Fri, 20 Aug 2010 23:40:30 -0700
6912064: type profiles need to be exploited more for dynamic language support
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2005, 2009, 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 "incls/_precompiled.incl"
26 #include "incls/_escape.cpp.incl"
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 }
37 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
40 _edges->remove(v);
41 }
43 #ifndef PRODUCT
44 static const char *node_type_names[] = {
45 "UnknownType",
46 "JavaObject",
47 "LocalVar",
48 "Field"
49 };
51 static const char *esc_names[] = {
52 "UnknownEscape",
53 "NoEscape",
54 "ArgEscape",
55 "GlobalEscape"
56 };
58 static const char *edge_type_suffix[] = {
59 "?", // UnknownEdge
60 "P", // PointsToEdge
61 "D", // DeferredEdge
62 "F" // FieldEdge
63 };
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
84 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
85 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
86 _processed(C->comp_arena()),
87 _collecting(true),
88 _compile(C),
89 _igvn(igvn),
90 _node_map(C->comp_arena()) {
92 _phantom_object = C->top()->_idx,
93 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
95 // Add ConP(#NULL) and ConN(#NULL) nodes.
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);
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 }
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);
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 }
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);
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 }
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 }
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);
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);
160 f->add_edge(to_i, PointsToNode::FieldEdge);
161 }
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 }
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);
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);
181 if (done)
182 _processed.set(n->_idx);
183 }
185 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
186 uint idx = n->_idx;
187 PointsToNode::EscapeState es;
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;
194 // if the node was created after the escape computation, return
195 // UnknownEscape
196 if (idx >= nodes_size())
197 return PointsToNode::UnknownEscape;
199 es = ptnode_adr(idx)->escape_state();
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;
206 // PointsTo() calls n->uncast() which can return a new ideal node.
207 if (n->uncast()->_idx >= nodes_size())
208 return PointsToNode::UnknownEscape;
210 PointsToNode::EscapeState orig_es = es;
212 // compute max escape state of anything this node could point to
213 VectorSet ptset(Thread::current()->resource_area());
214 PointsTo(ptset, n);
215 for(VectorSetI i(&ptset); i.test() && es != PointsToNode::GlobalEscape; ++i) {
216 uint pt = i.elem;
217 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
218 if (pes > es)
219 es = pes;
220 }
221 if (orig_es != es) {
222 // cache the computed escape state
223 assert(es != PointsToNode::UnknownEscape, "should have computed an escape state");
224 ptnode_adr(idx)->set_escape_state(es);
225 } // orig_es could be PointsToNode::UnknownEscape
226 return es;
227 }
229 void ConnectionGraph::PointsTo(VectorSet &ptset, Node * n) {
230 VectorSet visited(Thread::current()->resource_area());
231 GrowableArray<uint> worklist;
233 #ifdef ASSERT
234 Node *orig_n = n;
235 #endif
237 n = n->uncast();
238 PointsToNode* npt = ptnode_adr(n->_idx);
240 // If we have a JavaObject, return just that object
241 if (npt->node_type() == PointsToNode::JavaObject) {
242 ptset.set(n->_idx);
243 return;
244 }
245 #ifdef ASSERT
246 if (npt->_node == NULL) {
247 if (orig_n != n)
248 orig_n->dump();
249 n->dump();
250 assert(npt->_node != NULL, "unregistered node");
251 }
252 #endif
253 worklist.push(n->_idx);
254 while(worklist.length() > 0) {
255 int ni = worklist.pop();
256 if (visited.test_set(ni))
257 continue;
259 PointsToNode* pn = ptnode_adr(ni);
260 // ensure that all inputs of a Phi have been processed
261 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
263 int edges_processed = 0;
264 uint e_cnt = pn->edge_count();
265 for (uint e = 0; e < e_cnt; e++) {
266 uint etgt = pn->edge_target(e);
267 PointsToNode::EdgeType et = pn->edge_type(e);
268 if (et == PointsToNode::PointsToEdge) {
269 ptset.set(etgt);
270 edges_processed++;
271 } else if (et == PointsToNode::DeferredEdge) {
272 worklist.push(etgt);
273 edges_processed++;
274 } else {
275 assert(false,"neither PointsToEdge or DeferredEdge");
276 }
277 }
278 if (edges_processed == 0) {
279 // no deferred or pointsto edges found. Assume the value was set
280 // outside this method. Add the phantom object to the pointsto set.
281 ptset.set(_phantom_object);
282 }
283 }
284 }
286 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
287 // This method is most expensive during ConnectionGraph construction.
288 // Reuse vectorSet and an additional growable array for deferred edges.
289 deferred_edges->clear();
290 visited->Clear();
292 visited->set(ni);
293 PointsToNode *ptn = ptnode_adr(ni);
295 // Mark current edges as visited and move deferred edges to separate array.
296 for (uint i = 0; i < ptn->edge_count(); ) {
297 uint t = ptn->edge_target(i);
298 #ifdef ASSERT
299 assert(!visited->test_set(t), "expecting no duplications");
300 #else
301 visited->set(t);
302 #endif
303 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
304 ptn->remove_edge(t, PointsToNode::DeferredEdge);
305 deferred_edges->append(t);
306 } else {
307 i++;
308 }
309 }
310 for (int next = 0; next < deferred_edges->length(); ++next) {
311 uint t = deferred_edges->at(next);
312 PointsToNode *ptt = ptnode_adr(t);
313 uint e_cnt = ptt->edge_count();
314 for (uint e = 0; e < e_cnt; e++) {
315 uint etgt = ptt->edge_target(e);
316 if (visited->test_set(etgt))
317 continue;
319 PointsToNode::EdgeType et = ptt->edge_type(e);
320 if (et == PointsToNode::PointsToEdge) {
321 add_pointsto_edge(ni, etgt);
322 if(etgt == _phantom_object) {
323 // Special case - field set outside (globally escaping).
324 ptn->set_escape_state(PointsToNode::GlobalEscape);
325 }
326 } else if (et == PointsToNode::DeferredEdge) {
327 deferred_edges->append(etgt);
328 } else {
329 assert(false,"invalid connection graph");
330 }
331 }
332 }
333 }
336 // Add an edge to node given by "to_i" from any field of adr_i whose offset
337 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
338 // a pointsto edge is added if it is a JavaObject
340 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
341 PointsToNode* an = ptnode_adr(adr_i);
342 PointsToNode* to = ptnode_adr(to_i);
343 bool deferred = (to->node_type() == PointsToNode::LocalVar);
345 for (uint fe = 0; fe < an->edge_count(); fe++) {
346 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
347 int fi = an->edge_target(fe);
348 PointsToNode* pf = ptnode_adr(fi);
349 int po = pf->offset();
350 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
351 if (deferred)
352 add_deferred_edge(fi, to_i);
353 else
354 add_pointsto_edge(fi, to_i);
355 }
356 }
357 }
359 // Add a deferred edge from node given by "from_i" to any field of adr_i
360 // whose offset matches "offset".
361 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
362 PointsToNode* an = ptnode_adr(adr_i);
363 for (uint fe = 0; fe < an->edge_count(); fe++) {
364 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
365 int fi = an->edge_target(fe);
366 PointsToNode* pf = ptnode_adr(fi);
367 int po = pf->offset();
368 if (pf->edge_count() == 0) {
369 // we have not seen any stores to this field, assume it was set outside this method
370 add_pointsto_edge(fi, _phantom_object);
371 }
372 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
373 add_deferred_edge(from_i, fi);
374 }
375 }
376 }
378 // Helper functions
380 static Node* get_addp_base(Node *addp) {
381 assert(addp->is_AddP(), "must be AddP");
382 //
383 // AddP cases for Base and Address inputs:
384 // case #1. Direct object's field reference:
385 // Allocate
386 // |
387 // Proj #5 ( oop result )
388 // |
389 // CheckCastPP (cast to instance type)
390 // | |
391 // AddP ( base == address )
392 //
393 // case #2. Indirect object's field reference:
394 // Phi
395 // |
396 // CastPP (cast to instance type)
397 // | |
398 // AddP ( base == address )
399 //
400 // case #3. Raw object's field reference for Initialize node:
401 // Allocate
402 // |
403 // Proj #5 ( oop result )
404 // top |
405 // \ |
406 // AddP ( base == top )
407 //
408 // case #4. Array's element reference:
409 // {CheckCastPP | CastPP}
410 // | | |
411 // | AddP ( array's element offset )
412 // | |
413 // AddP ( array's offset )
414 //
415 // case #5. Raw object's field reference for arraycopy stub call:
416 // The inline_native_clone() case when the arraycopy stub is called
417 // after the allocation before Initialize and CheckCastPP nodes.
418 // Allocate
419 // |
420 // Proj #5 ( oop result )
421 // | |
422 // AddP ( base == address )
423 //
424 // case #6. Constant Pool, ThreadLocal, CastX2P or
425 // Raw object's field reference:
426 // {ConP, ThreadLocal, CastX2P, raw Load}
427 // top |
428 // \ |
429 // AddP ( base == top )
430 //
431 // case #7. Klass's field reference.
432 // LoadKlass
433 // | |
434 // AddP ( base == address )
435 //
436 // case #8. narrow Klass's field reference.
437 // LoadNKlass
438 // |
439 // DecodeN
440 // | |
441 // AddP ( base == address )
442 //
443 Node *base = addp->in(AddPNode::Base)->uncast();
444 if (base->is_top()) { // The AddP case #3 and #6.
445 base = addp->in(AddPNode::Address)->uncast();
446 while (base->is_AddP()) {
447 // Case #6 (unsafe access) may have several chained AddP nodes.
448 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
449 base = base->in(AddPNode::Address)->uncast();
450 }
451 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
452 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
453 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
454 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
455 }
456 return base;
457 }
459 static Node* find_second_addp(Node* addp, Node* n) {
460 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
462 Node* addp2 = addp->raw_out(0);
463 if (addp->outcnt() == 1 && addp2->is_AddP() &&
464 addp2->in(AddPNode::Base) == n &&
465 addp2->in(AddPNode::Address) == addp) {
467 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
468 //
469 // Find array's offset to push it on worklist first and
470 // as result process an array's element offset first (pushed second)
471 // to avoid CastPP for the array's offset.
472 // Otherwise the inserted CastPP (LocalVar) will point to what
473 // the AddP (Field) points to. Which would be wrong since
474 // the algorithm expects the CastPP has the same point as
475 // as AddP's base CheckCastPP (LocalVar).
476 //
477 // ArrayAllocation
478 // |
479 // CheckCastPP
480 // |
481 // memProj (from ArrayAllocation CheckCastPP)
482 // | ||
483 // | || Int (element index)
484 // | || | ConI (log(element size))
485 // | || | /
486 // | || LShift
487 // | || /
488 // | AddP (array's element offset)
489 // | |
490 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
491 // | / /
492 // AddP (array's offset)
493 // |
494 // Load/Store (memory operation on array's element)
495 //
496 return addp2;
497 }
498 return NULL;
499 }
501 //
502 // Adjust the type and inputs of an AddP which computes the
503 // address of a field of an instance
504 //
505 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
506 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
507 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
508 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
509 if (t == NULL) {
510 // We are computing a raw address for a store captured by an Initialize
511 // compute an appropriate address type (cases #3 and #5).
