Mon, 07 Nov 2011 14:33:57 -0800
7059047: EA: can't find initializing store with several CheckCastPP
Summary: Split adjust_escape_state() method into two methods to find initializing stores.
Reviewed-by: never
1 /*
2 * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "ci/bcEscapeAnalyzer.hpp"
27 #include "libadt/vectset.hpp"
28 #include "memory/allocation.hpp"
29 #include "opto/c2compiler.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/escape.hpp"
34 #include "opto/phaseX.hpp"
35 #include "opto/rootnode.hpp"
37 void PointsToNode::add_edge(uint targIdx, PointsToNode::EdgeType et) {
38 uint v = (targIdx << EdgeShift) + ((uint) et);
39 if (_edges == NULL) {
40 Arena *a = Compile::current()->comp_arena();
41 _edges = new(a) GrowableArray<uint>(a, INITIAL_EDGE_COUNT, 0, 0);
42 }
43 _edges->append_if_missing(v);
44 }
46 void PointsToNode::remove_edge(uint targIdx, PointsToNode::EdgeType et) {
47 uint v = (targIdx << EdgeShift) + ((uint) et);
49 _edges->remove(v);
50 }
52 #ifndef PRODUCT
53 static const char *node_type_names[] = {
54 "UnknownType",
55 "JavaObject",
56 "LocalVar",
57 "Field"
58 };
60 static const char *esc_names[] = {
61 "UnknownEscape",
62 "NoEscape",
63 "ArgEscape",
64 "GlobalEscape"
65 };
67 static const char *edge_type_suffix[] = {
68 "?", // UnknownEdge
69 "P", // PointsToEdge
70 "D", // DeferredEdge
71 "F" // FieldEdge
72 };
74 void PointsToNode::dump(bool print_state) const {
75 NodeType nt = node_type();
76 tty->print("%s ", node_type_names[(int) nt]);
77 if (print_state) {
78 EscapeState es = escape_state();
79 tty->print("%s %s ", esc_names[(int) es], _scalar_replaceable ? "":"NSR");
80 }
81 tty->print("[[");
82 for (uint i = 0; i < edge_count(); i++) {
83 tty->print(" %d%s", edge_target(i), edge_type_suffix[(int) edge_type(i)]);
84 }
85 tty->print("]] ");
86 if (_node == NULL)
87 tty->print_cr("<null>");
88 else
89 _node->dump();
90 }
91 #endif
93 ConnectionGraph::ConnectionGraph(Compile * C, PhaseIterGVN *igvn) :
94 _nodes(C->comp_arena(), C->unique(), C->unique(), PointsToNode()),
95 _processed(C->comp_arena()),
96 pt_ptset(C->comp_arena()),
97 pt_visited(C->comp_arena()),
98 pt_worklist(C->comp_arena(), 4, 0, 0),
99 _collecting(true),
100 _progress(false),
101 _compile(C),
102 _igvn(igvn),
103 _node_map(C->comp_arena()) {
105 _phantom_object = C->top()->_idx,
106 add_node(C->top(), PointsToNode::JavaObject, PointsToNode::GlobalEscape,true);
108 // Add ConP(#NULL) and ConN(#NULL) nodes.
109 Node* oop_null = igvn->zerocon(T_OBJECT);
110 _oop_null = oop_null->_idx;
111 assert(_oop_null < nodes_size(), "should be created already");
112 add_node(oop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
114 if (UseCompressedOops) {
115 Node* noop_null = igvn->zerocon(T_NARROWOOP);
116 _noop_null = noop_null->_idx;
117 assert(_noop_null < nodes_size(), "should be created already");
118 add_node(noop_null, PointsToNode::JavaObject, PointsToNode::NoEscape, true);
119 } else {
120 _noop_null = _oop_null; // Should be initialized
121 }
122 }
124 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
125 PointsToNode *f = ptnode_adr(from_i);
126 PointsToNode *t = ptnode_adr(to_i);
128 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
129 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
130 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
131 add_edge(f, to_i, PointsToNode::PointsToEdge);
132 }
134 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
135 PointsToNode *f = ptnode_adr(from_i);
136 PointsToNode *t = ptnode_adr(to_i);
138 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
139 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
140 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
141 // don't add a self-referential edge, this can occur during removal of
142 // deferred edges
143 if (from_i != to_i)
144 add_edge(f, to_i, PointsToNode::DeferredEdge);
145 }
147 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
148 const Type *adr_type = phase->type(adr);
149 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
150 adr->in(AddPNode::Address)->is_Proj() &&
151 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
152 // We are computing a raw address for a store captured by an Initialize
153 // compute an appropriate address type. AddP cases #3 and #5 (see below).
154 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
155 assert(offs != Type::OffsetBot ||
156 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
157 "offset must be a constant or it is initialization of array");
158 return offs;
159 }
160 const TypePtr *t_ptr = adr_type->isa_ptr();
161 assert(t_ptr != NULL, "must be a pointer type");
162 return t_ptr->offset();
163 }
165 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
166 PointsToNode *f = ptnode_adr(from_i);
167 PointsToNode *t = ptnode_adr(to_i);
169 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
170 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
171 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
172 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
173 t->set_offset(offset);
175 add_edge(f, to_i, PointsToNode::FieldEdge);
176 }
178 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
179 // Don't change non-escaping state of NULL pointer.
180 if (ni == _noop_null || ni == _oop_null)
181 return;
182 PointsToNode *npt = ptnode_adr(ni);
183 PointsToNode::EscapeState old_es = npt->escape_state();
184 if (es > old_es)
185 npt->set_escape_state(es);
186 }
188 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
189 PointsToNode::EscapeState es, bool done) {
190 PointsToNode* ptadr = ptnode_adr(n->_idx);
191 ptadr->_node = n;
192 ptadr->set_node_type(nt);
194 // inline set_escape_state(idx, es);
195 PointsToNode::EscapeState old_es = ptadr->escape_state();
196 if (es > old_es)
197 ptadr->set_escape_state(es);
199 if (done)
200 _processed.set(n->_idx);
201 }
203 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
204 uint idx = n->_idx;
205 PointsToNode::EscapeState es;
207 // If we are still collecting or there were no non-escaping allocations
208 // we don't know the answer yet
209 if (_collecting)
210 return PointsToNode::UnknownEscape;
212 // if the node was created after the escape computation, return
213 // UnknownEscape
214 if (idx >= nodes_size())
215 return PointsToNode::UnknownEscape;
217 es = ptnode_adr(idx)->escape_state();
219 // if we have already computed a value, return it
220 if (es != PointsToNode::UnknownEscape &&
221 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
222 return es;
224 // PointsTo() calls n->uncast() which can return a new ideal node.
225 if (n->uncast()->_idx >= nodes_size())
226 return PointsToNode::UnknownEscape;
228 PointsToNode::EscapeState orig_es = es;
230 // compute max escape state of anything this node could point to
231 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) {
232 uint pt = i.elem;
233 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
234 if (pes > es)
235 es = pes;
236 }
237 if (orig_es != es) {
238 // cache the computed escape state
239 assert(es > orig_es, "should have computed an escape state");
240 set_escape_state(idx, es);
241 } // orig_es could be PointsToNode::UnknownEscape
242 return es;
243 }
245 VectorSet* ConnectionGraph::PointsTo(Node * n) {
246 pt_ptset.Reset();
247 pt_visited.Reset();
248 pt_worklist.clear();
250 #ifdef ASSERT
251 Node *orig_n = n;
252 #endif
254 n = n->uncast();
255 PointsToNode* npt = ptnode_adr(n->_idx);
257 // If we have a JavaObject, return just that object
258 if (npt->node_type() == PointsToNode::JavaObject) {
259 pt_ptset.set(n->_idx);
260 return &pt_ptset;
261 }
262 #ifdef ASSERT
263 if (npt->_node == NULL) {
264 if (orig_n != n)
265 orig_n->dump();
266 n->dump();
267 assert(npt->_node != NULL, "unregistered node");
268 }
269 #endif
270 pt_worklist.push(n->_idx);
271 while(pt_worklist.length() > 0) {
272 int ni = pt_worklist.pop();
273 if (pt_visited.test_set(ni))
274 continue;
276 PointsToNode* pn = ptnode_adr(ni);
277 // ensure that all inputs of a Phi have been processed
278 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
280 int edges_processed = 0;
281 uint e_cnt = pn->edge_count();
282 for (uint e = 0; e < e_cnt; e++) {
283 uint etgt = pn->edge_target(e);
284 PointsToNode::EdgeType et = pn->edge_type(e);
285 if (et == PointsToNode::PointsToEdge) {
286 pt_ptset.set(etgt);
287 edges_processed++;
288 } else if (et == PointsToNode::DeferredEdge) {
289 pt_worklist.push(etgt);
290 edges_processed++;
291 } else {
292 assert(false,"neither PointsToEdge or DeferredEdge");
293 }
294 }
295 if (edges_processed == 0) {
296 // no deferred or pointsto edges found. Assume the value was set
297 // outside this method. Add the phantom object to the pointsto set.
298 pt_ptset.set(_phantom_object);
299 }
300 }
301 return &pt_ptset;
302 }
304 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
305 // This method is most expensive during ConnectionGraph construction.
306 // Reuse vectorSet and an additional growable array for deferred edges.
307 deferred_edges->clear();
308 visited->Reset();
310 visited->set(ni);
311 PointsToNode *ptn = ptnode_adr(ni);
313 // Mark current edges as visited and move deferred edges to separate array.
314 for (uint i = 0; i < ptn->edge_count(); ) {
315 uint t = ptn->edge_target(i);
316 #ifdef ASSERT
317 assert(!visited->test_set(t), "expecting no duplications");
318 #else
319 visited->set(t);
320 #endif
321 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
322 ptn->remove_edge(t, PointsToNode::DeferredEdge);
323 deferred_edges->append(t);
324 } else {
325 i++;
326 }
327 }
328 for (int next = 0; next < deferred_edges->length(); ++next) {
329 uint t = deferred_edges->at(next);
330 PointsToNode *ptt = ptnode_adr(t);
331 uint e_cnt = ptt->edge_count();
332 for (uint e = 0; e < e_cnt; e++) {
333 uint etgt = ptt->edge_target(e);
334 if (visited->test_set(etgt))
335 continue;
337 PointsToNode::EdgeType et = ptt->edge_type(e);
338 if (et == PointsToNode::PointsToEdge) {
339 add_pointsto_edge(ni, etgt);
340 if(etgt == _phantom_object) {
341 // Special case - field set outside (globally escaping).
342 set_escape_state(ni, PointsToNode::GlobalEscape);
343 }
344 } else if (et == PointsToNode::DeferredEdge) {
345 deferred_edges->append(etgt);
346 } else {
347 assert(false,"invalid connection graph");
348 }
349 }
350 }
351 }
354 // Add an edge to node given by "to_i" from any field of adr_i whose offset
355 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
356 // a pointsto edge is added if it is a JavaObject
358 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
359 PointsToNode* an = ptnode_adr(adr_i);
360 PointsToNode* to = ptnode_adr(to_i);
361 bool deferred = (to->node_type() == PointsToNode::LocalVar);
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 (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
369 if (deferred)
370 add_deferred_edge(fi, to_i);
371 else
372 add_pointsto_edge(fi, to_i);
373 }
374 }
375 }
377 // Add a deferred edge from node given by "from_i" to any field of adr_i
378 // whose offset matches "offset".
