproj_idx);
1876 }
1877 }
1878 }
1879
1880 // Populate Connection Graph with Ideal nodes and create simple
1881 // connection graph edges (do not need to check the node_type of inputs
1882 // or to call PointsTo() to walk the connection graph).
1883 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
1884 if (_processed.test(n->_idx))
1885 return; // No need to redefine node's state.
1886
1887 if (n->is_Call()) {
1888 // Arguments to allocation and locking don't escape.
1889 if (n->is_Allocate()) {
1890 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
1891 record_for_optimizer(n);
1892 } else if (n->is_Lock() || n->is_Unlock()) {
1893 // Put Lock and Unlock nodes on IGVN worklist to process them during
1894 // the first IGVN optimization when escape information is still available.
1895 record_for_optimizer(n);
1896 _processed.set(n->_idx);
1897 } else {
1898 // Have to process call's arguments first.
1899 PointsToNode::NodeType nt = PointsToNode::UnknownType;
1900
1901 // Check if a call returns an object.
1902 const TypeTuple *r = n->as_Call()->tf()->range();
1903 if (n->is_CallStaticJava() && r->cnt() > TypeFunc::Parms &&
1904 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
1905 // Note: use isa_ptr() instead of isa_oopptr() here because
1906 // the _multianewarray functions return a TypeRawPtr.
1907 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
1908 nt = PointsToNode::JavaObject;
1909 }
1910 }
1911 add_node(n, nt, PointsToNode::UnknownEscape, false);
1912 }
1913 return;
1914 }
1915
1916 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
1917 // ThreadLocal has RawPrt type.
1918 switch (n->Opcode()) {
1919 case Op_AddP:
1920 {
1921 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
1922 break;
1923 }
1924 case Op_CastX2P:
1925 { // "Unsafe" memory access.
1926 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1927 break;
1928 }
1929 case Op_CastPP:
1930 case Op_CheckCastPP:
1931 case Op_EncodeP:
1932 case Op_DecodeN:
1933 {
1934 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
1935 int ti = n->in(1)->_idx;
1936 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
1937 if (nt == PointsToNode::UnknownType) {
1938 _delayed_worklist.push(n); // Process it later.
1939 break;
1940 } else if (nt == PointsToNode::JavaObject) {
1941 add_pointsto_edge(n->_idx, ti);
1942 } else {
1943 add_deferred_edge(n->_idx, ti);
1944 }
1945 _processed.set(n->_idx);
1946 break;
1947 }
1948 case Op_ConP:
1949 {
1950 // assume all pointer constants globally escape except for null
1951 PointsToNode::EscapeState es;
1952 if (phase->type(n) == TypePtr::NULL_PTR)
1953 es = PointsToNode::NoEscape;
1954 else
1955 es = PointsToNode::GlobalEscape;
1956
1957 add_node(n, PointsToNode::JavaObject, es, true);
1958 break;
1959 }
1960 case Op_ConN:
1961 {
1962 // assume all narrow oop constants globally escape except for null
1963 PointsToNode::EscapeState es;
1964 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
1965 es = PointsToNode::NoEscape;
1966 else
1967 es = PointsToNode::GlobalEscape;
1968
1969 add_node(n, PointsToNode::JavaObject, es, true);
1970 break;
1971 }
1972 case Op_CreateEx:
1973 {
1974 // assume that all exception objects globally escape
1975 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1976 break;
1977 }
1978 case Op_LoadKlass:
1979 case Op_LoadNKlass:
1980 {
1981 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
1982 break;
1983 }
1984 case Op_LoadP:
1985 case Op_LoadN:
1986 {
1987 const Type *t = phase->type(n);
1988 if (t->make_ptr() == NULL) {
1989 _processed.set(n->_idx);
1990 return;
1991 }
1992 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
1993 break;
1994 }
1995 case Op_Parm:
1996 {
1997 _processed.set(n->_idx); // No need to redefine it state.