512 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
513 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
514 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
515 assert(offs != Type::OffsetBot, "offset must be a constant");
516 t = base_t->add_offset(offs)->is_oopptr();
517 }
518 int inst_id = base_t->instance_id();
519 assert(!t->is_known_instance() || t->instance_id() == inst_id,
520 "old type must be non-instance or match new type");
522 // The type 't' could be subclass of 'base_t'.
523 // As result t->offset() could be large then base_t's size and it will
524 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
525 // constructor verifies correctness of the offset.
526 //
527 // It could happened on subclass's branch (from the type profiling
528 // inlining) which was not eliminated during parsing since the exactness
529 // of the allocation type was not propagated to the subclass type check.
530 //
531 // Or the type 't' could be not related to 'base_t' at all.
532 // It could happened when CHA type is different from MDO type on a dead path
533 // (for example, from instanceof check) which is not collapsed during parsing.
534 //
535 // Do nothing for such AddP node and don't process its users since
536 // this code branch will go away.
537 //
538 if (!t->is_known_instance() &&
539 !base_t->klass()->is_subtype_of(t->klass())) {
540 return false; // bail out
541 }
543 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
544 // Do NOT remove the next line: ensure a new alias index is allocated
545 // for the instance type. Note: C++ will not remove it since the call
546 // has side effect.
547 int alias_idx = _compile->get_alias_index(tinst);
548 igvn->set_type(addp, tinst);
549 // record the allocation in the node map
550 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
551 set_map(addp->_idx, get_map(base->_idx));
553 // Set addp's Base and Address to 'base'.
554 Node *abase = addp->in(AddPNode::Base);
555 Node *adr = addp->in(AddPNode::Address);
556 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
557 adr->in(0)->_idx == (uint)inst_id) {
558 // Skip AddP cases #3 and #5.
559 } else {
560 assert(!abase->is_top(), "sanity"); // AddP case #3
561 if (abase != base) {
562 igvn->hash_delete(addp);
563 addp->set_req(AddPNode::Base, base);
564 if (abase == adr) {
565 addp->set_req(AddPNode::Address, base);
566 } else {
567 // AddP case #4 (adr is array's element offset AddP node)
568 #ifdef ASSERT
569 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
570 assert(adr->is_AddP() && atype != NULL &&
571 atype->instance_id() == inst_id, "array's element offset should be processed first");
572 #endif
573 }
574 igvn->hash_insert(addp);
575 }
576 }
577 // Put on IGVN worklist since at least addp's type was changed above.
578 record_for_optimizer(addp);
579 return true;
580 }
582 //
583 // Create a new version of orig_phi if necessary. Returns either the newly
584 // created phi or an existing phi. Sets create_new to indicate wheter a new
585 // phi was created. Cache the last newly created phi in the node map.
586 //
587 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
588 Compile *C = _compile;
589 new_created = false;
590 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
591 // nothing to do if orig_phi is bottom memory or matches alias_idx
592 if (phi_alias_idx == alias_idx) {
593 return orig_phi;
594 }
595 // Have we recently created a Phi for this alias index?
596 PhiNode *result = get_map_phi(orig_phi->_idx);
597 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
598 return result;
599 }
600 // Previous check may fail when the same wide memory Phi was split into Phis
601 // for different memory slices. Search all Phis for this region.
602 if (result != NULL) {
603 Node* region = orig_phi->in(0);
604 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
605 Node* phi = region->fast_out(i);
606 if (phi->is_Phi() &&
607 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
608 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
609 return phi->as_Phi();
610 }
611 }
612 }
613 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
614 if (C->do_escape_analysis() == true && !C->failing()) {
615 // Retry compilation without escape analysis.
616 // If this is the first failure, the sentinel string will "stick"
617 // to the Compile object, and the C2Compiler will see it and retry.
618 C->record_failure(C2Compiler::retry_no_escape_analysis());
619 }
620 return NULL;
621 }
622 orig_phi_worklist.append_if_missing(orig_phi);
623 const TypePtr *atype = C->get_adr_type(alias_idx);
624 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
625 C->copy_node_notes_to(result, orig_phi);
626 igvn->set_type(result, result->bottom_type());
627 record_for_optimizer(result);
629 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
630 assert(pn == NULL || pn == orig_phi, "wrong node");
631 set_map(orig_phi->_idx, result);
632 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
634 new_created = true;
635 return result;
636 }
638 //
639 // Return a new version of Memory Phi "orig_phi" with the inputs having the
640 // specified alias index.
641 //
642 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
644 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
645 Compile *C = _compile;
646 bool new_phi_created;
647 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
648 if (!new_phi_created) {
649 return result;
650 }
652 GrowableArray<PhiNode *> phi_list;
653 GrowableArray<uint> cur_input;
655 PhiNode *phi = orig_phi;
656 uint idx = 1;
657 bool finished = false;
658 while(!finished) {
659 while (idx < phi->req()) {
660 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
661 if (mem != NULL && mem->is_Phi()) {
662 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
663 if (new_phi_created) {
664 // found an phi for which we created a new split, push current one on worklist and begin
665 // processing new one
666 phi_list.push(phi);
667 cur_input.push(idx);
668 phi = mem->as_Phi();
669 result = newphi;
670 idx = 1;
671 continue;
672 } else {
673 mem = newphi;
674 }
675 }
676 if (C->failing()) {
677 return NULL;
678 }
679 result->set_req(idx++, mem);
680 }
681 #ifdef ASSERT
682 // verify that the new Phi has an input for each input of the original
683 assert( phi->req() == result->req(), "must have same number of inputs.");
684 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
685 #endif
686 // Check if all new phi's inputs have specified alias index.
687 // Otherwise use old phi.
688 for (uint i = 1; i < phi->req(); i++) {
689 Node* in = result->in(i);
690 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
691 }
692 // we have finished processing a Phi, see if there are any more to do
693 finished = (phi_list.length() == 0 );
694 if (!finished) {
695 phi = phi_list.pop();
696 idx = cur_input.pop();
697 PhiNode *prev_result = get_map_phi(phi->_idx);
698 prev_result->set_req(idx++, result);
699 result = prev_result;
700 }
701 }
702 return result;
703 }
706 //
707 // The next methods are derived from methods in MemNode.
708 //
709 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *tinst) {
710 Node *mem = mmem;
711 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
712 // means an array I have not precisely typed yet. Do not do any
713 // alias stuff with it any time soon.
714 if( tinst->base() != Type::AnyPtr &&
715 !(tinst->klass()->is_java_lang_Object() &&
716 tinst->offset() == Type::OffsetBot) ) {
717 mem = mmem->memory_at(alias_idx);
718 // Update input if it is progress over what we have now
719 }
720 return mem;
721 }
723 //
724 // Move memory users to their memory slices.
725 //
726 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
727 Compile* C = _compile;
729 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
730 assert(tp != NULL, "ptr type");
731 int alias_idx = C->get_alias_index(tp);
732 int general_idx = C->get_general_index(alias_idx);
734 // Move users first
735 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
736 Node* use = n->fast_out(i);
737 if (use->is_MergeMem()) {
738 MergeMemNode* mmem = use->as_MergeMem();
739 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
740 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
741 continue; // Nothing to do
742 }
743 // Replace previous general reference to mem node.
744 uint orig_uniq = C->unique();
745 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
746 assert(orig_uniq == C->unique(), "no new nodes");
747 mmem->set_memory_at(general_idx, m);
748 --imax;
749 --i;
750 } else if (use->is_MemBar()) {
751 assert(!use->is_Initialize(), "initializing stores should not be moved");
752 if (use->req() > MemBarNode::Precedent &&
753 use->in(MemBarNode::Precedent) == n) {
754 // Don't move related membars.
755 record_for_optimizer(use);
756 continue;
757 }
758 tp = use->as_MemBar()->adr_type()->isa_ptr();
759 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
760 alias_idx == general_idx) {
761 continue; // Nothing to do
762 }
763 // Move to general memory slice.
764 uint orig_uniq = C->unique();
765 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
766 assert(orig_uniq == C->unique(), "no new nodes");
767 igvn->hash_delete(use);
768 imax -= use->replace_edge(n, m);
769 igvn->hash_insert(use);
770 record_for_optimizer(use);
771 --i;
772 #ifdef ASSERT
773 } else if (use->is_Mem()) {
774 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
775 // Don't move related cardmark.
776 continue;
777 }
778 // Memory nodes should have new memory input.
779 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
780 assert(tp != NULL, "ptr type");
781 int idx = C->get_alias_index(tp);
782 assert(get_map(use->_idx) != NULL || idx == alias_idx,
783 "Following memory nodes should have new memory input or be on the same memory slice");
784 } else if (use->is_Phi()) {
785 // Phi nodes should be split and moved already.
786 tp = use->as_Phi()->adr_type()->isa_ptr();
787 assert(tp != NULL, "ptr type");
788 int idx = C->get_alias_index(tp);
789 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
790 } else {
791 use->dump();
792 assert(false, "should not be here");
793 #endif
794 }
795 }
796 }
798 //
799 // Search memory chain of "mem" to find a MemNode whose address
800 // is the specified alias index.
801 //
802 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
803 if (orig_mem == NULL)
804 return orig_mem;
805 Compile* C = phase->C;
806 const TypeOopPtr *tinst = C->get_adr_type(alias_idx)->isa_oopptr();
807 bool is_instance = (tinst != NULL) && tinst->is_known_instance();
808 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
809 Node *prev = NULL;
810 Node *result = orig_mem;
811 while (prev != result) {
812 prev = result;
813 if (result == start_mem)
814 break; // hit one of our sentinels
815 if (result->is_Mem()) {
816 const Type *at = phase->type(result->in(MemNode::Address));
817 if (at != Type::TOP) {
818 assert (at->isa_ptr() != NULL, "pointer type required.");
819 int idx = C->get_alias_index(at->is_ptr());
820 if (idx == alias_idx)
821 break;
822 }
823 result = result->in(MemNode::Memory);
824 }
825 if (!is_instance)
826 continue; // don't search further for non-instance types
827 // skip over a call which does not affect this memory slice
828 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
829 Node *proj_in = result->in(0);
830 if (proj_in->is_Allocate() && proj_in->_idx == (uint)tinst->instance_id()) {
831 break; // hit one of our sentinels
832 } else if (proj_in->is_Call()) {
833 CallNode *call = proj_in->as_Call();
834 if (!call->may_modify(tinst, phase)) {
835 result = call->in(TypeFunc::Memory);
836 }
837 } else if (proj_in->is_Initialize()) {
838 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
839 // Stop if this is the initialization for the object instance which
840 // which contains this memory slice, otherwise skip over it.