379 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
380 PointsToNode* an = ptnode_adr(adr_i);
381 bool is_alloc = an->_node->is_Allocate();
382 for (uint fe = 0; fe < an->edge_count(); fe++) {
383 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
384 int fi = an->edge_target(fe);
385 PointsToNode* pf = ptnode_adr(fi);
386 int offset = pf->offset();
387 if (!is_alloc) {
388 // Assume the field was set outside this method if it is not Allocation
389 add_pointsto_edge(fi, _phantom_object);
390 }
391 if (offset == offs || offset == Type::OffsetBot || offs == Type::OffsetBot) {
392 add_deferred_edge(from_i, fi);
393 }
394 }
395 }
397 // Helper functions
399 static Node* get_addp_base(Node *addp) {
400 assert(addp->is_AddP(), "must be AddP");
401 //
402 // AddP cases for Base and Address inputs:
403 // case #1. Direct object's field reference:
404 // Allocate
405 // |
406 // Proj #5 ( oop result )
407 // |
408 // CheckCastPP (cast to instance type)
409 // | |
410 // AddP ( base == address )
411 //
412 // case #2. Indirect object's field reference:
413 // Phi
414 // |
415 // CastPP (cast to instance type)
416 // | |
417 // AddP ( base == address )
418 //
419 // case #3. Raw object's field reference for Initialize node:
420 // Allocate
421 // |
422 // Proj #5 ( oop result )
423 // top |
424 // \ |
425 // AddP ( base == top )
426 //
427 // case #4. Array's element reference:
428 // {CheckCastPP | CastPP}
429 // | | |
430 // | AddP ( array's element offset )
431 // | |
432 // AddP ( array's offset )
433 //
434 // case #5. Raw object's field reference for arraycopy stub call:
435 // The inline_native_clone() case when the arraycopy stub is called
436 // after the allocation before Initialize and CheckCastPP nodes.
437 // Allocate
438 // |
439 // Proj #5 ( oop result )
440 // | |
441 // AddP ( base == address )
442 //
443 // case #6. Constant Pool, ThreadLocal, CastX2P or
444 // Raw object's field reference:
445 // {ConP, ThreadLocal, CastX2P, raw Load}
446 // top |
447 // \ |
448 // AddP ( base == top )
449 //
450 // case #7. Klass's field reference.
451 // LoadKlass
452 // | |
453 // AddP ( base == address )
454 //
455 // case #8. narrow Klass's field reference.
456 // LoadNKlass
457 // |
458 // DecodeN
459 // | |
460 // AddP ( base == address )
461 //
462 Node *base = addp->in(AddPNode::Base)->uncast();
463 if (base->is_top()) { // The AddP case #3 and #6.
464 base = addp->in(AddPNode::Address)->uncast();
465 while (base->is_AddP()) {
466 // Case #6 (unsafe access) may have several chained AddP nodes.
467 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
468 base = base->in(AddPNode::Address)->uncast();
469 }
470 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
471 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
472 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
473 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
474 }
475 return base;
476 }
478 static Node* find_second_addp(Node* addp, Node* n) {
479 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
481 Node* addp2 = addp->raw_out(0);
482 if (addp->outcnt() == 1 && addp2->is_AddP() &&
483 addp2->in(AddPNode::Base) == n &&
484 addp2->in(AddPNode::Address) == addp) {
486 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
487 //
488 // Find array's offset to push it on worklist first and
489 // as result process an array's element offset first (pushed second)
490 // to avoid CastPP for the array's offset.
491 // Otherwise the inserted CastPP (LocalVar) will point to what
492 // the AddP (Field) points to. Which would be wrong since
493 // the algorithm expects the CastPP has the same point as
494 // as AddP's base CheckCastPP (LocalVar).
495 //
496 // ArrayAllocation
497 // |
498 // CheckCastPP
499 // |
500 // memProj (from ArrayAllocation CheckCastPP)
501 // | ||
502 // | || Int (element index)
503 // | || | ConI (log(element size))
504 // | || | /
505 // | || LShift
506 // | || /
507 // | AddP (array's element offset)
508 // | |
509 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
510 // | / /
511 // AddP (array's offset)
512 // |
513 // Load/Store (memory operation on array's element)
514 //
515 return addp2;
516 }
517 return NULL;
518 }
520 //
521 // Adjust the type and inputs of an AddP which computes the
522 // address of a field of an instance
523 //
524 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
525 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
526 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
527 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
528 if (t == NULL) {
529 // We are computing a raw address for a store captured by an Initialize
530 // compute an appropriate address type (cases #3 and #5).
531 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
532 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
533 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
534 assert(offs != Type::OffsetBot, "offset must be a constant");
535 t = base_t->add_offset(offs)->is_oopptr();
536 }
537 int inst_id = base_t->instance_id();
538 assert(!t->is_known_instance() || t->instance_id() == inst_id,
539 "old type must be non-instance or match new type");
541 // The type 't' could be subclass of 'base_t'.
542 // As result t->offset() could be large then base_t's size and it will
543 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
544 // constructor verifies correctness of the offset.
545 //
546 // It could happened on subclass's branch (from the type profiling
547 // inlining) which was not eliminated during parsing since the exactness
548 // of the allocation type was not propagated to the subclass type check.
549 //
550 // Or the type 't' could be not related to 'base_t' at all.
551 // It could happened when CHA type is different from MDO type on a dead path
552 // (for example, from instanceof check) which is not collapsed during parsing.
553 //
554 // Do nothing for such AddP node and don't process its users since
555 // this code branch will go away.
556 //
557 if (!t->is_known_instance() &&
558 !base_t->klass()->is_subtype_of(t->klass())) {
559 return false; // bail out
560 }
562 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
563 // Do NOT remove the next line: ensure a new alias index is allocated
564 // for the instance type. Note: C++ will not remove it since the call
565 // has side effect.
566 int alias_idx = _compile->get_alias_index(tinst);
567 igvn->set_type(addp, tinst);
568 // record the allocation in the node map
569 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
570 set_map(addp->_idx, get_map(base->_idx));
572 // Set addp's Base and Address to 'base'.
573 Node *abase = addp->in(AddPNode::Base);
574 Node *adr = addp->in(AddPNode::Address);
575 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
576 adr->in(0)->_idx == (uint)inst_id) {
577 // Skip AddP cases #3 and #5.
578 } else {
579 assert(!abase->is_top(), "sanity"); // AddP case #3
580 if (abase != base) {
581 igvn->hash_delete(addp);
582 addp->set_req(AddPNode::Base, base);
583 if (abase == adr) {
584 addp->set_req(AddPNode::Address, base);
585 } else {
586 // AddP case #4 (adr is array's element offset AddP node)
587 #ifdef ASSERT
588 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
589 assert(adr->is_AddP() && atype != NULL &&
590 atype->instance_id() == inst_id, "array's element offset should be processed first");
591 #endif
592 }
593 igvn->hash_insert(addp);
594 }
595 }
596 // Put on IGVN worklist since at least addp's type was changed above.
597 record_for_optimizer(addp);
598 return true;
599 }
601 //
602 // Create a new version of orig_phi if necessary. Returns either the newly
603 // created phi or an existing phi. Sets create_new to indicate whether a new
604 // phi was created. Cache the last newly created phi in the node map.
605 //
606 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
607 Compile *C = _compile;
608 new_created = false;
609 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
610 // nothing to do if orig_phi is bottom memory or matches alias_idx
611 if (phi_alias_idx == alias_idx) {
612 return orig_phi;
613 }
614 // Have we recently created a Phi for this alias index?
615 PhiNode *result = get_map_phi(orig_phi->_idx);
616 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
617 return result;
618 }
619 // Previous check may fail when the same wide memory Phi was split into Phis
620 // for different memory slices. Search all Phis for this region.
621 if (result != NULL) {
622 Node* region = orig_phi->in(0);
623 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
624 Node* phi = region->fast_out(i);
625 if (phi->is_Phi() &&
626 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
627 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
628 return phi->as_Phi();
629 }
630 }
631 }
632 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
633 if (C->do_escape_analysis() == true && !C->failing()) {
634 // Retry compilation without escape analysis.
635 // If this is the first failure, the sentinel string will "stick"
636 // to the Compile object, and the C2Compiler will see it and retry.
637 C->record_failure(C2Compiler::retry_no_escape_analysis());
638 }
639 return NULL;
640 }
641 orig_phi_worklist.append_if_missing(orig_phi);
642 const TypePtr *atype = C->get_adr_type(alias_idx);
643 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
644 C->copy_node_notes_to(result, orig_phi);
645 igvn->set_type(result, result->bottom_type());
646 record_for_optimizer(result);
648 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
649 assert(pn == NULL || pn == orig_phi, "wrong node");
650 set_map(orig_phi->_idx, result);
651 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
653 new_created = true;
654 return result;
655 }
657 //
658 // Return a new version of Memory Phi "orig_phi" with the inputs having the
659 // specified alias index.
660 //
661 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
663 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
664 Compile *C = _compile;
665 bool new_phi_created;
666 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
667 if (!new_phi_created) {
668 return result;
669 }
671 GrowableArray<PhiNode *> phi_list;
672 GrowableArray<uint> cur_input;
674 PhiNode *phi = orig_phi;
675 uint idx = 1;
676 bool finished = false;
677 while(!finished) {
678 while (idx < phi->req()) {
679 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
680 if (mem != NULL && mem->is_Phi()) {
681 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
682 if (new_phi_created) {
683 // found an phi for which we created a new split, push current one on worklist and begin
684 // processing new one
685 phi_list.push(phi);
686 cur_input.push(idx);
687 phi = mem->as_Phi();
688 result = newphi;
689 idx = 1;
690 continue;
691 } else {
692 mem = newphi;
693 }
694 }
695 if (C->failing()) {
696 return NULL;
697 }
698 result->set_req(idx++, mem);
699 }
700 #ifdef ASSERT
701 // verify that the new Phi has an input for each input of the original
702 assert( phi->req() == result->req(), "must have same number of inputs.");
703 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
704 #endif
705 // Check if all new phi's inputs have specified alias index.
706 // Otherwise use old phi.
707 for (uint i = 1; i < phi->req(); i++) {
708 Node* in = result->in(i);
709 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
710 }
711 // we have finished processing a Phi, see if there are any more to do
712 finished = (phi_list.length() == 0 );
713 if (!finished) {
714 phi = phi_list.pop();
715 idx = cur_input.pop();
716 PhiNode *prev_result = get_map_phi(phi->_idx);
717 prev_result->set_req(idx++, result);
718 result = prev_result;
719 }
720 }
721 return result;
722 }
725 //
726 // The next methods are derived from methods in MemNode.
727 //
728 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
729 Node *mem = mmem;
730 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
731 // means an array I have not precisely typed yet. Do not do any
732 // alias stuff with it any time soon.
733 if( toop->base() != Type::AnyPtr &&
734 !(toop->klass() != NULL &&
735 toop->klass()->is_java_lang_Object() &&
736 toop->offset() == Type::OffsetBot) ) {
737 mem = mmem->memory_at(alias_idx);
738 // Update input if it is progress over what we have now
739 }
740 return mem;
741 }
743 //
744 // Move memory users to their memory slices.
745 //
746 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
747 Compile* C = _compile;
749 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
750 assert(tp != NULL, "ptr type");
751 int alias_idx = C->get_alias_index(tp);
752 int general_idx = C->get_general_index(alias_idx);
754 // Move users first
755 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
756 Node* use = n->fast_out(i);
757 if (use->is_MergeMem()) {
758 MergeMemNode* mmem = use->as_MergeMem();
759 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
760 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
761 continue; // Nothing to do
762 }
763 // Replace previous general reference to mem node.
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 mmem->set_memory_at(general_idx, m);
768 --imax;
769 --i;
770 } else if (use->is_MemBar()) {
771 assert(!use->is_Initialize(), "initializing stores should not be moved");
772 if (use->req() > MemBarNode::Precedent &&
773 use->in(MemBarNode::Precedent) == n) {
774 // Don't move related membars.
775 record_for_optimizer(use);
776 continue;
777 }
778 tp = use->as_MemBar()->adr_type()->isa_ptr();
779 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
780 alias_idx == general_idx) {
781 continue; // Nothing to do
782 }
783 // Move to general memory slice.
