Fri, 10 Feb 2012 12:53:43 -0800
7129284: +DoEscapeAnalysis regression w/ early build of 7u4 (HotSpot 23) on Linux
Summary: Removed code which tried to create edges from fields of destination objects of arraycopy to fields of source objects. Added 30 sec time limit for EA graph construction.
Reviewed-by: never
1 /*
2 * Copyright (c) 2005, 2012, 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 _pcmp_neq = NULL; // Should be initialized
123 _pcmp_eq = NULL;
124 }
126 void ConnectionGraph::add_pointsto_edge(uint from_i, uint to_i) {
127 PointsToNode *f = ptnode_adr(from_i);
128 PointsToNode *t = ptnode_adr(to_i);
130 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
131 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
132 assert(t->node_type() == PointsToNode::JavaObject, "invalid destination of PointsTo edge");
133 if (to_i == _phantom_object) { // Quick test for most common object
134 if (f->has_unknown_ptr()) {
135 return;
136 } else {
137 f->set_has_unknown_ptr();
138 }
139 }
140 add_edge(f, to_i, PointsToNode::PointsToEdge);
141 }
143 void ConnectionGraph::add_deferred_edge(uint from_i, uint to_i) {
144 PointsToNode *f = ptnode_adr(from_i);
145 PointsToNode *t = ptnode_adr(to_i);
147 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
148 assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
149 assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
150 // don't add a self-referential edge, this can occur during removal of
151 // deferred edges
152 if (from_i != to_i)
153 add_edge(f, to_i, PointsToNode::DeferredEdge);
154 }
156 int ConnectionGraph::address_offset(Node* adr, PhaseTransform *phase) {
157 const Type *adr_type = phase->type(adr);
158 if (adr->is_AddP() && adr_type->isa_oopptr() == NULL &&
159 adr->in(AddPNode::Address)->is_Proj() &&
160 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
161 // We are computing a raw address for a store captured by an Initialize
162 // compute an appropriate address type. AddP cases #3 and #5 (see below).
163 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
164 assert(offs != Type::OffsetBot ||
165 adr->in(AddPNode::Address)->in(0)->is_AllocateArray(),
166 "offset must be a constant or it is initialization of array");
167 return offs;
168 }
169 const TypePtr *t_ptr = adr_type->isa_ptr();
170 assert(t_ptr != NULL, "must be a pointer type");
171 return t_ptr->offset();
172 }
174 void ConnectionGraph::add_field_edge(uint from_i, uint to_i, int offset) {
175 // Don't add fields to NULL pointer.
176 if (is_null_ptr(from_i))
177 return;
178 PointsToNode *f = ptnode_adr(from_i);
179 PointsToNode *t = ptnode_adr(to_i);
181 assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
182 assert(f->node_type() == PointsToNode::JavaObject, "invalid destination of Field edge");
183 assert(t->node_type() == PointsToNode::Field, "invalid destination of Field edge");
184 assert (t->offset() == -1 || t->offset() == offset, "conflicting field offsets");
185 t->set_offset(offset);
187 add_edge(f, to_i, PointsToNode::FieldEdge);
188 }
190 void ConnectionGraph::set_escape_state(uint ni, PointsToNode::EscapeState es) {
191 // Don't change non-escaping state of NULL pointer.
192 if (is_null_ptr(ni))
193 return;
194 PointsToNode *npt = ptnode_adr(ni);
195 PointsToNode::EscapeState old_es = npt->escape_state();
196 if (es > old_es)
197 npt->set_escape_state(es);
198 }
200 void ConnectionGraph::add_node(Node *n, PointsToNode::NodeType nt,
201 PointsToNode::EscapeState es, bool done) {
202 PointsToNode* ptadr = ptnode_adr(n->_idx);
203 ptadr->_node = n;
204 ptadr->set_node_type(nt);
206 // inline set_escape_state(idx, es);
207 PointsToNode::EscapeState old_es = ptadr->escape_state();
208 if (es > old_es)
209 ptadr->set_escape_state(es);
211 if (done)
212 _processed.set(n->_idx);
213 }
215 PointsToNode::EscapeState ConnectionGraph::escape_state(Node *n) {
216 uint idx = n->_idx;
217 PointsToNode::EscapeState es;
219 // If we are still collecting or there were no non-escaping allocations
220 // we don't know the answer yet
221 if (_collecting)
222 return PointsToNode::UnknownEscape;
224 // if the node was created after the escape computation, return
225 // UnknownEscape
226 if (idx >= nodes_size())
227 return PointsToNode::UnknownEscape;
229 es = ptnode_adr(idx)->escape_state();
231 // if we have already computed a value, return it
232 if (es != PointsToNode::UnknownEscape &&
233 ptnode_adr(idx)->node_type() == PointsToNode::JavaObject)
234 return es;
236 // PointsTo() calls n->uncast() which can return a new ideal node.
237 if (n->uncast()->_idx >= nodes_size())
238 return PointsToNode::UnknownEscape;
240 PointsToNode::EscapeState orig_es = es;
242 // compute max escape state of anything this node could point to
243 for(VectorSetI i(PointsTo(n)); i.test() && es != PointsToNode::GlobalEscape; ++i) {
244 uint pt = i.elem;
245 PointsToNode::EscapeState pes = ptnode_adr(pt)->escape_state();
246 if (pes > es)
247 es = pes;
248 }
249 if (orig_es != es) {
250 // cache the computed escape state
251 assert(es > orig_es, "should have computed an escape state");
252 set_escape_state(idx, es);
253 } // orig_es could be PointsToNode::UnknownEscape
254 return es;
255 }
257 VectorSet* ConnectionGraph::PointsTo(Node * n) {
258 pt_ptset.Reset();
259 pt_visited.Reset();
260 pt_worklist.clear();
262 #ifdef ASSERT
263 Node *orig_n = n;
264 #endif
266 n = n->uncast();
267 PointsToNode* npt = ptnode_adr(n->_idx);
269 // If we have a JavaObject, return just that object
270 if (npt->node_type() == PointsToNode::JavaObject) {
271 pt_ptset.set(n->_idx);
272 return &pt_ptset;
273 }
274 #ifdef ASSERT
275 if (npt->_node == NULL) {
276 if (orig_n != n)
277 orig_n->dump();
278 n->dump();
279 assert(npt->_node != NULL, "unregistered node");
280 }
281 #endif
282 pt_worklist.push(n->_idx);
283 while(pt_worklist.length() > 0) {
284 int ni = pt_worklist.pop();
285 if (pt_visited.test_set(ni))
286 continue;
288 PointsToNode* pn = ptnode_adr(ni);
289 // ensure that all inputs of a Phi have been processed
290 assert(!_collecting || !pn->_node->is_Phi() || _processed.test(ni),"");
292 int edges_processed = 0;
293 uint e_cnt = pn->edge_count();
294 for (uint e = 0; e < e_cnt; e++) {
295 uint etgt = pn->edge_target(e);
296 PointsToNode::EdgeType et = pn->edge_type(e);
297 if (et == PointsToNode::PointsToEdge) {
298 pt_ptset.set(etgt);
299 edges_processed++;
300 } else if (et == PointsToNode::DeferredEdge) {
301 pt_worklist.push(etgt);
302 edges_processed++;
303 } else {
304 assert(false,"neither PointsToEdge or DeferredEdge");
305 }
306 }
307 if (edges_processed == 0) {
308 // no deferred or pointsto edges found. Assume the value was set
309 // outside this method. Add the phantom object to the pointsto set.
310 pt_ptset.set(_phantom_object);
311 }
312 }
313 return &pt_ptset;
314 }
316 void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
317 // This method is most expensive during ConnectionGraph construction.
318 // Reuse vectorSet and an additional growable array for deferred edges.
319 deferred_edges->clear();
320 visited->Reset();
322 visited->set(ni);
323 PointsToNode *ptn = ptnode_adr(ni);
324 assert(ptn->node_type() == PointsToNode::LocalVar ||
325 ptn->node_type() == PointsToNode::Field, "sanity");
326 assert(ptn->edge_count() != 0, "should have at least phantom_object");
328 // Mark current edges as visited and move deferred edges to separate array.
329 for (uint i = 0; i < ptn->edge_count(); ) {
330 uint t = ptn->edge_target(i);
331 #ifdef ASSERT
332 assert(!visited->test_set(t), "expecting no duplications");
333 #else
334 visited->set(t);
335 #endif
336 if (ptn->edge_type(i) == PointsToNode::DeferredEdge) {
337 ptn->remove_edge(t, PointsToNode::DeferredEdge);
338 deferred_edges->append(t);
339 } else {
340 i++;
341 }
342 }
343 for (int next = 0; next < deferred_edges->length(); ++next) {
344 uint t = deferred_edges->at(next);
345 PointsToNode *ptt = ptnode_adr(t);
346 uint e_cnt = ptt->edge_count();
347 assert(e_cnt != 0, "should have at least phantom_object");
348 for (uint e = 0; e < e_cnt; e++) {
349 uint etgt = ptt->edge_target(e);
350 if (visited->test_set(etgt))
351 continue;
353 PointsToNode::EdgeType et = ptt->edge_type(e);
354 if (et == PointsToNode::PointsToEdge) {
355 add_pointsto_edge(ni, etgt);
356 } else if (et == PointsToNode::DeferredEdge) {
357 deferred_edges->append(etgt);
358 } else {
359 assert(false,"invalid connection graph");
360 }
361 }
362 }
363 if (ptn->edge_count() == 0) {
364 // No pointsto edges found after deferred edges are removed.
365 // For example, in the next case where call is replaced
366 // with uncommon trap and as result array's load references
367 // itself through deferred edges:
368 //
369 // A a = b[i];
370 // if (c!=null) a = c.foo();
371 // b[i] = a;
372 //
373 // Assume the value was set outside this method and
374 // add edge to phantom object.
375 add_pointsto_edge(ni, _phantom_object);
376 }
377 }
380 // Add an edge to node given by "to_i" from any field of adr_i whose offset
381 // matches "offset" A deferred edge is added if to_i is a LocalVar, and
382 // a pointsto edge is added if it is a JavaObject
384 void ConnectionGraph::add_edge_from_fields(uint adr_i, uint to_i, int offs) {
385 // No fields for NULL pointer.
386 if (is_null_ptr(adr_i)) {
387 return;
388 }
389 PointsToNode* an = ptnode_adr(adr_i);
390 PointsToNode* to = ptnode_adr(to_i);
391 bool deferred = (to->node_type() == PointsToNode::LocalVar);
392 bool escaped = (to_i == _phantom_object) && (offs == Type::OffsetTop);
393 if (escaped) {
394 // Values in fields escaped during call.
395 assert(an->escape_state() >= PointsToNode::ArgEscape, "sanity");
396 offs = Type::OffsetBot;
397 }
398 for (uint fe = 0; fe < an->edge_count(); fe++) {
399 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
400 int fi = an->edge_target(fe);
401 if (escaped) {
402 set_escape_state(fi, PointsToNode::GlobalEscape);
403 }
404 PointsToNode* pf = ptnode_adr(fi);
405 int po = pf->offset();
406 if (po == offs || po == Type::OffsetBot || offs == Type::OffsetBot) {
407 if (deferred)
408 add_deferred_edge(fi, to_i);
409 else
410 add_pointsto_edge(fi, to_i);
411 }
412 }
413 }
415 // Add a deferred edge from node given by "from_i" to any field of adr_i
416 // whose offset matches "offset".
417 void ConnectionGraph::add_deferred_edge_to_fields(uint from_i, uint adr_i, int offs) {
418 // No fields for NULL pointer.
419 if (is_null_ptr(adr_i)) {
420 return;
421 }
422 if (adr_i == _phantom_object) {
423 // Add only one edge for unknown object.
424 add_pointsto_edge(from_i, _phantom_object);
425 return;
426 }
427 PointsToNode* an = ptnode_adr(adr_i);
428 bool is_alloc = an->_node->is_Allocate();
429 for (uint fe = 0; fe < an->edge_count(); fe++) {
430 assert(an->edge_type(fe) == PointsToNode::FieldEdge, "expecting a field edge");
431 int fi = an->edge_target(fe);
432 PointsToNode* pf = ptnode_adr(fi);
433 int offset = pf->offset();
434 if (!is_alloc) {
435 // Assume the field was set outside this method if it is not Allocation
436 add_pointsto_edge(fi, _phantom_object);
437 }
438 if (offset == offs || offset == Type::OffsetBot || offs == Type::OffsetBot) {
439 add_deferred_edge(from_i, fi);
440 }
441 }
442 // Some fields references (AddP) may still be missing
443 // until Connection Graph construction is complete.
444 // For example, loads from RAW pointers with offset 0
445 // which don't have AddP.
446 // A reference to phantom_object will be added if
447 // a field reference is still missing after completing
448 // Connection Graph (see remove_deferred()).
449 }
451 // Helper functions
453 static Node* get_addp_base(Node *addp) {
454 assert(addp->is_AddP(), "must be AddP");
455 //
456 // AddP cases for Base and Address inputs:
457 // case #1. Direct object's field reference:
458 // Allocate
459 // |
460 // Proj #5 ( oop result )
461 // |
462 // CheckCastPP (cast to instance type)
463 // | |
464 // AddP ( base == address )
465 //
466 // case #2. Indirect object's field reference:
467 // Phi
468 // |
469 // CastPP (cast to instance type)
470 // | |
471 // AddP ( base == address )
472 //
473 // case #3. Raw object's field reference for Initialize node:
474 // Allocate
475 // |
476 // Proj #5 ( oop result )
477 // top |
478 // \ |
479 // AddP ( base == top )
480 //
481 // case #4. Array's element reference:
482 // {CheckCastPP | CastPP}
483 // | | |
484 // | AddP ( array's element offset )
485 // | |
486 // AddP ( array's offset )
487 //
488 // case #5. Raw object's field reference for arraycopy stub call:
489 // The inline_native_clone() case when the arraycopy stub is called
490 // after the allocation before Initialize and CheckCastPP nodes.
491 // Allocate
492 // |
493 // Proj #5 ( oop result )
494 // | |
495 // AddP ( base == address )
496 //
497 // case #6. Constant Pool, ThreadLocal, CastX2P or
498 // Raw object's field reference:
499 // {ConP, ThreadLocal, CastX2P, raw Load}
500 // top |
501 // \ |
502 // AddP ( base == top )
503 //
504 // case #7. Klass's field reference.
505 // LoadKlass
506 // | |
507 // AddP ( base == address )
508 //
509 // case #8. narrow Klass's field reference.
510 // LoadNKlass
511 // |
512 // DecodeN
513 // | |
514 // AddP ( base == address )
515 //
516 Node *base = addp->in(AddPNode::Base)->uncast();
517 if (base->is_top()) { // The AddP case #3 and #6.
518 base = addp->in(AddPNode::Address)->uncast();
519 while (base->is_AddP()) {
520 // Case #6 (unsafe access) may have several chained AddP nodes.