841 if (alloc == NULL || alloc->_idx != (uint)tinst->instance_id()) {
842 result = proj_in->in(TypeFunc::Memory);
843 }
844 } else if (proj_in->is_MemBar()) {
845 result = proj_in->in(TypeFunc::Memory);
846 }
847 } else if (result->is_MergeMem()) {
848 MergeMemNode *mmem = result->as_MergeMem();
849 result = step_through_mergemem(mmem, alias_idx, tinst);
850 if (result == mmem->base_memory()) {
851 // Didn't find instance memory, search through general slice recursively.
852 result = mmem->memory_at(C->get_general_index(alias_idx));
853 result = find_inst_mem(result, alias_idx, orig_phis, phase);
854 if (C->failing()) {
855 return NULL;
856 }
857 mmem->set_memory_at(alias_idx, result);
858 }
859 } else if (result->is_Phi() &&
860 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
861 Node *un = result->as_Phi()->unique_input(phase);
862 if (un != NULL) {
863 orig_phis.append_if_missing(result->as_Phi());
864 result = un;
865 } else {
866 break;
867 }
868 } else if (result->is_ClearArray()) {
869 if (!ClearArrayNode::step_through(&result, (uint)tinst->instance_id(), phase)) {
870 // Can not bypass initialization of the instance
871 // we are looking for.
872 break;
873 }
874 // Otherwise skip it (the call updated 'result' value).
875 } else if (result->Opcode() == Op_SCMemProj) {
876 assert(result->in(0)->is_LoadStore(), "sanity");
877 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
878 if (at != Type::TOP) {
879 assert (at->isa_ptr() != NULL, "pointer type required.");
880 int idx = C->get_alias_index(at->is_ptr());
881 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
882 break;
883 }
884 result = result->in(0)->in(MemNode::Memory);
885 }
886 }
887 if (result->is_Phi()) {
888 PhiNode *mphi = result->as_Phi();
889 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
890 const TypePtr *t = mphi->adr_type();
891 if (C->get_alias_index(t) != alias_idx) {
892 // Create a new Phi with the specified alias index type.
893 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
894 } else if (!is_instance) {
895 // Push all non-instance Phis on the orig_phis worklist to update inputs
896 // during Phase 4 if needed.
897 orig_phis.append_if_missing(mphi);
898 }
899 }
900 // the result is either MemNode, PhiNode, InitializeNode.
901 return result;
902 }
904 //
905 // Convert the types of unescaped object to instance types where possible,
906 // propagate the new type information through the graph, and update memory
907 // edges and MergeMem inputs to reflect the new type.
908 //
909 // We start with allocations (and calls which may be allocations) on alloc_worklist.
910 // The processing is done in 4 phases:
911 //
912 // Phase 1: Process possible allocations from alloc_worklist. Create instance
913 // types for the CheckCastPP for allocations where possible.
914 // Propagate the the new types through users as follows:
915 // casts and Phi: push users on alloc_worklist
916 // AddP: cast Base and Address inputs to the instance type
917 // push any AddP users on alloc_worklist and push any memnode
918 // users onto memnode_worklist.
919 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
920 // search the Memory chain for a store with the appropriate type
921 // address type. If a Phi is found, create a new version with
922 // the appropriate memory slices from each of the Phi inputs.
923 // For stores, process the users as follows:
924 // MemNode: push on memnode_worklist
925 // MergeMem: push on mergemem_worklist
926 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
927 // moving the first node encountered of each instance type to the
928 // the input corresponding to its alias index.
929 // appropriate memory slice.
930 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
931 //
932 // In the following example, the CheckCastPP nodes are the cast of allocation
933 // results and the allocation of node 29 is unescaped and eligible to be an
934 // instance type.
935 //
936 // We start with:
937 //
938 // 7 Parm #memory
939 // 10 ConI "12"
940 // 19 CheckCastPP "Foo"
941 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
942 // 29 CheckCastPP "Foo"
943 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
944 //
945 // 40 StoreP 25 7 20 ... alias_index=4
946 // 50 StoreP 35 40 30 ... alias_index=4
947 // 60 StoreP 45 50 20 ... alias_index=4
948 // 70 LoadP _ 60 30 ... alias_index=4
949 // 80 Phi 75 50 60 Memory alias_index=4
950 // 90 LoadP _ 80 30 ... alias_index=4
951 // 100 LoadP _ 80 20 ... alias_index=4
952 //
953 //
954 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
955 // and creating a new alias index for node 30. This gives:
956 //
957 // 7 Parm #memory
958 // 10 ConI "12"
959 // 19 CheckCastPP "Foo"
960 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
961 // 29 CheckCastPP "Foo" iid=24
962 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
963 //
964 // 40 StoreP 25 7 20 ... alias_index=4
965 // 50 StoreP 35 40 30 ... alias_index=6
966 // 60 StoreP 45 50 20 ... alias_index=4
967 // 70 LoadP _ 60 30 ... alias_index=6
968 // 80 Phi 75 50 60 Memory alias_index=4
969 // 90 LoadP _ 80 30 ... alias_index=6
970 // 100 LoadP _ 80 20 ... alias_index=4
971 //
972 // In phase 2, new memory inputs are computed for the loads and stores,
973 // And a new version of the phi is created. In phase 4, the inputs to
974 // node 80 are updated and then the memory nodes are updated with the
975 // values computed in phase 2. This results in:
976 //
977 // 7 Parm #memory
978 // 10 ConI "12"
979 // 19 CheckCastPP "Foo"
980 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
981 // 29 CheckCastPP "Foo" iid=24
982 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
983 //
984 // 40 StoreP 25 7 20 ... alias_index=4
985 // 50 StoreP 35 7 30 ... alias_index=6
986 // 60 StoreP 45 40 20 ... alias_index=4
987 // 70 LoadP _ 50 30 ... alias_index=6
988 // 80 Phi 75 40 60 Memory alias_index=4
989 // 120 Phi 75 50 50 Memory alias_index=6
990 // 90 LoadP _ 120 30 ... alias_index=6
991 // 100 LoadP _ 80 20 ... alias_index=4
992 //
993 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
994 GrowableArray<Node *> memnode_worklist;
995 GrowableArray<PhiNode *> orig_phis;
997 PhaseGVN *igvn = _igvn;
998 uint new_index_start = (uint) _compile->num_alias_types();
999 Arena* arena = Thread::current()->resource_area();
1000 VectorSet visited(arena);
1001 VectorSet ptset(arena);
1004 // Phase 1: Process possible allocations from alloc_worklist.
1005 // Create instance types for the CheckCastPP for allocations where possible.
1006 //
1007 // (Note: don't forget to change the order of the second AddP node on
1008 // the alloc_worklist if the order of the worklist processing is changed,
1009 // see the comment in find_second_addp().)
1010 //
1011 while (alloc_worklist.length() != 0) {
1012 Node *n = alloc_worklist.pop();
1013 uint ni = n->_idx;
1014 const TypeOopPtr* tinst = NULL;
1015 if (n->is_Call()) {
1016 CallNode *alloc = n->as_Call();
1017 // copy escape information to call node
1018 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1019 PointsToNode::EscapeState es = escape_state(alloc);
1020 // We have an allocation or call which returns a Java object,
1021 // see if it is unescaped.
1022 if (es != PointsToNode::NoEscape || !ptn->_scalar_replaceable)
1023 continue;
1025 // Find CheckCastPP for the allocate or for the return value of a call
1026 n = alloc->result_cast();
1027 if (n == NULL) { // No uses except Initialize node
1028 if (alloc->is_Allocate()) {
1029 // Set the scalar_replaceable flag for allocation
1030 // so it could be eliminated if it has no uses.
1031 alloc->as_Allocate()->_is_scalar_replaceable = true;
1032 }
1033 continue;
1034 }
1035 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1036 assert(!alloc->is_Allocate(), "allocation should have unique type");
1037 continue;
1038 }
1040 // The inline code for Object.clone() casts the allocation result to
1041 // java.lang.Object and then to the actual type of the allocated
1042 // object. Detect this case and use the second cast.
1043 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1044 // the allocation result is cast to java.lang.Object and then
1045 // to the actual Array type.
1046 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1047 && (alloc->is_AllocateArray() ||
1048 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1049 Node *cast2 = NULL;
1050 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1051 Node *use = n->fast_out(i);
1052 if (use->is_CheckCastPP()) {
1053 cast2 = use;
1054 break;
1055 }
1056 }
1057 if (cast2 != NULL) {
1058 n = cast2;
1059 } else {
1060 // Non-scalar replaceable if the allocation type is unknown statically
1061 // (reflection allocation), the object can't be restored during
1062 // deoptimization without precise type.
1063 continue;
1064 }
1065 }
1066 if (alloc->is_Allocate()) {
1067 // Set the scalar_replaceable flag for allocation
1068 // so it could be eliminated.
1069 alloc->as_Allocate()->_is_scalar_replaceable = true;
1070 }
1071 set_escape_state(n->_idx, es);
1072 // in order for an object to be scalar-replaceable, it must be:
1073 // - a direct allocation (not a call returning an object)
1074 // - non-escaping
1075 // - eligible to be a unique type
1076 // - not determined to be ineligible by escape analysis
1077 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1078 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1079 set_map(alloc->_idx, n);
1080 set_map(n->_idx, alloc);
1081 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1082 if (t == NULL)
1083 continue; // not a TypeInstPtr
1084 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1085 igvn->hash_delete(n);
1086 igvn->set_type(n, tinst);
1087 n->raise_bottom_type(tinst);
1088 igvn->hash_insert(n);
1089 record_for_optimizer(n);
1090 if (alloc->is_Allocate() && ptn->_scalar_replaceable &&
1091 (t->isa_instptr() || t->isa_aryptr())) {
1093 // First, put on the worklist all Field edges from Connection Graph
1094 // which is more accurate then putting immediate users from Ideal Graph.
1095 for (uint e = 0; e < ptn->edge_count(); e++) {
1096 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1097 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1098 "only AddP nodes are Field edges in CG");
1099 if (use->outcnt() > 0) { // Don't process dead nodes
1100 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1101 if (addp2 != NULL) {
1102 assert(alloc->is_AllocateArray(),"array allocation was expected");
1103 alloc_worklist.append_if_missing(addp2);
1104 }
1105 alloc_worklist.append_if_missing(use);
1106 }
1107 }
1109 // An allocation may have an Initialize which has raw stores. Scan
1110 // the users of the raw allocation result and push AddP users
1111 // on alloc_worklist.