784 uint orig_uniq = C->unique();
785 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
786 assert(orig_uniq == C->unique(), "no new nodes");
787 igvn->hash_delete(use);
788 imax -= use->replace_edge(n, m);
789 igvn->hash_insert(use);
790 record_for_optimizer(use);
791 --i;
792 #ifdef ASSERT
793 } else if (use->is_Mem()) {
794 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
795 // Don't move related cardmark.
796 continue;
797 }
798 // Memory nodes should have new memory input.
799 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
800 assert(tp != NULL, "ptr type");
801 int idx = C->get_alias_index(tp);
802 assert(get_map(use->_idx) != NULL || idx == alias_idx,
803 "Following memory nodes should have new memory input or be on the same memory slice");
804 } else if (use->is_Phi()) {
805 // Phi nodes should be split and moved already.
806 tp = use->as_Phi()->adr_type()->isa_ptr();
807 assert(tp != NULL, "ptr type");
808 int idx = C->get_alias_index(tp);
809 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
810 } else {
811 use->dump();
812 assert(false, "should not be here");
813 #endif
814 }
815 }
816 }
818 //
819 // Search memory chain of "mem" to find a MemNode whose address
820 // is the specified alias index.
821 //
822 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
823 if (orig_mem == NULL)
824 return orig_mem;
825 Compile* C = phase->C;
826 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
827 bool is_instance = (toop != NULL) && toop->is_known_instance();
828 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
829 Node *prev = NULL;
830 Node *result = orig_mem;
831 while (prev != result) {
832 prev = result;
833 if (result == start_mem)
834 break; // hit one of our sentinels
835 if (result->is_Mem()) {
836 const Type *at = phase->type(result->in(MemNode::Address));
837 if (at == Type::TOP)
838 break; // Dead
839 assert (at->isa_ptr() != NULL, "pointer type required.");
840 int idx = C->get_alias_index(at->is_ptr());
841 if (idx == alias_idx)
842 break; // Found
843 if (!is_instance && (at->isa_oopptr() == NULL ||
844 !at->is_oopptr()->is_known_instance())) {
845 break; // Do not skip store to general memory slice.
846 }
847 result = result->in(MemNode::Memory);
848 }
849 if (!is_instance)
850 continue; // don't search further for non-instance types
851 // skip over a call which does not affect this memory slice
852 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
853 Node *proj_in = result->in(0);
854 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
855 break; // hit one of our sentinels
856 } else if (proj_in->is_Call()) {
857 CallNode *call = proj_in->as_Call();
858 if (!call->may_modify(toop, phase)) {
859 result = call->in(TypeFunc::Memory);
860 }
861 } else if (proj_in->is_Initialize()) {
862 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
863 // Stop if this is the initialization for the object instance which
864 // which contains this memory slice, otherwise skip over it.
865 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
866 result = proj_in->in(TypeFunc::Memory);
867 }
868 } else if (proj_in->is_MemBar()) {
869 result = proj_in->in(TypeFunc::Memory);
870 }
871 } else if (result->is_MergeMem()) {
872 MergeMemNode *mmem = result->as_MergeMem();
873 result = step_through_mergemem(mmem, alias_idx, toop);
874 if (result == mmem->base_memory()) {
875 // Didn't find instance memory, search through general slice recursively.
876 result = mmem->memory_at(C->get_general_index(alias_idx));
877 result = find_inst_mem(result, alias_idx, orig_phis, phase);
878 if (C->failing()) {
879 return NULL;
880 }
881 mmem->set_memory_at(alias_idx, result);
882 }
883 } else if (result->is_Phi() &&
884 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
885 Node *un = result->as_Phi()->unique_input(phase);
886 if (un != NULL) {
887 orig_phis.append_if_missing(result->as_Phi());
888 result = un;
889 } else {
890 break;
891 }
892 } else if (result->is_ClearArray()) {
893 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
894 // Can not bypass initialization of the instance
895 // we are looking for.
896 break;
897 }
898 // Otherwise skip it (the call updated 'result' value).
899 } else if (result->Opcode() == Op_SCMemProj) {
900 assert(result->in(0)->is_LoadStore(), "sanity");
901 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
902 if (at != Type::TOP) {
903 assert (at->isa_ptr() != NULL, "pointer type required.");
904 int idx = C->get_alias_index(at->is_ptr());
905 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
906 break;
907 }
908 result = result->in(0)->in(MemNode::Memory);
909 }
910 }
911 if (result->is_Phi()) {
912 PhiNode *mphi = result->as_Phi();
913 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
914 const TypePtr *t = mphi->adr_type();
915 if (!is_instance) {
916 // Push all non-instance Phis on the orig_phis worklist to update inputs
917 // during Phase 4 if needed.
918 orig_phis.append_if_missing(mphi);
919 } else if (C->get_alias_index(t) != alias_idx) {
920 // Create a new Phi with the specified alias index type.
921 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
922 }
923 }
924 // the result is either MemNode, PhiNode, InitializeNode.
925 return result;
926 }
928 //
929 // Convert the types of unescaped object to instance types where possible,
930 // propagate the new type information through the graph, and update memory
931 // edges and MergeMem inputs to reflect the new type.
932 //
933 // We start with allocations (and calls which may be allocations) on alloc_worklist.
934 // The processing is done in 4 phases:
935 //
936 // Phase 1: Process possible allocations from alloc_worklist. Create instance
937 // types for the CheckCastPP for allocations where possible.
938 // Propagate the the new types through users as follows:
939 // casts and Phi: push users on alloc_worklist
940 // AddP: cast Base and Address inputs to the instance type
941 // push any AddP users on alloc_worklist and push any memnode
942 // users onto memnode_worklist.
943 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
944 // search the Memory chain for a store with the appropriate type
945 // address type. If a Phi is found, create a new version with
946 // the appropriate memory slices from each of the Phi inputs.
947 // For stores, process the users as follows:
948 // MemNode: push on memnode_worklist
949 // MergeMem: push on mergemem_worklist
950 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
951 // moving the first node encountered of each instance type to the
952 // the input corresponding to its alias index.
953 // appropriate memory slice.
954 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
955 //
956 // In the following example, the CheckCastPP nodes are the cast of allocation
957 // results and the allocation of node 29 is unescaped and eligible to be an
958 // instance type.
959 //
960 // We start with:
961 //
962 // 7 Parm #memory
963 // 10 ConI "12"
964 // 19 CheckCastPP "Foo"
965 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
966 // 29 CheckCastPP "Foo"
967 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
968 //
969 // 40 StoreP 25 7 20 ... alias_index=4
970 // 50 StoreP 35 40 30 ... alias_index=4
971 // 60 StoreP 45 50 20 ... alias_index=4
972 // 70 LoadP _ 60 30 ... alias_index=4
973 // 80 Phi 75 50 60 Memory alias_index=4
974 // 90 LoadP _ 80 30 ... alias_index=4
975 // 100 LoadP _ 80 20 ... alias_index=4
976 //
977 //
978 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
979 // and creating a new alias index for node 30. This gives:
980 //
981 // 7 Parm #memory
982 // 10 ConI "12"
983 // 19 CheckCastPP "Foo"
984 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
985 // 29 CheckCastPP "Foo" iid=24
986 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
987 //
988 // 40 StoreP 25 7 20 ... alias_index=4
989 // 50 StoreP 35 40 30 ... alias_index=6
990 // 60 StoreP 45 50 20 ... alias_index=4
991 // 70 LoadP _ 60 30 ... alias_index=6
992 // 80 Phi 75 50 60 Memory alias_index=4
993 // 90 LoadP _ 80 30 ... alias_index=6
994 // 100 LoadP _ 80 20 ... alias_index=4
995 //
996 // In phase 2, new memory inputs are computed for the loads and stores,
997 // And a new version of the phi is created. In phase 4, the inputs to
998 // node 80 are updated and then the memory nodes are updated with the
999 // values computed in phase 2. This results in:
1000 //
1001 // 7 Parm #memory
1002 // 10 ConI "12"
1003 // 19 CheckCastPP "Foo"
1004 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
1005 // 29 CheckCastPP "Foo" iid=24
1006 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
1007 //
1008 // 40 StoreP 25 7 20 ... alias_index=4
1009 // 50 StoreP 35 7 30 ... alias_index=6
1010 // 60 StoreP 45 40 20 ... alias_index=4
1011 // 70 LoadP _ 50 30 ... alias_index=6
1012 // 80 Phi 75 40 60 Memory alias_index=4
1013 // 120 Phi 75 50 50 Memory alias_index=6
1014 // 90 LoadP _ 120 30 ... alias_index=6
1015 // 100 LoadP _ 80 20 ... alias_index=4
1016 //
1017 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1018 GrowableArray<Node *> memnode_worklist;
1019 GrowableArray<PhiNode *> orig_phis;
1021 PhaseIterGVN *igvn = _igvn;
1022 uint new_index_start = (uint) _compile->num_alias_types();
1023 Arena* arena = Thread::current()->resource_area();
1024 VectorSet visited(arena);
1027 // Phase 1: Process possible allocations from alloc_worklist.
1028 // Create instance types for the CheckCastPP for allocations where possible.
1029 //
1030 // (Note: don't forget to change the order of the second AddP node on
1031 // the alloc_worklist if the order of the worklist processing is changed,
1032 // see the comment in find_second_addp().)
1033 //
1034 while (alloc_worklist.length() != 0) {
1035 Node *n = alloc_worklist.pop();
1036 uint ni = n->_idx;
1037 const TypeOopPtr* tinst = NULL;
1038 if (n->is_Call()) {
1039 CallNode *alloc = n->as_Call();
1040 // copy escape information to call node
1041 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1042 PointsToNode::EscapeState es = escape_state(alloc);
1043 // We have an allocation or call which returns a Java object,
1044 // see if it is unescaped.
1045 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
1046 continue;
1048 // Find CheckCastPP for the allocate or for the return value of a call
1049 n = alloc->result_cast();
1050 if (n == NULL) { // No uses except Initialize node
1051 if (alloc->is_Allocate()) {
1052 // Set the scalar_replaceable flag for allocation
1053 // so it could be eliminated if it has no uses.
1054 alloc->as_Allocate()->_is_scalar_replaceable = true;
1055 }
1056 continue;
1057 }
1058 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1059 assert(!alloc->is_Allocate(), "allocation should have unique type");
1060 continue;
1061 }
1063 // The inline code for Object.clone() casts the allocation result to
1064 // java.lang.Object and then to the actual type of the allocated
1065 // object. Detect this case and use the second cast.
1066 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1067 // the allocation result is cast to java.lang.Object and then
1068 // to the actual Array type.
1069 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1070 && (alloc->is_AllocateArray() ||
1071 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1072 Node *cast2 = NULL;
1073 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1074 Node *use = n->fast_out(i);
1075 if (use->is_CheckCastPP()) {
1076 cast2 = use;
1077 break;
1078 }
1079 }
1080 if (cast2 != NULL) {
1081 n = cast2;
1082 } else {
1083 // Non-scalar replaceable if the allocation type is unknown statically
1084 // (reflection allocation), the object can't be restored during
1085 // deoptimization without precise type.
1086 continue;
1087 }
1088 }
1089 if (alloc->is_Allocate()) {
1090 // Set the scalar_replaceable flag for allocation
1091 // so it could be eliminated.
1092 alloc->as_Allocate()->_is_scalar_replaceable = true;
1093 }
1094 set_escape_state(n->_idx, es); // CheckCastPP escape state
1095 // in order for an object to be scalar-replaceable, it must be:
1096 // - a direct allocation (not a call returning an object)
1097 // - non-escaping
1098 // - eligible to be a unique type
1099 // - not determined to be ineligible by escape analysis
1100 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1101 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1102 set_map(alloc->_idx, n);
1103 set_map(n->_idx, alloc);
1104 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1105 if (t == NULL)
1106 continue; // not a TypeOopPtr
1107 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1108 igvn->hash_delete(n);
1109 igvn->set_type(n, tinst);
1110 n->raise_bottom_type(tinst);
1111 igvn->hash_insert(n);
1112 record_for_optimizer(n);
1113 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
1115 // First, put on the worklist all Field edges from Connection Graph
1116 // which is more accurate then putting immediate users from Ideal Graph.