521 assert(base->in(AddPNode::Base)->is_top(), "expected unsafe access address only");
522 base = base->in(AddPNode::Address)->uncast();
523 }
524 assert(base->Opcode() == Op_ConP || base->Opcode() == Op_ThreadLocal ||
525 base->Opcode() == Op_CastX2P || base->is_DecodeN() ||
526 (base->is_Mem() && base->bottom_type() == TypeRawPtr::NOTNULL) ||
527 (base->is_Proj() && base->in(0)->is_Allocate()), "sanity");
528 }
529 return base;
530 }
532 static Node* find_second_addp(Node* addp, Node* n) {
533 assert(addp->is_AddP() && addp->outcnt() > 0, "Don't process dead nodes");
535 Node* addp2 = addp->raw_out(0);
536 if (addp->outcnt() == 1 && addp2->is_AddP() &&
537 addp2->in(AddPNode::Base) == n &&
538 addp2->in(AddPNode::Address) == addp) {
540 assert(addp->in(AddPNode::Base) == n, "expecting the same base");
541 //
542 // Find array's offset to push it on worklist first and
543 // as result process an array's element offset first (pushed second)
544 // to avoid CastPP for the array's offset.
545 // Otherwise the inserted CastPP (LocalVar) will point to what
546 // the AddP (Field) points to. Which would be wrong since
547 // the algorithm expects the CastPP has the same point as
548 // as AddP's base CheckCastPP (LocalVar).
549 //
550 // ArrayAllocation
551 // |
552 // CheckCastPP
553 // |
554 // memProj (from ArrayAllocation CheckCastPP)
555 // | ||
556 // | || Int (element index)
557 // | || | ConI (log(element size))
558 // | || | /
559 // | || LShift
560 // | || /
561 // | AddP (array's element offset)
562 // | |
563 // | | ConI (array's offset: #12(32-bits) or #24(64-bits))
564 // | / /
565 // AddP (array's offset)
566 // |
567 // Load/Store (memory operation on array's element)
568 //
569 return addp2;
570 }
571 return NULL;
572 }
574 //
575 // Adjust the type and inputs of an AddP which computes the
576 // address of a field of an instance
577 //
578 bool ConnectionGraph::split_AddP(Node *addp, Node *base, PhaseGVN *igvn) {
579 const TypeOopPtr *base_t = igvn->type(base)->isa_oopptr();
580 assert(base_t != NULL && base_t->is_known_instance(), "expecting instance oopptr");
581 const TypeOopPtr *t = igvn->type(addp)->isa_oopptr();
582 if (t == NULL) {
583 // We are computing a raw address for a store captured by an Initialize
584 // compute an appropriate address type (cases #3 and #5).
585 assert(igvn->type(addp) == TypeRawPtr::NOTNULL, "must be raw pointer");
586 assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
587 intptr_t offs = (int)igvn->find_intptr_t_con(addp->in(AddPNode::Offset), Type::OffsetBot);
588 assert(offs != Type::OffsetBot, "offset must be a constant");
589 t = base_t->add_offset(offs)->is_oopptr();
590 }
591 int inst_id = base_t->instance_id();
592 assert(!t->is_known_instance() || t->instance_id() == inst_id,
593 "old type must be non-instance or match new type");
595 // The type 't' could be subclass of 'base_t'.
596 // As result t->offset() could be large then base_t's size and it will
597 // cause the failure in add_offset() with narrow oops since TypeOopPtr()
598 // constructor verifies correctness of the offset.
599 //
600 // It could happened on subclass's branch (from the type profiling
601 // inlining) which was not eliminated during parsing since the exactness
602 // of the allocation type was not propagated to the subclass type check.
603 //
604 // Or the type 't' could be not related to 'base_t' at all.
605 // It could happened when CHA type is different from MDO type on a dead path
606 // (for example, from instanceof check) which is not collapsed during parsing.
607 //
608 // Do nothing for such AddP node and don't process its users since
609 // this code branch will go away.
610 //
611 if (!t->is_known_instance() &&
612 !base_t->klass()->is_subtype_of(t->klass())) {
613 return false; // bail out
614 }
616 const TypeOopPtr *tinst = base_t->add_offset(t->offset())->is_oopptr();
617 // Do NOT remove the next line: ensure a new alias index is allocated
618 // for the instance type. Note: C++ will not remove it since the call
619 // has side effect.
620 int alias_idx = _compile->get_alias_index(tinst);
621 igvn->set_type(addp, tinst);
622 // record the allocation in the node map
623 assert(ptnode_adr(addp->_idx)->_node != NULL, "should be registered");
624 set_map(addp->_idx, get_map(base->_idx));
626 // Set addp's Base and Address to 'base'.
627 Node *abase = addp->in(AddPNode::Base);
628 Node *adr = addp->in(AddPNode::Address);
629 if (adr->is_Proj() && adr->in(0)->is_Allocate() &&
630 adr->in(0)->_idx == (uint)inst_id) {
631 // Skip AddP cases #3 and #5.
632 } else {
633 assert(!abase->is_top(), "sanity"); // AddP case #3
634 if (abase != base) {
635 igvn->hash_delete(addp);
636 addp->set_req(AddPNode::Base, base);
637 if (abase == adr) {
638 addp->set_req(AddPNode::Address, base);
639 } else {
640 // AddP case #4 (adr is array's element offset AddP node)
641 #ifdef ASSERT
642 const TypeOopPtr *atype = igvn->type(adr)->isa_oopptr();
643 assert(adr->is_AddP() && atype != NULL &&
644 atype->instance_id() == inst_id, "array's element offset should be processed first");
645 #endif
646 }
647 igvn->hash_insert(addp);
648 }
649 }
650 // Put on IGVN worklist since at least addp's type was changed above.
651 record_for_optimizer(addp);
652 return true;
653 }
655 //
656 // Create a new version of orig_phi if necessary. Returns either the newly
657 // created phi or an existing phi. Sets create_new to indicate whether a new
658 // phi was created. Cache the last newly created phi in the node map.
659 //
660 PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
661 Compile *C = _compile;
662 new_created = false;
663 int phi_alias_idx = C->get_alias_index(orig_phi->adr_type());
664 // nothing to do if orig_phi is bottom memory or matches alias_idx
665 if (phi_alias_idx == alias_idx) {
666 return orig_phi;
667 }
668 // Have we recently created a Phi for this alias index?
669 PhiNode *result = get_map_phi(orig_phi->_idx);
670 if (result != NULL && C->get_alias_index(result->adr_type()) == alias_idx) {
671 return result;
672 }
673 // Previous check may fail when the same wide memory Phi was split into Phis
674 // for different memory slices. Search all Phis for this region.
675 if (result != NULL) {
676 Node* region = orig_phi->in(0);
677 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
678 Node* phi = region->fast_out(i);
679 if (phi->is_Phi() &&
680 C->get_alias_index(phi->as_Phi()->adr_type()) == alias_idx) {
681 assert(phi->_idx >= nodes_size(), "only new Phi per instance memory slice");
682 return phi->as_Phi();
683 }
684 }
685 }
686 if ((int)C->unique() + 2*NodeLimitFudgeFactor > MaxNodeLimit) {
687 if (C->do_escape_analysis() == true && !C->failing()) {
688 // Retry compilation without escape analysis.
689 // If this is the first failure, the sentinel string will "stick"
690 // to the Compile object, and the C2Compiler will see it and retry.
691 C->record_failure(C2Compiler::retry_no_escape_analysis());
692 }
693 return NULL;
694 }
695 orig_phi_worklist.append_if_missing(orig_phi);
696 const TypePtr *atype = C->get_adr_type(alias_idx);
697 result = PhiNode::make(orig_phi->in(0), NULL, Type::MEMORY, atype);
698 C->copy_node_notes_to(result, orig_phi);
699 igvn->set_type(result, result->bottom_type());
700 record_for_optimizer(result);
702 debug_only(Node* pn = ptnode_adr(orig_phi->_idx)->_node;)
703 assert(pn == NULL || pn == orig_phi, "wrong node");
704 set_map(orig_phi->_idx, result);
705 ptnode_adr(orig_phi->_idx)->_node = orig_phi;
707 new_created = true;
708 return result;
709 }
711 //
712 // Return a new version of Memory Phi "orig_phi" with the inputs having the
713 // specified alias index.
714 //
715 PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
717 assert(alias_idx != Compile::AliasIdxBot, "can't split out bottom memory");
718 Compile *C = _compile;
719 bool new_phi_created;
720 PhiNode *result = create_split_phi(orig_phi, alias_idx, orig_phi_worklist, igvn, new_phi_created);
721 if (!new_phi_created) {
722 return result;
723 }
725 GrowableArray<PhiNode *> phi_list;
726 GrowableArray<uint> cur_input;
728 PhiNode *phi = orig_phi;
729 uint idx = 1;
730 bool finished = false;
731 while(!finished) {
732 while (idx < phi->req()) {
733 Node *mem = find_inst_mem(phi->in(idx), alias_idx, orig_phi_worklist, igvn);
734 if (mem != NULL && mem->is_Phi()) {
735 PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
736 if (new_phi_created) {
737 // found an phi for which we created a new split, push current one on worklist and begin
738 // processing new one
739 phi_list.push(phi);
740 cur_input.push(idx);
741 phi = mem->as_Phi();
742 result = newphi;
743 idx = 1;
744 continue;
745 } else {
746 mem = newphi;
747 }
748 }
749 if (C->failing()) {
750 return NULL;
751 }
752 result->set_req(idx++, mem);
753 }
754 #ifdef ASSERT
755 // verify that the new Phi has an input for each input of the original
756 assert( phi->req() == result->req(), "must have same number of inputs.");
757 assert( result->in(0) != NULL && result->in(0) == phi->in(0), "regions must match");
758 #endif
759 // Check if all new phi's inputs have specified alias index.
760 // Otherwise use old phi.
761 for (uint i = 1; i < phi->req(); i++) {
762 Node* in = result->in(i);
763 assert((phi->in(i) == NULL) == (in == NULL), "inputs must correspond.");
764 }
765 // we have finished processing a Phi, see if there are any more to do
766 finished = (phi_list.length() == 0 );
767 if (!finished) {
768 phi = phi_list.pop();
769 idx = cur_input.pop();
770 PhiNode *prev_result = get_map_phi(phi->_idx);
771 prev_result->set_req(idx++, result);
772 result = prev_result;
773 }
774 }
775 return result;
776 }
779 //
780 // The next methods are derived from methods in MemNode.
781 //
782 static Node *step_through_mergemem(MergeMemNode *mmem, int alias_idx, const TypeOopPtr *toop) {
783 Node *mem = mmem;
784 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
785 // means an array I have not precisely typed yet. Do not do any
786 // alias stuff with it any time soon.
787 if( toop->base() != Type::AnyPtr &&
788 !(toop->klass() != NULL &&
789 toop->klass()->is_java_lang_Object() &&
790 toop->offset() == Type::OffsetBot) ) {
791 mem = mmem->memory_at(alias_idx);
792 // Update input if it is progress over what we have now
793 }
794 return mem;
795 }
797 //
798 // Move memory users to their memory slices.
799 //
800 void ConnectionGraph::move_inst_mem(Node* n, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *igvn) {
801 Compile* C = _compile;
803 const TypePtr* tp = igvn->type(n->in(MemNode::Address))->isa_ptr();
804 assert(tp != NULL, "ptr type");
805 int alias_idx = C->get_alias_index(tp);
806 int general_idx = C->get_general_index(alias_idx);
808 // Move users first
809 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
810 Node* use = n->fast_out(i);
811 if (use->is_MergeMem()) {
812 MergeMemNode* mmem = use->as_MergeMem();
813 assert(n == mmem->memory_at(alias_idx), "should be on instance memory slice");
814 if (n != mmem->memory_at(general_idx) || alias_idx == general_idx) {
815 continue; // Nothing to do
816 }
817 // Replace previous general reference to mem node.
818 uint orig_uniq = C->unique();
819 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
820 assert(orig_uniq == C->unique(), "no new nodes");
821 mmem->set_memory_at(general_idx, m);
822 --imax;
823 --i;
824 } else if (use->is_MemBar()) {
825 assert(!use->is_Initialize(), "initializing stores should not be moved");
826 if (use->req() > MemBarNode::Precedent &&
827 use->in(MemBarNode::Precedent) == n) {
828 // Don't move related membars.
829 record_for_optimizer(use);
830 continue;
831 }
832 tp = use->as_MemBar()->adr_type()->isa_ptr();
833 if (tp != NULL && C->get_alias_index(tp) == alias_idx ||
834 alias_idx == general_idx) {
835 continue; // Nothing to do
836 }
837 // Move to general memory slice.
838 uint orig_uniq = C->unique();
839 Node* m = find_inst_mem(n, general_idx, orig_phis, igvn);
840 assert(orig_uniq == C->unique(), "no new nodes");
841 igvn->hash_delete(use);
842 imax -= use->replace_edge(n, m);
843 igvn->hash_insert(use);
844 record_for_optimizer(use);
845 --i;
846 #ifdef ASSERT
847 } else if (use->is_Mem()) {
848 if (use->Opcode() == Op_StoreCM && use->in(MemNode::OopStore) == n) {
849 // Don't move related cardmark.
850 continue;
851 }
852 // Memory nodes should have new memory input.
853 tp = igvn->type(use->in(MemNode::Address))->isa_ptr();
854 assert(tp != NULL, "ptr type");
855 int idx = C->get_alias_index(tp);
856 assert(get_map(use->_idx) != NULL || idx == alias_idx,
857 "Following memory nodes should have new memory input or be on the same memory slice");
858 } else if (use->is_Phi()) {
859 // Phi nodes should be split and moved already.
860 tp = use->as_Phi()->adr_type()->isa_ptr();
861 assert(tp != NULL, "ptr type");
862 int idx = C->get_alias_index(tp);
863 assert(idx == alias_idx, "Following Phi nodes should be on the same memory slice");
864 } else {
865 use->dump();
866 assert(false, "should not be here");
867 #endif
868 }
869 }
870 }
872 //
873 // Search memory chain of "mem" to find a MemNode whose address
874 // is the specified alias index.
875 //
876 Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
877 if (orig_mem == NULL)
878 return orig_mem;
879 Compile* C = phase->C;
880 const TypeOopPtr *toop = C->get_adr_type(alias_idx)->isa_oopptr();
881 bool is_instance = (toop != NULL) && toop->is_known_instance();
882 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
883 Node *prev = NULL;
884 Node *result = orig_mem;
885 while (prev != result) {
886 prev = result;
887 if (result == start_mem)
888 break; // hit one of our sentinels
889 if (result->is_Mem()) {
890 const Type *at = phase->type(result->in(MemNode::Address));
891 if (at == Type::TOP)
892 break; // Dead
893 assert (at->isa_ptr() != NULL, "pointer type required.");
894 int idx = C->get_alias_index(at->is_ptr());
895 if (idx == alias_idx)
896 break; // Found
897 if (!is_instance && (at->isa_oopptr() == NULL ||
898 !at->is_oopptr()->is_known_instance())) {
899 break; // Do not skip store to general memory slice.
900 }
901 result = result->in(MemNode::Memory);
902 }
903 if (!is_instance)
904 continue; // don't search further for non-instance types
905 // skip over a call which does not affect this memory slice
906 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
907 Node *proj_in = result->in(0);
908 if (proj_in->is_Allocate() && proj_in->_idx == (uint)toop->instance_id()) {
909 break; // hit one of our sentinels
910 } else if (proj_in->is_Call()) {
911 CallNode *call = proj_in->as_Call();
912 if (!call->may_modify(toop, phase)) {
913 result = call->in(TypeFunc::Memory);
914 }
915 } else if (proj_in->is_Initialize()) {
916 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
917 // Stop if this is the initialization for the object instance which
918 // which contains this memory slice, otherwise skip over it.