1112 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1113 assert (raw_result != NULL, "must have an allocation result");
1114 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1115 Node *use = raw_result->fast_out(i);
1116 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1117 Node* addp2 = find_second_addp(use, raw_result);
1118 if (addp2 != NULL) {
1119 assert(alloc->is_AllocateArray(),"array allocation was expected");
1120 alloc_worklist.append_if_missing(addp2);
1121 }
1122 alloc_worklist.append_if_missing(use);
1123 } else if (use->is_MemBar()) {
1124 memnode_worklist.append_if_missing(use);
1125 }
1126 }
1127 }
1128 } else if (n->is_AddP()) {
1129 ptset.Clear();
1130 PointsTo(ptset, get_addp_base(n));
1131 assert(ptset.Size() == 1, "AddP address is unique");
1132 uint elem = ptset.getelem(); // Allocation node's index
1133 if (elem == _phantom_object) {
1134 assert(false, "escaped allocation");
1135 continue; // Assume the value was set outside this method.
1136 }
1137 Node *base = get_map(elem); // CheckCastPP node
1138 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1139 tinst = igvn->type(base)->isa_oopptr();
1140 } else if (n->is_Phi() ||
1141 n->is_CheckCastPP() ||
1142 n->is_EncodeP() ||
1143 n->is_DecodeN() ||
1144 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1145 if (visited.test_set(n->_idx)) {
1146 assert(n->is_Phi(), "loops only through Phi's");
1147 continue; // already processed
1148 }
1149 ptset.Clear();
1150 PointsTo(ptset, n);
1151 if (ptset.Size() == 1) {
1152 uint elem = ptset.getelem(); // Allocation node's index
1153 if (elem == _phantom_object) {
1154 assert(false, "escaped allocation");
1155 continue; // Assume the value was set outside this method.
1156 }
1157 Node *val = get_map(elem); // CheckCastPP node
1158 TypeNode *tn = n->as_Type();
1159 tinst = igvn->type(val)->isa_oopptr();
1160 assert(tinst != NULL && tinst->is_known_instance() &&
1161 (uint)tinst->instance_id() == elem , "instance type expected.");
1163 const Type *tn_type = igvn->type(tn);
1164 const TypeOopPtr *tn_t;
1165 if (tn_type->isa_narrowoop()) {
1166 tn_t = tn_type->make_ptr()->isa_oopptr();
1167 } else {
1168 tn_t = tn_type->isa_oopptr();
1169 }
1171 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1172 if (tn_type->isa_narrowoop()) {
1173 tn_type = tinst->make_narrowoop();
1174 } else {
1175 tn_type = tinst;
1176 }
1177 igvn->hash_delete(tn);
1178 igvn->set_type(tn, tn_type);
1179 tn->set_type(tn_type);
1180 igvn->hash_insert(tn);
1181 record_for_optimizer(n);
1182 } else {
1183 assert(tn_type == TypePtr::NULL_PTR ||
1184 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1185 "unexpected type");
1186 continue; // Skip dead path with different type
1187 }
1188 }
1189 } else {
1190 debug_only(n->dump();)
1191 assert(false, "EA: unexpected node");
1192 continue;
1193 }
1194 // push allocation's users on appropriate worklist
1195 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1196 Node *use = n->fast_out(i);
1197 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1198 // Load/store to instance's field
1199 memnode_worklist.append_if_missing(use);
1200 } else if (use->is_MemBar()) {
1201 memnode_worklist.append_if_missing(use);
1202 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1203 Node* addp2 = find_second_addp(use, n);
1204 if (addp2 != NULL) {
1205 alloc_worklist.append_if_missing(addp2);
1206 }
1207 alloc_worklist.append_if_missing(use);
1208 } else if (use->is_Phi() ||
1209 use->is_CheckCastPP() ||
1210 use->is_EncodeP() ||
1211 use->is_DecodeN() ||
1212 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1213 alloc_worklist.append_if_missing(use);
1214 #ifdef ASSERT
1215 } else if (use->is_Mem()) {
1216 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1217 } else if (use->is_MergeMem()) {
1218 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1219 } else if (use->is_SafePoint()) {
1220 // Look for MergeMem nodes for calls which reference unique allocation
1221 // (through CheckCastPP nodes) even for debug info.
1222 Node* m = use->in(TypeFunc::Memory);
1223 if (m->is_MergeMem()) {
1224 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1225 }
1226 } else {
1227 uint op = use->Opcode();
1228 if (!(op == Op_CmpP || op == Op_Conv2B ||
1229 op == Op_CastP2X || op == Op_StoreCM ||
1230 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1231 op == Op_StrEquals || op == Op_StrIndexOf)) {
1232 n->dump();
1233 use->dump();
1234 assert(false, "EA: missing allocation reference path");
1235 }
1236 #endif
1237 }
1238 }
1240 }
1241 // New alias types were created in split_AddP().
1242 uint new_index_end = (uint) _compile->num_alias_types();
1244 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1245 // compute new values for Memory inputs (the Memory inputs are not
1246 // actually updated until phase 4.)
1247 if (memnode_worklist.length() == 0)
1248 return; // nothing to do
1250 while (memnode_worklist.length() != 0) {
1251 Node *n = memnode_worklist.pop();
1252 if (visited.test_set(n->_idx))
1253 continue;
1254 if (n->is_Phi() || n->is_ClearArray()) {
1255 // we don't need to do anything, but the users must be pushed
1256 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1257 // we don't need to do anything, but the users must be pushed
1258 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1259 if (n == NULL)
1260 continue;
1261 } else {
1262 assert(n->is_Mem(), "memory node required.");
1263 Node *addr = n->in(MemNode::Address);
1264 const Type *addr_t = igvn->type(addr);
1265 if (addr_t == Type::TOP)
1266 continue;
1267 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1268 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1269 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1270 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1271 if (_compile->failing()) {
1272 return;
1273 }
1274 if (mem != n->in(MemNode::Memory)) {
1275 // We delay the memory edge update since we need old one in
1276 // MergeMem code below when instances memory slices are separated.
1277 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1278 assert(pn == NULL || pn == n, "wrong node");
1279 set_map(n->_idx, mem);
1280 ptnode_adr(n->_idx)->_node = n;
1281 }
1282 if (n->is_Load()) {
1283 continue; // don't push users
1284 } else if (n->is_LoadStore()) {
1285 // get the memory projection
1286 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1287 Node *use = n->fast_out(i);
1288 if (use->Opcode() == Op_SCMemProj) {
1289 n = use;
1290 break;
1291 }
1292 }
1293 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1294 }
1295 }
1296 // push user on appropriate worklist
1297 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1298 Node *use = n->fast_out(i);
1299 if (use->is_Phi() || use->is_ClearArray()) {
1300 memnode_worklist.append_if_missing(use);
1301 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1302 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1303 continue;
1304 memnode_worklist.append_if_missing(use);
1305 } else if (use->is_MemBar()) {
1306 memnode_worklist.append_if_missing(use);
1307 #ifdef ASSERT
1308 } else if(use->is_Mem()) {
1309 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1310 } else if (use->is_MergeMem()) {
1311 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1312 } else {
1313 uint op = use->Opcode();
1314 if (!(op == Op_StoreCM ||
1315 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1316 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1317 op == Op_AryEq || op == Op_StrComp ||
1318 op == Op_StrEquals || op == Op_StrIndexOf)) {
1319 n->dump();
1320 use->dump();
1321 assert(false, "EA: missing memory path");
1322 }
1323 #endif
1324 }
1325 }
1326 }
1328 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1329 // Walk each memory slice moving the first node encountered of each
1330 // instance type to the the input corresponding to its alias index.
1331 uint length = _mergemem_worklist.length();
1332 for( uint next = 0; next < length; ++next ) {
1333 MergeMemNode* nmm = _mergemem_worklist.at(next);
1334 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1335 // Note: we don't want to use MergeMemStream here because we only want to
1336 // scan inputs which exist at the start, not ones we add during processing.
1337 // Note 2: MergeMem may already contains instance memory slices added
1338 // during find_inst_mem() call when memory nodes were processed above.
1339 igvn->hash_delete(nmm);
1340 uint nslices = nmm->req();
1341 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1342 Node* mem = nmm->in(i);
1343 Node* cur = NULL;
1344 if (mem == NULL || mem->is_top())
1345 continue;
1346 // First, update mergemem by moving memory nodes to corresponding slices
1347 // if their type became more precise since this mergemem was created.
1348 while (mem->is_Mem()) {
1349 const Type *at = igvn->type(mem->in(MemNode::Address));
1350 if (at != Type::TOP) {
1351 assert (at->isa_ptr() != NULL, "pointer type required.");
1352 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1353 if (idx == i) {
1354 if (cur == NULL)
1355 cur = mem;
1356 } else {
1357 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1358 nmm->set_memory_at(idx, mem);
1359 }
1360 }
1361 }
1362 mem = mem->in(MemNode::Memory);
1363 }
1364 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1365 // Find any instance of the current type if we haven't encountered
1366 // already a memory slice of the instance along the memory chain.
1367 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1368 if((uint)_compile->get_general_index(ni) == i) {
1369 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1370 if (nmm->is_empty_memory(m)) {
1371 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1372 if (_compile->failing()) {
1373 return;
1374 }
1375 nmm->set_memory_at(ni, result);
1376 }
1377 }
1378 }
1379 }
1380 // Find the rest of instances values
1381 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1382 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1383 Node* result = step_through_mergemem(nmm, ni, tinst);
1384 if (result == nmm->base_memory()) {
1385 // Didn't find instance memory, search through general slice recursively.
1386 result = nmm->memory_at(_compile->get_general_index(ni));
1387 result = find_inst_mem(result, ni, orig_phis, igvn);
1388 if (_compile->failing()) {
1389 return;
1390 }
1391 nmm->set_memory_at(ni, result);
1392 }
1393 }
1394 igvn->hash_insert(nmm);
1395 record_for_optimizer(nmm);
1396 }
1398 // Phase 4: Update the inputs of non-instance memory Phis and
1399 // the Memory input of memnodes
1400 // First update the inputs of any non-instance Phi's from
1401 // which we split out an instance Phi. Note we don't have
1402 // to recursively process Phi's encounted on the input memory
1403 // chains as is done in split_memory_phi() since they will
1404 // also be processed here.
1405 for (int j = 0; j < orig_phis.length(); j++) {
1406 PhiNode *phi = orig_phis.at(j);
1407 int alias_idx = _compile->get_alias_index(phi->adr_type());
1408 igvn->hash_delete(phi);
1409 for (uint i = 1; i < phi->req(); i++) {
1410 Node *mem = phi->in(i);
1411 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1412 if (_compile->failing()) {
1413 return;
1414 }
1415 if (mem != new_mem) {
1416 phi->set_req(i, new_mem);
1417 }
1418 }
1419 igvn->hash_insert(phi);
1420 record_for_optimizer(phi);
1421 }
1423 // Update the memory inputs of MemNodes with the value we computed
1424 // in Phase 2 and move stores memory users to corresponding memory slices.