1117 for (uint e = 0; e < ptn->edge_count(); e++) {
1118 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1119 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1120 "only AddP nodes are Field edges in CG");
1121 if (use->outcnt() > 0) { // Don't process dead nodes
1122 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1123 if (addp2 != NULL) {
1124 assert(alloc->is_AllocateArray(),"array allocation was expected");
1125 alloc_worklist.append_if_missing(addp2);
1126 }
1127 alloc_worklist.append_if_missing(use);
1128 }
1129 }
1131 // An allocation may have an Initialize which has raw stores. Scan
1132 // the users of the raw allocation result and push AddP users
1133 // on alloc_worklist.
1134 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1135 assert (raw_result != NULL, "must have an allocation result");
1136 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1137 Node *use = raw_result->fast_out(i);
1138 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1139 Node* addp2 = find_second_addp(use, raw_result);
1140 if (addp2 != NULL) {
1141 assert(alloc->is_AllocateArray(),"array allocation was expected");
1142 alloc_worklist.append_if_missing(addp2);
1143 }
1144 alloc_worklist.append_if_missing(use);
1145 } else if (use->is_MemBar()) {
1146 memnode_worklist.append_if_missing(use);
1147 }
1148 }
1149 }
1150 } else if (n->is_AddP()) {
1151 VectorSet* ptset = PointsTo(get_addp_base(n));
1152 assert(ptset->Size() == 1, "AddP address is unique");
1153 uint elem = ptset->getelem(); // Allocation node's index
1154 if (elem == _phantom_object) {
1155 assert(false, "escaped allocation");
1156 continue; // Assume the value was set outside this method.
1157 }
1158 Node *base = get_map(elem); // CheckCastPP node
1159 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1160 tinst = igvn->type(base)->isa_oopptr();
1161 } else if (n->is_Phi() ||
1162 n->is_CheckCastPP() ||
1163 n->is_EncodeP() ||
1164 n->is_DecodeN() ||
1165 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1166 if (visited.test_set(n->_idx)) {
1167 assert(n->is_Phi(), "loops only through Phi's");
1168 continue; // already processed
1169 }
1170 VectorSet* ptset = PointsTo(n);
1171 if (ptset->Size() == 1) {
1172 uint elem = ptset->getelem(); // Allocation node's index
1173 if (elem == _phantom_object) {
1174 assert(false, "escaped allocation");
1175 continue; // Assume the value was set outside this method.
1176 }
1177 Node *val = get_map(elem); // CheckCastPP node
1178 TypeNode *tn = n->as_Type();
1179 tinst = igvn->type(val)->isa_oopptr();
1180 assert(tinst != NULL && tinst->is_known_instance() &&
1181 (uint)tinst->instance_id() == elem , "instance type expected.");
1183 const Type *tn_type = igvn->type(tn);
1184 const TypeOopPtr *tn_t;
1185 if (tn_type->isa_narrowoop()) {
1186 tn_t = tn_type->make_ptr()->isa_oopptr();
1187 } else {
1188 tn_t = tn_type->isa_oopptr();
1189 }
1191 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1192 if (tn_type->isa_narrowoop()) {
1193 tn_type = tinst->make_narrowoop();
1194 } else {
1195 tn_type = tinst;
1196 }
1197 igvn->hash_delete(tn);
1198 igvn->set_type(tn, tn_type);
1199 tn->set_type(tn_type);
1200 igvn->hash_insert(tn);
1201 record_for_optimizer(n);
1202 } else {
1203 assert(tn_type == TypePtr::NULL_PTR ||
1204 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1205 "unexpected type");
1206 continue; // Skip dead path with different type
1207 }
1208 }
1209 } else {
1210 debug_only(n->dump();)
1211 assert(false, "EA: unexpected node");
1212 continue;
1213 }
1214 // push allocation's users on appropriate worklist
1215 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1216 Node *use = n->fast_out(i);
1217 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1218 // Load/store to instance's field
1219 memnode_worklist.append_if_missing(use);
1220 } else if (use->is_MemBar()) {
1221 memnode_worklist.append_if_missing(use);
1222 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1223 Node* addp2 = find_second_addp(use, n);
1224 if (addp2 != NULL) {
1225 alloc_worklist.append_if_missing(addp2);
1226 }
1227 alloc_worklist.append_if_missing(use);
1228 } else if (use->is_Phi() ||
1229 use->is_CheckCastPP() ||
1230 use->is_EncodeP() ||
1231 use->is_DecodeN() ||
1232 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1233 alloc_worklist.append_if_missing(use);
1234 #ifdef ASSERT
1235 } else if (use->is_Mem()) {
1236 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1237 } else if (use->is_MergeMem()) {
1238 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1239 } else if (use->is_SafePoint()) {
1240 // Look for MergeMem nodes for calls which reference unique allocation
1241 // (through CheckCastPP nodes) even for debug info.
1242 Node* m = use->in(TypeFunc::Memory);
1243 if (m->is_MergeMem()) {
1244 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1245 }
1246 } else {
1247 uint op = use->Opcode();
1248 if (!(op == Op_CmpP || op == Op_Conv2B ||
1249 op == Op_CastP2X || op == Op_StoreCM ||
1250 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1251 op == Op_StrEquals || op == Op_StrIndexOf)) {
1252 n->dump();
1253 use->dump();
1254 assert(false, "EA: missing allocation reference path");
1255 }
1256 #endif
1257 }
1258 }
1260 }
1261 // New alias types were created in split_AddP().
1262 uint new_index_end = (uint) _compile->num_alias_types();
1264 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1265 // compute new values for Memory inputs (the Memory inputs are not
1266 // actually updated until phase 4.)
1267 if (memnode_worklist.length() == 0)
1268 return; // nothing to do
1270 while (memnode_worklist.length() != 0) {
1271 Node *n = memnode_worklist.pop();
1272 if (visited.test_set(n->_idx))
1273 continue;
1274 if (n->is_Phi() || n->is_ClearArray()) {
1275 // we don't need to do anything, but the users must be pushed
1276 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1277 // we don't need to do anything, but the users must be pushed
1278 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1279 if (n == NULL)
1280 continue;
1281 } else {
1282 assert(n->is_Mem(), "memory node required.");
1283 Node *addr = n->in(MemNode::Address);
1284 const Type *addr_t = igvn->type(addr);
1285 if (addr_t == Type::TOP)
1286 continue;
1287 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1288 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1289 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1290 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1291 if (_compile->failing()) {
1292 return;
1293 }
1294 if (mem != n->in(MemNode::Memory)) {
1295 // We delay the memory edge update since we need old one in
1296 // MergeMem code below when instances memory slices are separated.
1297 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1298 assert(pn == NULL || pn == n, "wrong node");
1299 set_map(n->_idx, mem);
1300 ptnode_adr(n->_idx)->_node = n;
1301 }
1302 if (n->is_Load()) {
1303 continue; // don't push users
1304 } else if (n->is_LoadStore()) {
1305 // get the memory projection
1306 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1307 Node *use = n->fast_out(i);
1308 if (use->Opcode() == Op_SCMemProj) {
1309 n = use;
1310 break;
1311 }
1312 }
1313 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1314 }
1315 }
1316 // push user on appropriate worklist
1317 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1318 Node *use = n->fast_out(i);
1319 if (use->is_Phi() || use->is_ClearArray()) {
1320 memnode_worklist.append_if_missing(use);
1321 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1322 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1323 continue;
1324 memnode_worklist.append_if_missing(use);
1325 } else if (use->is_MemBar()) {
1326 memnode_worklist.append_if_missing(use);
1327 #ifdef ASSERT
1328 } else if(use->is_Mem()) {
1329 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1330 } else if (use->is_MergeMem()) {
1331 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1332 } else {
1333 uint op = use->Opcode();
1334 if (!(op == Op_StoreCM ||
1335 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1336 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1337 op == Op_AryEq || op == Op_StrComp ||
1338 op == Op_StrEquals || op == Op_StrIndexOf)) {
1339 n->dump();
1340 use->dump();
1341 assert(false, "EA: missing memory path");
1342 }
1343 #endif
1344 }
1345 }
1346 }
1348 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1349 // Walk each memory slice moving the first node encountered of each
1350 // instance type to the the input corresponding to its alias index.
1351 uint length = _mergemem_worklist.length();
1352 for( uint next = 0; next < length; ++next ) {
1353 MergeMemNode* nmm = _mergemem_worklist.at(next);
1354 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1355 // Note: we don't want to use MergeMemStream here because we only want to
1356 // scan inputs which exist at the start, not ones we add during processing.
1357 // Note 2: MergeMem may already contains instance memory slices added
1358 // during find_inst_mem() call when memory nodes were processed above.
1359 igvn->hash_delete(nmm);
1360 uint nslices = nmm->req();
1361 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1362 Node* mem = nmm->in(i);
1363 Node* cur = NULL;
1364 if (mem == NULL || mem->is_top())
1365 continue;
1366 // First, update mergemem by moving memory nodes to corresponding slices
1367 // if their type became more precise since this mergemem was created.
1368 while (mem->is_Mem()) {
1369 const Type *at = igvn->type(mem->in(MemNode::Address));
1370 if (at != Type::TOP) {
1371 assert (at->isa_ptr() != NULL, "pointer type required.");
1372 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1373 if (idx == i) {
1374 if (cur == NULL)
1375 cur = mem;
1376 } else {
1377 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1378 nmm->set_memory_at(idx, mem);
1379 }
1380 }
1381 }
1382 mem = mem->in(MemNode::Memory);
1383 }
1384 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1385 // Find any instance of the current type if we haven't encountered
1386 // already a memory slice of the instance along the memory chain.
1387 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1388 if((uint)_compile->get_general_index(ni) == i) {
1389 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1390 if (nmm->is_empty_memory(m)) {
1391 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1392 if (_compile->failing()) {
1393 return;
1394 }
1395 nmm->set_memory_at(ni, result);
1396 }
1397 }
1398 }
1399 }
1400 // Find the rest of instances values
1401 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1402 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1403 Node* result = step_through_mergemem(nmm, ni, tinst);
1404 if (result == nmm->base_memory()) {
1405 // Didn't find instance memory, search through general slice recursively.
1406 result = nmm->memory_at(_compile->get_general_index(ni));
1407 result = find_inst_mem(result, ni, orig_phis, igvn);
1408 if (_compile->failing()) {
1409 return;
1410 }
1411 nmm->set_memory_at(ni, result);
1412 }
1413 }
1414 igvn->hash_insert(nmm);
1415 record_for_optimizer(nmm);
1416 }
1418 // Phase 4: Update the inputs of non-instance memory Phis and
1419 // the Memory input of memnodes
1420 // First update the inputs of any non-instance Phi's from
1421 // which we split out an instance Phi. Note we don't have
1422 // to recursively process Phi's encounted on the input memory
1423 // chains as is done in split_memory_phi() since they will
1424 // also be processed here.
1425 for (int j = 0; j < orig_phis.length(); j++) {
1426 PhiNode *phi = orig_phis.at(j);
1427 int alias_idx = _compile->get_alias_index(phi->adr_type());
1428 igvn->hash_delete(phi);
1429 for (uint i = 1; i < phi->req(); i++) {
1430 Node *mem = phi->in(i);
1431 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1432 if (_compile->failing()) {
1433 return;
1434 }
1435 if (mem != new_mem) {
1436 phi->set_req(i, new_mem);
1437 }
1438 }
1439 igvn->hash_insert(phi);
1440 record_for_optimizer(phi);
1441 }
1443 // Update the memory inputs of MemNodes with the value we computed
1444 // in Phase 2 and move stores memory users to corresponding memory slices.
1446 // Disable memory split verification code until the fix for 6984348.