919 if (alloc == NULL || alloc->_idx != (uint)toop->instance_id()) {
920 result = proj_in->in(TypeFunc::Memory);
921 }
922 } else if (proj_in->is_MemBar()) {
923 result = proj_in->in(TypeFunc::Memory);
924 }
925 } else if (result->is_MergeMem()) {
926 MergeMemNode *mmem = result->as_MergeMem();
927 result = step_through_mergemem(mmem, alias_idx, toop);
928 if (result == mmem->base_memory()) {
929 // Didn't find instance memory, search through general slice recursively.
930 result = mmem->memory_at(C->get_general_index(alias_idx));
931 result = find_inst_mem(result, alias_idx, orig_phis, phase);
932 if (C->failing()) {
933 return NULL;
934 }
935 mmem->set_memory_at(alias_idx, result);
936 }
937 } else if (result->is_Phi() &&
938 C->get_alias_index(result->as_Phi()->adr_type()) != alias_idx) {
939 Node *un = result->as_Phi()->unique_input(phase);
940 if (un != NULL) {
941 orig_phis.append_if_missing(result->as_Phi());
942 result = un;
943 } else {
944 break;
945 }
946 } else if (result->is_ClearArray()) {
947 if (!ClearArrayNode::step_through(&result, (uint)toop->instance_id(), phase)) {
948 // Can not bypass initialization of the instance
949 // we are looking for.
950 break;
951 }
952 // Otherwise skip it (the call updated 'result' value).
953 } else if (result->Opcode() == Op_SCMemProj) {
954 assert(result->in(0)->is_LoadStore(), "sanity");
955 const Type *at = phase->type(result->in(0)->in(MemNode::Address));
956 if (at != Type::TOP) {
957 assert (at->isa_ptr() != NULL, "pointer type required.");
958 int idx = C->get_alias_index(at->is_ptr());
959 assert(idx != alias_idx, "Object is not scalar replaceable if a LoadStore node access its field");
960 break;
961 }
962 result = result->in(0)->in(MemNode::Memory);
963 }
964 }
965 if (result->is_Phi()) {
966 PhiNode *mphi = result->as_Phi();
967 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
968 const TypePtr *t = mphi->adr_type();
969 if (!is_instance) {
970 // Push all non-instance Phis on the orig_phis worklist to update inputs
971 // during Phase 4 if needed.
972 orig_phis.append_if_missing(mphi);
973 } else if (C->get_alias_index(t) != alias_idx) {
974 // Create a new Phi with the specified alias index type.
975 result = split_memory_phi(mphi, alias_idx, orig_phis, phase);
976 }
977 }
978 // the result is either MemNode, PhiNode, InitializeNode.
979 return result;
980 }
982 //
983 // Convert the types of unescaped object to instance types where possible,
984 // propagate the new type information through the graph, and update memory
985 // edges and MergeMem inputs to reflect the new type.
986 //
987 // We start with allocations (and calls which may be allocations) on alloc_worklist.
988 // The processing is done in 4 phases:
989 //
990 // Phase 1: Process possible allocations from alloc_worklist. Create instance
991 // types for the CheckCastPP for allocations where possible.
992 // Propagate the the new types through users as follows:
993 // casts and Phi: push users on alloc_worklist
994 // AddP: cast Base and Address inputs to the instance type
995 // push any AddP users on alloc_worklist and push any memnode
996 // users onto memnode_worklist.
997 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
998 // search the Memory chain for a store with the appropriate type
999 // address type. If a Phi is found, create a new version with
1000 // the appropriate memory slices from each of the Phi inputs.
1001 // For stores, process the users as follows:
1002 // MemNode: push on memnode_worklist
1003 // MergeMem: push on mergemem_worklist
1004 // Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
1005 // moving the first node encountered of each instance type to the
1006 // the input corresponding to its alias index.
1007 // appropriate memory slice.
1008 // Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
1009 //
1010 // In the following example, the CheckCastPP nodes are the cast of allocation
1011 // results and the allocation of node 29 is unescaped and eligible to be an
1012 // instance type.
1013 //
1014 // We start with:
1015 //
1016 // 7 Parm #memory
1017 // 10 ConI "12"
1018 // 19 CheckCastPP "Foo"
1019 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
1020 // 29 CheckCastPP "Foo"
1021 // 30 AddP _ 29 29 10 Foo+12 alias_index=4
1022 //
1023 // 40 StoreP 25 7 20 ... alias_index=4
1024 // 50 StoreP 35 40 30 ... alias_index=4
1025 // 60 StoreP 45 50 20 ... alias_index=4
1026 // 70 LoadP _ 60 30 ... alias_index=4
1027 // 80 Phi 75 50 60 Memory alias_index=4
1028 // 90 LoadP _ 80 30 ... alias_index=4
1029 // 100 LoadP _ 80 20 ... alias_index=4
1030 //
1031 //
1032 // Phase 1 creates an instance type for node 29 assigning it an instance id of 24
1033 // and creating a new alias index for node 30. This gives:
1034 //
1035 // 7 Parm #memory
1036 // 10 ConI "12"
1037 // 19 CheckCastPP "Foo"
1038 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
1039 // 29 CheckCastPP "Foo" iid=24
1040 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
1041 //
1042 // 40 StoreP 25 7 20 ... alias_index=4
1043 // 50 StoreP 35 40 30 ... alias_index=6
1044 // 60 StoreP 45 50 20 ... alias_index=4
1045 // 70 LoadP _ 60 30 ... alias_index=6
1046 // 80 Phi 75 50 60 Memory alias_index=4
1047 // 90 LoadP _ 80 30 ... alias_index=6
1048 // 100 LoadP _ 80 20 ... alias_index=4
1049 //
1050 // In phase 2, new memory inputs are computed for the loads and stores,
1051 // And a new version of the phi is created. In phase 4, the inputs to
1052 // node 80 are updated and then the memory nodes are updated with the
1053 // values computed in phase 2. This results in:
1054 //
1055 // 7 Parm #memory
1056 // 10 ConI "12"
1057 // 19 CheckCastPP "Foo"
1058 // 20 AddP _ 19 19 10 Foo+12 alias_index=4
1059 // 29 CheckCastPP "Foo" iid=24
1060 // 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
1061 //
1062 // 40 StoreP 25 7 20 ... alias_index=4
1063 // 50 StoreP 35 7 30 ... alias_index=6
1064 // 60 StoreP 45 40 20 ... alias_index=4
1065 // 70 LoadP _ 50 30 ... alias_index=6
1066 // 80 Phi 75 40 60 Memory alias_index=4
1067 // 120 Phi 75 50 50 Memory alias_index=6
1068 // 90 LoadP _ 120 30 ... alias_index=6
1069 // 100 LoadP _ 80 20 ... alias_index=4
1070 //
1071 void ConnectionGraph::split_unique_types(GrowableArray<Node *> &alloc_worklist) {
1072 GrowableArray<Node *> memnode_worklist;
1073 GrowableArray<PhiNode *> orig_phis;
1075 PhaseIterGVN *igvn = _igvn;
1076 uint new_index_start = (uint) _compile->num_alias_types();
1077 Arena* arena = Thread::current()->resource_area();
1078 VectorSet visited(arena);
1081 // Phase 1: Process possible allocations from alloc_worklist.
1082 // Create instance types for the CheckCastPP for allocations where possible.
1083 //
1084 // (Note: don't forget to change the order of the second AddP node on
1085 // the alloc_worklist if the order of the worklist processing is changed,
1086 // see the comment in find_second_addp().)
1087 //
1088 while (alloc_worklist.length() != 0) {
1089 Node *n = alloc_worklist.pop();
1090 uint ni = n->_idx;
1091 const TypeOopPtr* tinst = NULL;
1092 if (n->is_Call()) {
1093 CallNode *alloc = n->as_Call();
1094 // copy escape information to call node
1095 PointsToNode* ptn = ptnode_adr(alloc->_idx);
1096 PointsToNode::EscapeState es = escape_state(alloc);
1097 // We have an allocation or call which returns a Java object,
1098 // see if it is unescaped.
1099 if (es != PointsToNode::NoEscape || !ptn->scalar_replaceable())
1100 continue;
1102 // Find CheckCastPP for the allocate or for the return value of a call
1103 n = alloc->result_cast();
1104 if (n == NULL) { // No uses except Initialize node
1105 if (alloc->is_Allocate()) {
1106 // Set the scalar_replaceable flag for allocation
1107 // so it could be eliminated if it has no uses.
1108 alloc->as_Allocate()->_is_scalar_replaceable = true;
1109 }
1110 continue;
1111 }
1112 if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
1113 assert(!alloc->is_Allocate(), "allocation should have unique type");
1114 continue;
1115 }
1117 // The inline code for Object.clone() casts the allocation result to
1118 // java.lang.Object and then to the actual type of the allocated
1119 // object. Detect this case and use the second cast.
1120 // Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
1121 // the allocation result is cast to java.lang.Object and then
1122 // to the actual Array type.
1123 if (alloc->is_Allocate() && n->as_Type()->type() == TypeInstPtr::NOTNULL
1124 && (alloc->is_AllocateArray() ||
1125 igvn->type(alloc->in(AllocateNode::KlassNode)) != TypeKlassPtr::OBJECT)) {
1126 Node *cast2 = NULL;
1127 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1128 Node *use = n->fast_out(i);
1129 if (use->is_CheckCastPP()) {
1130 cast2 = use;
1131 break;
1132 }
1133 }
1134 if (cast2 != NULL) {
1135 n = cast2;
1136 } else {
1137 // Non-scalar replaceable if the allocation type is unknown statically
1138 // (reflection allocation), the object can't be restored during
1139 // deoptimization without precise type.
1140 continue;
1141 }
1142 }
1143 if (alloc->is_Allocate()) {
1144 // Set the scalar_replaceable flag for allocation
1145 // so it could be eliminated.
1146 alloc->as_Allocate()->_is_scalar_replaceable = true;
1147 }
1148 set_escape_state(n->_idx, es); // CheckCastPP escape state
1149 // in order for an object to be scalar-replaceable, it must be:
1150 // - a direct allocation (not a call returning an object)
1151 // - non-escaping
1152 // - eligible to be a unique type
1153 // - not determined to be ineligible by escape analysis
1154 assert(ptnode_adr(alloc->_idx)->_node != NULL &&
1155 ptnode_adr(n->_idx)->_node != NULL, "should be registered");
1156 set_map(alloc->_idx, n);
1157 set_map(n->_idx, alloc);
1158 const TypeOopPtr *t = igvn->type(n)->isa_oopptr();
1159 if (t == NULL)
1160 continue; // not a TypeOopPtr
1161 tinst = t->cast_to_exactness(true)->is_oopptr()->cast_to_instance_id(ni);
1162 igvn->hash_delete(n);
1163 igvn->set_type(n, tinst);
1164 n->raise_bottom_type(tinst);
1165 igvn->hash_insert(n);
1166 record_for_optimizer(n);
1167 if (alloc->is_Allocate() && (t->isa_instptr() || t->isa_aryptr())) {
1169 // First, put on the worklist all Field edges from Connection Graph
1170 // which is more accurate then putting immediate users from Ideal Graph.
1171 for (uint e = 0; e < ptn->edge_count(); e++) {
1172 Node *use = ptnode_adr(ptn->edge_target(e))->_node;
1173 assert(ptn->edge_type(e) == PointsToNode::FieldEdge && use->is_AddP(),
1174 "only AddP nodes are Field edges in CG");
1175 if (use->outcnt() > 0) { // Don't process dead nodes
1176 Node* addp2 = find_second_addp(use, use->in(AddPNode::Base));
1177 if (addp2 != NULL) {
1178 assert(alloc->is_AllocateArray(),"array allocation was expected");
1179 alloc_worklist.append_if_missing(addp2);
1180 }
1181 alloc_worklist.append_if_missing(use);
1182 }
1183 }
1185 // An allocation may have an Initialize which has raw stores. Scan
1186 // the users of the raw allocation result and push AddP users
1187 // on alloc_worklist.
1188 Node *raw_result = alloc->proj_out(TypeFunc::Parms);
1189 assert (raw_result != NULL, "must have an allocation result");
1190 for (DUIterator_Fast imax, i = raw_result->fast_outs(imax); i < imax; i++) {
1191 Node *use = raw_result->fast_out(i);
1192 if (use->is_AddP() && use->outcnt() > 0) { // Don't process dead nodes
1193 Node* addp2 = find_second_addp(use, raw_result);
1194 if (addp2 != NULL) {
1195 assert(alloc->is_AllocateArray(),"array allocation was expected");
1196 alloc_worklist.append_if_missing(addp2);
1197 }
1198 alloc_worklist.append_if_missing(use);
1199 } else if (use->is_MemBar()) {
1200 memnode_worklist.append_if_missing(use);
1201 }
1202 }
1203 }
1204 } else if (n->is_AddP()) {
1205 VectorSet* ptset = PointsTo(get_addp_base(n));
1206 assert(ptset->Size() == 1, "AddP address is unique");
1207 uint elem = ptset->getelem(); // Allocation node's index
1208 if (elem == _phantom_object) {
1209 assert(false, "escaped allocation");
1210 continue; // Assume the value was set outside this method.
1211 }
1212 Node *base = get_map(elem); // CheckCastPP node
1213 if (!split_AddP(n, base, igvn)) continue; // wrong type from dead path
1214 tinst = igvn->type(base)->isa_oopptr();
1215 } else if (n->is_Phi() ||
1216 n->is_CheckCastPP() ||
1217 n->is_EncodeP() ||
1218 n->is_DecodeN() ||
1219 (n->is_ConstraintCast() && n->Opcode() == Op_CastPP)) {
1220 if (visited.test_set(n->_idx)) {
1221 assert(n->is_Phi(), "loops only through Phi's");
1222 continue; // already processed
1223 }
1224 VectorSet* ptset = PointsTo(n);
1225 if (ptset->Size() == 1) {
1226 uint elem = ptset->getelem(); // Allocation node's index
1227 if (elem == _phantom_object) {
1228 assert(false, "escaped allocation");
1229 continue; // Assume the value was set outside this method.
1230 }
1231 Node *val = get_map(elem); // CheckCastPP node
1232 TypeNode *tn = n->as_Type();
1233 tinst = igvn->type(val)->isa_oopptr();
1234 assert(tinst != NULL && tinst->is_known_instance() &&
1235 (uint)tinst->instance_id() == elem , "instance type expected.");
1237 const Type *tn_type = igvn->type(tn);
1238 const TypeOopPtr *tn_t;
1239 if (tn_type->isa_narrowoop()) {
1240 tn_t = tn_type->make_ptr()->isa_oopptr();
1241 } else {
1242 tn_t = tn_type->isa_oopptr();
1243 }
1245 if (tn_t != NULL && tinst->klass()->is_subtype_of(tn_t->klass())) {
1246 if (tn_type->isa_narrowoop()) {
1247 tn_type = tinst->make_narrowoop();
1248 } else {
1249 tn_type = tinst;
1250 }
1251 igvn->hash_delete(tn);
1252 igvn->set_type(tn, tn_type);
1253 tn->set_type(tn_type);
1254 igvn->hash_insert(tn);
1255 record_for_optimizer(n);
1256 } else {
1257 assert(tn_type == TypePtr::NULL_PTR ||
1258 tn_t != NULL && !tinst->klass()->is_subtype_of(tn_t->klass()),
1259 "unexpected type");
1260 continue; // Skip dead path with different type
1261 }
1262 }
1263 } else {
1264 debug_only(n->dump();)
1265 assert(false, "EA: unexpected node");
1266 continue;
1267 }
1268 // push allocation's users on appropriate worklist
1269 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1270 Node *use = n->fast_out(i);
1271 if(use->is_Mem() && use->in(MemNode::Address) == n) {
1272 // Load/store to instance's field
1273 memnode_worklist.append_if_missing(use);
1274 } else if (use->is_MemBar()) {
1275 memnode_worklist.append_if_missing(use);
1276 } else if (use->is_AddP() && use->outcnt() > 0) { // No dead nodes
1277 Node* addp2 = find_second_addp(use, n);
1278 if (addp2 != NULL) {
1279 alloc_worklist.append_if_missing(addp2);
1280 }
1281 alloc_worklist.append_if_missing(use);
1282 } else if (use->is_Phi() ||
1283 use->is_CheckCastPP() ||
1284 use->is_EncodeP() ||
1285 use->is_DecodeN() ||
1286 (use->is_ConstraintCast() && use->Opcode() == Op_CastPP)) {
1287 alloc_worklist.append_if_missing(use);
1288 #ifdef ASSERT
1289 } else if (use->is_Mem()) {
1290 assert(use->in(MemNode::Address) != n, "EA: missing allocation reference path");
1291 } else if (use->is_MergeMem()) {
1292 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1293 } else if (use->is_SafePoint()) {
1294 // Look for MergeMem nodes for calls which reference unique allocation
1295 // (through CheckCastPP nodes) even for debug info.