1425 #ifdef ASSERT
1426 visited.Clear();
1427 Node_Stack old_mems(arena, _compile->unique() >> 2);
1428 #endif
1429 for (uint i = 0; i < nodes_size(); i++) {
1430 Node *nmem = get_map(i);
1431 if (nmem != NULL) {
1432 Node *n = ptnode_adr(i)->_node;
1433 assert(n != NULL, "sanity");
1434 if (n->is_Mem()) {
1435 #ifdef ASSERT
1436 Node* old_mem = n->in(MemNode::Memory);
1437 if (!visited.test_set(old_mem->_idx)) {
1438 old_mems.push(old_mem, old_mem->outcnt());
1439 }
1440 #endif
1441 assert(n->in(MemNode::Memory) != nmem, "sanity");
1442 if (!n->is_Load()) {
1443 // Move memory users of a store first.
1444 move_inst_mem(n, orig_phis, igvn);
1445 }
1446 // Now update memory input
1447 igvn->hash_delete(n);
1448 n->set_req(MemNode::Memory, nmem);
1449 igvn->hash_insert(n);
1450 record_for_optimizer(n);
1451 } else {
1452 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1453 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1454 }
1455 }
1456 }
1457 #ifdef ASSERT
1458 // Verify that memory was split correctly
1459 while (old_mems.is_nonempty()) {
1460 Node* old_mem = old_mems.node();
1461 uint old_cnt = old_mems.index();
1462 old_mems.pop();
1463 assert(old_cnt = old_mem->outcnt(), "old mem could be lost");
1464 }
1465 #endif
1466 }
1468 bool ConnectionGraph::has_candidates(Compile *C) {
1469 // EA brings benefits only when the code has allocations and/or locks which
1470 // are represented by ideal Macro nodes.
1471 int cnt = C->macro_count();
1472 for( int i=0; i < cnt; i++ ) {
1473 Node *n = C->macro_node(i);
1474 if ( n->is_Allocate() )
1475 return true;
1476 if( n->is_Lock() ) {
1477 Node* obj = n->as_Lock()->obj_node()->uncast();
1478 if( !(obj->is_Parm() || obj->is_Con()) )
1479 return true;
1480 }
1481 }
1482 return false;
1483 }
1485 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1486 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1487 // to create space for them in ConnectionGraph::_nodes[].
1488 Node* oop_null = igvn->zerocon(T_OBJECT);
1489 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1491 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1492 // Perform escape analysis
1493 if (congraph->compute_escape()) {
1494 // There are non escaping objects.
1495 C->set_congraph(congraph);
1496 }
1498 // Cleanup.
1499 if (oop_null->outcnt() == 0)
1500 igvn->hash_delete(oop_null);
1501 if (noop_null->outcnt() == 0)
1502 igvn->hash_delete(noop_null);
1503 }
1505 bool ConnectionGraph::compute_escape() {
1506 Compile* C = _compile;
1508 // 1. Populate Connection Graph (CG) with Ideal nodes.
1510 Unique_Node_List worklist_init;
1511 worklist_init.map(C->unique(), NULL); // preallocate space
1513 // Initialize worklist
1514 if (C->root() != NULL) {
1515 worklist_init.push(C->root());
1516 }
1518 GrowableArray<int> cg_worklist;
1519 PhaseGVN* igvn = _igvn;
1520 bool has_allocations = false;
1522 // Push all useful nodes onto CG list and set their type.
1523 for( uint next = 0; next < worklist_init.size(); ++next ) {
1524 Node* n = worklist_init.at(next);
1525 record_for_escape_analysis(n, igvn);
1526 // Only allocations and java static calls results are checked
1527 // for an escape status. See process_call_result() below.
1528 if (n->is_Allocate() || n->is_CallStaticJava() &&
1529 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1530 has_allocations = true;
1531 }
1532 if(n->is_AddP()) {
1533 // Collect address nodes which directly reference an allocation.
1534 // Use them during stage 3 below to build initial connection graph
1535 // field edges. Other field edges could be added after StoreP/LoadP
1536 // nodes are processed during stage 4 below.
1537 Node* base = get_addp_base(n);
1538 if(base->is_Proj() && base->in(0)->is_Allocate()) {
1539 cg_worklist.append(n->_idx);
1540 }
1541 } else if (n->is_MergeMem()) {
1542 // Collect all MergeMem nodes to add memory slices for
1543 // scalar replaceable objects in split_unique_types().
1544 _mergemem_worklist.append(n->as_MergeMem());
1545 }
1546 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1547 Node* m = n->fast_out(i); // Get user
1548 worklist_init.push(m);
1549 }
1550 }
1552 if (!has_allocations) {
1553 _collecting = false;
1554 return false; // Nothing to do.
1555 }
1557 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1558 uint delayed_size = _delayed_worklist.size();
1559 for( uint next = 0; next < delayed_size; ++next ) {
1560 Node* n = _delayed_worklist.at(next);
1561 build_connection_graph(n, igvn);
1562 }
1564 // 3. Pass to create fields edges (Allocate -F-> AddP).
1565 uint cg_length = cg_worklist.length();
1566 for( uint next = 0; next < cg_length; ++next ) {
1567 int ni = cg_worklist.at(next);
1568 build_connection_graph(ptnode_adr(ni)->_node, igvn);
1569 }
1571 cg_worklist.clear();
1572 cg_worklist.append(_phantom_object);
1574 // 4. Build Connection Graph which need
1575 // to walk the connection graph.
1576 for (uint ni = 0; ni < nodes_size(); ni++) {
1577 PointsToNode* ptn = ptnode_adr(ni);
1578 Node *n = ptn->_node;
1579 if (n != NULL) { // Call, AddP, LoadP, StoreP
1580 build_connection_graph(n, igvn);
1581 if (ptn->node_type() != PointsToNode::UnknownType)
1582 cg_worklist.append(n->_idx); // Collect CG nodes
1583 }
1584 }
1586 Arena* arena = Thread::current()->resource_area();
1587 VectorSet ptset(arena);
1588 GrowableArray<uint> deferred_edges;
1589 VectorSet visited(arena);
1591 // 5. Remove deferred edges from the graph and adjust
1592 // escape state of nonescaping objects.
1593 cg_length = cg_worklist.length();
1594 for( uint next = 0; next < cg_length; ++next ) {
1595 int ni = cg_worklist.at(next);
1596 PointsToNode* ptn = ptnode_adr(ni);
1597 PointsToNode::NodeType nt = ptn->node_type();
1598 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1599 remove_deferred(ni, &deferred_edges, &visited);
1600 Node *n = ptn->_node;
1601 if (n->is_AddP()) {
1602 // Search for objects which are not scalar replaceable
1603 // and adjust their escape state.
1604 verify_escape_state(ni, ptset, igvn);
1605 }
1606 }
1607 }
1609 // 6. Propagate escape states.
1610 GrowableArray<int> worklist;
1611 bool has_non_escaping_obj = false;
1613 // push all GlobalEscape nodes on the worklist
1614 for( uint next = 0; next < cg_length; ++next ) {
1615 int nk = cg_worklist.at(next);
1616 if (ptnode_adr(nk)->escape_state() == PointsToNode::GlobalEscape)
1617 worklist.push(nk);
1618 }
1619 // mark all nodes reachable from GlobalEscape nodes
1620 while(worklist.length() > 0) {
1621 PointsToNode* ptn = ptnode_adr(worklist.pop());
1622 uint e_cnt = ptn->edge_count();
1623 for (uint ei = 0; ei < e_cnt; ei++) {
1624 uint npi = ptn->edge_target(ei);
1625 PointsToNode *np = ptnode_adr(npi);
1626 if (np->escape_state() < PointsToNode::GlobalEscape) {
1627 np->set_escape_state(PointsToNode::GlobalEscape);
1628 worklist.push(npi);
1629 }
1630 }
1631 }
1633 // push all ArgEscape nodes on the worklist
1634 for( uint next = 0; next < cg_length; ++next ) {
1635 int nk = cg_worklist.at(next);
1636 if (ptnode_adr(nk)->escape_state() == PointsToNode::ArgEscape)
1637 worklist.push(nk);
1638 }
1639 // mark all nodes reachable from ArgEscape nodes
1640 while(worklist.length() > 0) {
1641 PointsToNode* ptn = ptnode_adr(worklist.pop());
1642 if (ptn->node_type() == PointsToNode::JavaObject)
1643 has_non_escaping_obj = true; // Non GlobalEscape
1644 uint e_cnt = ptn->edge_count();
1645 for (uint ei = 0; ei < e_cnt; ei++) {
1646 uint npi = ptn->edge_target(ei);
1647 PointsToNode *np = ptnode_adr(npi);
1648 if (np->escape_state() < PointsToNode::ArgEscape) {
1649 np->set_escape_state(PointsToNode::ArgEscape);
1650 worklist.push(npi);
1651 }
1652 }
1653 }
1655 GrowableArray<Node*> alloc_worklist;
1657 // push all NoEscape nodes on the worklist
1658 for( uint next = 0; next < cg_length; ++next ) {
1659 int nk = cg_worklist.at(next);
1660 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1661 worklist.push(nk);
1662 }
1663 // mark all nodes reachable from NoEscape nodes
1664 while(worklist.length() > 0) {
1665 PointsToNode* ptn = ptnode_adr(worklist.pop());
1666 if (ptn->node_type() == PointsToNode::JavaObject)
1667 has_non_escaping_obj = true; // Non GlobalEscape
1668 Node* n = ptn->_node;
1669 if (n->is_Allocate() && ptn->_scalar_replaceable ) {
1670 // Push scalar replaceable allocations on alloc_worklist
1671 // for processing in split_unique_types().
1672 alloc_worklist.append(n);
1673 }
1674 uint e_cnt = ptn->edge_count();
1675 for (uint ei = 0; ei < e_cnt; ei++) {
1676 uint npi = ptn->edge_target(ei);
1677 PointsToNode *np = ptnode_adr(npi);
1678 if (np->escape_state() < PointsToNode::NoEscape) {
1679 np->set_escape_state(PointsToNode::NoEscape);
1680 worklist.push(npi);
1681 }
1682 }
1683 }
1685 _collecting = false;
1686 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1688 #ifndef PRODUCT
1689 if (PrintEscapeAnalysis) {
1690 dump(); // Dump ConnectionGraph
1691 }
1692 #endif
1694 bool has_scalar_replaceable_candidates = alloc_worklist.length() > 0;
1695 if ( has_scalar_replaceable_candidates &&
1696 C->AliasLevel() >= 3 && EliminateAllocations ) {
1698 // Now use the escape information to create unique types for
1699 // scalar replaceable objects.