1998 uint con = n->as_Proj()->_con;
1999 if (con < TypeFunc::Parms)
2000 return;
2001 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2002 if (t->isa_ptr() == NULL)
2003 return;
2004 // We have to assume all input parameters globally escape
2005 // (Note: passing 'false' since _processed is already set).
2006 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2007 break;
2008 }
2009 case Op_Phi:
2010 {
2011 const Type *t = n->as_Phi()->type();
2012 if (t->make_ptr() == NULL) {
2013 // nothing to do if not an oop or narrow oop
2014 _processed.set(n->_idx);
2015 return;
2016 }
2017 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2018 uint i;
2019 for (i = 1; i < n->req() ; i++) {
2020 Node* in = n->in(i);
2021 if (in == NULL)
2022 continue; // ignore NULL
2023 in = in->uncast();
2024 if (in->is_top() || in == n)
2025 continue; // ignore top or inputs which go back this node
2026 int ti = in->_idx;
2027 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2028 if (nt == PointsToNode::UnknownType) {
2029 break;
2030 } else if (nt == PointsToNode::JavaObject) {
2031 add_pointsto_edge(n->_idx, ti);
2032 } else {
2033 add_deferred_edge(n->_idx, ti);
2034 }
2035 }
2036 if (i >= n->req())
2037 _processed.set(n->_idx);
2038 else
2039 _delayed_worklist.push(n);
2040 break;
2041 }
2042 case Op_Proj:
2043 {
2044 // we are only interested in the result projection from a call
2045 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2046 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2047 process_call_result(n->as_Proj(), phase);
2048 if (!_processed.test(n->_idx)) {
2049 // The call's result may need to be processed later if the call
2050 // returns it's argument and the argument is not processed yet.
2051 _delayed_worklist.push(n);
2052 }
2053 } else {
2054 _processed.set(n->_idx);
2055 }
2056 break;
2057 }
2058 case Op_Return:
2059 {
2060 if( n->req() > TypeFunc::Parms &&
2061 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2062 // Treat Return value as LocalVar with GlobalEscape escape state.
2063 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2064 int ti = n->in(TypeFunc::Parms)->_idx;
2065 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2066 if (nt == PointsToNode::UnknownType) {
2067 _delayed_worklist.push(n); // Process it later.
2068 break;
2069 } else if (nt == PointsToNode::JavaObject) {
2070 add_pointsto_edge(n->_idx, ti);
2071 } else {
2072 add_deferred_edge(n->_idx, ti);
2073 }
2074 }
2075 _processed.set(n->_idx);
2076 break;
2077 }
2078 case Op_StoreP:
2079 case Op_StoreN:
2080 {
2081 const Type *adr_type = phase->type(n->in(MemNode::Address));
2082 adr_type = adr_type->make_ptr();
2083 if (adr_type->isa_oopptr()) {
2084 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2085 } else {
2086 Node* adr = n->in(MemNode::Address);
2087 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2088 adr->in(AddPNode::Address)->is_Proj() &&
2089 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2090 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2091 // We are computing a raw address for a store captured
2092 // by an Initialize compute an appropriate address type.
2093 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2094 assert(offs != Type::OffsetBot, "offset must be a constant");
2095 } else {
2096 _processed.set(n->_idx);
2097 return;
2098 }
2099 }
2100 break;
2101 }
2102 case Op_StorePConditional:
2103 case Op_CompareAndSwapP:
2104 case Op_CompareAndSwapN:
2105 {
2106 const Type *adr_type = phase->type(n->in(MemNode::Address));
2107 adr_type = adr_type->make_ptr();
2108 if (adr_type->isa_oopptr()) {
2109 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2110 } else {
2111 _processed.set(n->_idx);
2112 return;
2113 }
2114 break;
2115 }
2116 case Op_ThreadLocal:
2117 {
2118 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2119 break;
2120 }
2121 default:
2122 ;
2123 // nothing to do
2124 }
2125 return;
2126 }
2127
2128 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2129 uint n_idx = n->_idx;
2130
2131 // Don't set processed bit for AddP, LoadP, StoreP since
2132 // they may need more then one pass to process.