1296 Node* m = use->in(TypeFunc::Memory);
1297 if (m->is_MergeMem()) {
1298 assert(_mergemem_worklist.contains(m->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1299 }
1300 } else {
1301 uint op = use->Opcode();
1302 if (!(op == Op_CmpP || op == Op_Conv2B ||
1303 op == Op_CastP2X || op == Op_StoreCM ||
1304 op == Op_FastLock || op == Op_AryEq || op == Op_StrComp ||
1305 op == Op_StrEquals || op == Op_StrIndexOf)) {
1306 n->dump();
1307 use->dump();
1308 assert(false, "EA: missing allocation reference path");
1309 }
1310 #endif
1311 }
1312 }
1314 }
1315 // New alias types were created in split_AddP().
1316 uint new_index_end = (uint) _compile->num_alias_types();
1318 // Phase 2: Process MemNode's from memnode_worklist. compute new address type and
1319 // compute new values for Memory inputs (the Memory inputs are not
1320 // actually updated until phase 4.)
1321 if (memnode_worklist.length() == 0)
1322 return; // nothing to do
1324 while (memnode_worklist.length() != 0) {
1325 Node *n = memnode_worklist.pop();
1326 if (visited.test_set(n->_idx))
1327 continue;
1328 if (n->is_Phi() || n->is_ClearArray()) {
1329 // we don't need to do anything, but the users must be pushed
1330 } else if (n->is_MemBar()) { // Initialize, MemBar nodes
1331 // we don't need to do anything, but the users must be pushed
1332 n = n->as_MemBar()->proj_out(TypeFunc::Memory);
1333 if (n == NULL)
1334 continue;
1335 } else {
1336 assert(n->is_Mem(), "memory node required.");
1337 Node *addr = n->in(MemNode::Address);
1338 const Type *addr_t = igvn->type(addr);
1339 if (addr_t == Type::TOP)
1340 continue;
1341 assert (addr_t->isa_ptr() != NULL, "pointer type required.");
1342 int alias_idx = _compile->get_alias_index(addr_t->is_ptr());
1343 assert ((uint)alias_idx < new_index_end, "wrong alias index");
1344 Node *mem = find_inst_mem(n->in(MemNode::Memory), alias_idx, orig_phis, igvn);
1345 if (_compile->failing()) {
1346 return;
1347 }
1348 if (mem != n->in(MemNode::Memory)) {
1349 // We delay the memory edge update since we need old one in
1350 // MergeMem code below when instances memory slices are separated.
1351 debug_only(Node* pn = ptnode_adr(n->_idx)->_node;)
1352 assert(pn == NULL || pn == n, "wrong node");
1353 set_map(n->_idx, mem);
1354 ptnode_adr(n->_idx)->_node = n;
1355 }
1356 if (n->is_Load()) {
1357 continue; // don't push users
1358 } else if (n->is_LoadStore()) {
1359 // get the memory projection
1360 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1361 Node *use = n->fast_out(i);
1362 if (use->Opcode() == Op_SCMemProj) {
1363 n = use;
1364 break;
1365 }
1366 }
1367 assert(n->Opcode() == Op_SCMemProj, "memory projection required");
1368 }
1369 }
1370 // push user on appropriate worklist
1371 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1372 Node *use = n->fast_out(i);
1373 if (use->is_Phi() || use->is_ClearArray()) {
1374 memnode_worklist.append_if_missing(use);
1375 } else if(use->is_Mem() && use->in(MemNode::Memory) == n) {
1376 if (use->Opcode() == Op_StoreCM) // Ignore cardmark stores
1377 continue;
1378 memnode_worklist.append_if_missing(use);
1379 } else if (use->is_MemBar()) {
1380 memnode_worklist.append_if_missing(use);
1381 #ifdef ASSERT
1382 } else if(use->is_Mem()) {
1383 assert(use->in(MemNode::Memory) != n, "EA: missing memory path");
1384 } else if (use->is_MergeMem()) {
1385 assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
1386 } else {
1387 uint op = use->Opcode();
1388 if (!(op == Op_StoreCM ||
1389 (op == Op_CallLeaf && use->as_CallLeaf()->_name != NULL &&
1390 strcmp(use->as_CallLeaf()->_name, "g1_wb_pre") == 0) ||
1391 op == Op_AryEq || op == Op_StrComp ||
1392 op == Op_StrEquals || op == Op_StrIndexOf)) {
1393 n->dump();
1394 use->dump();
1395 assert(false, "EA: missing memory path");
1396 }
1397 #endif
1398 }
1399 }
1400 }
1402 // Phase 3: Process MergeMem nodes from mergemem_worklist.
1403 // Walk each memory slice moving the first node encountered of each
1404 // instance type to the the input corresponding to its alias index.
1405 uint length = _mergemem_worklist.length();
1406 for( uint next = 0; next < length; ++next ) {
1407 MergeMemNode* nmm = _mergemem_worklist.at(next);
1408 assert(!visited.test_set(nmm->_idx), "should not be visited before");
1409 // Note: we don't want to use MergeMemStream here because we only want to
1410 // scan inputs which exist at the start, not ones we add during processing.
1411 // Note 2: MergeMem may already contains instance memory slices added
1412 // during find_inst_mem() call when memory nodes were processed above.
1413 igvn->hash_delete(nmm);
1414 uint nslices = nmm->req();
1415 for (uint i = Compile::AliasIdxRaw+1; i < nslices; i++) {
1416 Node* mem = nmm->in(i);
1417 Node* cur = NULL;
1418 if (mem == NULL || mem->is_top())
1419 continue;
1420 // First, update mergemem by moving memory nodes to corresponding slices
1421 // if their type became more precise since this mergemem was created.
1422 while (mem->is_Mem()) {
1423 const Type *at = igvn->type(mem->in(MemNode::Address));
1424 if (at != Type::TOP) {
1425 assert (at->isa_ptr() != NULL, "pointer type required.");
1426 uint idx = (uint)_compile->get_alias_index(at->is_ptr());
1427 if (idx == i) {
1428 if (cur == NULL)
1429 cur = mem;
1430 } else {
1431 if (idx >= nmm->req() || nmm->is_empty_memory(nmm->in(idx))) {
1432 nmm->set_memory_at(idx, mem);
1433 }
1434 }
1435 }
1436 mem = mem->in(MemNode::Memory);
1437 }
1438 nmm->set_memory_at(i, (cur != NULL) ? cur : mem);
1439 // Find any instance of the current type if we haven't encountered
1440 // already a memory slice of the instance along the memory chain.
1441 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1442 if((uint)_compile->get_general_index(ni) == i) {
1443 Node *m = (ni >= nmm->req()) ? nmm->empty_memory() : nmm->in(ni);
1444 if (nmm->is_empty_memory(m)) {
1445 Node* result = find_inst_mem(mem, ni, orig_phis, igvn);
1446 if (_compile->failing()) {
1447 return;
1448 }
1449 nmm->set_memory_at(ni, result);
1450 }
1451 }
1452 }
1453 }
1454 // Find the rest of instances values
1455 for (uint ni = new_index_start; ni < new_index_end; ni++) {
1456 const TypeOopPtr *tinst = _compile->get_adr_type(ni)->isa_oopptr();
1457 Node* result = step_through_mergemem(nmm, ni, tinst);
1458 if (result == nmm->base_memory()) {
1459 // Didn't find instance memory, search through general slice recursively.
1460 result = nmm->memory_at(_compile->get_general_index(ni));
1461 result = find_inst_mem(result, ni, orig_phis, igvn);
1462 if (_compile->failing()) {
1463 return;
1464 }
1465 nmm->set_memory_at(ni, result);
1466 }
1467 }
1468 igvn->hash_insert(nmm);
1469 record_for_optimizer(nmm);
1470 }
1472 // Phase 4: Update the inputs of non-instance memory Phis and
1473 // the Memory input of memnodes
1474 // First update the inputs of any non-instance Phi's from
1475 // which we split out an instance Phi. Note we don't have
1476 // to recursively process Phi's encounted on the input memory
1477 // chains as is done in split_memory_phi() since they will
1478 // also be processed here.
1479 for (int j = 0; j < orig_phis.length(); j++) {
1480 PhiNode *phi = orig_phis.at(j);
1481 int alias_idx = _compile->get_alias_index(phi->adr_type());
1482 igvn->hash_delete(phi);
1483 for (uint i = 1; i < phi->req(); i++) {
1484 Node *mem = phi->in(i);
1485 Node *new_mem = find_inst_mem(mem, alias_idx, orig_phis, igvn);
1486 if (_compile->failing()) {
1487 return;
1488 }
1489 if (mem != new_mem) {
1490 phi->set_req(i, new_mem);
1491 }
1492 }
1493 igvn->hash_insert(phi);
1494 record_for_optimizer(phi);
1495 }
1497 // Update the memory inputs of MemNodes with the value we computed
1498 // in Phase 2 and move stores memory users to corresponding memory slices.
1500 // Disable memory split verification code until the fix for 6984348.
1501 // Currently it produces false negative results since it does not cover all cases.
1502 #if 0 // ifdef ASSERT
1503 visited.Reset();
1504 Node_Stack old_mems(arena, _compile->unique() >> 2);
1505 #endif
1506 for (uint i = 0; i < nodes_size(); i++) {
1507 Node *nmem = get_map(i);
1508 if (nmem != NULL) {
1509 Node *n = ptnode_adr(i)->_node;
1510 assert(n != NULL, "sanity");
1511 if (n->is_Mem()) {
1512 #if 0 // ifdef ASSERT
1513 Node* old_mem = n->in(MemNode::Memory);
1514 if (!visited.test_set(old_mem->_idx)) {
1515 old_mems.push(old_mem, old_mem->outcnt());
1516 }
1517 #endif
1518 assert(n->in(MemNode::Memory) != nmem, "sanity");
1519 if (!n->is_Load()) {
1520 // Move memory users of a store first.
1521 move_inst_mem(n, orig_phis, igvn);
1522 }
1523 // Now update memory input
1524 igvn->hash_delete(n);
1525 n->set_req(MemNode::Memory, nmem);
1526 igvn->hash_insert(n);
1527 record_for_optimizer(n);
1528 } else {
1529 assert(n->is_Allocate() || n->is_CheckCastPP() ||
1530 n->is_AddP() || n->is_Phi(), "unknown node used for set_map()");
1531 }
1532 }
1533 }
1534 #if 0 // ifdef ASSERT
1535 // Verify that memory was split correctly
1536 while (old_mems.is_nonempty()) {
1537 Node* old_mem = old_mems.node();
1538 uint old_cnt = old_mems.index();
1539 old_mems.pop();
1540 assert(old_cnt == old_mem->outcnt(), "old mem could be lost");
1541 }
1542 #endif
1543 }
1545 bool ConnectionGraph::has_candidates(Compile *C) {
1546 // EA brings benefits only when the code has allocations and/or locks which
1547 // are represented by ideal Macro nodes.
1548 int cnt = C->macro_count();
1549 for( int i=0; i < cnt; i++ ) {
1550 Node *n = C->macro_node(i);
1551 if ( n->is_Allocate() )
1552 return true;
1553 if( n->is_Lock() ) {
1554 Node* obj = n->as_Lock()->obj_node()->uncast();
1555 if( !(obj->is_Parm() || obj->is_Con()) )
1556 return true;
1557 }
1558 }
1559 return false;
1560 }
1562 void ConnectionGraph::do_analysis(Compile *C, PhaseIterGVN *igvn) {
1563 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
1564 // to create space for them in ConnectionGraph::_nodes[].
1565 Node* oop_null = igvn->zerocon(T_OBJECT);
1566 Node* noop_null = igvn->zerocon(T_NARROWOOP);
1568 ConnectionGraph* congraph = new(C->comp_arena()) ConnectionGraph(C, igvn);
1569 // Perform escape analysis
1570 if (congraph->compute_escape()) {
1571 // There are non escaping objects.
1572 C->set_congraph(congraph);
1573 }
1575 // Cleanup.
1576 if (oop_null->outcnt() == 0)
1577 igvn->hash_delete(oop_null);
1578 if (noop_null->outcnt() == 0)
1579 igvn->hash_delete(noop_null);
1580 }
1582 bool ConnectionGraph::compute_escape() {
1583 Compile* C = _compile;
1585 // 1. Populate Connection Graph (CG) with Ideal nodes.
1587 Unique_Node_List worklist_init;
1588 worklist_init.map(C->unique(), NULL); // preallocate space
1590 // Initialize worklist
1591 if (C->root() != NULL) {
1592 worklist_init.push(C->root());
1593 }
1595 GrowableArray<Node*> alloc_worklist;
1596 GrowableArray<Node*> addp_worklist;
1597 GrowableArray<Node*> ptr_cmp_worklist;
1598 GrowableArray<Node*> storestore_worklist;
1599 PhaseGVN* igvn = _igvn;
1601 // Push all useful nodes onto CG list and set their type.
1602 for( uint next = 0; next < worklist_init.size(); ++next ) {
1603 Node* n = worklist_init.at(next);
1604 record_for_escape_analysis(n, igvn);
1605 // Only allocations and java static calls results are checked
1606 // for an escape status. See process_call_result() below.
1607 if (n->is_Allocate() || n->is_CallStaticJava() &&
1608 ptnode_adr(n->_idx)->node_type() == PointsToNode::JavaObject) {
1609 alloc_worklist.append(n);
1610 } else if(n->is_AddP()) {
1611 // Collect address nodes. Use them during stage 3 below
1612 // to build initial connection graph field edges.
1613 addp_worklist.append(n);
1614 } else if (n->is_MergeMem()) {
1615 // Collect all MergeMem nodes to add memory slices for
1616 // scalar replaceable objects in split_unique_types().
1617 _mergemem_worklist.append(n->as_MergeMem());
1618 } else if (OptimizePtrCompare && n->is_Cmp() &&
1619 (n->Opcode() == Op_CmpP || n->Opcode() == Op_CmpN)) {
1620 // Compare pointers nodes
1621 ptr_cmp_worklist.append(n);
1622 } else if (n->is_MemBarStoreStore()) {
1623 // Collect all MemBarStoreStore nodes so that depending on the
1624 // escape status of the associated Allocate node some of them
1625 // may be eliminated.