1700 split_unique_types(alloc_worklist);
1702 if (C->failing()) return false;
1704 C->print_method("After Escape Analysis", 2);
1706 #ifdef ASSERT
1707 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1708 tty->print("=== No allocations eliminated for ");
1709 C->method()->print_short_name();
1710 if(!EliminateAllocations) {
1711 tty->print(" since EliminateAllocations is off ===");
1712 } else if(!has_scalar_replaceable_candidates) {
1713 tty->print(" since there are no scalar replaceable candidates ===");
1714 } else if(C->AliasLevel() < 3) {
1715 tty->print(" since AliasLevel < 3 ===");
1716 }
1717 tty->cr();
1718 #endif
1719 }
1720 return has_non_escaping_obj;
1721 }
1723 // Search for objects which are not scalar replaceable.
1724 void ConnectionGraph::verify_escape_state(int nidx, VectorSet& ptset, PhaseTransform* phase) {
1725 PointsToNode* ptn = ptnode_adr(nidx);
1726 Node* n = ptn->_node;
1727 assert(n->is_AddP(), "Should be called for AddP nodes only");
1728 // Search for objects which are not scalar replaceable.
1729 // Mark their escape state as ArgEscape to propagate the state
1730 // to referenced objects.
1731 // Note: currently there are no difference in compiler optimizations
1732 // for ArgEscape objects and NoEscape objects which are not
1733 // scalar replaceable.
1735 Compile* C = _compile;
1737 int offset = ptn->offset();
1738 Node* base = get_addp_base(n);
1739 ptset.Clear();
1740 PointsTo(ptset, base);
1741 int ptset_size = ptset.Size();
1743 // Check if a oop field's initializing value is recorded and add
1744 // a corresponding NULL field's value if it is not recorded.
1745 // Connection Graph does not record a default initialization by NULL
1746 // captured by Initialize node.
1747 //
1748 // Note: it will disable scalar replacement in some cases:
1749 //
1750 // Point p[] = new Point[1];
1751 // p[0] = new Point(); // Will be not scalar replaced
1752 //
1753 // but it will save us from incorrect optimizations in next cases:
1754 //
1755 // Point p[] = new Point[1];
1756 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1757 //
1758 // Do a simple control flow analysis to distinguish above cases.
1759 //
1760 if (offset != Type::OffsetBot && ptset_size == 1) {
1761 uint elem = ptset.getelem(); // Allocation node's index
1762 // It does not matter if it is not Allocation node since
1763 // only non-escaping allocations are scalar replaced.
1764 if (ptnode_adr(elem)->_node->is_Allocate() &&
1765 ptnode_adr(elem)->escape_state() == PointsToNode::NoEscape) {
1766 AllocateNode* alloc = ptnode_adr(elem)->_node->as_Allocate();
1767 InitializeNode* ini = alloc->initialization();
1769 // Check only oop fields.
1770 const Type* adr_type = n->as_AddP()->bottom_type();
1771 BasicType basic_field_type = T_INT;
1772 if (adr_type->isa_instptr()) {
1773 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1774 if (field != NULL) {
1775 basic_field_type = field->layout_type();
1776 } else {
1777 // Ignore non field load (for example, klass load)
1778 }
1779 } else if (adr_type->isa_aryptr()) {
1780 const Type* elemtype = adr_type->isa_aryptr()->elem();
1781 basic_field_type = elemtype->array_element_basic_type();
1782 } else {
1783 // Raw pointers are used for initializing stores so skip it.
1784 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1785 (base->in(0) == alloc),"unexpected pointer type");
1786 }
1787 if (basic_field_type == T_OBJECT ||
1788 basic_field_type == T_NARROWOOP ||
1789 basic_field_type == T_ARRAY) {
1790 Node* value = NULL;
1791 if (ini != NULL) {
1792 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1793 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1794 if (store != NULL && store->is_Store()) {
1795 value = store->in(MemNode::ValueIn);
1796 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1797 // Check for a store which follows allocation without branches.
1798 // For example, a volatile field store is not collected
1799 // by Initialize node. TODO: it would be nice to use idom() here.
1800 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1801 store = n->fast_out(i);
1802 if (store->is_Store() && store->in(0) != NULL) {
1803 Node* ctrl = store->in(0);
1804 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1805 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1806 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1807 ctrl = ctrl->in(0);
1808 }
1809 if (ctrl == ini || ctrl == alloc) {
1810 value = store->in(MemNode::ValueIn);
1811 break;
1812 }
1813 }
1814 }
1815 }
1816 }
1817 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1818 // A field's initializing value was not recorded. Add NULL.
1819 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1820 add_pointsto_edge(nidx, null_idx);
1821 }
1822 }
1823 }
1824 }
1826 // An object is not scalar replaceable if the field which may point
1827 // to it has unknown offset (unknown element of an array of objects).
1828 //
1829 if (offset == Type::OffsetBot) {
1830 uint e_cnt = ptn->edge_count();
1831 for (uint ei = 0; ei < e_cnt; ei++) {
1832 uint npi = ptn->edge_target(ei);
1833 set_escape_state(npi, PointsToNode::ArgEscape);
1834 ptnode_adr(npi)->_scalar_replaceable = false;
1835 }
1836 }
1838 // Currently an object is not scalar replaceable if a LoadStore node
1839 // access its field since the field value is unknown after it.
1840 //
1841 bool has_LoadStore = false;
1842 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1843 Node *use = n->fast_out(i);
1844 if (use->is_LoadStore()) {
1845 has_LoadStore = true;
1846 break;
1847 }
1848 }
1849 // An object is not scalar replaceable if the address points
1850 // to unknown field (unknown element for arrays, offset is OffsetBot).
1851 //
1852 // Or the address may point to more then one object. This may produce
1853 // the false positive result (set scalar_replaceable to false)
1854 // since the flow-insensitive escape analysis can't separate
1855 // the case when stores overwrite the field's value from the case
1856 // when stores happened on different control branches.
1857 //
1858 if (ptset_size > 1 || ptset_size != 0 &&
1859 (has_LoadStore || offset == Type::OffsetBot)) {
1860 for( VectorSetI j(&ptset); j.test(); ++j ) {
1861 set_escape_state(j.elem, PointsToNode::ArgEscape);
1862 ptnode_adr(j.elem)->_scalar_replaceable = false;
1863 }
1864 }
1865 }
1867 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
1869 switch (call->Opcode()) {
1870 #ifdef ASSERT
1871 case Op_Allocate:
1872 case Op_AllocateArray:
1873 case Op_Lock:
1874 case Op_Unlock:
1875 assert(false, "should be done already");
1876 break;
1877 #endif
1878 case Op_CallLeaf:
1879 case Op_CallLeafNoFP:
1880 {
1881 // Stub calls, objects do not escape but they are not scale replaceable.
1882 // Adjust escape state for outgoing arguments.
1883 const TypeTuple * d = call->tf()->domain();
1884 VectorSet ptset(Thread::current()->resource_area());
1885 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1886 const Type* at = d->field_at(i);
1887 Node *arg = call->in(i)->uncast();
1888 const Type *aat = phase->type(arg);
1889 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
1890 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
1892 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
1893 aat->isa_ptr() != NULL, "expecting an Ptr");
1894 #ifdef ASSERT
1895 if (!(call->Opcode() == Op_CallLeafNoFP &&
1896 call->as_CallLeaf()->_name != NULL &&
1897 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
1898 call->as_CallLeaf()->_name != NULL &&
1899 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
1900 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
1901 ) {
1902 call->dump();
1903 assert(false, "EA: unexpected CallLeaf");
1904 }
1905 #endif
1906 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1907 if (arg->is_AddP()) {
1908 //
1909 // The inline_native_clone() case when the arraycopy stub is called
1910 // after the allocation before Initialize and CheckCastPP nodes.
1911 //
1912 // Set AddP's base (Allocate) as not scalar replaceable since
1913 // pointer to the base (with offset) is passed as argument.
1914 //
1915 arg = get_addp_base(arg);
1916 }
1917 ptset.Clear();
1918 PointsTo(ptset, arg);
1919 for( VectorSetI j(&ptset); j.test(); ++j ) {
1920 uint pt = j.elem;
1921 set_escape_state(pt, PointsToNode::ArgEscape);
1922 }
1923 }
1924 }
1925 break;
1926 }
1928 case Op_CallStaticJava:
1929 // For a static call, we know exactly what method is being called.
1930 // Use bytecode estimator to record the call's escape affects
1931 {
1932 ciMethod *meth = call->as_CallJava()->method();
1933 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
1934 // fall-through if not a Java method or no analyzer information
1935 if (call_analyzer != NULL) {
1936 const TypeTuple * d = call->tf()->domain();
1937 VectorSet ptset(Thread::current()->resource_area());
1938 bool copy_dependencies = false;
1939 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1940 const Type* at = d->field_at(i);
1941 int k = i - TypeFunc::Parms;
1942 Node *arg = call->in(i)->uncast();
1944 if (at->isa_oopptr() != NULL &&
1945 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
1947 bool global_escapes = false;
1948 bool fields_escapes = false;
1949 if (!call_analyzer->is_arg_stack(k)) {
1950 // The argument global escapes, mark everything it could point to
1951 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1952 global_escapes = true;
1953 } else {
1954 if (!call_analyzer->is_arg_local(k)) {
1955 // The argument itself doesn't escape, but any fields might
1956 fields_escapes = true;
1957 }
1958 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
1959 copy_dependencies = true;
1960 }
1962 ptset.Clear();
1963 PointsTo(ptset, arg);
1964 for( VectorSetI j(&ptset); j.test(); ++j ) {
1965 uint pt = j.elem;
1966 if (global_escapes) {
1967 //The argument global escapes, mark everything it could point to
1968 set_escape_state(pt, PointsToNode::GlobalEscape);
1969 } else {
1970 if (fields_escapes) {
1971 // The argument itself doesn't escape, but any fields might
1972 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
1973 }
1974 set_escape_state(pt, PointsToNode::ArgEscape);
1975 }
1976 }
1977 }
1978 }
1979 if (copy_dependencies)
1980 call_analyzer->copy_dependencies(_compile->dependencies());
1981 break;
1982 }
1983 }
1985 default:
1986 // Fall-through here if not a Java method or no analyzer information
1987 // or some other type of call, assume the worst case: all arguments
1988 // globally escape.