1447 // Currently it produces false negative results since it does not cover all cases.
1448 #if 0 // ifdef ASSERT
1449 visited.Reset();
1450 Node_Stack old_mems(arena, _compile->unique() >> 2);
1451 #endif
1452 for (uint i = 0; i < nodes_size(); i++) {
1453 Node *nmem = get_map(i);
1454 if (nmem != NULL) {
1455 Node *n = ptnode_adr(i)->_node;
1456 assert(n != NULL, "sanity");
1457 if (n->is_Mem()) {
1458 #if 0 // ifdef ASSERT
1459 Node* old_mem = n->in(MemNode::Memory);
1460 if (!visited.test_set(old_mem->_idx)) {
1461 old_mems.push(old_mem, old_mem->outcnt());
1462 }
1463 #endif
1464 assert(n->in(MemNode::Memory) != nmem, "sanity");
1465 if (!n->is_Load()) {
1466 // Move memory users of a store first.
1467 move_inst_mem(n, orig_phis, igvn);
1468 }
1469 // Now update memory input
1470 igvn->hash_delete(n);
1471 n->set_req(MemNode::Memory, nmem);
1472 igvn->hash_insert(n);
1473 record_for_optimizer(n);
1474 } else {
1475 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1476 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1477 }
1478 }
1479 }
1480 #if 0 // ifdef ASSERT
1481 // Verify that memory was split correctly
1482 while (old_mems.is_nonempty()) {
1483 Node* old_mem = old_mems.node();
1484 uint old_cnt = old_mems.index();
1485 old_mems.pop();
1486 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
1487 }
1488 #endif
1489 }
1491 bool ConnectionGraph::has_candidates(Compile *C) {
1492 // EA brings benefits only when the code has allocations and/or locks which
1493 // are represented by ideal Macro nodes.
1494 int cnt = C->macro_count();
1495 for( int i=0; i < cnt; i++ ) {
1496 Node *n = C->macro_node(i);
1497 if ( n->is_Allocate() )
1498 return true;
1499 if( n->is_Lock() ) {
1500 Node* obj = n->as_Lock()->obj_node()->uncast();
1501 if( !(obj->is_Parm() || obj->is_Con()) )
1502 return true;
1503 }
1504 }
1505 return false;
1506 }
1508 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1509 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1510 // to create space for them in ConnectionGraph::_nodes[].
1511 Node* oop_null = igvn->zerocon(T_OBJECT);
1512 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1514 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1515 // Perform escape analysis
1516 if (congraph->compute_escape()) {
1517 // There are non escaping objects.
1518 C->set_congraph(congraph);
1519 }
1521 // Cleanup.
1522 if (oop_null->outcnt() == 0)
1523 igvn->hash_delete(oop_null);
1524 if (noop_null->outcnt() == 0)
1525 igvn->hash_delete(noop_null);
1526 }
1528 bool ConnectionGraph::compute_escape() {
1529 Compile* C = _compile;
1531 // 1. Populate Connection Graph (CG) with Ideal nodes.
1533 Unique_Node_List worklist_init;
1534 worklist_init.map(C->unique(), NULL); // preallocate space
1536 // Initialize worklist
1537 if (C->root() != NULL) {
1538 worklist_init.push(C->root());
1539 }
1541 GrowableArray<Node*> alloc_worklist;
1542 GrowableArray<Node*> addp_worklist;
1543 PhaseGVN* igvn = _igvn;
1544 bool has_allocations = false;
1546 // Push all useful nodes onto CG list and set their type.
1547 for( uint next = 0; next < worklist_init.size(); ++next ) {
1548 Node* n = worklist_init.at(next);
1549 record_for_escape_analysis(n, igvn);
1550 // Only allocations and java static calls results are checked
1551 // for an escape status. See process_call_result() below.
1552 if (n->is_Allocate() || n->is_CallStaticJava() &&
1553 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1554 has_allocations = true;
1555 if (n->is_Allocate())
1556 alloc_worklist.append(n);
1557 }
1558 if(n->is_AddP()) {
1559 // Collect address nodes. Use them during stage 3 below
1560 // to build initial connection graph field edges.
1561 addp_worklist.append(n);
1562 } else if (n->is_MergeMem()) {
1563 // Collect all MergeMem nodes to add memory slices for
1564 // scalar replaceable objects in split_unique_types().
1565 _mergemem_worklist.append(n->as_MergeMem());
1566 }
1567 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1568 Node* m = n->fast_out(i); // Get user
1569 worklist_init.push(m);
1570 }
1571 }
1573 if (!has_allocations) {
1574 _collecting = false;
1575 return false; // Nothing to do.
1576 }
1578 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1579 uint delayed_size = _delayed_worklist.size();
1580 for( uint next = 0; next < delayed_size; ++next ) {
1581 Node* n = _delayed_worklist.at(next);
1582 build_connection_graph(n, igvn);
1583 }
1585 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1586 // to reduce number of iterations during stage 4 below.
1587 uint addp_length = addp_worklist.length();
1588 for( uint next = 0; next < addp_length; ++next ) {
1589 Node* n = addp_worklist.at(next);
1590 Node* base = get_addp_base(n);
1591 if (base->is_Proj())
1592 base = base->in(0);
1593 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1594 if (nt == PointsToNode::JavaObject) {
1595 build_connection_graph(n, igvn);
1596 }
1597 }
1599 GrowableArray<int> cg_worklist;
1600 cg_worklist.append(_phantom_object);
1601 GrowableArray<uint> worklist;
1603 // 4. Build Connection Graph which need
1604 // to walk the connection graph.
1605 _progress = false;
1606 for (uint ni = 0; ni < nodes_size(); ni++) {
1607 PointsToNode* ptn = ptnode_adr(ni);
1608 Node *n = ptn->_node;
1609 if (n != NULL) { // Call, AddP, LoadP, StoreP
1610 build_connection_graph(n, igvn);
1611 if (ptn->node_type() != PointsToNode::UnknownType)
1612 cg_worklist.append(n->_idx); // Collect CG nodes
1613 if (!_processed.test(n->_idx))
1614 worklist.append(n->_idx); // Collect C/A/L/S nodes
1615 }
1616 }
1618 // After IGVN user nodes may have smaller _idx than
1619 // their inputs so they will be processed first in
1620 // previous loop. Because of that not all Graph
1621 // edges will be created. Walk over interesting
1622 // nodes again until no new edges are created.
1623 //
1624 // Normally only 1-3 passes needed to build
1625 // Connection Graph depending on graph complexity.
1626 // Observed 8 passes in jvm2008 compiler.compiler.
1627 // Set limit to 20 to catch situation when something
1628 // did go wrong and recompile the method without EA.
1630 #define CG_BUILD_ITER_LIMIT 20
1632 uint length = worklist.length();
1633 int iterations = 0;
1634 while(_progress && (iterations++ < CG_BUILD_ITER_LIMIT)) {
1635 _progress = false;
1636 for( uint next = 0; next < length; ++next ) {
1637 int ni = worklist.at(next);
1638 PointsToNode* ptn = ptnode_adr(ni);
1639 Node* n = ptn->_node;
1640 assert(n != NULL, "should be known node");
1641 build_connection_graph(n, igvn);
1642 }
1643 }
1644 if (iterations >= CG_BUILD_ITER_LIMIT) {
1645 assert(iterations < CG_BUILD_ITER_LIMIT,
1646 err_msg("infinite EA connection graph build with %d nodes and worklist size %d",
1647 nodes_size(), length));
1648 // Possible infinite build_connection_graph loop,
1649 // retry compilation without escape analysis.
1650 C->record_failure(C2Compiler::retry_no_escape_analysis());
1651 _collecting = false;
1652 return false;
1653 }
1654 #undef CG_BUILD_ITER_LIMIT
1656 Arena* arena = Thread::current()->resource_area();
1657 VectorSet visited(arena);
1659 // 5. Find fields initializing values for not escaped allocations
1660 uint alloc_length = alloc_worklist.length();
1661 for (uint next = 0; next < alloc_length; ++next) {
1662 Node* n = alloc_worklist.at(next);
1663 if (ptnode_adr(n->_idx)->escape_state() == PointsToNode::NoEscape) {
1664 find_init_values(n, &visited, igvn);
1665 }
1666 }
1668 worklist.clear();
1670 // 6. Remove deferred edges from the graph.
1671 uint cg_length = cg_worklist.length();
1672 for (uint next = 0; next < cg_length; ++next) {
1673 int ni = cg_worklist.at(next);
1674 PointsToNode* ptn = ptnode_adr(ni);
1675 PointsToNode::NodeType nt = ptn->node_type();
1676 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1677 remove_deferred(ni, &worklist, &visited);
1678 Node *n = ptn->_node;
1679 }
1680 }
1682 // 7. Adjust escape state of nonescaping objects.
1683 for (uint next = 0; next < addp_length; ++next) {
1684 Node* n = addp_worklist.at(next);
1685 adjust_escape_state(n);
1686 }
1688 // 8. Propagate escape states.
1689 worklist.clear();
1691 // mark all nodes reachable from GlobalEscape nodes
1692 (void)propagate_escape_state(&cg_worklist, &worklist, PointsToNode::GlobalEscape);
1694 // mark all nodes reachable from ArgEscape nodes
1695 bool has_non_escaping_obj = propagate_escape_state(&cg_worklist, &worklist, PointsToNode::ArgEscape);
1697 // push all NoEscape nodes on the worklist
1698 for( uint next = 0; next < cg_length; ++next ) {
1699 int nk = cg_worklist.at(next);
1700 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape)
1701 worklist.push(nk);
1702 }
1703 alloc_worklist.clear();
1704 // mark all nodes reachable from NoEscape nodes
1705 while(worklist.length() > 0) {
1706 uint nk = worklist.pop();
1707 PointsToNode* ptn = ptnode_adr(nk);
1708 if (ptn->node_type() == PointsToNode::JavaObject &&
1709 !(nk == _noop_null || nk == _oop_null))
1710 has_non_escaping_obj = true; // Non Escape
1711 Node* n = ptn->_node;
1712 bool scalar_replaceable = ptn->scalar_replaceable();
1713 if (n->is_Allocate() && scalar_replaceable) {
1714 // Push scalar replaceable allocations on alloc_worklist
1715 // for processing in split_unique_types(). Note,
1716 // following code may change scalar_replaceable value.
1717 alloc_worklist.append(n);
1718 }
1719 uint e_cnt = ptn->edge_count();
1720 for (uint ei = 0; ei < e_cnt; ei++) {
1721 uint npi = ptn->edge_target(ei);
1722 PointsToNode *np = ptnode_adr(npi);
1723 if (np->escape_state() < PointsToNode::NoEscape) {
1724 set_escape_state(npi, PointsToNode::NoEscape);
1725 if (!scalar_replaceable) {
1726 np->set_scalar_replaceable(false);
1727 }
1728 worklist.push(npi);
1729 } else if (np->scalar_replaceable() && !scalar_replaceable) {
1730 // Propagate scalar_replaceable value.
1731 np->set_scalar_replaceable(false);
1732 worklist.push(npi);
1733 }
1734 }
1735 }
1737 _collecting = false;
1738 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1740 assert(ptnode_adr(_oop_null)->escape_state() == PointsToNode::NoEscape, "sanity");
1741 if (UseCompressedOops) {
1742 assert(ptnode_adr(_noop_null)->escape_state() == PointsToNode::NoEscape, "sanity");
1743 }
1745 if (EliminateLocks && has_non_escaping_obj) {
1746 // Mark locks before changing ideal graph.