1626 storestore_worklist.append(n);
1627 }
1628 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1629 Node* m = n->fast_out(i); // Get user
1630 worklist_init.push(m);
1631 }
1632 }
1634 if (alloc_worklist.length() == 0) {
1635 _collecting = false;
1636 return false; // Nothing to do.
1637 }
1639 // 2. First pass to create simple CG edges (doesn't require to walk CG).
1640 uint delayed_size = _delayed_worklist.size();
1641 for( uint next = 0; next < delayed_size; ++next ) {
1642 Node* n = _delayed_worklist.at(next);
1643 build_connection_graph(n, igvn);
1644 }
1646 // 3. Pass to create initial fields edges (JavaObject -F-> AddP)
1647 // to reduce number of iterations during stage 4 below.
1648 uint addp_length = addp_worklist.length();
1649 for( uint next = 0; next < addp_length; ++next ) {
1650 Node* n = addp_worklist.at(next);
1651 Node* base = get_addp_base(n);
1652 if (base->is_Proj() && base->in(0)->is_Call())
1653 base = base->in(0);
1654 PointsToNode::NodeType nt = ptnode_adr(base->_idx)->node_type();
1655 if (nt == PointsToNode::JavaObject) {
1656 build_connection_graph(n, igvn);
1657 }
1658 }
1660 GrowableArray<int> cg_worklist;
1661 cg_worklist.append(_phantom_object);
1662 GrowableArray<uint> worklist;
1664 // 4. Build Connection Graph which need
1665 // to walk the connection graph.
1666 _progress = false;
1667 for (uint ni = 0; ni < nodes_size(); ni++) {
1668 PointsToNode* ptn = ptnode_adr(ni);
1669 Node *n = ptn->_node;
1670 if (n != NULL) { // Call, AddP, LoadP, StoreP
1671 build_connection_graph(n, igvn);
1672 if (ptn->node_type() != PointsToNode::UnknownType)
1673 cg_worklist.append(n->_idx); // Collect CG nodes
1674 if (!_processed.test(n->_idx))
1675 worklist.append(n->_idx); // Collect C/A/L/S nodes
1676 }
1677 }
1679 // After IGVN user nodes may have smaller _idx than
1680 // their inputs so they will be processed first in
1681 // previous loop. Because of that not all Graph
1682 // edges will be created. Walk over interesting
1683 // nodes again until no new edges are created.
1684 //
1685 // Normally only 1-3 passes needed to build
1686 // Connection Graph depending on graph complexity.
1687 // Observed 8 passes in jvm2008 compiler.compiler.
1688 // Set limit to 20 to catch situation when something
1689 // did go wrong and recompile the method without EA.
1690 // Also limit build time to 30 sec (60 in debug VM).
1692 #define CG_BUILD_ITER_LIMIT 20
1694 #ifdef ASSERT
1695 #define CG_BUILD_TIME_LIMIT 60.0
1696 #else
1697 #define CG_BUILD_TIME_LIMIT 30.0
1698 #endif
1700 uint length = worklist.length();
1701 int iterations = 0;
1702 elapsedTimer time;
1703 while(_progress &&
1704 (iterations++ < CG_BUILD_ITER_LIMIT) &&
1705 (time.seconds() < CG_BUILD_TIME_LIMIT)) {
1706 time.start();
1707 _progress = false;
1708 for( uint next = 0; next < length; ++next ) {
1709 int ni = worklist.at(next);
1710 PointsToNode* ptn = ptnode_adr(ni);
1711 Node* n = ptn->_node;
1712 assert(n != NULL, "should be known node");
1713 build_connection_graph(n, igvn);
1714 }
1715 time.stop();
1716 }
1717 if ((iterations >= CG_BUILD_ITER_LIMIT) ||
1718 (time.seconds() >= CG_BUILD_TIME_LIMIT)) {
1719 assert(false, err_msg("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
1720 time.seconds(), iterations, nodes_size(), length));
1721 // Possible infinite build_connection_graph loop,
1722 // bailout (no changes to ideal graph were made).
1723 _collecting = false;
1724 return false;
1725 }
1726 #undef CG_BUILD_ITER_LIMIT
1727 #undef CG_BUILD_TIME_LIMIT
1729 // 5. Propagate escaped states.
1730 worklist.clear();
1732 // mark all nodes reachable from GlobalEscape nodes
1733 (void)propagate_escape_state(&cg_worklist, &worklist, PointsToNode::GlobalEscape);
1735 // mark all nodes reachable from ArgEscape nodes
1736 bool has_non_escaping_obj = propagate_escape_state(&cg_worklist, &worklist, PointsToNode::ArgEscape);
1738 Arena* arena = Thread::current()->resource_area();
1739 VectorSet visited(arena);
1741 // 6. Find fields initializing values for not escaped allocations
1742 uint alloc_length = alloc_worklist.length();
1743 for (uint next = 0; next < alloc_length; ++next) {
1744 Node* n = alloc_worklist.at(next);
1745 PointsToNode::EscapeState es = ptnode_adr(n->_idx)->escape_state();
1746 if (es == PointsToNode::NoEscape) {
1747 has_non_escaping_obj = true;
1748 if (n->is_Allocate()) {
1749 find_init_values(n, &visited, igvn);
1750 // The object allocated by this Allocate node will never be
1751 // seen by an other thread. Mark it so that when it is
1752 // expanded no MemBarStoreStore is added.
1753 n->as_Allocate()->initialization()->set_does_not_escape();
1754 }
1755 } else if ((es == PointsToNode::ArgEscape) && n->is_Allocate()) {
1756 // Same as above. Mark this Allocate node so that when it is
1757 // expanded no MemBarStoreStore is added.
1758 n->as_Allocate()->initialization()->set_does_not_escape();
1759 }
1760 }
1762 uint cg_length = cg_worklist.length();
1764 // Skip the rest of code if all objects escaped.
1765 if (!has_non_escaping_obj) {
1766 cg_length = 0;
1767 addp_length = 0;
1768 }
1770 for (uint next = 0; next < cg_length; ++next) {
1771 int ni = cg_worklist.at(next);
1772 PointsToNode* ptn = ptnode_adr(ni);
1773 PointsToNode::NodeType nt = ptn->node_type();
1774 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1775 if (ptn->edge_count() == 0) {
1776 // No values were found. Assume the value was set
1777 // outside this method - add edge to phantom object.
1778 add_pointsto_edge(ni, _phantom_object);
1779 }
1780 }
1781 }
1783 // 7. Remove deferred edges from the graph.
1784 for (uint next = 0; next < cg_length; ++next) {
1785 int ni = cg_worklist.at(next);
1786 PointsToNode* ptn = ptnode_adr(ni);
1787 PointsToNode::NodeType nt = ptn->node_type();
1788 if (nt == PointsToNode::LocalVar || nt == PointsToNode::Field) {
1789 remove_deferred(ni, &worklist, &visited);
1790 }
1791 }
1793 // 8. Adjust escape state of nonescaping objects.
1794 for (uint next = 0; next < addp_length; ++next) {
1795 Node* n = addp_worklist.at(next);
1796 adjust_escape_state(n);
1797 }
1799 // push all NoEscape nodes on the worklist
1800 worklist.clear();
1801 for( uint next = 0; next < cg_length; ++next ) {
1802 int nk = cg_worklist.at(next);
1803 if (ptnode_adr(nk)->escape_state() == PointsToNode::NoEscape &&
1804 !is_null_ptr(nk))
1805 worklist.push(nk);
1806 }
1808 alloc_worklist.clear();
1809 // Propagate scalar_replaceable value.
1810 while(worklist.length() > 0) {
1811 uint nk = worklist.pop();
1812 PointsToNode* ptn = ptnode_adr(nk);
1813 Node* n = ptn->_node;
1814 bool scalar_replaceable = ptn->scalar_replaceable();
1815 if (n->is_Allocate() && scalar_replaceable) {
1816 // Push scalar replaceable allocations on alloc_worklist
1817 // for processing in split_unique_types(). Note,
1818 // following code may change scalar_replaceable value.
1819 alloc_worklist.append(n);
1820 }
1821 uint e_cnt = ptn->edge_count();
1822 for (uint ei = 0; ei < e_cnt; ei++) {
1823 uint npi = ptn->edge_target(ei);
1824 if (is_null_ptr(npi))
1825 continue;
1826 PointsToNode *np = ptnode_adr(npi);
1827 if (np->escape_state() < PointsToNode::NoEscape) {
1828 set_escape_state(npi, PointsToNode::NoEscape);
1829 if (!scalar_replaceable) {
1830 np->set_scalar_replaceable(false);
1831 }
1832 worklist.push(npi);
1833 } else if (np->scalar_replaceable() && !scalar_replaceable) {
1834 np->set_scalar_replaceable(false);
1835 worklist.push(npi);
1836 }
1837 }
1838 }
1840 _collecting = false;
1841 assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
1843 assert(ptnode_adr(_oop_null)->escape_state() == PointsToNode::NoEscape &&
1844 ptnode_adr(_oop_null)->edge_count() == 0, "sanity");
1845 if (UseCompressedOops) {
1846 assert(ptnode_adr(_noop_null)->escape_state() == PointsToNode::NoEscape &&
1847 ptnode_adr(_noop_null)->edge_count() == 0, "sanity");
1848 }
1850 if (EliminateLocks && has_non_escaping_obj) {
1851 // Mark locks before changing ideal graph.
1852 int cnt = C->macro_count();
1853 for( int i=0; i < cnt; i++ ) {
1854 Node *n = C->macro_node(i);
1855 if (n->is_AbstractLock()) { // Lock and Unlock nodes
1856 AbstractLockNode* alock = n->as_AbstractLock();
1857 if (!alock->is_non_esc_obj()) {
1858 PointsToNode::EscapeState es = escape_state(alock->obj_node());
1859 assert(es != PointsToNode::UnknownEscape, "should know");
1860 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1861 assert(!alock->is_eliminated() || alock->is_coarsened(), "sanity");
1862 // The lock could be marked eliminated by lock coarsening
1863 // code during first IGVN before EA. Replace coarsened flag
1864 // to eliminate all associated locks/unlocks.
1865 alock->set_non_esc_obj();
1866 }
1867 }
1868 }
1869 }
1870 }
1872 if (OptimizePtrCompare && has_non_escaping_obj) {
1873 // Add ConI(#CC_GT) and ConI(#CC_EQ).
1874 _pcmp_neq = igvn->makecon(TypeInt::CC_GT);
1875 _pcmp_eq = igvn->makecon(TypeInt::CC_EQ);
1876 // Optimize objects compare.
1877 while (ptr_cmp_worklist.length() != 0) {
1878 Node *n = ptr_cmp_worklist.pop();
1879 Node *res = optimize_ptr_compare(n);
1880 if (res != NULL) {
1881 #ifndef PRODUCT
1882 if (PrintOptimizePtrCompare) {
1883 tty->print_cr("++++ Replaced: %d %s(%d,%d) --> %s", n->_idx, (n->Opcode() == Op_CmpP ? "CmpP" : "CmpN"), n->in(1)->_idx, n->in(2)->_idx, (res == _pcmp_eq ? "EQ" : "NotEQ"));
1884 if (Verbose) {
1885 n->dump(1);
1886 }
1887 }
1888 #endif
1889 _igvn->replace_node(n, res);
1890 }
1891 }
1892 // cleanup
1893 if (_pcmp_neq->outcnt() == 0)
1894 igvn->hash_delete(_pcmp_neq);
1895 if (_pcmp_eq->outcnt() == 0)
1896 igvn->hash_delete(_pcmp_eq);
1897 }
1899 // For MemBarStoreStore nodes added in library_call.cpp, check
1900 // escape status of associated AllocateNode and optimize out
1901 // MemBarStoreStore node if the allocated object never escapes.
1902 while (storestore_worklist.length() != 0) {
1903 Node *n = storestore_worklist.pop();
1904 MemBarStoreStoreNode *storestore = n ->as_MemBarStoreStore();
1905 Node *alloc = storestore->in(MemBarNode::Precedent)->in(0);
1906 assert (alloc->is_Allocate(), "storestore should point to AllocateNode");
1907 PointsToNode::EscapeState es = ptnode_adr(alloc->_idx)->escape_state();
1908 if (es == PointsToNode::NoEscape || es == PointsToNode::ArgEscape) {
1909 MemBarNode* mb = MemBarNode::make(C, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1910 mb->init_req(TypeFunc::Memory, storestore->in(TypeFunc::Memory));
1911 mb->init_req(TypeFunc::Control, storestore->in(TypeFunc::Control));
1913 _igvn->register_new_node_with_optimizer(mb);
1914 _igvn->replace_node(storestore, mb);
1915 }
1916 }
1918 #ifndef PRODUCT
1919 if (PrintEscapeAnalysis) {
1920 dump(); // Dump ConnectionGraph
1921 }
1922 #endif
1924 bool has_scalar_replaceable_candidates = false;
1925 alloc_length = alloc_worklist.length();
1926 for (uint next = 0; next < alloc_length; ++next) {
1927 Node* n = alloc_worklist.at(next);
1928 PointsToNode* ptn = ptnode_adr(n->_idx);
1929 assert(ptn->escape_state() == PointsToNode::NoEscape, "sanity");
1930 if (ptn->scalar_replaceable()) {
1931 has_scalar_replaceable_candidates = true;
1932 break;
1933 }
1934 }
1936 if ( has_scalar_replaceable_candidates &&
1937 C->AliasLevel() >= 3 && EliminateAllocations ) {
1939 // Now use the escape information to create unique types for
1940 // scalar replaceable objects.
1941 split_unique_types(alloc_worklist);
1943 if (C->failing()) return false;
1945 C->print_method("After Escape Analysis", 2);
1947 #ifdef ASSERT
1948 } else if (Verbose && (PrintEscapeAnalysis || PrintEliminateAllocations)) {
1949 tty->print("=== No allocations eliminated for ");
1950 C->method()->print_short_name();
1951 if(!EliminateAllocations) {
1952 tty->print(" since EliminateAllocations is off ===");
1953 } else if(!has_scalar_replaceable_candidates) {
1954 tty->print(" since there are no scalar replaceable candidates ===");
1955 } else if(C->AliasLevel() < 3) {
1956 tty->print(" since AliasLevel < 3 ===");
1957 }
1958 tty->cr();
1959 #endif
1960 }
1961 return has_non_escaping_obj;
1962 }
1964 // Find fields initializing values for allocations.
1965 void ConnectionGraph::find_init_values(Node* alloc, VectorSet* visited, PhaseTransform* phase) {
1966 assert(alloc->is_Allocate(), "Should be called for Allocate nodes only");
1967 PointsToNode* pta = ptnode_adr(alloc->_idx);
1968 assert(pta->escape_state() == PointsToNode::NoEscape, "Not escaped Allocate nodes only");
1969 InitializeNode* ini = alloc->as_Allocate()->initialization();
1971 Compile* C = _compile;
1972 visited->Reset();
1973 // Check if a oop field's initializing value is recorded and add
1974 // a corresponding NULL field's value if it is not recorded.
1975 // Connection Graph does not record a default initialization by NULL
1976 // captured by Initialize node.