1989 {
1990 // adjust escape state for outgoing arguments
1991 const TypeTuple * d = call->tf()->domain();
1992 VectorSet ptset(Thread::current()->resource_area());
1993 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1994 const Type* at = d->field_at(i);
1995 if (at->isa_oopptr() != NULL) {
1996 Node *arg = call->in(i)->uncast();
1997 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
1998 ptset.Clear();
1999 PointsTo(ptset, arg);
2000 for( VectorSetI j(&ptset); j.test(); ++j ) {
2001 uint pt = j.elem;
2002 set_escape_state(pt, PointsToNode::GlobalEscape);
2003 }
2004 }
2005 }
2006 }
2007 }
2008 }
2009 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2010 CallNode *call = resproj->in(0)->as_Call();
2011 uint call_idx = call->_idx;
2012 uint resproj_idx = resproj->_idx;
2014 switch (call->Opcode()) {
2015 case Op_Allocate:
2016 {
2017 Node *k = call->in(AllocateNode::KlassNode);
2018 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2019 assert(kt != NULL, "TypeKlassPtr required.");
2020 ciKlass* cik = kt->klass();
2022 PointsToNode::EscapeState es;
2023 uint edge_to;
2024 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2025 !cik->is_instance_klass() || // StressReflectiveCode
2026 cik->as_instance_klass()->has_finalizer()) {
2027 es = PointsToNode::GlobalEscape;
2028 edge_to = _phantom_object; // Could not be worse
2029 } else {
2030 es = PointsToNode::NoEscape;
2031 edge_to = call_idx;
2032 }
2033 set_escape_state(call_idx, es);
2034 add_pointsto_edge(resproj_idx, edge_to);
2035 _processed.set(resproj_idx);
2036 break;
2037 }
2039 case Op_AllocateArray:
2040 {
2042 Node *k = call->in(AllocateNode::KlassNode);
2043 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2044 assert(kt != NULL, "TypeKlassPtr required.");
2045 ciKlass* cik = kt->klass();
2047 PointsToNode::EscapeState es;
2048 uint edge_to;
2049 if (!cik->is_array_klass()) { // StressReflectiveCode
2050 es = PointsToNode::GlobalEscape;
2051 edge_to = _phantom_object;
2052 } else {
2053 es = PointsToNode::NoEscape;
2054 edge_to = call_idx;
2055 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2056 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2057 // Not scalar replaceable if the length is not constant or too big.
2058 ptnode_adr(call_idx)->_scalar_replaceable = false;
2059 }
2060 }
2061 set_escape_state(call_idx, es);
2062 add_pointsto_edge(resproj_idx, edge_to);
2063 _processed.set(resproj_idx);
2064 break;
2065 }
2067 case Op_CallStaticJava:
2068 // For a static call, we know exactly what method is being called.
2069 // Use bytecode estimator to record whether the call's return value escapes
2070 {
2071 bool done = true;
2072 const TypeTuple *r = call->tf()->range();
2073 const Type* ret_type = NULL;
2075 if (r->cnt() > TypeFunc::Parms)
2076 ret_type = r->field_at(TypeFunc::Parms);
2078 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2079 // _multianewarray functions return a TypeRawPtr.
2080 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2081 _processed.set(resproj_idx);
2082 break; // doesn't return a pointer type
2083 }
2084 ciMethod *meth = call->as_CallJava()->method();
2085 const TypeTuple * d = call->tf()->domain();
2086 if (meth == NULL) {
2087 // not a Java method, assume global escape
2088 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2089 add_pointsto_edge(resproj_idx, _phantom_object);
2090 } else {
2091 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2092 bool copy_dependencies = false;
2094 if (call_analyzer->is_return_allocated()) {
2095 // Returns a newly allocated unescaped object, simply
2096 // update dependency information.
2097 // Mark it as NoEscape so that objects referenced by
2098 // it's fields will be marked as NoEscape at least.
2099 set_escape_state(call_idx, PointsToNode::NoEscape);
2100 add_pointsto_edge(resproj_idx, call_idx);
2101 copy_dependencies = true;
2102 } else if (call_analyzer->is_return_local()) {
2103 // determine whether any arguments are returned
2104 set_escape_state(call_idx, PointsToNode::NoEscape);
2105 bool ret_arg = false;
2106 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2107 const Type* at = d->field_at(i);
2109 if (at->isa_oopptr() != NULL) {
2110 Node *arg = call->in(i)->uncast();
2112 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2113 ret_arg = true;
2114 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2115 if (arg_esp->node_type() == PointsToNode::UnknownType)
2116 done = false;
2117 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2118 add_pointsto_edge(resproj_idx, arg->_idx);
2119 else
2120 add_deferred_edge(resproj_idx, arg->_idx);
2121 arg_esp->_hidden_alias = true;
2122 }
2123 }
2124 }
2125 if (done && !ret_arg) {
2126 // Returns unknown object.
2127 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2128 add_pointsto_edge(resproj_idx, _phantom_object);
2129 }
2130 copy_dependencies = true;
2131 } else {
2132 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2133 add_pointsto_edge(resproj_idx, _phantom_object);
2134 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2135 const Type* at = d->field_at(i);
2136 if (at->isa_oopptr() != NULL) {
2137 Node *arg = call->in(i)->uncast();
2138 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2139 arg_esp->_hidden_alias = true;
2140 }
2141 }
2142 }
2143 if (copy_dependencies)
2144 call_analyzer->copy_dependencies(_compile->dependencies());
2145 }
2146 if (done)
2147 _processed.set(resproj_idx);
2148 break;
2149 }
2151 default:
2152 // Some other type of call, assume the worst case that the
2153 // returned value, if any, globally escapes.
2154 {
2155 const TypeTuple *r = call->tf()->range();
2156 if (r->cnt() > TypeFunc::Parms) {
2157 const Type* ret_type = r->field_at(TypeFunc::Parms);
2159 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2160 // _multianewarray functions return a TypeRawPtr.
2161 if (ret_type->isa_ptr() != NULL) {
2162 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2163 add_pointsto_edge(resproj_idx, _phantom_object);
2164 }
2165 }
2166 _processed.set(resproj_idx);
2167 }
2168 }
2169 }
2171 // Populate Connection Graph with Ideal nodes and create simple
2172 // connection graph edges (do not need to check the node_type of inputs
2173 // or to call PointsTo() to walk the connection graph).
2174 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2175 if (_processed.test(n->_idx))
2176 return; // No need to redefine node's state.
2178 if (n->is_Call()) {
2179 // Arguments to allocation and locking don't escape.
2180 if (n->is_Allocate()) {
2181 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2182 record_for_optimizer(n);
2183 } else if (n->is_Lock() || n->is_Unlock()) {
2184 // Put Lock and Unlock nodes on IGVN worklist to process them during
2185 // the first IGVN optimization when escape information is still available.
2186 record_for_optimizer(n);
2187 _processed.set(n->_idx);
2188 } else {
2189 // Don't mark as processed since call's arguments have to be processed.
2190 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2191 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2193 // Check if a call returns an object.
2194 const TypeTuple *r = n->as_Call()->tf()->range();
2195 if (r->cnt() > TypeFunc::Parms &&
2196 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2197 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2198 nt = PointsToNode::JavaObject;
2199 if (!n->is_CallStaticJava()) {
2200 // Since the called mathod is statically unknown assume
2201 // the worst case that the returned value globally escapes.
2202 es = PointsToNode::GlobalEscape;
2203 }
2204 }
2205 add_node(n, nt, es, false);
2206 }
2207 return;
2208 }
2210 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2211 // ThreadLocal has RawPrt type.
2212 switch (n->Opcode()) {
2213 case Op_AddP:
2214 {
2215 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2216 break;
2217 }
2218 case Op_CastX2P:
2219 { // "Unsafe" memory access.
2220 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2221 break;
2222 }
2223 case Op_CastPP:
2224 case Op_CheckCastPP:
2225 case Op_EncodeP:
2226 case Op_DecodeN:
2227 {
2228 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2229 int ti = n->in(1)->_idx;
2230 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2231 if (nt == PointsToNode::UnknownType) {
2232 _delayed_worklist.push(n); // Process it later.
2233 break;
2234 } else if (nt == PointsToNode::JavaObject) {
2235 add_pointsto_edge(n->_idx, ti);
2236 } else {
2237 add_deferred_edge(n->_idx, ti);
2238 }
2239 _processed.set(n->_idx);
2240 break;
2241 }
2242 case Op_ConP:
2243 {
2244 // assume all pointer constants globally escape except for null
2245 PointsToNode::EscapeState es;
2246 if (phase->type(n) == TypePtr::NULL_PTR)
2247 es = PointsToNode::NoEscape;
2248 else
2249 es = PointsToNode::GlobalEscape;
2251 add_node(n, PointsToNode::JavaObject, es, true);
2252 break;
2253 }
2254 case Op_ConN:
2255 {
2256 // assume all narrow oop constants globally escape except for null
2257 PointsToNode::EscapeState es;
2258 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2259 es = PointsToNode::NoEscape;
2260 else
2261 es = PointsToNode::GlobalEscape;
2263 add_node(n, PointsToNode::JavaObject, es, true);
2264 break;
2265 }
2266 case Op_CreateEx:
2267 {
2268 // assume that all exception objects globally escape
2269 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2270 break;
2271 }
2272 case Op_LoadKlass:
2273 case Op_LoadNKlass:
2274 {
2275 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2276 break;
2277 }
2278 case Op_LoadP:
2279 case Op_LoadN:
2280 {
2281 const Type *t = phase->type(n);
2282 if (t->make_ptr() == NULL) {
2283 _processed.set(n->_idx);
2284 return;
2285 }
2286 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2287 break;
2288 }
2289 case Op_Parm:
2290 {
2291 _processed.set(n->_idx); // No need to redefine it state.
2292 uint con = n->as_Proj()->_con;
2293 if (con < TypeFunc::Parms)
2294 return;
2295 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2296 if (t->isa_ptr() == NULL)
2297 return;
2298 // We have to assume all input parameters globally escape
2299 // (Note: passing 'false' since _processed is already set).
2300 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2301 break;
2302 }
2303 case Op_Phi:
2304 {
2305 const Type *t = n->as_Phi()->type();
2306 if (t->make_ptr() == NULL) {
2307 // nothing to do if not an oop or narrow oop
2308 _processed.set(n->_idx);
2309 return;
2310 }
2311 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2312 uint i;
2313 for (i = 1; i < n->req() ; i++) {
2314 Node* in = n->in(i);
2315 if (in == NULL)
2316 continue; // ignore NULL
2317 in = in->uncast();
2318 if (in->is_top() || in == n)
2319 continue; // ignore top or inputs which go back this node
2320 int ti = in->_idx;
2321 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2322 if (nt == PointsToNode::UnknownType) {
2323 break;
2324 } else if (nt == PointsToNode::JavaObject) {
2325 add_pointsto_edge(n->_idx, ti);
2326 } else {
2327 add_deferred_edge(n->_idx, ti);
2328 }
2329 }
2330 if (i >= n->req())
2331 _processed.set(n->_idx);
2332 else
2333 _delayed_worklist.push(n);
2334 break;
2335 }
2336 case Op_Proj:
2337 {
2338 // we are only interested in the oop result projection from a call
2339 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2340 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2341 assert(r->cnt() > TypeFunc::Parms, "sanity");
2342 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2343 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2344 int ti = n->in(0)->_idx;
2345 // The call may not be registered yet (since not all its inputs are registered)
2346 // if this is the projection from backbranch edge of Phi.
2347 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2348 process_call_result(n->as_Proj(), phase);
2349 }
2350 if (!_processed.test(n->_idx)) {
2351 // The call's result may need to be processed later if the call
2352 // returns it's argument and the argument is not processed yet.