1747 int cnt = C->macro_count();
1748 for( int i=0; i < cnt; i++ ) {
1749 Node *n = C->macro_node(i);
1750 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1751 AbstractLockNode* alock = n->as_AbstractLock();
1752 if (!alock->is_eliminated()) {
1753 PointsToNode::EscapeState es = escape_state(alock->obj_node());
1754 assert(es != PointsToNode::UnknownEscape, "should know");
1755 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1756 // Mark it eliminated
1757 alock->set_eliminated();
1758 }
1759 }
1760 }
1761 }
1762 }
1764 #ifndef PRODUCT
1765 if (PrintEscapeAnalysis) {
1766 dump(); // Dump ConnectionGraph
1767 }
1768 #endif
1770 bool has_scalar_replaceable_candidates = false;
1771 alloc_length = alloc_worklist.length();
1772 for (uint next = 0; next < alloc_length; ++next) {
1773 Node* n = alloc_worklist.at(next);
1774 PointsToNode* ptn = ptnode_adr(n->_idx);
1775 assert(ptn->escape_state() == PointsToNode::NoEscape, "sanity");
1776 if (ptn->scalar_replaceable()) {
1777 has_scalar_replaceable_candidates = true;
1778 break;
1779 }
1780 }
1782 if ( has_scalar_replaceable_candidates &&
1783 C->AliasLevel() >= 3 && EliminateAllocations ) {
1785 // Now use the escape information to create unique types for
1786 // scalar replaceable objects.
1787 split_unique_types(alloc_worklist);
1789 if (C->failing()) return false;
1791 C->print_method("After Escape Analysis", 2);
1793 #ifdef ASSERT
1794 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1795 tty->print("=== No allocations eliminated for ");
1796 C->method()->print_short_name();
1797 if(!EliminateAllocations) {
1798 tty->print(" since EliminateAllocations is off ===");
1799 } else if(!has_scalar_replaceable_candidates) {
1800 tty->print(" since there are no scalar replaceable candidates ===");
1801 } else if(C->AliasLevel() < 3) {
1802 tty->print(" since AliasLevel < 3 ===");
1803 }
1804 tty->cr();
1805 #endif
1806 }
1807 return has_non_escaping_obj;
1808 }
1810 // Find fields initializing values for allocations.
1811 void ConnectionGraph::find_init_values(Node* alloc, VectorSet* visited, PhaseTransform* phase) {
1812 assert(alloc->is_Allocate(), "Should be called for Allocate nodes only");
1813 PointsToNode* pta = ptnode_adr(alloc->_idx);
1814 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1815 InitializeNode* ini = alloc->as_Allocate()->initialization();
1817 Compile* C = _compile;
1818 visited->Reset();
1819 // Check if a oop field's initializing value is recorded and add
1820 // a corresponding NULL field's value if it is not recorded.
1821 // Connection Graph does not record a default initialization by NULL
1822 // captured by Initialize node.
1823 //
1824 uint ae_cnt = pta->edge_count();
1825 for (uint ei = 0; ei < ae_cnt; ei++) {
1826 uint nidx = pta->edge_target(ei); // Field (AddP)
1827 PointsToNode* ptn = ptnode_adr(nidx);
1828 assert(ptn->_node->is_AddP(), "Should be AddP nodes only");
1829 int offset = ptn->offset();
1830 if (offset != Type::OffsetBot &&
1831 offset != oopDesc::klass_offset_in_bytes() &&
1832 !visited->test_set(offset)) {
1834 // Check only oop fields.
1835 const Type* adr_type = ptn->_node->as_AddP()->bottom_type();
1836 BasicType basic_field_type = T_INT;
1837 if (adr_type->isa_instptr()) {
1838 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
1839 if (field != NULL) {
1840 basic_field_type = field->layout_type();
1841 } else {
1842 // Ignore non field load (for example, klass load)
1843 }
1844 } else if (adr_type->isa_aryptr()) {
1845 if (offset != arrayOopDesc::length_offset_in_bytes()) {
1846 const Type* elemtype = adr_type->isa_aryptr()->elem();
1847 basic_field_type = elemtype->array_element_basic_type();
1848 } else {
1849 // Ignore array length load
1850 }
1851 #ifdef ASSERT
1852 } else {
1853 // Raw pointers are used for initializing stores so skip it
1854 // since it should be recorded already
1855 Node* base = get_addp_base(ptn->_node);
1856 assert(adr_type->isa_rawptr() && base->is_Proj() &&
1857 (base->in(0) == alloc),"unexpected pointer type");
1858 #endif
1859 }
1860 if (basic_field_type == T_OBJECT ||
1861 basic_field_type == T_NARROWOOP ||
1862 basic_field_type == T_ARRAY) {
1863 Node* value = NULL;
1864 if (ini != NULL) {
1865 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
1866 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
1867 if (store != NULL && store->is_Store()) {
1868 value = store->in(MemNode::ValueIn);
1869 } else if (ptn->edge_count() > 0) { // Are there oop stores?
1870 // Check for a store which follows allocation without branches.
1871 // For example, a volatile field store is not collected
1872 // by Initialize node. TODO: it would be nice to use idom() here.
1873 //
1874 // Search all references to the same field which use different
1875 // AddP nodes, for example, in the next case:
1876 //
1877 // Point p[] = new Point[1];
1878 // if ( x ) { p[0] = new Point(); p[0].x = x; }
1879 // if ( p[0] != null ) { y = p[0].x; } // has CastPP
1880 //
1881 for (uint next = ei; (next < ae_cnt) && (value == NULL); next++) {
1882 uint fpi = pta->edge_target(next); // Field (AddP)
1883 PointsToNode *ptf = ptnode_adr(fpi);
1884 if (ptf->offset() == offset) {
1885 Node* nf = ptf->_node;
1886 for (DUIterator_Fast imax, i = nf->fast_outs(imax); i < imax; i++) {
1887 store = nf->fast_out(i);
1888 if (store->is_Store() && store->in(0) != NULL) {
1889 Node* ctrl = store->in(0);
1890 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
1891 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
1892 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
1893 ctrl = ctrl->in(0);
1894 }
1895 if (ctrl == ini || ctrl == alloc) {
1896 value = store->in(MemNode::ValueIn);
1897 break;
1898 }
1899 }
1900 }
1901 }
1902 }
1903 }
1904 }
1905 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
1906 // A field's initializing value was not recorded. Add NULL.
1907 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1908 add_edge_from_fields(alloc->_idx, null_idx, offset);
1909 }
1910 }
1911 }
1912 }
1913 }
1915 // Adjust escape state after Connection Graph is built.
1916 void ConnectionGraph::adjust_escape_state(Node* n) {
1917 PointsToNode* ptn = ptnode_adr(n->_idx);
1918 assert(n->is_AddP(), "Should be called for AddP nodes only");
1919 // Search for objects which are not scalar replaceable
1920 // and mark them to propagate the state to referenced objects.
1921 //
1923 int offset = ptn->offset();
1924 Node* base = get_addp_base(n);
1925 VectorSet* ptset = PointsTo(base);
1926 int ptset_size = ptset->Size();
1928 // An object is not scalar replaceable if the field which may point
1929 // to it has unknown offset (unknown element of an array of objects).
1930 //
1932 if (offset == Type::OffsetBot) {
1933 uint e_cnt = ptn->edge_count();
1934 for (uint ei = 0; ei < e_cnt; ei++) {
1935 uint npi = ptn->edge_target(ei);
1936 ptnode_adr(npi)->set_scalar_replaceable(false);
1937 }
1938 }
1940 // Currently an object is not scalar replaceable if a LoadStore node
1941 // access its field since the field value is unknown after it.
1942 //
1943 bool has_LoadStore = false;
1944 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1945 Node *use = n->fast_out(i);
1946 if (use->is_LoadStore()) {
1947 has_LoadStore = true;
1948 break;
1949 }
1950 }
1951 // An object is not scalar replaceable if the address points
1952 // to unknown field (unknown element for arrays, offset is OffsetBot).
1953 //
1954 // Or the address may point to more then one object. This may produce
1955 // the false positive result (set not scalar replaceable)
1956 // since the flow-insensitive escape analysis can't separate
1957 // the case when stores overwrite the field's value from the case
1958 // when stores happened on different control branches.
1959 //
1960 // Note: it will disable scalar replacement in some cases:
1961 //
1962 // Point p[] = new Point[1];
1963 // p[0] = new Point(); // Will be not scalar replaced
1964 //
1965 // but it will save us from incorrect optimizations in next cases:
1966 //
1967 // Point p[] = new Point[1];
1968 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
1969 //
1970 if (ptset_size > 1 || ptset_size != 0 &&
1971 (has_LoadStore || offset == Type::OffsetBot)) {
1972 for( VectorSetI j(ptset); j.test(); ++j ) {
1973 ptnode_adr(j.elem)->set_scalar_replaceable(false);
1974 }
1975 }
1976 }
1978 // Propagate escape states to referenced nodes.
1979 bool ConnectionGraph::propagate_escape_state(GrowableArray<int>* cg_worklist,
1980 GrowableArray<uint>* worklist,
1981 PointsToNode::EscapeState esc_state) {
1982 bool has_java_obj = false;
1984 // push all nodes with the same escape state on the worklist
1985 uint cg_length = cg_worklist->length();
1986 for (uint next = 0; next < cg_length; ++next) {
1987 int nk = cg_worklist->at(next);
1988 if (ptnode_adr(nk)->escape_state() == esc_state)
1989 worklist->push(nk);
1990 }
1991 // mark all reachable nodes
1992 while (worklist->length() > 0) {
1993 PointsToNode* ptn = ptnode_adr(worklist->pop());
1994 if (ptn->node_type() == PointsToNode::JavaObject) {
1995 has_java_obj = true;
1996 }
1997 uint e_cnt = ptn->edge_count();
1998 for (uint ei = 0; ei < e_cnt; ei++) {
1999 uint npi = ptn->edge_target(ei);
2000 PointsToNode *np = ptnode_adr(npi);
2001 if (np->escape_state() < esc_state) {
2002 set_escape_state(npi, esc_state);
2003 worklist->push(npi);
2004 }
2005 }
2006 }
2007 // Has not escaping java objects
2008 return has_java_obj && (esc_state < PointsToNode::GlobalEscape);
2009 }
2011 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
2013 switch (call->Opcode()) {
2014 #ifdef ASSERT
2015 case Op_Allocate:
2016 case Op_AllocateArray:
2017 case Op_Lock:
2018 case Op_Unlock:
2019 assert(false, "should be done already");
2020 break;
2021 #endif
2022 case Op_CallLeaf:
2023 case Op_CallLeafNoFP:
2024 {
2025 // Stub calls, objects do not escape but they are not scale replaceable.
2026 // Adjust escape state for outgoing arguments.
2027 const TypeTuple * d = call->tf()->domain();
2028 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2029 const Type* at = d->field_at(i);
2030 Node *arg = call->in(i)->uncast();
2031 const Type *aat = phase->type(arg);
2032 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
2033 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::ArgEscape) {
2035 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2036 aat->isa_ptr() != NULL, "expecting an Ptr");
2037 #ifdef ASSERT
2038 if (!(call->Opcode() == Op_CallLeafNoFP &&
2039 call->as_CallLeaf()->_name != NULL &&
2040 (strstr(call->as_CallLeaf()->_name, "arraycopy") != 0) ||
2041 call->as_CallLeaf()->_name != NULL &&
2042 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
2043 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
2044 ) {
2045 call->dump();
2046 assert(false, "EA: unexpected CallLeaf");
2047 }
2048 #endif
2049 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2050 if (arg->is_AddP()) {
2051 //
2052 // The inline_native_clone() case when the arraycopy stub is called
2053 // after the allocation before Initialize and CheckCastPP nodes.
2054 //
2055 // Set AddP's base (Allocate) as not scalar replaceable since
2056 // pointer to the base (with offset) is passed as argument.
2057 //
2058 arg = get_addp_base(arg);
2059 }
2060 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2061 uint pt = j.elem;
2062 set_escape_state(pt, PointsToNode::ArgEscape);
2063 }
2064 }
2065 }
2066 break;
2067 }
2069 case Op_CallStaticJava:
2070 // For a static call, we know exactly what method is being called.