1977 //
1978 uint null_idx = UseCompressedOops ? _noop_null : _oop_null;
1979 uint ae_cnt = pta->edge_count();
1980 bool visited_bottom_offset = false;
1981 for (uint ei = 0; ei < ae_cnt; ei++) {
1982 uint nidx = pta->edge_target(ei); // Field (AddP)
1983 PointsToNode* ptn = ptnode_adr(nidx);
1984 assert(ptn->_node->is_AddP(), "Should be AddP nodes only");
1985 int offset = ptn->offset();
1986 if (offset == Type::OffsetBot) {
1987 if (!visited_bottom_offset) {
1988 visited_bottom_offset = true;
1989 // Check only oop fields.
1990 const Type* adr_type = ptn->_node->as_AddP()->bottom_type();
1991 if (!adr_type->isa_aryptr() ||
1992 (adr_type->isa_aryptr()->klass() == NULL) ||
1993 adr_type->isa_aryptr()->klass()->is_obj_array_klass()) {
1994 // OffsetBot is used to reference array's element,
1995 // always add reference to NULL since we don't
1996 // known which element is referenced.
1997 add_edge_from_fields(alloc->_idx, null_idx, offset);
1998 }
1999 }
2000 } else if (offset != oopDesc::klass_offset_in_bytes() &&
2001 !visited->test_set(offset)) {
2003 // Check only oop fields.
2004 const Type* adr_type = ptn->_node->as_AddP()->bottom_type();
2005 BasicType basic_field_type = T_INT;
2006 if (adr_type->isa_instptr()) {
2007 ciField* field = C->alias_type(adr_type->isa_instptr())->field();
2008 if (field != NULL) {
2009 basic_field_type = field->layout_type();
2010 } else {
2011 // Ignore non field load (for example, klass load)
2012 }
2013 } else if (adr_type->isa_aryptr()) {
2014 if (offset != arrayOopDesc::length_offset_in_bytes()) {
2015 const Type* elemtype = adr_type->isa_aryptr()->elem();
2016 basic_field_type = elemtype->array_element_basic_type();
2017 } else {
2018 // Ignore array length load
2019 }
2020 #ifdef ASSERT
2021 } else {
2022 // Raw pointers are used for initializing stores so skip it
2023 // since it should be recorded already
2024 Node* base = get_addp_base(ptn->_node);
2025 assert(adr_type->isa_rawptr() && base->is_Proj() &&
2026 (base->in(0) == alloc),"unexpected pointer type");
2027 #endif
2028 }
2029 if (basic_field_type == T_OBJECT ||
2030 basic_field_type == T_NARROWOOP ||
2031 basic_field_type == T_ARRAY) {
2032 Node* value = NULL;
2033 if (ini != NULL) {
2034 BasicType ft = UseCompressedOops ? T_NARROWOOP : T_OBJECT;
2035 Node* store = ini->find_captured_store(offset, type2aelembytes(ft), phase);
2036 if (store != NULL && store->is_Store()) {
2037 value = store->in(MemNode::ValueIn);
2038 } else if (ptn->edge_count() > 0) { // Are there oop stores?
2039 // Check for a store which follows allocation without branches.
2040 // For example, a volatile field store is not collected
2041 // by Initialize node. TODO: it would be nice to use idom() here.
2042 //
2043 // Search all references to the same field which use different
2044 // AddP nodes, for example, in the next case:
2045 //
2046 // Point p[] = new Point[1];
2047 // if ( x ) { p[0] = new Point(); p[0].x = x; }
2048 // if ( p[0] != null ) { y = p[0].x; } // has CastPP
2049 //
2050 for (uint next = ei; (next < ae_cnt) && (value == NULL); next++) {
2051 uint fpi = pta->edge_target(next); // Field (AddP)
2052 PointsToNode *ptf = ptnode_adr(fpi);
2053 if (ptf->offset() == offset) {
2054 Node* nf = ptf->_node;
2055 for (DUIterator_Fast imax, i = nf->fast_outs(imax); i < imax; i++) {
2056 store = nf->fast_out(i);
2057 if (store->is_Store() && store->in(0) != NULL) {
2058 Node* ctrl = store->in(0);
2059 while(!(ctrl == ini || ctrl == alloc || ctrl == NULL ||
2060 ctrl == C->root() || ctrl == C->top() || ctrl->is_Region() ||
2061 ctrl->is_IfTrue() || ctrl->is_IfFalse())) {
2062 ctrl = ctrl->in(0);
2063 }
2064 if (ctrl == ini || ctrl == alloc) {
2065 value = store->in(MemNode::ValueIn);
2066 break;
2067 }
2068 }
2069 }
2070 }
2071 }
2072 }
2073 }
2074 if (value == NULL || value != ptnode_adr(value->_idx)->_node) {
2075 // A field's initializing value was not recorded. Add NULL.
2076 add_edge_from_fields(alloc->_idx, null_idx, offset);
2077 }
2078 }
2079 }
2080 }
2081 }
2083 // Adjust escape state after Connection Graph is built.
2084 void ConnectionGraph::adjust_escape_state(Node* n) {
2085 PointsToNode* ptn = ptnode_adr(n->_idx);
2086 assert(n->is_AddP(), "Should be called for AddP nodes only");
2087 // Search for objects which are not scalar replaceable
2088 // and mark them to propagate the state to referenced objects.
2089 //
2091 int offset = ptn->offset();
2092 Node* base = get_addp_base(n);
2093 VectorSet* ptset = PointsTo(base);
2094 int ptset_size = ptset->Size();
2096 // An object is not scalar replaceable if the field which may point
2097 // to it has unknown offset (unknown element of an array of objects).
2098 //
2100 if (offset == Type::OffsetBot) {
2101 uint e_cnt = ptn->edge_count();
2102 for (uint ei = 0; ei < e_cnt; ei++) {
2103 uint npi = ptn->edge_target(ei);
2104 ptnode_adr(npi)->set_scalar_replaceable(false);
2105 }
2106 }
2108 // Currently an object is not scalar replaceable if a LoadStore node
2109 // access its field since the field value is unknown after it.
2110 //
2111 bool has_LoadStore = false;
2112 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2113 Node *use = n->fast_out(i);
2114 if (use->is_LoadStore()) {
2115 has_LoadStore = true;
2116 break;
2117 }
2118 }
2119 // An object is not scalar replaceable if the address points
2120 // to unknown field (unknown element for arrays, offset is OffsetBot).
2121 //
2122 // Or the address may point to more then one object. This may produce
2123 // the false positive result (set not scalar replaceable)
2124 // since the flow-insensitive escape analysis can't separate
2125 // the case when stores overwrite the field's value from the case
2126 // when stores happened on different control branches.
2127 //
2128 // Note: it will disable scalar replacement in some cases:
2129 //
2130 // Point p[] = new Point[1];
2131 // p[0] = new Point(); // Will be not scalar replaced
2132 //
2133 // but it will save us from incorrect optimizations in next cases:
2134 //
2135 // Point p[] = new Point[1];
2136 // if ( x ) p[0] = new Point(); // Will be not scalar replaced
2137 //
2138 if (ptset_size > 1 || ptset_size != 0 &&
2139 (has_LoadStore || offset == Type::OffsetBot)) {
2140 for( VectorSetI j(ptset); j.test(); ++j ) {
2141 ptnode_adr(j.elem)->set_scalar_replaceable(false);
2142 }
2143 }
2144 }
2146 // Propagate escape states to referenced nodes.
2147 bool ConnectionGraph::propagate_escape_state(GrowableArray<int>* cg_worklist,
2148 GrowableArray<uint>* worklist,
2149 PointsToNode::EscapeState esc_state) {
2150 bool has_java_obj = false;
2152 // push all nodes with the same escape state on the worklist
2153 uint cg_length = cg_worklist->length();
2154 for (uint next = 0; next < cg_length; ++next) {
2155 int nk = cg_worklist->at(next);
2156 if (ptnode_adr(nk)->escape_state() == esc_state)
2157 worklist->push(nk);
2158 }
2159 // mark all reachable nodes
2160 while (worklist->length() > 0) {
2161 int pt = worklist->pop();
2162 PointsToNode* ptn = ptnode_adr(pt);
2163 if (ptn->node_type() == PointsToNode::JavaObject &&
2164 !is_null_ptr(pt)) {
2165 has_java_obj = true;
2166 if (esc_state > PointsToNode::NoEscape) {
2167 // fields values are unknown if object escapes
2168 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2169 }
2170 }
2171 uint e_cnt = ptn->edge_count();
2172 for (uint ei = 0; ei < e_cnt; ei++) {
2173 uint npi = ptn->edge_target(ei);
2174 if (is_null_ptr(npi))
2175 continue;
2176 PointsToNode *np = ptnode_adr(npi);
2177 if (np->escape_state() < esc_state) {
2178 set_escape_state(npi, esc_state);
2179 worklist->push(npi);
2180 }
2181 }
2182 }
2183 // Has not escaping java objects
2184 return has_java_obj && (esc_state < PointsToNode::GlobalEscape);
2185 }
2187 // Optimize objects compare.
2188 Node* ConnectionGraph::optimize_ptr_compare(Node* n) {
2189 assert(OptimizePtrCompare, "sanity");
2190 // Clone returned Set since PointsTo() returns pointer
2191 // to the same structure ConnectionGraph.pt_ptset.
2192 VectorSet ptset1 = *PointsTo(n->in(1));
2193 VectorSet ptset2 = *PointsTo(n->in(2));
2195 // Check simple cases first.
2196 if (ptset1.Size() == 1) {
2197 uint pt1 = ptset1.getelem();
2198 PointsToNode* ptn1 = ptnode_adr(pt1);
2199 if (ptn1->escape_state() == PointsToNode::NoEscape) {
2200 if (ptset2.Size() == 1 && ptset2.getelem() == pt1) {
2201 // Comparing the same not escaping object.
2202 return _pcmp_eq;
2203 }
2204 Node* obj = ptn1->_node;
2205 // Comparing not escaping allocation.
2206 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
2207 !ptset2.test(pt1)) {
2208 return _pcmp_neq; // This includes nullness check.
2209 }
2210 }
2211 } else if (ptset2.Size() == 1) {
2212 uint pt2 = ptset2.getelem();
2213 PointsToNode* ptn2 = ptnode_adr(pt2);
2214 if (ptn2->escape_state() == PointsToNode::NoEscape) {
2215 Node* obj = ptn2->_node;
2216 // Comparing not escaping allocation.
2217 if ((obj->is_Allocate() || obj->is_CallStaticJava()) &&
2218 !ptset1.test(pt2)) {
2219 return _pcmp_neq; // This includes nullness check.
2220 }
2221 }
2222 }
2224 if (!ptset1.disjoint(ptset2)) {
2225 return NULL; // Sets are not disjoint
2226 }
2228 // Sets are disjoint.
2229 bool set1_has_unknown_ptr = ptset1.test(_phantom_object) != 0;
2230 bool set2_has_unknown_ptr = ptset2.test(_phantom_object) != 0;
2231 bool set1_has_null_ptr = (ptset1.test(_oop_null) | ptset1.test(_noop_null)) != 0;
2232 bool set2_has_null_ptr = (ptset2.test(_oop_null) | ptset2.test(_noop_null)) != 0;
2234 if (set1_has_unknown_ptr && set2_has_null_ptr ||
2235 set2_has_unknown_ptr && set1_has_null_ptr) {
2236 // Check nullness of unknown object.
2237 return NULL;
2238 }
2240 // Disjointness by itself is not sufficient since
2241 // alias analysis is not complete for escaped objects.
2242 // Disjoint sets are definitely unrelated only when
2243 // at least one set has only not escaping objects.
2244 if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
2245 bool has_only_non_escaping_alloc = true;
2246 for (VectorSetI i(&ptset1); i.test(); ++i) {
2247 uint pt = i.elem;
2248 PointsToNode* ptn = ptnode_adr(pt);
2249 Node* obj = ptn->_node;
2250 if (ptn->escape_state() != PointsToNode::NoEscape ||
2251 !(obj->is_Allocate() || obj->is_CallStaticJava())) {
2252 has_only_non_escaping_alloc = false;
2253 break;
2254 }
2255 }
2256 if (has_only_non_escaping_alloc) {
2257 return _pcmp_neq;
2258 }
2259 }
2260 if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
2261 bool has_only_non_escaping_alloc = true;
2262 for (VectorSetI i(&ptset2); i.test(); ++i) {
2263 uint pt = i.elem;
2264 PointsToNode* ptn = ptnode_adr(pt);
2265 Node* obj = ptn->_node;
2266 if (ptn->escape_state() != PointsToNode::NoEscape ||
2267 !(obj->is_Allocate() || obj->is_CallStaticJava())) {
2268 has_only_non_escaping_alloc = false;
2269 break;
2270 }
2271 }
2272 if (has_only_non_escaping_alloc) {
2273 return _pcmp_neq;
2274 }
2275 }
2276 return NULL;
2277 }
2279 void ConnectionGraph::process_call_arguments(CallNode *call, PhaseTransform *phase) {
2280 bool is_arraycopy = false;
2281 switch (call->Opcode()) {
2282 #ifdef ASSERT
2283 case Op_Allocate:
2284 case Op_AllocateArray:
2285 case Op_Lock:
2286 case Op_Unlock:
2287 assert(false, "should be done already");
2288 break;
2289 #endif
2290 case Op_CallLeafNoFP:
2291 is_arraycopy = (call->as_CallLeaf()->_name != NULL &&
2292 strstr(call->as_CallLeaf()->_name, "arraycopy") != 0);
2293 // fall through
2294 case Op_CallLeaf:
2295 {
2296 // Stub calls, objects do not escape but they are not scale replaceable.
2297 // Adjust escape state for outgoing arguments.
2298 const TypeTuple * d = call->tf()->domain();
2299 bool src_has_oops = false;
2300 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2301 const Type* at = d->field_at(i);
2302 Node *arg = call->in(i)->uncast();
2303 const Type *aat = phase->type(arg);
2304 PointsToNode::EscapeState arg_esc = ptnode_adr(arg->_idx)->escape_state();
2305 if (!arg->is_top() && at->isa_ptr() && aat->isa_ptr() &&
2306 (is_arraycopy || arg_esc < PointsToNode::ArgEscape)) {
2307 #ifdef ASSERT
2308 assert(aat == Type::TOP || aat == TypePtr::NULL_PTR ||
2309 aat->isa_ptr() != NULL, "expecting an Ptr");
2310 if (!(is_arraycopy ||
2311 call->as_CallLeaf()->_name != NULL &&
2312 (strcmp(call->as_CallLeaf()->_name, "g1_wb_pre") == 0 ||
2313 strcmp(call->as_CallLeaf()->_name, "g1_wb_post") == 0 ))
2314 ) {
2315 call->dump();
2316 assert(false, "EA: unexpected CallLeaf");
2317 }
2318 #endif
2319 if (arg_esc < PointsToNode::ArgEscape) {
2320 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2321 Node* arg_base = arg;
2322 if (arg->is_AddP()) {
2323 //
2324 // The inline_native_clone() case when the arraycopy stub is called
2325 // after the allocation before Initialize and CheckCastPP nodes.
2326 // Or normal arraycopy for object arrays case.
2327 //
2328 // Set AddP's base (Allocate) as not scalar replaceable since
2329 // pointer to the base (with offset) is passed as argument.