2353 _delayed_worklist.push(n);
2354 }
2355 break;
2356 }
2357 }
2358 _processed.set(n->_idx);
2359 break;
2360 }
2361 case Op_Return:
2362 {
2363 if( n->req() > TypeFunc::Parms &&
2364 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2365 // Treat Return value as LocalVar with GlobalEscape escape state.
2366 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2367 int ti = n->in(TypeFunc::Parms)->_idx;
2368 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2369 if (nt == PointsToNode::UnknownType) {
2370 _delayed_worklist.push(n); // Process it later.
2371 break;
2372 } else if (nt == PointsToNode::JavaObject) {
2373 add_pointsto_edge(n->_idx, ti);
2374 } else {
2375 add_deferred_edge(n->_idx, ti);
2376 }
2377 }
2378 _processed.set(n->_idx);
2379 break;
2380 }
2381 case Op_StoreP:
2382 case Op_StoreN:
2383 {
2384 const Type *adr_type = phase->type(n->in(MemNode::Address));
2385 adr_type = adr_type->make_ptr();
2386 if (adr_type->isa_oopptr()) {
2387 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2388 } else {
2389 Node* adr = n->in(MemNode::Address);
2390 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2391 adr->in(AddPNode::Address)->is_Proj() &&
2392 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2393 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2394 // We are computing a raw address for a store captured
2395 // by an Initialize compute an appropriate address type.
2396 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2397 assert(offs != Type::OffsetBot, "offset must be a constant");
2398 } else {
2399 _processed.set(n->_idx);
2400 return;
2401 }
2402 }
2403 break;
2404 }
2405 case Op_StorePConditional:
2406 case Op_CompareAndSwapP:
2407 case Op_CompareAndSwapN:
2408 {
2409 const Type *adr_type = phase->type(n->in(MemNode::Address));
2410 adr_type = adr_type->make_ptr();
2411 if (adr_type->isa_oopptr()) {
2412 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2413 } else {
2414 _processed.set(n->_idx);
2415 return;
2416 }
2417 break;
2418 }
2419 case Op_AryEq:
2420 case Op_StrComp:
2421 case Op_StrEquals:
2422 case Op_StrIndexOf:
2423 {
2424 // char[] arrays passed to string intrinsics are not scalar replaceable.
2425 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2426 break;
2427 }
2428 case Op_ThreadLocal:
2429 {
2430 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2431 break;
2432 }
2433 default:
2434 ;
2435 // nothing to do
2436 }
2437 return;
2438 }
2440 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2441 uint n_idx = n->_idx;
2442 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2444 // Don't set processed bit for AddP, LoadP, StoreP since
2445 // they may need more then one pass to process.
2446 if (_processed.test(n_idx))
2447 return; // No need to redefine node's state.
2449 if (n->is_Call()) {
2450 CallNode *call = n->as_Call();
2451 process_call_arguments(call, phase);
2452 _processed.set(n_idx);
2453 return;
2454 }
2456 switch (n->Opcode()) {
2457 case Op_AddP:
2458 {
2459 Node *base = get_addp_base(n);
2460 // Create a field edge to this node from everything base could point to.
2461 VectorSet ptset(Thread::current()->resource_area());
2462 PointsTo(ptset, base);
2463 for( VectorSetI i(&ptset); i.test(); ++i ) {
2464 uint pt = i.elem;
2465 add_field_edge(pt, n_idx, address_offset(n, phase));
2466 }
2467 break;
2468 }
2469 case Op_CastX2P:
2470 {
2471 assert(false, "Op_CastX2P");
2472 break;
2473 }
2474 case Op_CastPP:
2475 case Op_CheckCastPP:
2476 case Op_EncodeP:
2477 case Op_DecodeN:
2478 {
2479 int ti = n->in(1)->_idx;
2480 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2481 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2482 add_pointsto_edge(n_idx, ti);
2483 } else {
2484 add_deferred_edge(n_idx, ti);
2485 }
2486 _processed.set(n_idx);
2487 break;
2488 }
2489 case Op_ConP:
2490 {
2491 assert(false, "Op_ConP");
2492 break;
2493 }
2494 case Op_ConN:
2495 {
2496 assert(false, "Op_ConN");
2497 break;
2498 }
2499 case Op_CreateEx:
2500 {
2501 assert(false, "Op_CreateEx");
2502 break;
2503 }
2504 case Op_LoadKlass:
2505 case Op_LoadNKlass:
2506 {
2507 assert(false, "Op_LoadKlass");
2508 break;
2509 }
2510 case Op_LoadP:
2511 case Op_LoadN:
2512 {
2513 const Type *t = phase->type(n);
2514 #ifdef ASSERT
2515 if (t->make_ptr() == NULL)
2516 assert(false, "Op_LoadP");
2517 #endif
2519 Node* adr = n->in(MemNode::Address)->uncast();
2520 Node* adr_base;
2521 if (adr->is_AddP()) {
2522 adr_base = get_addp_base(adr);
2523 } else {
2524 adr_base = adr;
2525 }
2527 // For everything "adr_base" could point to, create a deferred edge from
2528 // this node to each field with the same offset.
2529 VectorSet ptset(Thread::current()->resource_area());
2530 PointsTo(ptset, adr_base);
2531 int offset = address_offset(adr, phase);
2532 for( VectorSetI i(&ptset); i.test(); ++i ) {
2533 uint pt = i.elem;
2534 add_deferred_edge_to_fields(n_idx, pt, offset);
2535 }
2536 break;
2537 }
2538 case Op_Parm:
2539 {
2540 assert(false, "Op_Parm");
2541 break;
2542 }
2543 case Op_Phi:
2544 {
2545 #ifdef ASSERT
2546 const Type *t = n->as_Phi()->type();
2547 if (t->make_ptr() == NULL)
2548 assert(false, "Op_Phi");
2549 #endif
2550 for (uint i = 1; i < n->req() ; i++) {
2551 Node* in = n->in(i);
2552 if (in == NULL)
2553 continue; // ignore NULL
2554 in = in->uncast();
2555 if (in->is_top() || in == n)
2556 continue; // ignore top or inputs which go back this node
2557 int ti = in->_idx;
2558 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2559 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2560 if (nt == PointsToNode::JavaObject) {
2561 add_pointsto_edge(n_idx, ti);
2562 } else {
2563 add_deferred_edge(n_idx, ti);
2564 }
2565 }
2566 _processed.set(n_idx);
2567 break;
2568 }
2569 case Op_Proj:
2570 {
2571 // we are only interested in the oop result projection from a call
2572 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2573 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2574 "all nodes should be registered");
2575 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2576 assert(r->cnt() > TypeFunc::Parms, "sanity");
2577 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2578 process_call_result(n->as_Proj(), phase);
2579 assert(_processed.test(n_idx), "all call results should be processed");
2580 break;
2581 }
2582 }
2583 assert(false, "Op_Proj");
2584 break;
2585 }
2586 case Op_Return:
2587 {
2588 #ifdef ASSERT
2589 if( n->req() <= TypeFunc::Parms ||
2590 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2591 assert(false, "Op_Return");
2592 }
2593 #endif
2594 int ti = n->in(TypeFunc::Parms)->_idx;
2595 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2596 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2597 add_pointsto_edge(n_idx, ti);
2598 } else {
2599 add_deferred_edge(n_idx, ti);
2600 }
2601 _processed.set(n_idx);
2602 break;
2603 }
2604 case Op_StoreP:
2605 case Op_StoreN:
2606 case Op_StorePConditional:
2607 case Op_CompareAndSwapP:
2608 case Op_CompareAndSwapN:
2609 {
2610 Node *adr = n->in(MemNode::Address);
2611 const Type *adr_type = phase->type(adr)->make_ptr();
2612 #ifdef ASSERT
2613 if (!adr_type->isa_oopptr())
2614 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2615 #endif
2617 assert(adr->is_AddP(), "expecting an AddP");
2618 Node *adr_base = get_addp_base(adr);
2619 Node *val = n->in(MemNode::ValueIn)->uncast();
2620 // For everything "adr_base" could point to, create a deferred edge
2621 // to "val" from each field with the same offset.
2622 VectorSet ptset(Thread::current()->resource_area());
2623 PointsTo(ptset, adr_base);
2624 for( VectorSetI i(&ptset); i.test(); ++i ) {
2625 uint pt = i.elem;
2626 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2627 }
2628 break;
2629 }
2630 case Op_AryEq:
2631 case Op_StrComp:
2632 case Op_StrEquals:
2633 case Op_StrIndexOf:
2634 {
2635 // char[] arrays passed to string intrinsic do not escape but
2636 // they are not scalar replaceable. Adjust escape state for them.
2637 // Start from in(2) edge since in(1) is memory edge.
2638 for (uint i = 2; i < n->req(); i++) {
2639 Node* adr = n->in(i)->uncast();
2640 const Type *at = phase->type(adr);
2641 if (!adr->is_top() && at->isa_ptr()) {
2642 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2643 at->isa_ptr() != NULL, "expecting an Ptr");
2644 if (adr->is_AddP()) {
2645 adr = get_addp_base(adr);
2646 }
2647 // Mark as ArgEscape everything "adr" could point to.
2648 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2649 }
2650 }
2651 _processed.set(n_idx);
2652 break;
2653 }
2654 case Op_ThreadLocal:
2655 {
2656 assert(false, "Op_ThreadLocal");
2657 break;
2658 }
2659 default:
2660 // This method should be called only for EA specific nodes.
2661 ShouldNotReachHere();
2662 }
2663 }
2665 #ifndef PRODUCT
2666 void ConnectionGraph::dump() {
2667 bool first = true;
2669 uint size = nodes_size();
2670 for (uint ni = 0; ni < size; ni++) {
2671 PointsToNode *ptn = ptnode_adr(ni);
2672 PointsToNode::NodeType ptn_type = ptn->node_type();
2674 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2675 continue;
2676 PointsToNode::EscapeState es = escape_state(ptn->_node);
2677 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2678 if (first) {
2679 tty->cr();
2680 tty->print("======== Connection graph for ");
2681 _compile->method()->print_short_name();
2682 tty->cr();
2683 first = false;
2684 }
2685 tty->print("%6d ", ni);
2686 ptn->dump();
2687 // Print all locals which reference this allocation
2688 for (uint li = ni; li < size; li++) {
2689 PointsToNode *ptn_loc = ptnode_adr(li);
2690 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2691 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2692 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2693 ptnode_adr(li)->dump(false);
2694 }
2695 }
2696 if (Verbose) {
2697 // Print all fields which reference this allocation
2698 for (uint i = 0; i < ptn->edge_count(); i++) {
2699 uint ei = ptn->edge_target(i);
2700 ptnode_adr(ei)->dump(false);
2701 }
2702 }
2703 tty->cr();
2704 }
2705 }
2706 }
2707 #endif