2071 // Use bytecode estimator to record the call's escape affects
2072 {
2073 ciMethod *meth = call->as_CallJava()->method();
2074 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
2075 // fall-through if not a Java method or no analyzer information
2076 if (call_analyzer != NULL) {
2077 const TypeTuple * d = call->tf()->domain();
2078 bool copy_dependencies = false;
2079 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2080 const Type* at = d->field_at(i);
2081 int k = i - TypeFunc::Parms;
2082 Node *arg = call->in(i)->uncast();
2084 if (at->isa_oopptr() != NULL &&
2085 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
2087 bool global_escapes = false;
2088 bool fields_escapes = false;
2089 if (!call_analyzer->is_arg_stack(k)) {
2090 // The argument global escapes, mark everything it could point to
2091 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2092 global_escapes = true;
2093 } else {
2094 if (!call_analyzer->is_arg_local(k)) {
2095 // The argument itself doesn't escape, but any fields might
2096 fields_escapes = true;
2097 }
2098 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2099 copy_dependencies = true;
2100 }
2102 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2103 uint pt = j.elem;
2104 if (global_escapes) {
2105 //The argument global escapes, mark everything it could point to
2106 set_escape_state(pt, PointsToNode::GlobalEscape);
2107 } else {
2108 if (fields_escapes) {
2109 // The argument itself doesn't escape, but any fields might
2110 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2111 }
2112 set_escape_state(pt, PointsToNode::ArgEscape);
2113 }
2114 }
2115 }
2116 }
2117 if (copy_dependencies)
2118 call_analyzer->copy_dependencies(_compile->dependencies());
2119 break;
2120 }
2121 }
2123 default:
2124 // Fall-through here if not a Java method or no analyzer information
2125 // or some other type of call, assume the worst case: all arguments
2126 // globally escape.
2127 {
2128 // adjust escape state for outgoing arguments
2129 const TypeTuple * d = call->tf()->domain();
2130 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2131 const Type* at = d->field_at(i);
2132 if (at->isa_oopptr() != NULL) {
2133 Node *arg = call->in(i)->uncast();
2134 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2135 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2136 uint pt = j.elem;
2137 set_escape_state(pt, PointsToNode::GlobalEscape);
2138 }
2139 }
2140 }
2141 }
2142 }
2143 }
2144 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2145 CallNode *call = resproj->in(0)->as_Call();
2146 uint call_idx = call->_idx;
2147 uint resproj_idx = resproj->_idx;
2149 switch (call->Opcode()) {
2150 case Op_Allocate:
2151 {
2152 Node *k = call->in(AllocateNode::KlassNode);
2153 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2154 assert(kt != NULL, "TypeKlassPtr required.");
2155 ciKlass* cik = kt->klass();
2157 PointsToNode::EscapeState es;
2158 uint edge_to;
2159 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2160 !cik->is_instance_klass() || // StressReflectiveCode
2161 cik->as_instance_klass()->has_finalizer()) {
2162 es = PointsToNode::GlobalEscape;
2163 edge_to = _phantom_object; // Could not be worse
2164 } else {
2165 es = PointsToNode::NoEscape;
2166 edge_to = call_idx;
2167 assert(ptnode_adr(call_idx)->scalar_replaceable(), "sanity");
2168 }
2169 set_escape_state(call_idx, es);
2170 add_pointsto_edge(resproj_idx, edge_to);
2171 _processed.set(resproj_idx);
2172 break;
2173 }
2175 case Op_AllocateArray:
2176 {
2178 Node *k = call->in(AllocateNode::KlassNode);
2179 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2180 assert(kt != NULL, "TypeKlassPtr required.");
2181 ciKlass* cik = kt->klass();
2183 PointsToNode::EscapeState es;
2184 uint edge_to;
2185 if (!cik->is_array_klass()) { // StressReflectiveCode
2186 es = PointsToNode::GlobalEscape;
2187 edge_to = _phantom_object;
2188 } else {
2189 es = PointsToNode::NoEscape;
2190 edge_to = call_idx;
2191 assert(ptnode_adr(call_idx)->scalar_replaceable(), "sanity");
2192 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2193 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2194 // Not scalar replaceable if the length is not constant or too big.
2195 ptnode_adr(call_idx)->set_scalar_replaceable(false);
2196 }
2197 }
2198 set_escape_state(call_idx, es);
2199 add_pointsto_edge(resproj_idx, edge_to);
2200 _processed.set(resproj_idx);
2201 break;
2202 }
2204 case Op_CallStaticJava:
2205 // For a static call, we know exactly what method is being called.
2206 // Use bytecode estimator to record whether the call's return value escapes
2207 {
2208 bool done = true;
2209 const TypeTuple *r = call->tf()->range();
2210 const Type* ret_type = NULL;
2212 if (r->cnt() > TypeFunc::Parms)
2213 ret_type = r->field_at(TypeFunc::Parms);
2215 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2216 // _multianewarray functions return a TypeRawPtr.
2217 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2218 _processed.set(resproj_idx);
2219 break; // doesn't return a pointer type
2220 }
2221 ciMethod *meth = call->as_CallJava()->method();
2222 const TypeTuple * d = call->tf()->domain();
2223 if (meth == NULL) {
2224 // not a Java method, assume global escape
2225 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2226 add_pointsto_edge(resproj_idx, _phantom_object);
2227 } else {
2228 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2229 bool copy_dependencies = false;
2231 if (call_analyzer->is_return_allocated()) {
2232 // Returns a newly allocated unescaped object, simply
2233 // update dependency information.
2234 // Mark it as NoEscape so that objects referenced by
2235 // it's fields will be marked as NoEscape at least.
2236 set_escape_state(call_idx, PointsToNode::NoEscape);
2237 ptnode_adr(call_idx)->set_scalar_replaceable(false);
2238 add_pointsto_edge(resproj_idx, call_idx);
2239 copy_dependencies = true;
2240 } else if (call_analyzer->is_return_local()) {
2241 // determine whether any arguments are returned
2242 set_escape_state(call_idx, PointsToNode::ArgEscape);
2243 bool ret_arg = false;
2244 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2245 const Type* at = d->field_at(i);
2247 if (at->isa_oopptr() != NULL) {
2248 Node *arg = call->in(i)->uncast();
2250 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2251 ret_arg = true;
2252 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2253 if (arg_esp->node_type() == PointsToNode::UnknownType)
2254 done = false;
2255 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2256 add_pointsto_edge(resproj_idx, arg->_idx);
2257 else
2258 add_deferred_edge(resproj_idx, arg->_idx);
2259 }
2260 }
2261 }
2262 if (done && !ret_arg) {
2263 // Returns unknown object.
2264 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2265 add_pointsto_edge(resproj_idx, _phantom_object);
2266 }
2267 if (done) {
2268 copy_dependencies = true;
2269 }
2270 } else {
2271 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2272 add_pointsto_edge(resproj_idx, _phantom_object);
2273 }
2274 if (copy_dependencies)
2275 call_analyzer->copy_dependencies(_compile->dependencies());
2276 }
2277 if (done)
2278 _processed.set(resproj_idx);
2279 break;
2280 }
2282 default:
2283 // Some other type of call, assume the worst case that the
2284 // returned value, if any, globally escapes.
2285 {
2286 const TypeTuple *r = call->tf()->range();
2287 if (r->cnt() > TypeFunc::Parms) {
2288 const Type* ret_type = r->field_at(TypeFunc::Parms);
2290 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2291 // _multianewarray functions return a TypeRawPtr.
2292 if (ret_type->isa_ptr() != NULL) {
2293 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2294 add_pointsto_edge(resproj_idx, _phantom_object);
2295 }
2296 }
2297 _processed.set(resproj_idx);
2298 }
2299 }
2300 }
2302 // Populate Connection Graph with Ideal nodes and create simple
2303 // connection graph edges (do not need to check the node_type of inputs
2304 // or to call PointsTo() to walk the connection graph).
2305 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2306 if (_processed.test(n->_idx))
2307 return; // No need to redefine node's state.
2309 if (n->is_Call()) {
2310 // Arguments to allocation and locking don't escape.
2311 if (n->is_Allocate()) {
2312 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2313 record_for_optimizer(n);
2314 } else if (n->is_Lock() || n->is_Unlock()) {
2315 // Put Lock and Unlock nodes on IGVN worklist to process them during
2316 // the first IGVN optimization when escape information is still available.
2317 record_for_optimizer(n);
2318 _processed.set(n->_idx);
2319 } else {
2320 // Don't mark as processed since call's arguments have to be processed.
2321 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2322 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2324 // Check if a call returns an object.
2325 const TypeTuple *r = n->as_Call()->tf()->range();
2326 if (r->cnt() > TypeFunc::Parms &&
2327 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2328 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2329 nt = PointsToNode::JavaObject;
2330 if (!n->is_CallStaticJava()) {
2331 // Since the called mathod is statically unknown assume
2332 // the worst case that the returned value globally escapes.
2333 es = PointsToNode::GlobalEscape;
2334 }
2335 }
2336 add_node(n, nt, es, false);
2337 }
2338 return;
2339 }
2341 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2342 // ThreadLocal has RawPrt type.
2343 switch (n->Opcode()) {
2344 case Op_AddP:
2345 {
2346 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2347 break;
2348 }
2349 case Op_CastX2P:
2350 { // "Unsafe" memory access.
2351 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2352 break;
2353 }
2354 case Op_CastPP:
2355 case Op_CheckCastPP:
2356 case Op_EncodeP:
2357 case Op_DecodeN:
2358 {
2359 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2360 int ti = n->in(1)->_idx;
2361 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2362 if (nt == PointsToNode::UnknownType) {
2363 _delayed_worklist.push(n); // Process it later.
2364 break;
2365 } else if (nt == PointsToNode::JavaObject) {
2366 add_pointsto_edge(n->_idx, ti);
2367 } else {
2368 add_deferred_edge(n->_idx, ti);
2369 }
2370 _processed.set(n->_idx);
2371 break;
2372 }
2373 case Op_ConP:
2374 {
2375 // assume all pointer constants globally escape except for null
2376 PointsToNode::EscapeState es;
2377 if (phase->type(n) == TypePtr::NULL_PTR)
2378 es = PointsToNode::NoEscape;
2379 else
2380 es = PointsToNode::GlobalEscape;
2382 add_node(n, PointsToNode::JavaObject, es, true);
2383 break;
2384 }
2385 case Op_ConN:
2386 {
2387 // assume all narrow oop constants globally escape except for null
2388 PointsToNode::EscapeState es;
2389 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2390 es = PointsToNode::NoEscape;
2391 else
2392 es = PointsToNode::GlobalEscape;
2394 add_node(n, PointsToNode::JavaObject, es, true);
2395 break;
2396 }
2397 case Op_CreateEx:
2398 {
2399 // assume that all exception objects globally escape
2400 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2401 break;
2402 }
2403 case Op_LoadKlass:
2404 case Op_LoadNKlass:
2405 {
2406 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2407 break;
2408 }
2409 case Op_LoadP:
2410 case Op_LoadN:
2411 {
2412 const Type *t = phase->type(n);
2413 if (t->make_ptr() == NULL) {
2414 _processed.set(n->_idx);
2415 return;
2416 }
2417 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2418 break;
2419 }
2420 case Op_Parm:
2421 {
2422 _processed.set(n->_idx); // No need to redefine it state.
2423 uint con = n->as_Proj()->_con;
2424 if (con < TypeFunc::Parms)
2425 return;
2426 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2427 if (t->isa_ptr() == NULL)
2428 return;
2429 // We have to assume all input parameters globally escape
2430 // (Note: passing 'false' since _processed is already set).