2330 //
2331 arg_base = get_addp_base(arg);
2332 set_escape_state(arg_base->_idx, PointsToNode::ArgEscape);
2333 }
2334 }
2336 bool arg_has_oops = aat->isa_oopptr() &&
2337 (aat->isa_oopptr()->klass() == NULL || aat->isa_instptr() ||
2338 (aat->isa_aryptr() && aat->isa_aryptr()->klass()->is_obj_array_klass()));
2339 if (i == TypeFunc::Parms) {
2340 src_has_oops = arg_has_oops;
2341 }
2342 //
2343 // src or dst could be j.l.Object when other is basic type array:
2344 //
2345 // arraycopy(char[],0,Object*,0,size);
2346 // arraycopy(Object*,0,char[],0,size);
2347 //
2348 // Do nothing special in such cases.
2349 //
2350 if (is_arraycopy && (i > TypeFunc::Parms) &&
2351 src_has_oops && arg_has_oops) {
2352 // Destination object's fields reference an unknown object.
2353 Node* arg_base = arg;
2354 if (arg->is_AddP()) {
2355 arg_base = get_addp_base(arg);
2356 }
2357 for (VectorSetI s(PointsTo(arg_base)); s.test(); ++s) {
2358 uint ps = s.elem;
2359 set_escape_state(ps, PointsToNode::ArgEscape);
2360 add_edge_from_fields(ps, _phantom_object, Type::OffsetBot);
2361 }
2362 // Conservatively all values in source object fields globally escape
2363 // since we don't know if values in destination object fields
2364 // escape (it could be traced but it is too expensive).
2365 Node* src = call->in(TypeFunc::Parms)->uncast();
2366 Node* src_base = src;
2367 if (src->is_AddP()) {
2368 src_base = get_addp_base(src);
2369 }
2370 for (VectorSetI s(PointsTo(src_base)); s.test(); ++s) {
2371 uint ps = s.elem;
2372 set_escape_state(ps, PointsToNode::ArgEscape);
2373 // Use OffsetTop to indicate fields global escape.
2374 add_edge_from_fields(ps, _phantom_object, Type::OffsetTop);
2375 }
2376 }
2377 }
2378 }
2379 break;
2380 }
2382 case Op_CallStaticJava:
2383 // For a static call, we know exactly what method is being called.
2384 // Use bytecode estimator to record the call's escape affects
2385 {
2386 ciMethod *meth = call->as_CallJava()->method();
2387 BCEscapeAnalyzer *call_analyzer = (meth !=NULL) ? meth->get_bcea() : NULL;
2388 // fall-through if not a Java method or no analyzer information
2389 if (call_analyzer != NULL) {
2390 const TypeTuple * d = call->tf()->domain();
2391 bool copy_dependencies = false;
2392 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2393 const Type* at = d->field_at(i);
2394 int k = i - TypeFunc::Parms;
2395 Node *arg = call->in(i)->uncast();
2397 if (at->isa_oopptr() != NULL &&
2398 ptnode_adr(arg->_idx)->escape_state() < PointsToNode::GlobalEscape) {
2400 bool global_escapes = false;
2401 bool fields_escapes = false;
2402 if (!call_analyzer->is_arg_stack(k)) {
2403 // The argument global escapes, mark everything it could point to
2404 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2405 global_escapes = true;
2406 } else {
2407 if (!call_analyzer->is_arg_local(k)) {
2408 // The argument itself doesn't escape, but any fields might
2409 fields_escapes = true;
2410 }
2411 set_escape_state(arg->_idx, PointsToNode::ArgEscape);
2412 copy_dependencies = true;
2413 }
2415 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2416 uint pt = j.elem;
2417 if (global_escapes) {
2418 // The argument global escapes, mark everything it could point to
2419 set_escape_state(pt, PointsToNode::GlobalEscape);
2420 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2421 } else {
2422 set_escape_state(pt, PointsToNode::ArgEscape);
2423 if (fields_escapes) {
2424 // The argument itself doesn't escape, but any fields might.
2425 // Use OffsetTop to indicate such case.
2426 add_edge_from_fields(pt, _phantom_object, Type::OffsetTop);
2427 }
2428 }
2429 }
2430 }
2431 }
2432 if (copy_dependencies)
2433 call_analyzer->copy_dependencies(_compile->dependencies());
2434 break;
2435 }
2436 }
2438 default:
2439 // Fall-through here if not a Java method or no analyzer information
2440 // or some other type of call, assume the worst case: all arguments
2441 // globally escape.
2442 {
2443 // adjust escape state for outgoing arguments
2444 const TypeTuple * d = call->tf()->domain();
2445 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2446 const Type* at = d->field_at(i);
2447 if (at->isa_oopptr() != NULL) {
2448 Node *arg = call->in(i)->uncast();
2449 set_escape_state(arg->_idx, PointsToNode::GlobalEscape);
2450 for( VectorSetI j(PointsTo(arg)); j.test(); ++j ) {
2451 uint pt = j.elem;
2452 set_escape_state(pt, PointsToNode::GlobalEscape);
2453 add_edge_from_fields(pt, _phantom_object, Type::OffsetBot);
2454 }
2455 }
2456 }
2457 }
2458 }
2459 }
2460 void ConnectionGraph::process_call_result(ProjNode *resproj, PhaseTransform *phase) {
2461 CallNode *call = resproj->in(0)->as_Call();
2462 uint call_idx = call->_idx;
2463 uint resproj_idx = resproj->_idx;
2465 switch (call->Opcode()) {
2466 case Op_Allocate:
2467 {
2468 Node *k = call->in(AllocateNode::KlassNode);
2469 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2470 assert(kt != NULL, "TypeKlassPtr required.");
2471 ciKlass* cik = kt->klass();
2473 PointsToNode::EscapeState es;
2474 uint edge_to;
2475 if (cik->is_subclass_of(_compile->env()->Thread_klass()) ||
2476 !cik->is_instance_klass() || // StressReflectiveCode
2477 cik->as_instance_klass()->has_finalizer()) {
2478 es = PointsToNode::GlobalEscape;
2479 edge_to = _phantom_object; // Could not be worse
2480 } else {
2481 es = PointsToNode::NoEscape;
2482 edge_to = call_idx;
2483 assert(ptnode_adr(call_idx)->scalar_replaceable(), "sanity");
2484 }
2485 set_escape_state(call_idx, es);
2486 add_pointsto_edge(resproj_idx, edge_to);
2487 _processed.set(resproj_idx);
2488 break;
2489 }
2491 case Op_AllocateArray:
2492 {
2494 Node *k = call->in(AllocateNode::KlassNode);
2495 const TypeKlassPtr *kt = k->bottom_type()->isa_klassptr();
2496 assert(kt != NULL, "TypeKlassPtr required.");
2497 ciKlass* cik = kt->klass();
2499 PointsToNode::EscapeState es;
2500 uint edge_to;
2501 if (!cik->is_array_klass()) { // StressReflectiveCode
2502 es = PointsToNode::GlobalEscape;
2503 edge_to = _phantom_object;
2504 } else {
2505 es = PointsToNode::NoEscape;
2506 edge_to = call_idx;
2507 assert(ptnode_adr(call_idx)->scalar_replaceable(), "sanity");
2508 int length = call->in(AllocateNode::ALength)->find_int_con(-1);
2509 if (length < 0 || length > EliminateAllocationArraySizeLimit) {
2510 // Not scalar replaceable if the length is not constant or too big.
2511 ptnode_adr(call_idx)->set_scalar_replaceable(false);
2512 }
2513 }
2514 set_escape_state(call_idx, es);
2515 add_pointsto_edge(resproj_idx, edge_to);
2516 _processed.set(resproj_idx);
2517 break;
2518 }
2520 case Op_CallStaticJava:
2521 // For a static call, we know exactly what method is being called.
2522 // Use bytecode estimator to record whether the call's return value escapes
2523 {
2524 bool done = true;
2525 const TypeTuple *r = call->tf()->range();
2526 const Type* ret_type = NULL;
2528 if (r->cnt() > TypeFunc::Parms)
2529 ret_type = r->field_at(TypeFunc::Parms);
2531 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2532 // _multianewarray functions return a TypeRawPtr.
2533 if (ret_type == NULL || ret_type->isa_ptr() == NULL) {
2534 _processed.set(resproj_idx);
2535 break; // doesn't return a pointer type
2536 }
2537 ciMethod *meth = call->as_CallJava()->method();
2538 const TypeTuple * d = call->tf()->domain();
2539 if (meth == NULL) {
2540 // not a Java method, assume global escape
2541 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2542 add_pointsto_edge(resproj_idx, _phantom_object);
2543 } else {
2544 BCEscapeAnalyzer *call_analyzer = meth->get_bcea();
2545 bool copy_dependencies = false;
2547 if (call_analyzer->is_return_allocated()) {
2548 // Returns a newly allocated unescaped object, simply
2549 // update dependency information.
2550 // Mark it as NoEscape so that objects referenced by
2551 // it's fields will be marked as NoEscape at least.
2552 set_escape_state(call_idx, PointsToNode::NoEscape);
2553 ptnode_adr(call_idx)->set_scalar_replaceable(false);
2554 // Fields values are unknown
2555 add_edge_from_fields(call_idx, _phantom_object, Type::OffsetBot);
2556 add_pointsto_edge(resproj_idx, call_idx);
2557 copy_dependencies = true;
2558 } else {
2559 // determine whether any arguments are returned
2560 set_escape_state(call_idx, PointsToNode::ArgEscape);
2561 bool ret_arg = false;
2562 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
2563 const Type* at = d->field_at(i);
2564 if (at->isa_oopptr() != NULL) {
2565 Node *arg = call->in(i)->uncast();
2567 if (call_analyzer->is_arg_returned(i - TypeFunc::Parms)) {
2568 ret_arg = true;
2569 PointsToNode *arg_esp = ptnode_adr(arg->_idx);
2570 if (arg_esp->node_type() == PointsToNode::UnknownType)
2571 done = false;
2572 else if (arg_esp->node_type() == PointsToNode::JavaObject)
2573 add_pointsto_edge(resproj_idx, arg->_idx);
2574 else
2575 add_deferred_edge(resproj_idx, arg->_idx);
2576 }
2577 }
2578 }
2579 if (done) {
2580 copy_dependencies = true;
2581 // is_return_local() is true when only arguments are returned.
2582 if (!ret_arg || !call_analyzer->is_return_local()) {
2583 // Returns unknown object.
2584 add_pointsto_edge(resproj_idx, _phantom_object);
2585 }
2586 }
2587 }
2588 if (copy_dependencies)
2589 call_analyzer->copy_dependencies(_compile->dependencies());
2590 }
2591 if (done)
2592 _processed.set(resproj_idx);
2593 break;
2594 }
2596 default:
2597 // Some other type of call, assume the worst case that the
2598 // returned value, if any, globally escapes.
2599 {
2600 const TypeTuple *r = call->tf()->range();
2601 if (r->cnt() > TypeFunc::Parms) {
2602 const Type* ret_type = r->field_at(TypeFunc::Parms);
2604 // Note: we use isa_ptr() instead of isa_oopptr() here because the
2605 // _multianewarray functions return a TypeRawPtr.
2606 if (ret_type->isa_ptr() != NULL) {
2607 set_escape_state(call_idx, PointsToNode::GlobalEscape);
2608 add_pointsto_edge(resproj_idx, _phantom_object);
2609 }
2610 }
2611 _processed.set(resproj_idx);
2612 }
2613 }
2614 }
2616 // Populate Connection Graph with Ideal nodes and create simple
2617 // connection graph edges (do not need to check the node_type of inputs
2618 // or to call PointsTo() to walk the connection graph).
2619 void ConnectionGraph::record_for_escape_analysis(Node *n, PhaseTransform *phase) {
2620 if (_processed.test(n->_idx))
2621 return; // No need to redefine node's state.
2623 if (n->is_Call()) {
2624 // Arguments to allocation and locking don't escape.
2625 if (n->is_Allocate()) {
2626 add_node(n, PointsToNode::JavaObject, PointsToNode::UnknownEscape, true);
2627 record_for_optimizer(n);
2628 } else if (n->is_Lock() || n->is_Unlock()) {
2629 // Put Lock and Unlock nodes on IGVN worklist to process them during
2630 // the first IGVN optimization when escape information is still available.
2631 record_for_optimizer(n);
2632 _processed.set(n->_idx);
2633 } else {
2634 // Don't mark as processed since call's arguments have to be processed.
2635 PointsToNode::NodeType nt = PointsToNode::UnknownType;
2636 PointsToNode::EscapeState es = PointsToNode::UnknownEscape;
2638 // Check if a call returns an object.
2639 const TypeTuple *r = n->as_Call()->tf()->range();
2640 if (r->cnt() > TypeFunc::Parms &&
2641 r->field_at(TypeFunc::Parms)->isa_ptr() &&
2642 n->as_Call()->proj_out(TypeFunc::Parms) != NULL) {
2643 nt = PointsToNode::JavaObject;
2644 if (!n->is_CallStaticJava()) {
2645 // Since the called mathod is statically unknown assume
2646 // the worst case that the returned value globally escapes.
2647 es = PointsToNode::GlobalEscape;
2648 }
2649 }
2650 add_node(n, nt, es, false);
2651 }
2652 return;
2653 }
2655 // Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
2656 // ThreadLocal has RawPrt type.
2657 switch (n->Opcode()) {
2658 case Op_AddP:
2659 {
2660 add_node(n, PointsToNode::Field, PointsToNode::UnknownEscape, false);
2661 break;
2662 }
2663 case Op_CastX2P:
2664 { // "Unsafe" memory access.
2665 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2666 break;
2667 }
2668 case Op_CastPP:
2669 case Op_CheckCastPP:
2670 case Op_EncodeP:
2671 case Op_DecodeN:
2672 {
2673 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2674 int ti = n->in(1)->_idx;
2675 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2676 if (nt == PointsToNode::UnknownType) {
2677 _delayed_worklist.push(n); // Process it later.
2678 break;
2679 } else if (nt == PointsToNode::JavaObject) {
2680 add_pointsto_edge(n->_idx, ti);
2681 } else {
2682 add_deferred_edge(n->_idx, ti);
2683 }
2684 _processed.set(n->_idx);
2685 break;
2686 }
2687 case Op_ConP:
2688 {
2689 // assume all pointer constants globally escape except for null
2690 PointsToNode::EscapeState es;
2691 if (phase->type(n) == TypePtr::NULL_PTR)
2692 es = PointsToNode::NoEscape;
2693 else
2694 es = PointsToNode::GlobalEscape;
2696 add_node(n, PointsToNode::JavaObject, es, true);
2697 break;
2698 }
2699 case Op_ConN:
2700 {
2701 // assume all narrow oop constants globally escape except for null
2702 PointsToNode::EscapeState es;
2703 if (phase->type(n) == TypeNarrowOop::NULL_PTR)
2704 es = PointsToNode::NoEscape;
2705 else
2706 es = PointsToNode::GlobalEscape;
2708 add_node(n, PointsToNode::JavaObject, es, true);
2709 break;
2710 }
2711 case Op_CreateEx:
2712 {
2713 // assume that all exception objects globally escape
2714 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2715 break;
2716 }
2717 case Op_LoadKlass:
2718 case Op_LoadNKlass:
2719 {
2720 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, true);
2721 break;
2722 }
2723 case Op_LoadP:
2724 case Op_LoadN:
2725 {
2726 const Type *t = phase->type(n);
2727 if (t->make_ptr() == NULL) {
2728 _processed.set(n->_idx);
2729 return;
2730 }
2731 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2732 break;
2733 }
2734 case Op_Parm:
2735 {
2736 _processed.set(n->_idx); // No need to redefine it state.