2133 if (_processed.test(n_idx))
2134 return; // No need to redefine node's state.
2135
2136 if (n->is_Call()) {
2137 CallNode *call = n->as_Call();
2138 process_call_arguments(call, phase);
2139 _processed.set(n_idx);
2140 return;
2141 }
2142
2143 switch (n->Opcode()) {
2144 case Op_AddP:
2145 {
2146 Node *base = get_addp_base(n);
2147 // Create a field edge to this node from everything base could point to.
2148 VectorSet ptset(Thread::current()->resource_area());
2149 PointsTo(ptset, base, phase);
2150 for( VectorSetI i(&ptset); i.test(); ++i ) {
2151 uint pt = i.elem;
2152 add_field_edge(pt, n_idx, address_offset(n, phase));
2153 }
2154 break;
2155 }
2156 case Op_CastX2P:
2157 {
2158 assert(false, "Op_CastX2P");
2159 break;
2160 }
2161 case Op_CastPP:
2162 case Op_CheckCastPP:
2163 case Op_EncodeP:
2164 case Op_DecodeN:
2165 {
2166 int ti = n->in(1)->_idx;
2167 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2168 add_pointsto_edge(n_idx, ti);
2169 } else {
2170 add_deferred_edge(n_idx, ti);
2171 }
2172 _processed.set(n_idx);
2173 break;
2174 }
2175 case Op_ConP:
2176 {
2177 assert(false, "Op_ConP");
2178 break;
2179 }
2180 case Op_ConN:
2181 {
2182 assert(false, "Op_ConN");
2183 break;
2184 }
2185 case Op_CreateEx:
2186 {
2187 assert(false, "Op_CreateEx");
2188 break;
2189 }
2190 case Op_LoadKlass:
2191 case Op_LoadNKlass:
2192 {
2193 assert(false, "Op_LoadKlass");
2194 break;
2195 }
2196 case Op_LoadP:
2197 case Op_LoadN:
2198 {
2199 const Type *t = phase->type(n);
2200 #ifdef ASSERT
2201 if (t->make_ptr() == NULL)
2202 assert(false, "Op_LoadP");
2203 #endif
2204
2205 Node* adr = n->in(MemNode::Address)->uncast();
2206 const Type *adr_type = phase->type(adr);
2207 Node* adr_base;
2208 if (adr->is_AddP()) {
2209 adr_base = get_addp_base(adr);
2210 } else {
2211 adr_base = adr;
2212 }
2213
2214 // For everything "adr_base" could point to, create a deferred edge from
2215 // this node to each field with the same offset.