2431 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2432 break;
2433 }
2434 case Op_Phi:
2435 {
2436 const Type *t = n->as_Phi()->type();
2437 if (t->make_ptr() == NULL) {
2438 // nothing to do if not an oop or narrow oop
2439 _processed.set(n->_idx);
2440 return;
2441 }
2442 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2443 uint i;
2444 for (i = 1; i < n->req() ; i++) {
2445 Node* in = n->in(i);
2446 if (in == NULL)
2447 continue; // ignore NULL
2448 in = in->uncast();
2449 if (in->is_top() || in == n)
2450 continue; // ignore top or inputs which go back this node
2451 int ti = in->_idx;
2452 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2453 if (nt == PointsToNode::UnknownType) {
2454 break;
2455 } else if (nt == PointsToNode::JavaObject) {
2456 add_pointsto_edge(n->_idx, ti);
2457 } else {
2458 add_deferred_edge(n->_idx, ti);
2459 }
2460 }
2461 if (i >= n->req())
2462 _processed.set(n->_idx);
2463 else
2464 _delayed_worklist.push(n);
2465 break;
2466 }
2467 case Op_Proj:
2468 {
2469 // we are only interested in the oop result projection from a call
2470 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2471 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2472 assert(r->cnt() > TypeFunc::Parms, "sanity");
2473 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2474 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2475 int ti = n->in(0)->_idx;
2476 // The call may not be registered yet (since not all its inputs are registered)
2477 // if this is the projection from backbranch edge of Phi.
2478 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2479 process_call_result(n->as_Proj(), phase);
2480 }
2481 if (!_processed.test(n->_idx)) {
2482 // The call's result may need to be processed later if the call
2483 // returns it's argument and the argument is not processed yet.
2484 _delayed_worklist.push(n);
2485 }
2486 break;
2487 }
2488 }
2489 _processed.set(n->_idx);
2490 break;
2491 }
2492 case Op_Return:
2493 {
2494 if( n->req() > TypeFunc::Parms &&
2495 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2496 // Treat Return value as LocalVar with GlobalEscape escape state.
2497 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2498 int ti = n->in(TypeFunc::Parms)->_idx;
2499 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2500 if (nt == PointsToNode::UnknownType) {
2501 _delayed_worklist.push(n); // Process it later.
2502 break;
2503 } else if (nt == PointsToNode::JavaObject) {
2504 add_pointsto_edge(n->_idx, ti);
2505 } else {
2506 add_deferred_edge(n->_idx, ti);
2507 }
2508 }
2509 _processed.set(n->_idx);
2510 break;
2511 }
2512 case Op_StoreP:
2513 case Op_StoreN:
2514 {
2515 const Type *adr_type = phase->type(n->in(MemNode::Address));
2516 adr_type = adr_type->make_ptr();
2517 if (adr_type->isa_oopptr()) {
2518 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2519 } else {
2520 Node* adr = n->in(MemNode::Address);
2521 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2522 adr->in(AddPNode::Address)->is_Proj() &&
2523 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2524 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2525 // We are computing a raw address for a store captured
2526 // by an Initialize compute an appropriate address type.
2527 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2528 assert(offs != Type::OffsetBot, "offset must be a constant");
2529 } else {
2530 _processed.set(n->_idx);
2531 return;
2532 }
2533 }
2534 break;
2535 }
2536 case Op_StorePConditional:
2537 case Op_CompareAndSwapP:
2538 case Op_CompareAndSwapN:
2539 {
2540 const Type *adr_type = phase->type(n->in(MemNode::Address));
2541 adr_type = adr_type->make_ptr();
2542 if (adr_type->isa_oopptr()) {
2543 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2544 } else {
2545 _processed.set(n->_idx);
2546 return;
2547 }
2548 break;
2549 }
2550 case Op_AryEq:
2551 case Op_StrComp:
2552 case Op_StrEquals:
2553 case Op_StrIndexOf:
2554 {
2555 // char[] arrays passed to string intrinsics are not scalar replaceable.
2556 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2557 break;
2558 }
2559 case Op_ThreadLocal:
2560 {
2561 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2562 break;
2563 }
2564 default:
2565 ;
2566 // nothing to do
2567 }
2568 return;
2569 }
2571 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2572 uint n_idx = n->_idx;
2573 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2575 // Don't set processed bit for AddP, LoadP, StoreP since
2576 // they may need more then one pass to process.
2577 // Also don't mark as processed Call nodes since their
2578 // arguments may need more then one pass to process.
2579 if (_processed.test(n_idx))
2580 return; // No need to redefine node's state.
2582 if (n->is_Call()) {
2583 CallNode *call = n->as_Call();
2584 process_call_arguments(call, phase);
2585 return;
2586 }
2588 switch (n->Opcode()) {
2589 case Op_AddP:
2590 {
2591 Node *base = get_addp_base(n);
2592 // Create a field edge to this node from everything base could point to.
2593 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) {
2594 uint pt = i.elem;
2595 add_field_edge(pt, n_idx, address_offset(n, phase));
2596 }
2597 break;
2598 }
2599 case Op_CastX2P:
2600 {
2601 assert(false, "Op_CastX2P");
2602 break;
2603 }
2604 case Op_CastPP:
2605 case Op_CheckCastPP:
2606 case Op_EncodeP:
2607 case Op_DecodeN:
2608 {
2609 int ti = n->in(1)->_idx;
2610 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2611 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2612 add_pointsto_edge(n_idx, ti);
2613 } else {
2614 add_deferred_edge(n_idx, ti);
2615 }
2616 _processed.set(n_idx);
2617 break;
2618 }
2619 case Op_ConP:
2620 {
2621 assert(false, "Op_ConP");
2622 break;
2623 }
2624 case Op_ConN:
2625 {
2626 assert(false, "Op_ConN");
2627 break;
2628 }
2629 case Op_CreateEx:
2630 {
2631 assert(false, "Op_CreateEx");
2632 break;
2633 }
2634 case Op_LoadKlass:
2635 case Op_LoadNKlass:
2636 {
2637 assert(false, "Op_LoadKlass");
2638 break;
2639 }
2640 case Op_LoadP:
2641 case Op_LoadN:
2642 {
2643 const Type *t = phase->type(n);
2644 #ifdef ASSERT
2645 if (t->make_ptr() == NULL)
2646 assert(false, "Op_LoadP");
2647 #endif
2649 Node* adr = n->in(MemNode::Address)->uncast();
2650 Node* adr_base;
2651 if (adr->is_AddP()) {
2652 adr_base = get_addp_base(adr);
2653 } else {
2654 adr_base = adr;
2655 }
2657 // For everything "adr_base" could point to, create a deferred edge from
2658 // this node to each field with the same offset.
2659 int offset = address_offset(adr, phase);
2660 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2661 uint pt = i.elem;
2662 add_deferred_edge_to_fields(n_idx, pt, offset);
2663 }
2664 break;
2665 }
2666 case Op_Parm:
2667 {
2668 assert(false, "Op_Parm");
2669 break;
2670 }
2671 case Op_Phi:
2672 {
2673 #ifdef ASSERT
2674 const Type *t = n->as_Phi()->type();
2675 if (t->make_ptr() == NULL)
2676 assert(false, "Op_Phi");
2677 #endif
2678 for (uint i = 1; i < n->req() ; i++) {
2679 Node* in = n->in(i);
2680 if (in == NULL)
2681 continue; // ignore NULL
2682 in = in->uncast();
2683 if (in->is_top() || in == n)
2684 continue; // ignore top or inputs which go back this node
2685 int ti = in->_idx;
2686 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2687 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
2688 if (nt == PointsToNode::JavaObject) {
2689 add_pointsto_edge(n_idx, ti);
2690 } else {
2691 add_deferred_edge(n_idx, ti);
2692 }
2693 }
2694 _processed.set(n_idx);
2695 break;
2696 }
2697 case Op_Proj:
2698 {
2699 // we are only interested in the oop result projection from a call
2700 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2701 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
2702 "all nodes should be registered");
2703 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2704 assert(r->cnt() > TypeFunc::Parms, "sanity");
2705 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2706 process_call_result(n->as_Proj(), phase);
2707 assert(_processed.test(n_idx), "all call results should be processed");
2708 break;
2709 }
2710 }
2711 assert(false, "Op_Proj");
2712 break;
2713 }
2714 case Op_Return:
2715 {
2716 #ifdef ASSERT
2717 if( n->req() <= TypeFunc::Parms ||
2718 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2719 assert(false, "Op_Return");
2720 }
2721 #endif
2722 int ti = n->in(TypeFunc::Parms)->_idx;
2723 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
2724 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2725 add_pointsto_edge(n_idx, ti);
2726 } else {
2727 add_deferred_edge(n_idx, ti);
2728 }
2729 _processed.set(n_idx);
2730 break;
2731 }
2732 case Op_StoreP:
2733 case Op_StoreN:
2734 case Op_StorePConditional:
2735 case Op_CompareAndSwapP:
2736 case Op_CompareAndSwapN:
2737 {
2738 Node *adr = n->in(MemNode::Address);
2739 const Type *adr_type = phase->type(adr)->make_ptr();
2740 #ifdef ASSERT
2741 if (!adr_type->isa_oopptr())
2742 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
2743 #endif
2745 assert(adr->is_AddP(), "expecting an AddP");
2746 Node *adr_base = get_addp_base(adr);
2747 Node *val = n->in(MemNode::ValueIn)->uncast();
2748 // For everything "adr_base" could point to, create a deferred edge
2749 // to "val" from each field with the same offset.
2750 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2751 uint pt = i.elem;
2752 add_edge_from_fields(pt, val->_idx, address_offset(adr, phase));
2753 }
2754 break;
2755 }
2756 case Op_AryEq:
2757 case Op_StrComp:
2758 case Op_StrEquals:
2759 case Op_StrIndexOf:
2760 {
2761 // char[] arrays passed to string intrinsic do not escape but
2762 // they are not scalar replaceable. Adjust escape state for them.
2763 // Start from in(2) edge since in(1) is memory edge.
2764 for (uint i = 2; i < n->req(); i++) {
2765 Node* adr = n->in(i)->uncast();
2766 const Type *at = phase->type(adr);
2767 if (!adr->is_top() && at->isa_ptr()) {
2768 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
2769 at->isa_ptr() != NULL, "expecting an Ptr");
2770 if (adr->is_AddP()) {
2771 adr = get_addp_base(adr);
2772 }
2773 // Mark as ArgEscape everything "adr" could point to.
2774 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
2775 }
2776 }
2777 _processed.set(n_idx);
2778 break;
2779 }
2780 case Op_ThreadLocal:
2781 {
2782 assert(false, "Op_ThreadLocal");
2783 break;
2784 }
2785 default:
2786 // This method should be called only for EA specific nodes.
2787 ShouldNotReachHere();
2788 }
2789 }
2791 #ifndef PRODUCT
2792 void ConnectionGraph::dump() {
2793 bool first = true;
2795 uint size = nodes_size();
2796 for (uint ni = 0; ni < size; ni++) {
2797 PointsToNode *ptn = ptnode_adr(ni);
2798 PointsToNode::NodeType ptn_type = ptn->node_type();
2800 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
2801 continue;
2802 PointsToNode::EscapeState es = escape_state(ptn->_node);
2803 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
2804 if (first) {
2805 tty->cr();
2806 tty->print("======== Connection graph for ");
2807 _compile->method()->print_short_name();
2808 tty->cr();
2809 first = false;
2810 }
2811 tty->print("%6d ", ni);
2812 ptn->dump();
2813 // Print all locals which reference this allocation
2814 for (uint li = ni; li < size; li++) {
2815 PointsToNode *ptn_loc = ptnode_adr(li);
2816 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
2817 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
2818 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
2819 ptnode_adr(li)->dump(false);
2820 }
2821 }
2822 if (Verbose) {
2823 // Print all fields which reference this allocation
2824 for (uint i = 0; i < ptn->edge_count(); i++) {
2825 uint ei = ptn->edge_target(i);
2826 ptnode_adr(ei)->dump(false);
2827 }
2828 }
2829 tty->cr();
2830 }
2831 }
2832 }
2833 #endif