2737 uint con = n->as_Proj()->_con;
2738 if (con < TypeFunc::Parms)
2739 return;
2740 const Type *t = n->in(0)->as_Start()->_domain->field_at(con);
2741 if (t->isa_ptr() == NULL)
2742 return;
2743 // We have to assume all input parameters globally escape
2744 // (Note: passing 'false' since _processed is already set).
2745 add_node(n, PointsToNode::JavaObject, PointsToNode::GlobalEscape, false);
2746 break;
2747 }
2748 case Op_PartialSubtypeCheck:
2749 { // Produces Null or notNull and is used in CmpP.
2750 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2751 break;
2752 }
2753 case Op_Phi:
2754 {
2755 const Type *t = n->as_Phi()->type();
2756 if (t->make_ptr() == NULL) {
2757 // nothing to do if not an oop or narrow oop
2758 _processed.set(n->_idx);
2759 return;
2760 }
2761 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2762 uint i;
2763 for (i = 1; i < n->req() ; i++) {
2764 Node* in = n->in(i);
2765 if (in == NULL)
2766 continue; // ignore NULL
2767 in = in->uncast();
2768 if (in->is_top() || in == n)
2769 continue; // ignore top or inputs which go back this node
2770 int ti = in->_idx;
2771 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2772 if (nt == PointsToNode::UnknownType) {
2773 break;
2774 } else if (nt == PointsToNode::JavaObject) {
2775 add_pointsto_edge(n->_idx, ti);
2776 } else {
2777 add_deferred_edge(n->_idx, ti);
2778 }
2779 }
2780 if (i >= n->req())
2781 _processed.set(n->_idx);
2782 else
2783 _delayed_worklist.push(n);
2784 break;
2785 }
2786 case Op_Proj:
2787 {
2788 // we are only interested in the oop result projection from a call
2789 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
2790 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
2791 assert(r->cnt() > TypeFunc::Parms, "sanity");
2792 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
2793 add_node(n, PointsToNode::LocalVar, PointsToNode::UnknownEscape, false);
2794 int ti = n->in(0)->_idx;
2795 // The call may not be registered yet (since not all its inputs are registered)
2796 // if this is the projection from backbranch edge of Phi.
2797 if (ptnode_adr(ti)->node_type() != PointsToNode::UnknownType) {
2798 process_call_result(n->as_Proj(), phase);
2799 }
2800 if (!_processed.test(n->_idx)) {
2801 // The call's result may need to be processed later if the call
2802 // returns it's argument and the argument is not processed yet.
2803 _delayed_worklist.push(n);
2804 }
2805 break;
2806 }
2807 }
2808 _processed.set(n->_idx);
2809 break;
2810 }
2811 case Op_Return:
2812 {
2813 if( n->req() > TypeFunc::Parms &&
2814 phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
2815 // Treat Return value as LocalVar with GlobalEscape escape state.
2816 add_node(n, PointsToNode::LocalVar, PointsToNode::GlobalEscape, false);
2817 int ti = n->in(TypeFunc::Parms)->_idx;
2818 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
2819 if (nt == PointsToNode::UnknownType) {
2820 _delayed_worklist.push(n); // Process it later.
2821 break;
2822 } else if (nt == PointsToNode::JavaObject) {
2823 add_pointsto_edge(n->_idx, ti);
2824 } else {
2825 add_deferred_edge(n->_idx, ti);
2826 }
2827 }
2828 _processed.set(n->_idx);
2829 break;
2830 }
2831 case Op_StoreP:
2832 case Op_StoreN:
2833 {
2834 const Type *adr_type = phase->type(n->in(MemNode::Address));
2835 adr_type = adr_type->make_ptr();
2836 if (adr_type->isa_oopptr()) {
2837 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2838 } else {
2839 Node* adr = n->in(MemNode::Address);
2840 if (adr->is_AddP() && phase->type(adr) == TypeRawPtr::NOTNULL &&
2841 adr->in(AddPNode::Address)->is_Proj() &&
2842 adr->in(AddPNode::Address)->in(0)->is_Allocate()) {
2843 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2844 // We are computing a raw address for a store captured
2845 // by an Initialize compute an appropriate address type.
2846 int offs = (int)phase->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
2847 assert(offs != Type::OffsetBot, "offset must be a constant");
2848 } else {
2849 _processed.set(n->_idx);
2850 return;
2851 }
2852 }
2853 break;
2854 }
2855 case Op_StorePConditional:
2856 case Op_CompareAndSwapP:
2857 case Op_CompareAndSwapN:
2858 {
2859 const Type *adr_type = phase->type(n->in(MemNode::Address));
2860 adr_type = adr_type->make_ptr();
2861 if (adr_type->isa_oopptr()) {
2862 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2863 } else {
2864 _processed.set(n->_idx);
2865 return;
2866 }
2867 break;
2868 }
2869 case Op_AryEq:
2870 case Op_StrComp:
2871 case Op_StrEquals:
2872 case Op_StrIndexOf:
2873 {
2874 // char[] arrays passed to string intrinsics are not scalar replaceable.
2875 add_node(n, PointsToNode::UnknownType, PointsToNode::UnknownEscape, false);
2876 break;
2877 }
2878 case Op_ThreadLocal:
2879 {
2880 add_node(n, PointsToNode::JavaObject, PointsToNode::ArgEscape, true);
2881 break;
2882 }
2883 default:
2884 ;
2885 // nothing to do
2886 }
2887 return;
2888 }
2890 void ConnectionGraph::build_connection_graph(Node *n, PhaseTransform *phase) {
2891 uint n_idx = n->_idx;
2892 assert(ptnode_adr(n_idx)->_node != NULL, "node should be registered");
2894 // Don't set processed bit for AddP, LoadP, StoreP since
2895 // they may need more then one pass to process.
2896 // Also don't mark as processed Call nodes since their
2897 // arguments may need more then one pass to process.
2898 if (_processed.test(n_idx))
2899 return; // No need to redefine node's state.
2901 if (n->is_Call()) {
2902 CallNode *call = n->as_Call();
2903 process_call_arguments(call, phase);
2904 return;
2905 }
2907 switch (n->Opcode()) {
2908 case Op_AddP:
2909 {
2910 Node *base = get_addp_base(n);
2911 int offset = address_offset(n, phase);
2912 // Create a field edge to this node from everything base could point to.
2913 for( VectorSetI i(PointsTo(base)); i.test(); ++i ) {
2914 uint pt = i.elem;
2915 add_field_edge(pt, n_idx, offset);
2916 }
2917 break;
2918 }
2919 case Op_CastX2P:
2920 {
2921 assert(false, "Op_CastX2P");
2922 break;
2923 }
2924 case Op_CastPP:
2925 case Op_CheckCastPP:
2926 case Op_EncodeP:
2927 case Op_DecodeN:
2928 {
2929 int ti = n->in(1)->_idx;
2930 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "all nodes should be registered");
2931 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
2932 add_pointsto_edge(n_idx, ti);
2933 } else {
2934 add_deferred_edge(n_idx, ti);
2935 }
2936 _processed.set(n_idx);
2937 break;
2938 }
2939 case Op_ConP:
2940 {
2941 assert(false, "Op_ConP");
2942 break;
2943 }
2944 case Op_ConN:
2945 {
2946 assert(false, "Op_ConN");
2947 break;
2948 }
2949 case Op_CreateEx:
2950 {
2951 assert(false, "Op_CreateEx");
2952 break;
2953 }
2954 case Op_LoadKlass:
2955 case Op_LoadNKlass:
2956 {
2957 assert(false, "Op_LoadKlass");
2958 break;
2959 }
2960 case Op_LoadP:
2961 case Op_LoadN:
2962 {
2963 const Type *t = phase->type(n);
2964 #ifdef ASSERT
2965 if (t->make_ptr() == NULL)
2966 assert(false, "Op_LoadP");
2967 #endif
2969 Node* adr = n->in(MemNode::Address)->uncast();
2970 Node* adr_base;
2971 if (adr->is_AddP()) {
2972 adr_base = get_addp_base(adr);
2973 } else {
2974 adr_base = adr;
2975 }
2977 // For everything "adr_base" could point to, create a deferred edge from
2978 // this node to each field with the same offset.
2979 int offset = address_offset(adr, phase);
2980 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
2981 uint pt = i.elem;
2982 if (adr->is_AddP()) {
2983 // Add field edge if it is missing.
2984 add_field_edge(pt, adr->_idx, offset);
2985 }
2986 add_deferred_edge_to_fields(n_idx, pt, offset);
2987 }
2988 break;
2989 }
2990 case Op_Parm:
2991 {
2992 assert(false, "Op_Parm");
2993 break;
2994 }
2995 case Op_PartialSubtypeCheck:
2996 {
2997 assert(false, "Op_PartialSubtypeCheck");
2998 break;
2999 }
3000 case Op_Phi:
3001 {
3002 #ifdef ASSERT
3003 const Type *t = n->as_Phi()->type();
3004 if (t->make_ptr() == NULL)
3005 assert(false, "Op_Phi");
3006 #endif
3007 for (uint i = 1; i < n->req() ; i++) {
3008 Node* in = n->in(i);
3009 if (in == NULL)
3010 continue; // ignore NULL
3011 in = in->uncast();
3012 if (in->is_top() || in == n)
3013 continue; // ignore top or inputs which go back this node
3014 int ti = in->_idx;
3015 PointsToNode::NodeType nt = ptnode_adr(ti)->node_type();
3016 assert(nt != PointsToNode::UnknownType, "all nodes should be known");
3017 if (nt == PointsToNode::JavaObject) {
3018 add_pointsto_edge(n_idx, ti);
3019 } else {
3020 add_deferred_edge(n_idx, ti);
3021 }
3022 }
3023 _processed.set(n_idx);
3024 break;
3025 }
3026 case Op_Proj:
3027 {
3028 // we are only interested in the oop result projection from a call
3029 if (n->as_Proj()->_con == TypeFunc::Parms && n->in(0)->is_Call() ) {
3030 assert(ptnode_adr(n->in(0)->_idx)->node_type() != PointsToNode::UnknownType,
3031 "all nodes should be registered");
3032 const TypeTuple *r = n->in(0)->as_Call()->tf()->range();
3033 assert(r->cnt() > TypeFunc::Parms, "sanity");
3034 if (r->field_at(TypeFunc::Parms)->isa_ptr() != NULL) {
3035 process_call_result(n->as_Proj(), phase);
3036 assert(_processed.test(n_idx), "all call results should be processed");
3037 break;
3038 }
3039 }
3040 assert(false, "Op_Proj");
3041 break;
3042 }
3043 case Op_Return:
3044 {
3045 #ifdef ASSERT
3046 if( n->req() <= TypeFunc::Parms ||
3047 !phase->type(n->in(TypeFunc::Parms))->isa_oopptr() ) {
3048 assert(false, "Op_Return");
3049 }
3050 #endif
3051 int ti = n->in(TypeFunc::Parms)->_idx;
3052 assert(ptnode_adr(ti)->node_type() != PointsToNode::UnknownType, "node should be registered");
3053 if (ptnode_adr(ti)->node_type() == PointsToNode::JavaObject) {
3054 add_pointsto_edge(n_idx, ti);
3055 } else {
3056 add_deferred_edge(n_idx, ti);
3057 }
3058 _processed.set(n_idx);
3059 break;
3060 }
3061 case Op_StoreP:
3062 case Op_StoreN:
3063 case Op_StorePConditional:
3064 case Op_CompareAndSwapP:
3065 case Op_CompareAndSwapN:
3066 {
3067 Node *adr = n->in(MemNode::Address);
3068 const Type *adr_type = phase->type(adr)->make_ptr();
3069 #ifdef ASSERT
3070 if (!adr_type->isa_oopptr())
3071 assert(phase->type(adr) == TypeRawPtr::NOTNULL, "Op_StoreP");
3072 #endif
3074 assert(adr->is_AddP(), "expecting an AddP");
3075 Node *adr_base = get_addp_base(adr);
3076 Node *val = n->in(MemNode::ValueIn)->uncast();
3077 int offset = address_offset(adr, phase);
3078 // For everything "adr_base" could point to, create a deferred edge
3079 // to "val" from each field with the same offset.
3080 for( VectorSetI i(PointsTo(adr_base)); i.test(); ++i ) {
3081 uint pt = i.elem;
3082 // Add field edge if it is missing.
3083 add_field_edge(pt, adr->_idx, offset);
3084 add_edge_from_fields(pt, val->_idx, offset);
3085 }
3086 break;
3087 }
3088 case Op_AryEq:
3089 case Op_StrComp:
3090 case Op_StrEquals:
3091 case Op_StrIndexOf:
3092 {
3093 // char[] arrays passed to string intrinsic do not escape but
3094 // they are not scalar replaceable. Adjust escape state for them.
3095 // Start from in(2) edge since in(1) is memory edge.
3096 for (uint i = 2; i < n->req(); i++) {
3097 Node* adr = n->in(i)->uncast();
3098 const Type *at = phase->type(adr);
3099 if (!adr->is_top() && at->isa_ptr()) {
3100 assert(at == Type::TOP || at == TypePtr::NULL_PTR ||
3101 at->isa_ptr() != NULL, "expecting an Ptr");
3102 if (adr->is_AddP()) {
3103 adr = get_addp_base(adr);
3104 }
3105 // Mark as ArgEscape everything "adr" could point to.
3106 set_escape_state(adr->_idx, PointsToNode::ArgEscape);
3107 }
3108 }
3109 _processed.set(n_idx);
3110 break;
3111 }
3112 case Op_ThreadLocal:
3113 {
3114 assert(false, "Op_ThreadLocal");
3115 break;
3116 }
3117 default:
3118 // This method should be called only for EA specific nodes.
3119 ShouldNotReachHere();
3120 }
3121 }
3123 #ifndef PRODUCT
3124 void ConnectionGraph::dump() {
3125 bool first = true;
3127 uint size = nodes_size();
3128 for (uint ni = 0; ni < size; ni++) {
3129 PointsToNode *ptn = ptnode_adr(ni);
3130 PointsToNode::NodeType ptn_type = ptn->node_type();
3132 if (ptn_type != PointsToNode::JavaObject || ptn->_node == NULL)
3133 continue;
3134 PointsToNode::EscapeState es = escape_state(ptn->_node);
3135 if (ptn->_node->is_Allocate() && (es == PointsToNode::NoEscape || Verbose)) {
3136 if (first) {
3137 tty->cr();
3138 tty->print("======== Connection graph for ");
3139 _compile->method()->print_short_name();
3140 tty->cr();
3141 first = false;
3142 }
3143 tty->print("%6d ", ni);
3144 ptn->dump();
3145 // Print all locals which reference this allocation
3146 for (uint li = ni; li < size; li++) {
3147 PointsToNode *ptn_loc = ptnode_adr(li);
3148 PointsToNode::NodeType ptn_loc_type = ptn_loc->node_type();
3149 if ( ptn_loc_type == PointsToNode::LocalVar && ptn_loc->_node != NULL &&
3150 ptn_loc->edge_count() == 1 && ptn_loc->edge_target(0) == ni ) {
3151 ptnode_adr(li)->dump(false);
3152 }
3153 }
3154 if (Verbose) {
3155 // Print all fields which reference this allocation
3156 for (uint i = 0; i < ptn->edge_count(); i++) {
3157 uint ei = ptn->edge_target(i);
3158 ptnode_adr(ei)->dump(false);
3159 }
3160 }
3161 tty->cr();
3162 }
3163 }
3164 }
3165 #endif