2216 VectorSet ptset(Thread::current()->resource_area());
2217 PointsTo(ptset, adr_base, phase);
2218 int offset = address_offset(adr, phase);
2219 for( VectorSetI i(&ptset); i.test(); ++i ) {
2220 uint pt = i.elem;
2221 add_deferred_edge_to_fields(n_idx, pt, offset);
2222 }
2223 break;
2224 }
2225 case Op_Parm:
2226 {
2227 assert(false, "Op_Parm");
2228 break;
2229 }
2230 case Op_Phi:
2231 {
2232 #ifdef ASSERT
2233 const Type *t = n->as_Phi()->type();
2234 if (t->make_ptr() == NULL)
2235 assert(false, "Op_Phi");
2236 #endif
2237 for (uint i = 1; i < n->req() ; i++) {
2238 Node* in = n->in(i);
2239 if (in == NULL)
2240 continue; // ignore NULL
2241 in = in->uncast();
2242 if (in->is_top() || in == n)
2243 continue; // ignore top or inputs which go back this node
2244 int ti = in->_idx;
2245 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2246 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2247 if (nt == PointsToNode::JavaObject) {
2248 add_pointsto_edge(n_idx, ti);
2249 } else {
2250 add_deferred_edge(n_idx, ti);
2251 }
2252 }
2253 _processed.set(n_idx);
2254 break;
2255 }
2256 case Op_Proj:
2257 {
2258 // we are only interested in the result projection from a call
2259 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2260 process_call_result(n->as_Proj(), phase);
2261 assert(_processed.test(n_idx), "all call results should be processed");
2262 } else {
2263 assert(false, "Op_Proj");
2264 }
2265 break;
2266 }
2267 case Op_Return:
2268 {
2269 #ifdef ASSERT
2270 if( n->req() <= TypeFunc::Parms ||
2271 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2272 assert(false, "Op_Return");
2273 }
2274 #endif
2275 int ti = n->in(TypeFunc::Parms)->_idx;
2276 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2277 add_pointsto_edge(n_idx, ti);
2278 } else {
2279 add_deferred_edge(n_idx, ti);
2280 }
2281 _processed.set(n_idx);
2282 break;
2283 }
2284 case Op_StoreP:
2285 case Op_StoreN:
2286 case Op_StorePConditional:
2287 case Op_CompareAndSwapP:
2288 case Op_CompareAndSwapN:
2289 {
2290 Node *adr = n->in(MemNode::Address);
2291 const Type *adr_type = phase->type(adr)->make_ptr();
2292 #ifdef ASSERT
2293 if (!adr_type->isa_oopptr())
2294 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2295 #endif
2296
2297 assert(adr->is_AddP(), "expecting an AddP");
2298 Node *adr_base = get_addp_base(adr);
2299 Node *val = n->in(MemNode::ValueIn)->uncast();
2300 // For everything "adr_base" could point to, create a deferred edge
2301 // to "val" from each field with the same offset.
2302 VectorSet ptset(Thread::current()->resource_area());
2303 PointsTo(ptset, adr_base, phase);
2304 for( VectorSetI i(&ptset); i.test(); ++i ) {
2305 uint pt = i.elem;
2306 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2307 }
2308 break;
2309 }
2310 case Op_ThreadLocal:
2311 {
2312 assert(false, "Op_ThreadLocal");
2313 break;
2314 }
2315 default:
2316 ;
2317 // nothing to do
2318 }
2319 }
2320
2321 #ifndef PRODUCT
2322 void ConnectionGraph::dump() {
2323 PhaseGVN *igvn = _compile->initial_gvn();
2324 bool first = true;
2325
2326 uint size = nodes_size();
2327 for (uint ni = 0; ni < size; ni++) {
2328 PointsToNode *ptn = ptnode_adr(ni);
2329 PointsToNode::NodeType ptn_type = ptn->node_type();
2330
2331 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2332 continue;
2333 PointsToNode::EscapeState es = escape_state(ptn->_node, igvn);
2334 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2335 if (first) {
2336 tty->cr();
2337 tty->print("======== Connection graph for ");
2338 _compile->method()->print_short_name();
2339 tty->cr();
2340 first = false;
2341 }
2342 tty->print("%6d ", ni);
2343 ptn->dump();
2344 // Print all locals which reference this allocation
2345 for (uint li = ni; li < size; li++) {
2346 PointsToNode *ptn_loc = ptnode_adr(li);
2347 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2348 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2349 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2350 ptnode_adr(li)->dump(false);
2351 }
2352 }
2353 if (Verbose) {
2354 // Print all fields which reference this allocation
2355 for (uint i = 0; i < ptn->edge_count(); i++) {
2356 uint ei = ptn->edge_target(i);
2357 ptnode_adr(ei)->dump(false);
2358 }
2359 }
2360 tty->cr();
2361 }
2362 }
2363 }
2364 #endif